Lecture Notes in Artificial Intelligence 5855

Edited by R. Goebel, J. Siekmann, and W. Wahlster

Subseries of Lecture Notes in Computer Science

Hepu Deng Lanzhou Wang
Fu Lee Wang Jingsheng Lei (Eds.)

Artificial Intelligence
and Computational
Intelligence
International Conference, AICI 2009
Shanghai, China, November 7-8, 2009
Proceedings

13
Series Editors
Randy Goebel, University of Alberta, Edmonton, Canada
Jörg Siekmann, University of Saarland, Saarbrücken, Germany
Wolfgang Wahlster, DFKI and University of Saarland, Saarbrücken, Germany

Volume Editors
Hepu Deng
RMIT University, School of Business Information Technology
City Campus, 124 La Trobe Street, Melbourne, Victoria 3000, Australia
E-mail: hepu.deng@rmit.edu.au

Lanzhou Wang
China Jiliang University, College of Metrological Technology and Engineering
Hangzhou 310018, Zhejiang, China
E-mail: lzwang@cjlu.edu.cn

Fu Lee Wang
City University of Hong Kong, Department of Computer Science
83 Tat Chee Avenue, Kowloon Tong, Hong Kong, China
E-mail: flwang@cityu.edu.hk

Jingsheng Lei
Hainan University, College of Information Science and Technology
Haikou 570228, China
E-mail: jshlei@hainu.edu.cn

Library of Congress Control Number: 2009936961

CR Subject Classification (1998): I.2, F.2, I.5, I.4, J.3, J.2

LNCS Sublibrary: SL 7 – Artificial Intelligence

ISSN 0302-9743
ISBN-10 3-642-05252-5 Springer Berlin Heidelberg New York
ISBN-13 978-3-642-05252-1 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer. Violations are liable
to prosecution under the German Copyright Law.
springer.com
© Springer-Verlag Berlin Heidelberg 2009
Printed in Germany
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper SPIN: 12781256 06/3180 543210
 Preface

The 2009 International Conference on Artificial Intelligence and Computational Intel-

ligence (AICI 2009) was held during November 7–8, 2009 in Shanghai, China. The
technical program of the conference reflects the tremendous growth in the fields of
artificial intelligence and computational intelligence with contributions from a large
number of participants around the world.
AICI 2009 received 1,203 submissions from 20 countries and regions. After rigor-
ous reviews, 79 high-quality papers were selected for this volume, representing an
acceptance rate of 6.6%. These selected papers cover many new developments and
their applications in the fields of artificial intelligence and computational intelligence.
Their publications reflect a sustainable interest from the wide academic community
worldwide in tirelessly pursuing new solutions through effective utilizations of artifi-
cial intelligence and computational intelligence to real-world problems.
We would like to specially thank all the committee members and reviewers, without
whose timely help it would have been impossible to review all the submitted papers to
assemble this program. We also would like take this opportunity to express our heartfelt
appreciation for all those who worked together in organizing this conference, establish-
ing the technical programs and running the conference meetings. We greatly appreciate
the authors, speakers, invited session organizers, session Chairs, and others who made
this conference possible. Lastly, we would like to express our gratitude to the Shanghai
University of Electric Power for the sponsorship and support of the conference.
We trust that the participants of this conference had opportunities to meet both old
and new friends from all over the world, to exchange ideas, to initiate collaborations
in research, to develop lasting friendships, as well as to visit places in Shanghai and
China.

November 2009 Hepu Deng

Lanzhou Wang
Fu Lee Wang
Jingsheng Lei
 Organization

Organizing Committee

General Co-chairs

Jialin Cao Shanghai University of Electric Power, China

Jingsheng Lei Hainan University, China

Program Committee
Co-chairs

Hepu Deng RMIT University, Australia

Lanzhou Wang China JILIANG University, China

Local Arrangements

Chair

Hao Zhang Shanghai University of Electric Power, China

Proceedings

Co-chairs

Fu Lee Wang City University of Hong Kong, Hong Kong

Jun Yang Shanghai University of Electric Power, China

Publicity

Chair

Tim Kovacs University of Bristol, UK

Sponsorship

Chair

Zhiyu Zhou Zhejiang Sci-Tech University, China

VIII Organization

Program Committee
Ahmad Abareshi RMIT University, Australia
Stephen Burgess Victoria University, Australia
Jennie Carroll RMIT University, Australia
Eng Chew University of Technology Sydney, Australia
Vanessa Cooper RMIT University, Australia
Minxia Luo China Jiliang University, China
Tayyab Maqsood RMIT University, Australia
Ravi Mayasandra RMIT University, Australia
Elspeth McKay RMIT University, Australia
Alemayehu Molla RMIT University, Australia
Konrad Peszynski RMIT University, Australia
Siddhi Pittayachawan RMIT University, Australia
Ian Sadler Victoria University, Australia
Pradip Sarkar RMIT University, Australia
Carmine Sellitto Victoria University, Australia
Peter Shackleton Victoria University, Australia
Sitalakshmi Venkatraman University of Ballarat, Australia
Leslie Young RMIT University, Australia
Adil Bagirov University of Ballarat, Australia
Philip Branch Swinburne University of Technology, Australia
Feilong Cao China Jiliang University, China
Maple Carsten University of Bedfordshire, UK
Caroline Chan Deakin University, Australia
Jinjun Chen Swinburne University of Technology, Australia
Richard Dazeley University of Ballarat, Australia
Yi-Hua Fan Chung Yuan Christian University Taiwan, Taiwan
Richter Hendrik HTWK Leipzig, Germany
Furutani Hiroshi University of Miyazaki, Japan
Bae Hyeon Pusan National University, Korea
Tae-Ryong Jeon Pusan National University, Korea
Sungshin Kim Pusan National University, Korea
Wei Lai Swinburne University of Technology, Australia
Edmonds Lau Swinburne University of Technology, Australia
Qiang Li University of Calgary, Canada
Xiaodong Li RMIT University, Australia
Kuoming Lin Kainan University, Taiwan
YangCheng Lin National Dong Hwa University, Taiwan
An-Feng Liu Central South University, China
Liping Ma University of Ballarat, Australia
Costa Marly Federal University of the Amazon, Brazil
Jamie Mustard Deakin University, Australia
Syed Nasirin Brunel University, UK
Lemai Nguyen Deakin University, Australia
Heping Pan University of Ballarat, Australia
 Organization IX

Craig Parker Deakin University, Australia

Wei Peng RMIT University, Australia
Adi Prananto Swinburne University of Technology, Australia
Takahashi Ryouei Hachinohe Institute of Technology, Japan
Cheol Park Soon Chonbuk National University, Korea
Andrew Stranier University of Ballarat, Australia
Zhaohao Sun University of Ballarat, Australia
Sleesongsom Suwin Chiangrai College, Thailand
Arthur Tatnall Victoria University, Australia
Luba Torline Deakin University, Australia
Schetinin Vitaly University of Bedfordshire, UK
Zhichun Wang Tianjin University, China
Shengxiang Yang University of Leicester, UK
Chung-Hsing Yeh Monash University, Australia
Yubo Yuan China Jiliang University, China
Hossein Zadeh RMIT University, Australia
Dengsheng Zhang Monash University, Australia
Irene Zhang Victoria University, Australia
Jianying Zhang Cancer Victoria, Australia
Tony Zhang Qingdao Univesity, China
Weijian Zhao China Jiliang University, China
Yan-Gang Zhao Nagoya Institute of Technology, Japan
Xiaohui Zhao Swinburne University of Technology, Australia

Reviewers
Adil Bagirov Chen Lifei Du Junping
Ahmad Abareshi Chen Xiang Du Xufeng
Bai ShiZhong Chen Ling Duan Yong
Bi Shuoben Chen Dongming Duan Rongxing
Bo Rui-Feng Chen Ming Fan Jiancong
Cai Weihua Chen Ting Fan Xikun
Cai Zhihua Chen Yuquan Fan Shurui
Cai Kun Chen Ailing Fang Zhimin
Cao Yang Chen Haizhen Fang Gu
Cao Qingnian Chu Fenghong Fuzhen Huang
Cao Guang-Yi Chung-Hsing Yeh Gan Zhaohui
Cao Jianrong Congping Chen Gan Rongwei
Carmine Sellitto Cui Shigang Gao chunrong
Chai Zhonglin Cui Mingyi Gao Xiaoqiang
Chang Chunguang DeJun Mu Gao Jiaquan
Chen Yong Liang Deng Liguo Gao Jun
Chen Haiming Deng-ao Li Gao Boyong
Chen Yong Ding Yi-jie Gu Xuejing
Chen Pengnian Dingjun Chen Gu Tao
X Organization

Guang Yang Li Xiang Peng Qinke

Guo Li Jun Li Ming Perng Chyuan
Guo Yecai Li Guangwen Phaisangittisagul Ekachai
Han Fei Li Guohou Qiang Niu
Hao Xiuqing Li Hongjiao Qiang Li
He Hongjiang Liang Benliang Qiao Mei
He Shengping Lin Yuesong Qiu Jiqing
He Bingwei Lin Dong Mei Rao Congjun
He Xingshi Lin Chen Ren Li
He Zhiwei Lin Jipeng Ren Kaijun
He Yue Liu Jun Ren Chunyu
Hendrik Richter Liu Zhijie Richard Dazeley
Ho Zih-Ping Liu Xiaoji Ruiming Jia
Hu Fu Yuan Liu Hui Sang Yan-fang
Huang Pei-Hwa Liu Anping Sanyou Zeng
Huang Zhiwei Liu Zhide Shen Xiaojing
Huang Rongbo Liu Jihong Shen Xiaohua
Huang Nantian Liu Heng Sheng Zhongqi
Ji Genlin Liu Jin Sheng Xueli
Ji Tingwei Liu Wenyu Shi Qian
Ji Xia Liu Dianting Shi Guoliang
Ji Chang-Ming Liu Lei Shi Xuerong
Jia Qiuling Liu Jin Shuai Ren
Jia Fang Liu Jinglei Si Chun-di
Jiang Jiafu Liu Hui Sitalakshmi Venkatraman
Jianying Zhang Liu Guangli Song Xiaoyu
Jiao Feng Liu Yang Su H.T.
Jiechun Chen Liu Ying Sun Zhanquan
Jingbo Xia Liu Hao Sun Yanxia
Jingfang Liu Liu Jian Sun Quanping
Jin-min Wang Liu Ying-chun Sun Feixian
Ju-min Zhao Lu Yaosheng Sun Ziwen
Koga Takanori Lu Bin Suwin Sleesongsom
Kwok Ngai Ming Lu Yuzheng Tan Junshan
Lai Jiajun Lu Huijuan Tan Chunqiao
Lai Lien-Fu Ma Xianmin Tang Xianghong
Lai-Cheng Cao Matsuhisa Takashi Tao Luo
Lei Zhang Meng Jianliang Tao Zhengru
Lei Xiujuan Miao Xuna Tao Zhang
Leng Cuiping Ni Jian-jun Tian Fu-qing
Li Yu Nie Qing Tong Zhi-Wei
Li Xinyun Ning Ning Uchino Eiji
Li Xiao Niu Junyu Ullrich Carsten
Li Shun-Ming Ouyang Zhonghui Wan Li
Li Yun Pan Zhenghua Wang Cheng
Li Haiming Pan Jingchang Wang Hao
Li Chengjia Pan Defeng Wang Li-dong
 Organization XI

Wang Hao Wu Xiaoqin Ye Feiyue

Wang Li-dong Wu Zhangjun Yi-Hua Fan
Wang Jie-sheng Wu Da-sheng Yin Chunxia
Wang Guoqiang Xi Liang Yong Yang
Wang Caihong Xia Qingjun Yong Qidong
Wang Jingmin Xiao Xiaohong You Xiao-ming
Wang Shuangcheng Xiao Hongying You-hua Jiang
Wang Chujiao Xie Yanmin Yu Xinqiao
Wang Junwei Xie Zhenping Yuhui Zhao
Wang XiaoMing Xie Jianhong Yuyi Zhai
Wang Xin Xiuli Yu Zhan Yong-zhao
Wang Feng Xiumin Yang Zhang Huaixiang
Wang Yuanqing Xiuxia Yang Zhang Meng
Wang Yong Xizhong Shen Zhang Sumin
Wang Xiang Xu Jie Zhang Yizhuo
Wang Yu-Jie Xu Hailong Zhang Erhu
Wang Yufei Xu Hua Zhang Jianxin
Wang Hongsheng Xu Xinsheng Zhang Caiqing
Wang Xiuqing Xu Yan Zhang Jian
Wang Wenchuan Xu Zhangyan Zhang Yanjie
Wang Dong Xue Heru Zhang Dabin
Wang Haibo Xue Yang Zhang Xiang
Wang Dawei Xue Heng Zhang Jingjun
Wang Hui Xueqin Lu Zhang Zhenya
Wang Junnian Yan Ping Zhang Ke
Wang Xuhong Yan Wang Zhang Xuejun
Wang Xiaolu Yan Tie Zhang Xuehua
Wan-gan Song Yang Genghuang Zhang Junran
Wei Peng Yang Ming Zhang Xueming
Wen Mi Yang Jiadong Zhang Changjiang
Wenjun Liu Yang Xiaobing Zhao Ping
Wu Xia Yang Jun Zhao Fu
Wu Hongli Yang Liping Zhao Liaoying
Wu Xiaojun Yang-Cheng Lin Zhao Hui
Wu Xiaojin Yao Zhijun
 Table of Contents

Support Vector Machine

An Experimental Research on Vector Control of Induction Motor Based
on Simple Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Yinhai Zhang, Jinfa Ge, Weixia Liu, and Qin Wang

Rough Set Theory

An United Extended Rough Set Model Based on Developed Set Pair
Analysis Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Xia Ji, Longshu Li, Shengbing Chen, and Yi Xu

Learning Rules from Pairwise Comparison Table . . . . . . . . . . . . . . . . . . . . . 18

Liping An and Lingyun Tong

Particle Swarm Optimization

A Novel Hybrid Particle Swarm Optimization for Multi-Objective
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Siwei Jiang and Zhihua Cai

Research on Constrained Layout Optimization Problem Using

Multi-adaptive Strategies Particle Swarm Optimizer . . . . . . . . . . . . . . . . . . 38
Kaiyou Lei
ARIMA Model Estimated by Particle Swarm Optimization Algorithm
for Consumer Price Index Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Hongjie Wang and Weigang Zhao

Neural Computation
A Uniform Solution to HPP in Terms of Membrane Computing . . . . . . . . 59
Haizhu Chen and Zhongshi He

Self-organizing Quantum Evolutionary Algorithm Based on Quantum

Dynamic Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Sheng Liu and Xiaoming You

Intelligent Agents and Systems

Can Moral Hazard Be Resolved by Common-Knowledge in
S4n-Knowledge? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Takashi Matsuhisa
XIV Table of Contents

Casuist BDI-Agent: A New Extended BDI Architecture with the

Capability of Ethical Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Ali Reza Honarvar and Nasser Ghasem-Aghaee

Information Security
Research of Trust Chain of Operating System . . . . . . . . . . . . . . . . . . . . . . . 96
Hongjiao Li and Xiuxia Tian
A Novel Application for Text Watermarking in Digital Reading . . . . . . . . 103
Jin Zhang, Qing-cheng Li, Cong Wang, and Ji Fang

Immune Computation
Optimization of Real-Valued Self Set for Anomaly Detection Using
Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Liang Xi, Fengbin Zhang, and Dawei Wang

Genetic Algorithms
A GP Process Mining Approach from a Structural Perspective . . . . . . . . . 121
Anhua Wang, Weidong Zhao, Chongchen Chen, and Haifeng Wu
Effects of Diversity on Optimality in GA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Glen MacDonald and Gu Fang
Dynamic Crossover and Mutation Genetic Algorithm Based on
Expansion Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Min Dong and Yan Wu
Multidisciplinary Optimization of Airborne Radome Using Genetic
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Xinggang Tang, Weihong Zhang, and Jihong Zhu
Global Similarity and Local Variance in Human Gene Coexpression
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Ivan Krivosheev, Lei Du, Hongzhi Wang, Shaojun Zhang,
Yadong Wang, and Xia Li
A Grid Based Cooperative Co-evolutionary Multi-Objective
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Sepehr Meshkinfam Fard, Ali Hamzeh, and Koorush Ziarati

Fuzzy Computation
Fuzzy Modeling for Analysis of Load Curve in Power System . . . . . . . . . . 176
Pei-Hwa Huang, Ta-Hsiu Tseng, Chien-Heng Liu, and
Guang-Zhong Fan
 Table of Contents XV

Biological Computing
Honey Bee Mating Optimization Vector Quantization Scheme in Image
Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Ming-Huwi Horng

Towards a Population Dynamics Theory for Evolutionary Computing:

Learning from Biological Population Dynamics in Nature . . . . . . . . . . . . . 195
Zhanshan (Sam) Ma

Applications of Computational Intelligence

Application of Improved Particle Swarm Optimization Algorithm in
UCAV Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Qianzhi Ma and Xiujuan Lei

Modal Analysis for Connecting Rod of Reciprocating Mud Pump . . . . . . 215

Zhiwei Tong, Hao Liu, and Fengxia Zhu

A Time-History Analysis Algorithm of a Non-viscously Damped System

Using Gauss Precise Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Hongyu Shen and Zhongdong Duan

Application of LSSVM-PSO to Load Identification in Frequency

Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Dike Hu, Wentao Mao, Jinwei Zhao, and Guirong Yan

Local Model Networks for the Optimization of a Tablet Production

Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Benjamin Hartmann, Oliver Nelles, Aleš Belič, and
Damjana Zupančič–Božič

Retracted: Implementation of On/Off Controller for Automation of

Greenhouse Using LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
R. Alimardani, P. Javadikia, A. Tabatabaeefar, M. Omid, and
M. Fathi

Ant Colony Algorithm

Structure Design of the 3-D Braided Composite Based on a Hybrid
Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Ke Zhang

Robotics
Robot Virtual Assembly Based on Collision Detection in Java3D . . . . . . . 270
Peihua Chen, Qixin Cao, Charles Lo, Zhen Zhang, and Yang Yang
XVI Table of Contents

Robust Mobile Robot Localization by Tracking Natural Landmarks . . . . 278

Xiaowei Feng, Shuai Guo, Xianhua Li, and Yongyi He
Multi-Robot Dynamic Task Allocation Using Modified Ant Colony
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
Zhenzhen Xu, Feng Xia, and Xianchao Zhang

Pattern Recognition
A Novel Character Recognition Algorithm Based on Hidden Markov
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Yu Wang, Xueye Wei, Lei Han, and Xiaojin Wu
New Algorithms for Complex Fiber Image Recognition . . . . . . . . . . . . . . . 306
Yan Ma and Shun-bao Li
Laplacian Discriminant Projection Based on Affinity Propagation . . . . . . 313
Xueping Chang and Zhonglong Zheng
An Improved Fast ICA Algorithm for IR Objects Recognition . . . . . . . . . 322
Jin Liu and Hong Bing Ji
Facial Feature Extraction Based on Wavelet Transform . . . . . . . . . . . . . . . 330
Nguyen Viet Hung
Fast Iris Segmentation by Rotation Average Analysis of
Intensity-Inversed Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Wei Li and Lin-Hua Jiang

Neural Networks
A New Criterion for Global Asymptotic Stability of Multi-delayed
Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Kaiyu Liu and Hongqiang Zhang
An Improved Approach Combining Random PSO with BP for
Feedforward Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Yu Cui, Shi-Guang Ju, Fei Han, and Tong-Yue Gu
Fuzzy Multiresolution Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Li Ying, Shang Qigang, and Lei Na
Research on Nonlinear Time Series Forecasting of Time-Delay NN
Embedded with Bayesian Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Weijin Jiang, Yusheng Xu, Yuhui Xu, and Jianmin Wang
An Adaptive Learning Algorithm for Supervised Neural Network with
Contour Preserving Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Piyabute Fuangkhon and Thitipong Tanprasert
 Table of Contents XVII

Natural Language Processing

Application Study of Hidden Markov Model and Maximum Entropy in
Text Information Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Rong Li, Li-ying Liu, He-fang Fu, and Jia-heng Zheng

Automatic Expansion of Chinese Abbreviations by Web Mining . . . . . . . . 408

Hui Liu, Yuquan Chen, and Lei Liu

Adaptive Maximum Marginal Relevance Based Multi-email

Summarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Baoxun Wang, Bingquan Liu, Chengjie Sun, Xiaolong Wang, and
Bo Li

Semantic Relation Extraction by Automatically Constructed Rules . . . . . 425

Rongfeng Huang

Machine Vision
Object Recognition Based on Efficient Sub-window Search . . . . . . . . . . . . 435
Qing Nie, Shouyi Zhan, and Weiming Li

Machine Learning
A Multi-Scale Algorithm for Graffito Advertisement Detection from
Images of Real Estate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
Jun Yang and Shi-jiao Zhu

An Improved Path-Based Transductive Support Vector Machines

Algorithm for Blind Steganalysis Classification . . . . . . . . . . . . . . . . . . . . . . . 453
Xue Zhang and ShangPing Zhong

Logic Reasoning and Theorem-Proving

Resolution with Limited Factoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Dafa Li

Knowledge Representation and Acquisition

Formal Analysis of an Airplane Accident in N Σ-Labeled Calculus . . . . . . 469
Tetsuya Mizutani, Shigeru Igarashi, Yasuwo Ikeda, and
Masayuki Shio

Using Concept Space to Verify Hyponymy in Building a Hyponymy

Lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Lei Liu, Sen Zhang, Lu Hong Diao, Shu Ying Yan, and Cun Gen Cao
XVIII Table of Contents

Advanced Self-adaptation Learning and Inference Techniques for Fuzzy

Petri Net Expert System Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
Zipeng Zhang, Shuqing Wang, and Xiaohui Yuan

Intelligent Signal Processing

MIMO Instantaneous Blind Identification Based on Second-Order
Temporal Structure and Newton’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Xizhong Shen and Guang Meng

Seed Point Detection of Multiple Cancers Based on Empirical Domain

Knowledge and K-means in Ultrasound Breast Image . . . . . . . . . . . . . . . . . 507
Lock-Jo Koo, Min-Suk Ko, Hee-Won Jo, Sang-Chul Park, and
Gi-Nam Wang

Intelligent Scheduling
A Controlled Scheduling Algorithm Decreasing the Incidence of
Starvation in Grid Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Minna Liu, Kairi Ou, Yuelong Zhao, and Tian Sun

A Space Allocation Algorithm for Minimal Makespan in Space

Scheduling Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
Chyuan Perng, Yi-Chiuan Lai, Zih-Ping Ho, and Chin-Lun Ouyang

Intelligent Information Retrieval

A New Approach for Chest CT Image Retrieval . . . . . . . . . . . . . . . . . . . . . . 535
Li-dong Wang and Zhou-xiang Shou

Intelligent Information Fusion

Depth Extension for Multiple Focused Images by Adaptive Point
Spread Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
Yuanqing Wang and Yijun Mei

Intelligent Image Processing

Image Annotation Using Sub-block Energy of Color Correlograms . . . . . . 555
Jingsheng Lei

Heuristic Searching Methods

Optimal Evolutionary-Based Deployment of Mobile Sensor Networks . . . 563
Niusha Shafiabady and Mohammad Ali Nekoui
 Table of Contents XIX

Fuzzy Logic and Soft Computing

Research of Current Control Strategy of Hybrid Active Power Filter
Based on Fuzzy Recursive Integral PI Control Algorithm . . . . . . . . . . . . . . 568
Zhong Tang and Daifa Liao

A Fuzzy Query Mechanism for Human Resource Websites . . . . . . . . . . . . . 579

Lien-Fu Lai, Chao-Chin Wu, Liang-Tsung Huang, and
Jung-Chih Kuo

Selecting Cooperative Enterprise in Dynamic Enterprise Alliance Based

on Fuzzy Comprehensive Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
Yan Li, Guoxing Zhang, and Jun Liu

Distributed AI and Agents

An Evolutionary Solution for Cooperative and Competitive Mobile
Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
Jiancong Fan, Jiuhong Ruan, and Yongquan Liang

Towards an Extended Evolutionary Game Theory with Survival

Analysis and Agreement Algorithms for Modeling Uncertainty,
Vulnerability, and Deception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
Zhanshan (Sam) Ma

Data Mining and Knowledge Discovering

Uncovering Overlap Community Structure in Complex Networks Using
Particle Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619
Fabricio Breve, Liang Zhao, and Marcos Quiles

Semi-supervised Classification Based on Clustering Ensembles . . . . . . . . . 629

Si Chen, Gongde Guo, and Lifei Chen

Research on a Novel Data Mining Method Based on the Rough Sets

and Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639
Weijin Jiang, Yusheng Xu, Jing He, Dejia Shi, and Yuhui Xu

Applications of Artificial Intelligence

The Techno-Economic Analysis of Reducing NOx of Industrial Boiler
by Recycled Flue Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Caiqing Zhang, Wenpei Hu, Yu Sun, and Bo Yang

Urban Traffic Flow Forecasting Based on Adaptive Hinging

Hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
Yang Lu, Jianming Hu, Jun Xu, and Shuning Wang
XX Table of Contents

Turbine Fault Diagnosis Based on Fuzzy Theory and SVM . . . . . . . . . . . . 668

Fei Xia, Hao Zhang, Daogang Peng, Hui Li, and Yikang Su

Architecture of Multiple Algorithm Integration for Real-Time Image

Understanding Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Bobo Duan, Chunyang Yang, Wei Liu, and Huai Yuan

Others
Formalizing the Modeling Process of Physical Systems in MBD . . . . . . . . 685
Nan Wang, Dantong OuYang, Shanwu Sun, and Chengli Zhao

Study on Stochastic Programming Methods Based on Synthesizing

Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
FaChao Li, XianLei Liu, and ChenXia Jin

Rate Control, Routing Algorithm and Scheduling for Multicast with

Network Coding in Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
Xu Na Miao and Xian Wei Zhou

Design and Experimental Study on Spinning Solid Rocket Motor . . . . . . . 715

Heng Xue, Chunlan Jiang, and Zaicheng Wang

An Energy-Based Method for Computing Radial Stiffness of Single

Archimedes Spiral Plane Supporting Spring . . . . . . . . . . . . . . . . . . . . . . . . . 724
Enlai Zheng, Fang Jia, Changhui Lu, He Chen, and Xin Ji

Erratum
Implementation of On/Off Controller for Automation of
Greenhouse Using LabVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E1
R. Alimardani, P. Javadikia, A. Tabatabaeefar, M. Omid, and
M. Fathi

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735

 An Experimental Research on Vector Control of
Induction Motor Based on Simple Model

Yinhai Zhang, Jinfa Ge, Weixia Liu, and Qin Wang

College of Informatics & Electronics

Zhejiang Sci-Tech University
Hangzhou City, Zhejiang Province, China
nikon_nike@163.com, gejinfa@yahoo.com.cn

Abstract. Given the heavy computation, easy saturation and cumulate errors of
conventional direct vector control, the vector control of induction motor based on
simple model is studied and the detailed scheme is described on the basis of the
decomposing and approximating the rotor flux. Because of the direct closed-loop
control of the magnetizing current and the torque current and the complex current
regulator is completed by PI regulator, so the direct vector control of induction
motor is simplified. The experimental results show that the proposed method is
effective in decreasing the dynamic disturbance and has the advantages of the
simplicity of the code program, rare saturation and shocks.

Keywords: vector control, induction motor, rotor flux, decoupler.

1 Introduction
The vector control can make the induction motor gain the excellent dynamic per-
formance that is similar to the one of the DC motor. Now, the vector control that based
on the rotor field orientated control is paid enough attention because it easily realizes
the absolute decoupling between the magnetizing current and the torque current [1-2].
The direct vector control and the indirect vector control are the two most commonly
used in the vector control of induction motor system. The former has the speed
closed-loop control which includes torque closed-loop and the flux closed-loop control
system, the latter is an open-loop flux control system. There are a large number of
operations such as the differential and the product in the direct vector control. Fur-
thermore, the complex processes of current regulator in the voltage-type inverter bring
the heavy computation [3-5], easy saturation, cumulate errors, and other uncertain
disturbance, resulting in deterioration of system performance.
Therefore, in order to overcome the above-mentioned disadvantages of the direct
vector control and maintain its advantages of performance, the vector control of in-
duction motor based on simple model is studied and the detailed scheme is described in
this paper. After the decomposing and approximating the rotor flux, the magnetizing
current and the torque current are directly close-looped and the complex process of
current regulator in the voltage-type inverter is completed by PI regulator to achieve the
regulation without static error. Thus, the proposed method simplifies the direct vector

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 1–8, 2009.
© Springer-Verlag Berlin Heidelberg 2009
2 Y. Zhang et al.

control of induction motor, and makes the direct vector control system easy, clear, less
code program, rare saturation and shocks.

2 Conventional Direct Vector Control

Fig.1 shows the system of the conventional direct vector control of induction motor.
This system which is built based on the rotor field orientated control includes the speed
control subsystem and the flux control subsystem, and the inner loop of the speed
control subsystem is the torque closed-loop control system. There are the speed regu-
lator (ASR), the torque regulator (ATR) and the flux regulator (AΨR) in the direct
vector control system of induction motor. The current transform includes the Clarke
and the Park conversion.

Park −1t.
ω * Te* i s*t ust* us*α
m, t SA
− ˆ
− Te SB
ωr ψ r* *
ism *
usm
α,β
us*β SC
− ψˆ r
θˆs

iA
pn Lm iB
Lr i iC
st

Fig. 1. The conventional direct vector control system of induction motor

The following are the main formulas that used in the conventional direct vector
control system of the induction motor.
Rotor flux is expressed as follows
Lm
ψr = ism (1)
Tr p + 1
The feedback of the torque calculation equation is
pn Lm
Te = ist ψ r (2)
Lr
The angular speed of synchronous rotation is given by t-he following formula
Lm ist
ωs = ωr + ωsl = ωr + (3)
Trψ r
In the most example applications, the voltage-type inverter is used widely, so it is
necessary to change the current to voltage. This process is called current regulator. And
the equations of the conversion are
 An Experimental Research on Vector Control of Induction Motor 3

⎧ Ls σTr p + 1
⎪usm = Rs (1 + R p T p + 1 )ism
⎪ s r
⎪ Tr p + 1 ist
⎪ − σLs (ωr + )ist
⎪ Tr ism
⎨ (4)
⎪u = [ R ( σLs p + 1) + Ls (σT p + 1)]i
⎪ st s
Rs Tr
r st
⎪
⎪ σT p + 1
+ ωr Ls r )ism
⎪ Tr p + 1
⎩
The parameters of equation (1) to (4) are described as follows:
L2m
σ =1−
Ls Lr
Lm — Mutual inductance
Lr — Rotor inductance
Rs — Stator resistance
Ls — Stator inductance
Tr — Rotor time constant
p — Differential operator
ωr — Rotor angular speed
ωsl — Slip angular speed
pn — The number of pole pairs
ism — M-axis stator current
ist — T-axis stator current
usm — M-axis stator voltage
ust — T-axis stator voltage

From the equation (1) to (3), we can see that the accuracy of estimation of equation (2)
and (3) depend on (1). It shows that the importance of the accuracy of estimation of the
rotor flux. Because of the existence of differential part, the transient drift will appear in
the part of direct current following the transient changes of the digital operation when
the rectangular discrete integrator is used, thus making the value of estimated rotor flux
inaccurate and then reducing the calculation accuracy of the torque and the synchro-
nous speed. Furthermore, the existence of a large number of the differential and product
operations in the equations described by (4) will bring the heavy computation, easy
saturation and cumulate errors. Especially, in the low speed region, the torque is very
prone to oscillate.

3 Vector Control System Based on Simple Model

Fig. 2 gives the vector control system based on simple model. In this paper, the rotor
flux is decomposed and approximated. Then the magnetizing current and the torque
current are for closed-loop control directly in order to overcome these shortcomings of
4 Y. Zhang et al.

Park −1t.
ω* i s*t u st* u s*α
m, t SA
− −i SB
ωr α,β
st

i* *
usm u*sβ SC
sm
−i sm

θˆs
isα
m, t α,β iA
i st

α , β isβ a, b, c iB
ωr ism
Park t. Clarke t.

Fig. 2. The vector control system of induction motor based on simple model

the conventional direct vector control. The system includes the speed control subsys-
tem and the torque current control subsystem. And the inner loop of the speed control
subsystem is the torque current closed-loop control subsystem. The system is com-
posed of three PI regulators, the speed regulator, the torque current regulator and the
magnetizing current regulator.
The process of the main formulas derivation that used in the vector control system
based on simple model are as follows, in which they are discretized to fit DSP computing.
The equation (1) can be rewrote as
Tr pψ r + ψ r = Lm ism (5)
And the rotor flux equation that based on MT reference frame is
ψ r = Lmism + Lr irm (6)
Lσ r is rotor leakage inductance and it is small enough compared with the mutual in-
ductance Lm. So we can consider that Lm and Lr are approximately equal in value ac-
cording to L r = L m + L σ r . Furthermore, in the rotor field orientated control system,
the rotor flux must maintain constant in order to achieve an ideal speed-adjusting
performance under the rating frequency. And above the rating frequency, the flux
weakening control is adopted and is usually achieved through the look-up table. Lmism is
regard as constant and contained by Lrirm . So the equation (6) can be simplified as
ψ r = Lm irm (7)
In equation (7), irm is the M-axis component of rotor current, and applying equation (7)
to (5), we can obtain
dirm
Tr + irm = ism (8)
dt
After discrete differential, equation (8) can be rewrote as
 An Experimental Research on Vector Control of Induction Motor 5

Ts
irm (k + 1) = irm (k ) + (ism (k ) − irm (k )) (9)
Tr
Where Ts is the switching period. And as the same method, equation (3) can be modi-
fied as
1 ist (k )
ωs (k + 1) = ωr (k + 1) + (10)
Tr irm (k + 1)
The equation (9) and (10) are derived based on the decomposition and approximation
of the rotor flux, which are considered as crucial observation equations of the system
showed in Fig. 2, being used to calculate the angular speed of synchronous rotation.
Comparing to the conventional direct vector control system, we can find that in the
simplified vector control system it is not to calculate the angular speed of synchronous
rotation by calculating the rotor flux, but to obtain directly through the rotor magnet-
izing current. Then the heavy computational processes are ignored, so the saturation
drift and cumulative errors are avoided. And we will have a good foundation during the
study of the real-time high-performance vector control system.
It can be seen from Fig. 2 that the magnetizing current and the torque current are for
closed-loop control directly rather than the torque and flux in the simplified vector
control system that compared to the conventional vector control. So the calculation of
torque feedback is avoided, and the whole system can regulate the torque and the flux
more quickly and efficiently. In addition, the complex process of current regulator is
included by PI regulator to achieve the regulation without static error. So it make the
system simple, clear and easy computation, and more suitable for real-time
high-performance control.

4 Experiment Analysis
The hardware and software environments of the experiment: the power inverter part is
designed with IGBT (IRGB15B60KD) which is produced by IR. The control system
part adopts the TI 2407DSP EVM by Wintech. The current sensor is LA28-NP pro-
duced by the company of LEM. The resolution of the speed encoder is 1000P/R, and
the type of digital oscilloscope is TektronixTDS2002. Switching frequency is 16 KHz,
switching period Ts = 62.5µs. PI parameters: in the current loop, P=1.12, I=0.0629 and
in the speed loop, P= 4.89, I=0.0131.
The parameters of the induction motor: the rated power: Pn =500W, the rated cur-
rent: In =2.9A, the rated voltage: Vn=220V, the rotor resistance: Rr =5.486Ω, the stator
resistance: Rs =4.516Ω, the mutual inductance: Lm =152mH, Lσr =13mH, Lr =165mH,
pn=2, Ls =168mH.
Fig. 3 gives the trajectory of the total flux circle when the operation frequency f
=5Hz in the vector control system of induction motor based on simple model. Ac-
cording to the experimental waveform, we can find that the trajectory of the total flux
circle is near round and that the system has a good running status when the system is
running at low speed.
6 Y. Zhang et al.

Ψβ (pu)

Ψα (pu)

Fig. 3. The total flux circle with the operation frequency f =5Hz
/450rpm/div

/450rpm/div

ωr
ωr
iA /1.5A/div

ist (pu)

t/50ms/div t/100ms/div

(a) rotor speed and phase current (b) rotor speed and torque current
Fig. 4. The start-up of induction motor with no-load

Fig. 4 represents the experimental waveforms in the process of induction motor

starting up with no-load to running at the set speed of f =30Hz by the simplified vector
control system. Fig. 4a gives the waveform of the rotor speed and the stator phase
current. Fig. 4b shows the waveform of the rotor speed and the torque current. In Fig. 4,
the peak to peak value of phase current is less than 2A, and the adjusting time is 200ms,
and the maximum overshoot of the speed is about 13%. During the increasing speed, ist
reaches the biggest value to meet the requirements of increasing speed quickly and
when the speed reach to a certain degree, the torque current gradually decreased until
the speed reach the feed-forward speed.
 An Experimental Research on Vector Control of Induction Motor 7
/300rpm/div

ism (pu)
ωr
ist (pu)

ist (pu)
t/100ms/div t/500ms/div

(a) rotor speed and torque current (b) magnetizing current and torque current

Fig. 5. The experimental wave with random load at f =10Hz

The experimental waveforms that under the random load when the motor runs at the
steady state of f =10Hz is represented by Fig. 5. Fig. 5a gives the waveforms of the rotor
speed and torque current, and we can see that the maximum dynamic speed downhill is
about 16% and the recovery time is less than the time of disturbance operation. Fig. 5b
shows the waveforms of the magnetizing current and torque current, and we can find
that the magnetizing current ism always remain unchanged and be able to get ideal de-
coupling with the torque current. So the vector control of induction motor based on
simple model works well in the anti-disturbance performance.

5 Conclusion
In this paper, the vector control of induction motor based on simple model is studied
and the detailed scheme is described through the decomposing and approximating the
rotor flux. The proposed method makes the system easy and has less code program,
shorter execution time. The experimental results show that the anti-disturbance per-
formance of the simplified vector control system is reliable and efficient and suitable
for real-time high-performance control system.

References
1. Noguchi, T., Kondo, S., Takahashi, L.: Field-oriented control of an induction motor with
robust on-line tuning of its parameters. IEEE Trans. Industry Applications 33(1), 35–42
(1997)
2. Telford, D., Dunnigan, M.W.: Online identification of induction machine electrical parame-
ters for vector control loop tuning. IEEE Trans. Industrial Electronics 50(2), 253–261 (2003)
8 Y. Zhang et al.

3. Zhang, Y., Chen, J., Bao, J.: Implementation of closed-loop vector control system of induc-
tion motor without voltage feedforward decoupler. In: IEEE 21st Annual Applied Power
Electronics Conference and Exposition, APEC, pp. 1631–1633 (2006)
4. Lorenz, Lawson, R.D., Donald, B.: Performance of feedforward current regulators for
field-oriented induction machine controllers. IEEE Trans. Industry Applications 23(1.4),
597–602 (1987)
5. Del Blanco, F.B., Degner, M.W., Lorenz, R.D.: Dynamic analysis of current regulators for
AC motors using complex vectors. IEEE Trans. Industry Applications 35(6), 1424–1432
(1999)
 An United Extended Rough Set Model Based on
Developed Set Pair Analysis Method

Xia Ji, Longshu Li, Shengbing Chen, and Yi Xu

Key Lab of Ministry of Education for CI & SP,

Anhui University,
Hefei Anhui, China
jixia1983@163.com

Abstract. Different with traditional set pair analysis method, a new method to
micro-decompound the discrepancy degree is proposed in line with the distrib-
uting actuality of missing attributes. Integrated with rough set theory, advances
the united set pair tolerance relation and gives corresponding extended rough set
model. Different values of identity degree and discrepancy degree can modulate
the performance of this model and extend it's application range. Then expound
some existed extended rough set model are the subschema of it. At last simula-
tion experiments and conclusion are given, which validate the united set pair
tolerance relation model can improve classification capability.

Keywords: Set pair analysis, Otherness, Rough set, United set pair tolerance
relation.

1 Introduction

Classic rough set theory is based on the equivalence relation, and for a complete
information system. But in real life, acknowledge obtain always face the incomplete
information system for the reason of data measure error or the limitation of data
acquisition. There are two kind of null value: (1)omit but exist;[2]; (2)lost and not
allowed to be compared. Kryszkiewicz[3] established tolerance relationship for
type(1). Stefanowski[4] and others built similar relationship for type(2). Wang GY[5]
proposed the limited tolerance relation by in-depth study of tolerance relations and
similar relation.
The incomplete degree and the distribution of loss data in various systems are dif-
ferent from each other. And any extended rough set model can achieve a good effect
sometimes but not always. Some quantified extended rough set model, like quantified
tolerance relation and quantified limited tolerance relation[6], has a good effect , but the
process of quantifying costs large calculation.
The set pair analysis (Set Pair Analysis, SPA) theory was formally proposed by
China scholars ZHAO Ke-qin in 1989, which is used to study the relationship between
two set [7].It uses "a + bi + cj "as an associated number to deal with uncertain sys-
tem,such as fuzzy, stochastic or intermediary. It studies the uncertain of two sets from

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 9–17, 2009.
© Springer-Verlag Berlin Heidelberg 2009
10 X. Ji et al.

three aspects: identity, difference and opposition. At present, the set pair analysis the-
ory is widely used in artificial intelligence, system control, management deci-
sion-making and other fields [7].
In this paper, we expand the set pair analysis theory firstly, then put forward a united
extended rough set model based on it.This model can convert to different existed
models with different identity and discrepancy thresholds, including tolerance relation,
similarity relation and limited tolerance relation. Experiments on the UCI data sets
show validity ,rationality and effectiveness of this model.

2 Set Pair Analysis Theory and Rough Set Theory

Set Pair is composed of two sets A and B ,namely H=(A,B).Under general circum-
stances (W), it’s a formula:

u W (A , B ) =
S F P
+ i+ j (1)
N N N

Here N is the total number of features, S is the number of identity features, and P is the
number of contrary features, of the set pair discussed. F = N-S-P is the number of
features of the two sets that are neither identity nor contrary. S/N, F/N, and P/N are
called identity degree, discrepancy degree, and contrary degree of the two sets under
certain circumstances respectively. In order to clarify formula(1),we set a=S/N, b=F/N,
c=P/N, and then uW(A,B) can be rewritten as:
u = a + bi + cj (2)

It is obvious that 0 ≤ a, b, c ≤ 1, and the “a”, “b”, and “c” satisfy the equation a + b + c = 1.
An incomplete information system is the following tuple: I=(U,A,F), where U is a
finite nonempty set of objects, A is a finite nonempty set of attributes, Va is a nonempty
∈
set of values of a A, and F={Fl: U→ρ(Va)} is an information function that maps an
∈ ∈
object in U to exactly one value in Va. For every al A and every xi U, if Fl(xi)is a
single point set, then (U,A,F)is a complete information system. If there is some al ∈ A
∈
and some xi U makes Fl(xi) is not a single point set, then (U,A,F)is an incomplete
information system. An incomplete information system is the special case of a com-
plete information system.
Let I=(U,A,F) is an incomplete information system. For any B ⊆ A,the tolerance
relation TR(B) proposed by M.Kryskiewcz is defined as follows:[3]
TR(B)={(x,y) ∈ U×U| ∀ a ∈ A,a(x)=a(y) ∪ a(x)=* ∪ a(y)=*}
Let I=(U,A,F) is an incomplete information system. For any B ⊆ A ,the similarity
relation SR(B) proposed by Stefanowski is defined as follows:[4]

SR(B)={(x,y)∈U×U|∀b∈B,b(x)=b(y)orb(x)=*}
An United Extended Rough Set Model Based on Developed Set Pair Analysis Method 11

Let I=(U,A,F) is an incomplete information system. For any B ⊆ A, the limited toler-
ance relation LTR(B) proposed by WANG GY is defined as follows:[5]

PB(x)={b|b∈B and b(x)≠*}

LTR (B) = ((x, y) ∈U×U|∀b∈B,b(x)=b(y)=*or((P (x)∩P (y)≠Φ)

B B

and (b(x)≠*&b(y)≠*)→(b (x) = b (y )))}.

3 A Unified Extended Rough Set Model Based on Expanded Set

Pair Analysis Theory

A New Expanded Level Decomposition of Set Pair Analysis

The connection degree u=a +bi +cj can be expanded breadthwise and lengthways ac-
cording to the demand of research The transverse expanding form is:

(a1+a2+... +an)+(blil+b2i2+.. .+ bnin)+(c1jl+c2j2+... +cnjn)

The lengthways expanding form is (see Fig. l):

u(x,y)=a + bi + cj

a1 b1 c1

Fig. 1. The traditional micro-de-compound method of discrepancy degree

On incomplete information system S = (U, A, V),"bi" is also deserved to be studied.

Here N is the total number of attributes, S is the number of identity attributes, and P is the
number of contrary attributes, F = N-S-P is the number of missing attributes. Different
from traditional set pair analysis theory, we de-compound "bi" from another angle: dis-
tribution of missing attributes. Specific decomposition shown in Fig. 2.

u(x,y)=a + bi + cj

b1 b2 b3

Fig. 2. Developed micro-de-compound method of discrepancy degree

12 X. Ji et al.

Here for any a in attribute set A, b1=|{a| a(x)=*&a(y)≠*}|, b2=|{a| a(x)≠*&a(y)=*}|,

。
b3=|{a| a(x)=*&a(y)=*}| The "b1" represents that attribute of x is missing. The "b2"

represents that attribute of y is missing. And "b3" represents attribute of x and y are all
missing.
The traditional set pair analysis theory decompose the discrepancy term "bi" ac-
cording to possible value of missing attribute. While in this paper we from the angle of
distribution of missing attributes to decompose discrepancy term "bi".In the later sec-
tion we'll prove that the symmetry and transitivity of binary relation depend on the
threshold of b1 and b2.

B Extended Rough Set Model Based on United Set Pair Tolerance Relation

The connection degree between two objects is composed of three parts: the ratio of
known and same attributes (the identity degree a), the ratio of known but different at-
tributes (contrary degree cj) and the ratio of unknown attributes (discrepancy degree b ).The
contribution of contrary degree to similarity is negative, without consideration to deal
with noise data, so we set contrary degree be 0,that is c=0.According to expanded set pair
analysis theory, we give the united set pair tolerance relation.

Definition 1. Let I= (U, A, V)is an incomplete information system, ∀ x, y ∈ U, B ⊆ A,

define unified set pair tolerance relation (USPTR) as follows:

USPTR(B)={(x, y)| uB(x, y)=a+(b1+b2+b3)i,a+b1+b2+b3=1,a≥α,bi≤βi}

Here α+β1+β2+β3=1.The united set pair tolerance relation class of x is defined as :

∈
USPTRB(x)={y| y USPTR (B)}.

Definition 2. Given incomplete information system S = (U, A, V), B ⊆ A, X ⊆ U, the

upper and lower approximation of X are defined respectively as follows:

（
USPTR B X ) = {x | USPTRB ( x) ∩ X ≠ Φ}

USPTRB ( X ) = {x | USPTRB ( x) ⊆ X }

When B=A, USPTR B ( X ) and USPTR B ( X ) can be abridged as USPTR ( X ) and

USPTR(X ) .

Definition 3. Given incomplete information system S = (U, A, V),X ⊆ U, the positive

domain, negative domain, boundary region, roughness of X in universe U are defined
as follows:
An United Extended Rough Set Model Based on Developed Set Pair Analysis Method 13

Positive Domain: pos ( X ) = USPTR ( X )

Negative Domain: NEG ( X ) = U − pos ( X )
Boundary Domain : BND( X ) = USPTR( X ) − USPTR( X )
Roughness: ρ ( X ) = 1 − USPTR ( X ) / USPTR ( X )

C Performance Analysis of Extended Rough Set Model Based on United Set Pair
Tolerance Relation

Unified set pair tolerance relation is reflexive,but whether it satisfy symmetry and
transitivity is not a conclusion. When the identity degree and discrepancy degree
thresholds are changed, this extended model will have different performance. We have
three theorems as follows.
∈
Let I= (U,A,V)is an incomplete information system, ∀ x, y U, B ⊆ A,α is threshold
of identity degree a, β1,β2 and β3 are thresholds of discrepancy degree bi respectly.

Theorem 1. If threshold of b1 doesn't equal to threshold of b2, that is β1≠β2,

USPTR(B)doesn't satisfy symmetry.

∈
Proof. Let y USPTRB(x), that is u(x,y)=a+(b1+b2+b3)i, b1≤β1 and b2≤β2. Because
β1≠β2, there may be some y in USPTRB(x) make b1>β2 or b2>β1.
u(y,x)=a+(b2+b1+b3)i, so x ∉ USPTRB(y).

Theorem 2. If β1=0 or β2=0,USPTR(B) satisfies transitivity.

∈
Proof. Firstly prove the situation of β1=0. Let y USPTRB(x) and z ∈USPTR (y),
∈A, if a(x)≠*, then a(y)=a(x). Simul-
B

then (x,y)=a+(b2+b3)i, u(y,z)=a+(b2+b3)i. ∀ a

taneously a(y)≠* and a(z)=a(y). So ∀ a ∈A, if a(x)≠*, we can get a(z)≠*,that is
u(x,z)=a(b2+b3)i, z ∈USPTR (x). The transitivity of USPTR(B)is proven. The situa-
B

tion of β2=0 is similar, here we doesn't give the details.

Theorem 3. If β1=0 and β2=0,USPTR(B) satisfies symmetry and transitivity.

Proof. According to Theorem1 and Theorem2, the proof is easy, so here we doesn't give
unnecessary details.

The performance of united set pair tolerance relation will change with the modification
of identity and discrepancy thresholds, and some representative existed extended rough
set models are its subschema., see Table 1.
14 X. Ji et al.

Table 1. The relationship between united set pair tolerance relation and some existed models

Thresholds Combination Performance Existed Rough Set Models

α+β1+β2+β3=1 reflexivity , symmetry Tolerance Relation

α+β2+β3=1,β1=0 reflexivity , transitivity Similarity Relation

β3=1 or α>0 reflexivity , symmetry Limited Tolerance Relation

α=1 reflexivity , symmetry , transitivity Equivalence Relation

4 Simulation Experiments

In this section, we select three test databases in UCI machine learning repository, Iris
and Zoo and cmc-data. Details see in Tbale2 .We get nine incomplete databases though
getting rid of data by random function, which are I-5%, I-10%, I-30%, Z-5%, Z-10%
,Z-30%,C-5%,C-10% and C-30%.

Table 2. Experiment Database

Database Instance Number Attribute Number

Iris 4 150
Zoo 101 17
cmc-data 1473 10

Let U be test database, E(xi) be the complete equivalent class of xi ,and R(xi) be the
tolerance class of xi under different extended rough set models. We do ten times ran-
dom experiments on every database ,and take their average as final result. The com-
parability( u ) of test result can be computed through formula(1) and formula(2).
|U|
| E(xi ) ∩ R(xi ) |
∑| E(x ) | + | R(x ) | − | E(x ) ∩ R(x ) |
i =1
u= i i i i (3)
|U |

10

∑µ t
(4)
µ= t =1

10

The bigger u , the higher comparability, the better classification effect.

An United Extended Rough Set Model Based on Developed Set Pair Analysis Method 15

Table 3. Comparability( u ) Comparison of Four Extended Rough Set Models

Tolerance Similarity Limited Tolerance United Set Pair

Database
Relation Relation Relation Tolerance Relation
Z-5% 0.821467 0.708342 0.862431 0.894354
Z-10% 0.809237 0.636242 0.811237 0.873016
Z-30% 0.462345 0.613456 0.473458 0.691325
I-5% 0.932402 0.883422 0.943679 0.988792
I-10% 0.899306 0.825033 0.917963 0.984485
I-30% 0.459135 0.524384 0.564023 0.899621
C-5% 0.563526 0.612432 0.642135 0.734956
C-10% 0.431256 0.516243 0.542251 0.660125
C-30% 0.339534 0.504135 0.462131 0.617243

(1) β1=β2

(2) β1≠β2

Fig. 3. The influence of identity and discrepancy to comparability ( u )

16 X. Ji et al.

We do two groups of experiment. Table3 gives experiment result comparison of

several extended rough set models. The u of every extended model decreases as the
degree of loss data increases, whatever in Zoo, Iris or cmc-data. While u decreases
degree of united set pair tolerance relation is lower than other three models. Meanwhile
united set pair tolerance relation has more steady performances on three databases. The
tolerance relation and limited tolerance relation are disagree with cmc-data, similarity
relation is disagree with Zoo. Totally speaking, united set pair tolerance relation has
better classification effects in nine situations of three databases.
According to the threshold determination difficulty, the second experiment try to
narrow down its range. In Figue1, identity degree α value range is[0.3,0.9] and
，
discrepancy degree β is separated into two situations which are β1=β2 and β1≠β2.
Figue1shows that when α is taken small value, like 0.1and 0.3,the classification per-
formance is very bad. When α arrives 0.9,the performance is unstable and decrease
∈
somewhat. When α [0.5,0.7],the new extended model has a steady and good per-
formance. On all conditions, the classification performance of β1=β2 is better than that
of β1≠β2.In other words, the united set pair tolerance relation with symmetry has a
better performance in classification.

5 Conclusions

In this paper, we firstly develop set pair analysis method, give a developed mi-
cro-de-compound method of discrepancy degree, then propose united set pair tolerance
relation and give corresponding extended rough set model. The threshold α, β1 and β2
can adjust according to the incomplete degree of databases. So this model is more
objective, flexible and effective. The experiments on three databases of UCI, which are
Iris, Zoo and cmc-data, can clearly manifest the advantage of new methods in this
paper. But how to fast calculate the identity and discrepancy degree thresholds is still a
problem need to solve.

Acknowledgement

This work is supported by the National Natural Science Foundation of China

(NO.60273043), the Natural Science Foundation of Anhui(NO.050420204), the Edu-
cation Burea Natural Science Foundation of Anhui(NO.2006KJ098B) and High School
Surpass Talent Foundation of Anhui(NO.05025102).

References

1. Pawlak, Z.: Rough set. International Journal of Computer and Information Sciences 11(5),
341–356 (1982)
2. Grzymala Busse, J.W.: On the unknown attribute values in learning from examples. In: Raś,
Z.W., Zemankova, M. (eds.) ISMIS 1991. LNCS (LNAI), vol. 542, pp. 368–377. Springer,
Heidelberg (1991)
An United Extended Rough Set Model Based on Developed Set Pair Analysis Method 17

3. Stefanowski, J., Tsoukias: A incomplete information tables and rough classification. Com-
putational Intelligence 17, 545–566 (2001)
4. Kryszkiewicz, M.: Rough Set Approach to Incomplete Information System. Information
Sciences (S0020-0255) 112(1/4), 39–49 (1998)
5. Wang, G.-y.: The extension of rough set theory in incomplete information system. Com-
puter Research and Development 39(10), 1238–1243 (2002)
6. Sun, C.-m., Liu, D.-y., Sun, S.-y.: Research of rough set method oriented to incomplete
information system. Mini-micro Computer System 10(10), 1869–1873 (2007)
7. Zhao, K.-q.: Set pair and its prilimitary applications. Hangzhou Zhejiang Science Publishers
(2000)
8. Xu, Y., Li, L.-s., Li, X.-j.: Extended rough set model based on set pair power. Journal of
System Simulation 20(6), 1515–1522 (2008)
9. Stefanowski, J., Tsoukias, A.: On the Extension of Rough Sets under Incomplete Informa-
tion. In: Zhong, N., Skowron, A., Ohsuga, S. (eds.) RSFDGrC 1999. LNCS (LNAI),
vol. 1711, pp. 73–82. Springer, Heidelberg (1999)
10. Lei, Z., Lan, S.: Rough Set Model Based on New Set Pair Analysis. Fuzzy Systems and
Mathmatics (S1001-7402) 20(4), 111–116 (2006)
 Learning Rules from Pairwise Comparison Table

Liping An1 and Lingyun Tong2

1
Business School, Nankai University, Tianjin, 300071, China
anliping2000@sina.com
2
School of Management, Hebei University of Technology, Tianjin, 300130, China
tonglingyun2008@sina.com

Abstract. Multicriteria choice problems deal with a selection of a small subset

of best actions from a larger set and multicriteria ranking problems aim to ob-
tain an ordering of actions from a set from the best to the worst. In order to con-
struct a comprehensive decision-rule preference model for choice and ranking,
we propose the concepts of decision matrices and decision functions to generate
the minimal decision rules from a pairwise comparison table. Decision rules de-
rived can be used to obtain a recommendation in multicriteria choice and rank-
ing problems.

Keywords: Multiple criteria decision analysis, rough sets, choice, ranking,

decision rules.

1 Introduction
Multiple Criteria Decision Analysis (MCDA) aims at helping a decision maker (DM)
to prepare and make a decision where more than one point of view has to be consid-
ered. There are three major models used until now in MCDA: (1) the functional
model expressed in terms of a utility function within multiple attribute utility theory
[1]; (2)the relational model expressed in the form of an outranking relation [2] and a
fuzzy relation [3]; (3) the function-free model expressed in terms of symbolic forms,
like “if… then…” decision rules [6] or decision trees, or in a sub-symbolic form, like
artificial neural nets [4].
Both functional and relational models require that the DM gives some preference
information, such as importance weights, substitution ratios and various thresholds
on particular criteria, which is often quite difficult for DMs not acquainted with the
MCDA methodology [4]. According to Slovic [5], people make decisions by search-
ing for rules that provide good justification of their choices. The decision rule ap-
proach [6] follows the paradigm of artificial intelligence and inductive learning.
These decision rules are induced from preference information supplied by the DM in
terms of some decision examples. Therefore, the decision rules are intelligible and
speak the language of the DM [7].
Roy [8] has stated that the objective of an MCDA is to solve one of the following
five typologies of problems: classification, sorting, choice, ranking and description.
Classification concerns an assignment of a set of actions to a set of pre-defined
classes. The actions are described by a set of regular attributes and the classes are not

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 18–27, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Learning Rules from Pairwise Comparison Table 19

necessarily ordered. Sorting regards a set of actions evaluated by criteria, i.e.

attributes with preference-ordered domains. In this problem, the classes are also pref-
erence-ordered. Choice deals with a selection of a small subset of best actions, evalu-
ated by attributes and criteria, from a larger set. Ranking aims to obtain an ordering of
actions, evaluated by attributes and criteria, from a set from the best to the worst.
Description identifies the major distinguishing features of the actions and performs
their description based on these features.
Recently, a complete and well-axiomatized methodology for constructing decision
rule preference models from decision examples based on the rough sets theory [9] has
been proposed. Several attempts have already been made to use this theory to decision
analysis [10]. However, the classical rough set theory is based on the indiscernibility
relation that is not able to deal with preference-ordered attribute domains and decision
classes. Therefore, the use of classical rough sets has been limited to problems of
multiattribute classification and description only. For this reason, Greco et al. [6, 11]
have proposed an extension of the rough sets approach, called Dominance-based
Rough Set Approach (DRSA). DRSA is able to deal with sorting problems involving
both criteria and regular attributes. This innovation is mainly based on substitution of
the indiscernibility relation by a dominance relation in the rough approximation. In
order to construct a comprehensive preference model that could be used to support the
sorting task, Greco et al. [11] distinguish between qualitative attributes and quantita-
tive attributes. The rules are derived from rough approximations of decision classes.
They satisfy conditions of completeness and dominance, and manage with possible
inconsistencies in the set of examples.
In the case of multicriteria choice and ranking problems, the original version of
rough set theory has been extended by Greco et al. [7, 12, 13] in two ways: substitut-
ing the classical indiscernibility relation by a dominance relation; and substituting the
information table by a pairwise comparison table, which permits approximation of a
comprehensive preference relation in multicriteria choice and ranking problems.
In this paper, the decision matrix of a pairwise comparison table is defined to con-
struct the decision functions which are Boolean functions. The set of ‘‘if … then …’’
decision rules are decoded from prime implicants of the Boolean functions.

2 The Pairwise Comparison Table and Multigraded Dominance

In this section, we recall some basic concepts from [7, 12, 13] to be used in this paper.

2.1 The Pairwise Comparison Table

Let C be the set of criteria used for evaluation of actions from A. For any criterion
q∈C, let Tq be a finite set of binary relations defined on A on the basis of the evalua-
tions of actions from A with respect to the considered criterion q, such that for every
(x, y)∈A×A exactly one binary relation t∈Tq is verified.
The preferential information has the form of pairwise comparisons of reference ac-
tions from B⊆A, considered as exemplary decisions. The pairwise comparison table
(PCT) is defined as data table SPCT =(B, C∪{d}, TC∪Td, g), where B⊆B×B is a non-
empty set of exemplary pairwise comparisons of reference actions, TC= ∪ q∈C Tq , d is a
20 L. An and L. Tong

decision corresponding to the comprehensive pairwise comparison, and g:

B×(C∪{d})→TC∪Td is a total function such that g[(x, y), q]∈Tq for every (x, y)∈A×A
and for each q∈C, and g[(x, y), d]∈Td for every (x, y)∈B.
It follows that for any pair of reference actions (x, y)∈B there is verified one and
only one binary relation t∈Td. Thus, Td induces a partition of B. The data table SPCT
can be seen as decision table, since the set of considered criteria C and decision d are
distinguished.

2.2 The Definition of Multigraded Dominance

Let CO be the set of criteria expressing preferences on an ordinal scale, and CN, the set
of criteria expressing preferences on a quantitative scale or a numerical nonquantita-
tive scale, such that CO∪CN=C and CO∩CN=∅. Moreover, for each P⊆C, let PO be the
subset of P composed of criteria expressing preferences on an ordinal scale, i.e.
PO=P∩CO, and PN, the subset of P composed of criteria expressing preferences on a
quantitative scale or a numerical non-quantitative scale, i.e. PN=P∩CN. For each P⊆C,
we have P=PO∪PN and PO∩PN=∅. The following three situations are considered:
(1) P=PN and PO=∅.
The exemplary pairwise comparisons made by the DM can be represented in terms
of graded preference relations Pqh : for each q∈C and for every (x, y)∈A×A, Tq={ Pqh :
h∈Hq}, where Hq is a particular subset of the relative integers and

• x Pqh y, h>0, means that action x is preferred to action y by degree h with respect
to the criterion q,
• x Pqh y, h<0, means that action x is not preferred to action y by degree h with re-
spect to the criterion q,
• x Pq0 y means that action x is similar to action y with respect to the criterion q.

For each q∈C and for every (x, y)∈A×A, [x Pqh y, h>0]⇒[y Pqk x, k≤0] and [x Pqh y,
h<0]⇒[y Pqk x, k≥0].
Given PN⊆C (PN≠∅), (x, y), (w, z)∈A×A, the pair of actions (x, y) is said to domi-
nate (w, z), taking into account the criteria from PN, denoted by (x, y) D P N (w, z), if x
is preferred to y at least as strongly as w is preferred to z with respect to each q∈PN.
Precisely, “at least as strongly as” means “by at least the same degree”, i.e. hq≥kq,
where hq, kq∈Hq, x Pq q y and w Pq q z, for each q∈P.
h k

The upward cumulated preferences, denoted by Pqfh , and downward cumulated

preferences, denoted by Pqph , having the following interpretation:

• x Pqfh y means “x is preferred to y with respect to q by at least degree h”,

• x Pqph y means “x is preferred to y with respect to q by at most degree h”.
 Learning Rules from Pairwise Comparison Table 21

Exact definition of the cumulated preferences, for each (x, y)∈A×A, q∈C and
h∈Hq, is the following:

• x Pqfh y if x Pqk y, where k∈Hq and k≥h,

• x Pqph y if x Pqk y, where k∈Hq and k≤h.

(2) P=PO and PN=∅.

In this paper, we modify the original definition of multigraded dominance with re-
spect to P=PO and PN=∅ [12] as follows:
Given (x, y), (w, z)∈A×A, the pair (x, y) is said to dominate the pair (w, z) with re-
spect to PO if, for each q∈PO, (1) cq(x) f cq(y) and cq(z) f cq(w), denoted by (x,
y) D{1q} (w, z), or (2) cq(x) f cq(w) and cq(z) f cq(y) for cq(x) f cq(y) and cq(w) f cq(z),
denoted by (x, y) D{2q} (w, z).

(3) PN≠∅ and PO≠∅.

Given (x, y), (w, z)∈A×A, the pair (x, y) is said to dominate the pair (w, z) with re-
spect to criteria from P, if (x, y) dominates (w, z) with respect to both PN and PO.

2.3 Lower and Upper Approximations

Given P⊆C and (x, y)∈A×A, the P-dominating set and the P-dominated set are de-
fined, respectively, as:
• a set of pairs of actions dominating (x, y), DP+ (x, y)={(w, z)∈A×A: (w, z)DP(x, y)},
• a set of pairs of actions dominated by (x, y), DP− (x, y)={(w, z)∈A×A: (x, y)DP(w, z)}.

In a PCT, the set Td is composed of two binary relations defined on A:

(1) x outranks y (denoted by xSy or (x, y)∈S), where (x, y)∈B and “x outranks y”
means “x is at least as good as y”,
(2) x does not outrank y (denoted by xSCy or (x, y)∈SC), where (x, y)∈B and
S∪SC=B.
The P-dominating sets and the P-dominated sets defined on B for all given pairs of
reference actions from B are “granules of knowledge” that can be used to express P-
lower and P-upper approximations of comprehensive outranking relations S and Sc,
respectively:
P (S)={(x, y)∈B: DP+ (x, y)⊆S}, P (S)= ∪ DP+ (x, y).
( x , y )∈S
−
P (S )={(x, y)∈B: D (x, y)⊆S }, P (S )= ∪
c
P
c c
DP− (x, y).
( x , y )∈S c

The P-boundaries (P-doubtful regions) of S and Sc are defined as

BnP(S)= P (S)− P (S), BnP(Sc)= P (Sc)− P (Sc).
From the above it follows that BnP(S)=BnP(Sc).
22 L. An and L. Tong

2.4 Decision Rules

Using the approximations of S and Sc based on the dominance relation defined above, it
is possible to induce a generalized description of the available preferential information
in terms of decision rules. The decision rules have in this case the following syntax:
(1) Df -decision rules
IF x Pqf1 h ( q1 ) y and x Pqf2h ( q2 ) y and… x Pqfeh ( qe ) y and cqe +1 (x) f rqe +1 and cqe +1 (x) p sqe +1
and … cq p (x) f rq p and cq p (x) p s q p , THEN xSy,
where P={ql, q2, …, qp}⊆C, PN={ql, q2, …, qe}, PO={qe+l, qe+2, …, qp}, (h(ql), h(q2),
…, h(qp))∈ H q1 × H q2 ×…× H q p and ( rqe +1 , …, rq p ), ( sqe +1 , …, s q p )∈ C qe +1 ×…× C q p .
These rules are supported by pairs of actions from the P-lower approximation of S
only.
(2) Dp -decision rules
IF x Pqp1 h ( q1 ) y and x Pqp2h ( q2 ) y and… x Pqpeh ( qe ) y and cqe +1 (x) p rqe +1 and cqe +1 (x) f sqe +1
and … cq p (x) f rq p and cq p (x) p s q p , THEN xScy,
where P={ql, q2, …, qp}⊆C, PN={ql, q2, …, qe}, PO={qe+l, qe+2, …, qp}, (h(ql), h(q2),
…, h(qp))∈ H q1 × H q2 ×…× H q p and ( rqe +1 , …, rq p ), ( sq e +1 , …,

sq p )∈ Cqe +1 ×…× Cq p .
These rules are supported by pairs of actions from the P-lower approximation of Sc
only.
(3) Df p -decision rules
ph ( qe+1 )
IF x Pqf1 h ( q1 ) y and x Pqf2h ( q2 ) y and… x Pqfeh ( qe ) y and x Pqe+1 y and x Pqpeh+ 2( qe+ 2 ) y and…
ph ( q f )
x Pq f y cq f +1 (x) f rq f +1 and cq f +1 (y) p sq f +1 and … cqg (x) f rq g and cqg (x) p sqg
and cq g +1 (x) p rqg +1 and cq g +1 (y) f s qg +1 and … cq p (x) p rq p and cq p (x) f s q p , THEN
xSy or xScy,
where O′={ql, q2, …, qe}⊆C, O″={qe+l, qe+2, …, qf}⊆C, PN=O′∪O″, O′ and O″ not
necessarily disjoint, PO={qf+l, qf+2, …, qf}, (h(ql), h(q2), …,
h(qf))∈ H q1 × H q2 ×…× H q f and ( rq f +1 , …, rq p ), ( sq f +1 , …, s q p )∈ C q f +1 ×…× C q p .
These rules are supported by pairs of actions from the P-boundary of S and Sc only.
Applying the decision rules induced from a given SPCT, a final recommendation for
choice or ranking can be obtained upon a suitable exploitation of this structure [12].

3 Decision Matrix and Decision Functions of a PCT

In this section, by analogy with the indiscernibility matrix and its Boolean functions
[14] to generate all minimal decision rules in classical rough set theory, we propose
 Learning Rules from Pairwise Comparison Table 23

the concepts of decision matrices and decision functions to generate the minimal
decision rules.

Definition 1. Let SPCT =(B, C∪{d}, TC∪Td, g) be a pairwise comparison table,

h k
B=S∪SC, (x, y)∈S, (w, z)∈SC, pi∈PN, qj∈PO. x Ppi pi y and w Ppi pi z, for each pi∈PN, h pi ,
k pi ∈ H pi . The decision matrix of S in SPCT, denoted by M(S), is defined as:

M(S)=m[(x, y), (w, z)] S × SC

,

where m[(x, y), (w, z)]={pi, qj: (x, y) D{ pi } (w, z), h pi > k pi , (x, y) D{1q j } (w, z) and

cq j (x)≠ cq j (y) or cq j (w)≠ cq j (z), (x, y) D{2q j } (w, z) and cq j (x)≠ cq j (w) or
cq j (y)≠ cq j (z)}.

Definition 2. Let M(S) be the decision matrix of S in SPCT. The decision function of
(x, y) with respect to M(S), is defined as:

fS[(x, y)]= ∧ { (∨ pi* ) ∨ (∨ q *j ) : pi, qj∈m[(x, y), (w, z)] and m[(x, y), (w, z)]≠∅}.
( w, z ) i j

where pi* and q *j are corresponding to the attribute pi and qj, respectively.
The decision function fS[(x, y)] is a Boolean function that expresses how a pair (x,
y)∈S can be discerned from all of the pairs (w, z)∈SC. Turning fS[(x, y)] into disjunc-
tive normal form, the prime implicants of fS[(x, y)] reveal the minimal subsets of P
that are needed to discern the pair (x, y) from the pair in SC and correspond to the
minimal Df -decision rules.
Similarly, we can define the decision matrix of SC, M(SC), and the decision func-
tion with respect to M(SC).
An example. Let us consider the example used in [6]. The students are evaluated
according to the level in Mathes, Phys and Lit. Marks are given on a scale from 0 to
20. Three students presented in Table 1 are considered.

Table 1. Students’ evaluation table

Student Maths Phys Lit Comprehensive valuation

a 18 16 10 13.9
b 10 12 18 13.6
c 14 15 15 14.9

DM’s preference-order of the students is represented in terms of pairwise evalua-

tions in Table 2.
The relation (x, y)D{Maths}(w, z) is true if and only if
(1) cMaths(x) f cMaths(y) and cMaths(z) f cMaths(w), or (2) cMaths(x) f cMaths(w) and
cMaths(z) f cMaths(y), where c q j (x) f cq j (y) and cq j (w) f cq j (z).
24 L. An and L. Tong

Table 2. Pairwise comparison table

Comprehensive
Pair of students Maths Phys Lit
outranking relation
(a, a) 18, 18 16, 16 10, 10 S
(a, b) 18, 10 16, 12 10, 18 S
(a, c) 18, 14 16, 15 10, 15 Sc
(b, a) 10, 18 12, 16 18, 10 Sc
(b, b) 10, 10 12, 12 18, 18 S
(b, c) 10, 14 12, 15 18, 15 Sc
(c, a) 14, 18 15, 16 15, 10 S
(c, b) 14, 10 15, 12 15, 18 S
(c, c) 14, 14 15, 15 15, 15 S

The analogous partial preorders can be induced using dominance relations on Phys
and on Lit. The entities to compute the lower and upper approximations of S are
shown in Table 3.

Table 3. The entities to compute the lower and upper approximations of S

Pair of
D{+Maths} (x, y) D{+Physics } (x, y) D{+Literature} (x, y) D P+ (x, y)
students
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c), (b, (a, a), (b, a), (b, b), (a, a), (b,
(a, a)
(b, b), (c, b), (c, c) b), (c, b), (c, c) (b, c), (c, a), (c, c) b), (c, c)
(a, a), (a, b), (a, c),
(a, b) (a, b) (a, b) (b, a), (b, b), (b, c), (a, b)
(c, a), (c, b), (c, c)
(a, a), (a, b), (a, c),
(a, c) (a, b), (a, c) (a, b), (a, c) (b, a), (b, b), (b, c), (a, c)
(c, a), (c, b), (c, c)
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c),
(b, a) (b, a), (b, b), (b, c), (b, a), (b, b), (b, c), (b, a) (b, a)
(c, a), (c, b), (c, c) (c, a), (c, b), (c, c)
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c), (a, a), (b, a), (b, b), (a, a), (b,
(b, b)
(b, b), (c, b), (c, c) (b, b), (c, b), (c, c) (b, c), (c, a), (c, c) b), (c, c)
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c),
(b, c) (b, b), (b, c), (c, b), (b, b), (b, c), (c, b), (b, a), (b, c) (b, c)
(c, c) (c, c)
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c),
(c, a) (b, b), (c, a), (c, b), (b, b), (c, a), (c, b), (b, a), (c, a) (c, a)
(c, c) (c, c)
(a, a), (b, a), (b, b),
(c, b) (a, b), (c, b) (a, b), (c, b) (b, c), (c, a), (c, b), (c, b)
(c, c)
(a, a), (a, b), (a, c), (b, (a, a), (a, b), (a, c), (a, a), (b, a), (b, b), (a, a), (b,
(c, c)
b), (c, b), (c, c) (b, b), (c, b), (c, c) (b, c), (c, a), (c, c) b), (c, c)

P (S)={(a, a), (a, b), (b, b), (c, a), (c, b), (c, c)},
P (S)={(a, a), (a, b), (b, b), (c, a), (c, b), (c, c)},
BnP(S)=∅.
 Learning Rules from Pairwise Comparison Table 25

The entities to compute the lower and upper approximations of SC are shown in
Table 4.

Table 4. The entities to compute the lower and upper approximations of SC

Pair of
D{−Maths} (x, y) D{−Physics } (x, y) D{−Literature} (x, y) D P− (x, y)
students
(a, a), (b, a), (b, b), (a, a), (b, a), (b, b), (a, a), (a, b), (a, c), (a, a), (b,
(a, a)
(b, c), (c, a), (c, c) (b, c), (c, a), (c, c) (b, b), (c, b), (c, c) b), (c, c)
(a, a), (a, b), (a, c), (a, a), (a, b), (a, c),
(a, b) (b, a), (b, b), (b, c), (b, a), (b, b), (b, c), (a, b) (a, b)
(c, a), (c, b), (c, c) (c, a), (c, b), (c, c)
(a, a), (a, c), (b, a), (a, a), (a, c), (b, a),
(a, c) (b, b), (b, c), (c, a), (b, b), (b, c), (c, a), (a, b), (a, c) (a, c)
(c, c) (c, c)
(a, a), (a, b), (a, c),
(b, a) (b, a) (b, a) (b, a), (b, b), (b, c), (b, a)
(c, a), (c, b), (c, c)
(a, a), (b, a), (b, b), (a, a), (b, a), (b, b), (a, a), (a, b), (a, c), (a, a), (b,
(b, b)
(b, c), (c, c) (b, c), (c, c) (b, b), (c, b), (c, c) b), (c, c)
(a, a), (a, b), (a, c),
(b, c) (b, a), (b, c) (b, a), (b, c) (b, b), (b, c), (c, b), (b, c)
(c, c)
(a, a), (a, b), (a, c),
(c, a) (b, a), (c, a) (b, a), (c, a) (b, b), (c, a), (c, b), (c, a)
(c, c)
(b, a), (b, b), (b, c), (b, a), (b, b), (b, c),
(c, b) (a, b), (c, b) (c, b)
(c, a), (c, b), (c, c) (c, a), (c, b), (c, c)
(a, a), (b, a), (b, b), (a, a), (b, a), (b, b), (a, a), (a, b), (a, c), (a, a), (b,
(c, c)
(b, c), (c, a), (c, c) (b, c), (c, a), (c, c) (b, b), (c, b), (c, c) b), (c, c)

P (Sc)={(a, c), (b, a), (b, c)}, P (Sc)={(a, c), (b, a), (b, c)}, BnP(S)=∅.
The decision matrix of S in SPCT,
( a, c ) (b, a) (b, c)
( a, a ) Lit Maths, Phys Maths, Phys
( a, b) Maths, Phys Maths, Phys Maths, Phys
M(S)= (b, b) Lit Maths, Phys Maths, Phys
( c, a ) Lit Maths, Phys ∅
( c, b ) ∅ Maths, Phys Maths, Phys
(c , c ) Lit Maths, Phys Maths, Phys

The decision functions of (x, y) with respect to M(S):

fS[(a, a)]=Lit∧(Maths∨Phys)∧(Maths∨Phys)=(Lit∧Maths)∨(Lit∧Phys), similarly,

fS[(a, b)]=Maths∨Phys,
fS[(b, b)]=(Lit∧Maths)∨(Lit∧Phys),
fS[(c, a)]=(Lit∧Maths)∨(Lit∧Phys),
26 L. An and L. Tong

fS[(c, b)]=Maths∨Phys,
fS[(c, c)]=(Lit∧Maths)∨(Lit∧Phys).
Then, the following minimal Df -decision rules can be obtained:
IF (cMaths(x) f 18 and cMaths(y) p 18) and (cLit(x) f 10 and cLit(y) p 10) THEN xSy,
(a, a),
IF (cPhys(x) f 16 and cPhys(y) p 16) and (cLit(x) f 10 and cLit(y) p 10) THEN xSy, (a, a),
IF cMaths(x) f 18 and cMaths(y) p 10 THEN xSy, (a, b),
IF cPhys(x) f 16 and cPhys(y) p 12 THEN xSy, (a, b),
IF (cMaths(x) f 10 and cMaths(y) p 10) and (cLit(x) f 18 and cLit(y) p 18) THEN xSy,
(b, b),
IF (cPhys(x) f 12 and cPhys(y) p 12) and (cLit(x) f 18 and cLit(y) p 18) THEN xSy, (b, b),
IF (cMaths(x) f 14 and cMaths(y) p 18) and (cLit(x) f 15 and cLit(y) p 10) THEN xSy,
(c, a),
IF (cPhys(x) f 15 and cPhys(y) p 16) and (cLit(x) f 15 and cLit(y) p 10) THEN xSy, (c, a),
IF cMaths(x) f 14 and cMaths(y) p 10 THEN xSy, (c, b),
IF cPhys(x) f 15 and cPhys(y) p 12 THEN xSy, (c, b),
IF (cMaths(x) f 14 and cMaths(y) p 14) and (cLit(x) f 15 and cLit(y) p 15) THEN xSy,
(c, c),
IF (cPhys(x) f 15 and cPhys(y) p 15) and (cLit(x) f 15 and cLit(y) p 15) THEN xSy, (c, c).
Similarly, we can obtain the Dp -decision rules as follows.
IF (cMaths(x) p 18 and cMaths(y) f 14) and (cLit(x) p 10 and cLit(y) f 15) THEN xSCy,
(a, c),
IF (cPhys(x) p 16 and cPhys(y) f 15) and (cLit(x) p 10 and cLit(y) f 15) THEN xSCy, (a, c),
IF cMaths(x) p 10 and cMaths(y) f 18 THEN xSCy, (b, a),
IF cPhys(x) p 18 and cPhys(y) f 10 THEN xSCy, (b, a),
IF cMaths(x) p 10 and cMaths(y) f 14 THEN xSCy, (c, b),
IF cPhys(x) p 18 and cPhys(y) f 15 THEN xSCy, (c, b).

4 Conclusions
Learning decision rules from preference-ordered data differs from usual machine learn-
ing, since the former involves preference orders in domains of attributes and in the set of
decision classes. This requires that a knowledge discovery method applied to prefer-
ence-ordered data respects the dominance principle, which is addressed in the Domi-
nance-Based Rough Set Approach. This approach enables us to apply a rough set
approach to multicriteria choice and ranking. In this paper, we propose the concepts of
decision matrices and decision functions to generate the minimal decision rules from a
pairwise comparison table. Then, the decision rules can be used to obtain a recommen-
dation in multicriteria choice and ranking problems. We will further present some
extensions of the approach that make it a useful tool for other practical applications.
 Learning Rules from Pairwise Comparison Table 27

Acknowledgement. It is a project supported by National Natural Science Foundation

of China (No. 70601013).

References
1. Keeney, R.L., Raiffa, H.: Decision with Multiple Objectives-Preferences and Value Trade-
offs. Wiley, New York (1976)
2. Roy, B.: The Outranking Approach and the Foundation of ELECTRE Methods. Theory
and Decision 31(1), 49–73 (1991)
3. Fodor, J., Roubens, M.: Fuzzy Preference Modelling and Multicriteria Decision Support.
Kluwer, Dordrecht (1994)
4. Zopounidis, C., Doumpos, M.: Multicriteria Classification and Sorting Methods: A litera-
ture Review. European Journal of Operational Research 138(2), 229–246 (2002)
5. Slovic, P.: Choice between Equally-valued Alternatives. Journal of Experimental Psychol-
ogy: Human Perception Performance 1, 280–287 (1975)
6. Greco, S., Matarazzo, B., Slowinski, R.: Rough Set Theory for Multicriteria Decision
Analysis. European Journal of Operational Research 129(1), 1–47 (2001)
7. Fortemps, P., Greco, S., Slowinski, R.: Multicriteria Decision Support Using Rules That
Represent Rough-Graded Preference Relations. European Journal of Operational Re-
search 188(1), 206–223 (2008)
8. Roy, B.: Méthodologie multicritère d’aide à la décision. Economica, Paris (1985)
9. Pawlak, Z.: Rough Sets. International Journal of Computer and Information
Sciences 11(5), 341–356 (1982)
10. Pawlak, Z., Slowinski, R.: Rough Set Approach to Multi-attribute Decision Analysis.
European Journal of Operational Research 72(3), 443–459 (1994)
11. Greco, S., Matarazzo, B., Slowinski, R.: Rough Sets Methodology for Sorting Problems in
Presence of Multiple Attributes and Criteria. European Journal of Operational Re-
search 138(2), 247–259 (2002)
12. Greco, S., Matarazzo, B., Slowinski, R.: Extension of the Rough Set Approach to Multicri-
teria Decision Support. INFOR 38(3), 161–193 (2000)
13. Greco, S., Matarazzo, B., Slowinski, R.: Rough Approximation of a Preference Relation by
Dominance Relations. European Journal of Operational Research 117(1), 63–83 (1999)
14. Skowron, A., Rauszer, C.: The Discernibility Matrices and Functions in Information Table.
In: Intelligent Decision Support: Handbook of Applications and Advances of the Rough
Set Theory, pp. 331–362. Kluwer Academic Publishers, Dordrecht (1991)
 A Novel Hybrid Particle Swarm Optimization
for Multi-Objective Problems

Siwei Jiang and Zhihua Cai

School of Computer Science, China University of Geosciences, Wuhan 430074, China

amosonic@gmail.com,zhcai@cug.edu.cn

Abstract. To solve the multi-objective problems, a novel hybrid parti-

cle swarm optimization algorithm is proposed(called HPSODE). The new
algorithm includes three major improvement: (I)Population initialization
is constructed by statistical method Uniform Design, (II)Regeneration
method has two phases: the first phase is particles updated by adap-
tive PSO model with constriction factor χ, the second phase is Differ-
ential Evolution operator with archive, (III)A new accept rule called
Distance/volume fitness is designed to update archive. Experiment on
ZDTx and DTLZx problems by jMetal 2.1, the results show that the new
hybrid algorithm significant outperforms OMOPSO, SMPSO in terms
of additive Epsilon, HyperVolume, Genetic Distance, Inverted Genetic
Distance.

Keywords: Multi-Objective Optimization, Uniform Design, Particle

Swam OPtimization, Differential Evolution, Minimum Reduce Hyper-
volume, Spread.

1 Introduction
Multi-objective Evolutionary Algorithms(MOEAs) are powerful tools to solve
multi-Objective optimize problems(MOPs), it gains popularity in recent years
[1]. Recently, some elitism algorithms are proposed as: NSGA-II [2], SPEA2 [3],
GDE3 [4], MOPSO [5, 6, 7].
NSGA-II adopt a fast non-dominated sorting approach to reduce computer
burden, it uses Ranking and Crowding Distance to choose the candidate solutions
[2]. SPEA2 proposes a fitness assignment with cluster technique, it designs a
truncation operator based on the nearest Neighbor Density Estimation metric [3].
GDE3 is a developed version of Differential Evolution, which is suited for global
optimization with an arbitrary number of objectives and constraints [4].
MOPSO is proposed by Coello et al, it adopts swarm intelligence to optimize
MOPs, and it uses the Pareto-optimal set to guide the particles flight [5]. Sierra
adopt Crowding Distance to filter the leaders, the different mutation methods
are acted on divisions particles [6]; Mostaghim introduces a new Sigam-method
to find local best information to guide particle [7].


The Project was supported by the Research Foundation for Outstanding Young
Teachers, China University of Geosciences(Wuhan)( No:CUGQNL0911).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 28–37, 2009.


c Springer-Verlag Berlin Heidelberg 2009
 A Novel Hybrid Particle Swarm Optimization 29

Population initialization is an important part in EAs, but it has been long

ignored. Orthogonal design and Uniform Design belong to a sophisticated branch
of statistics [9,10]. Zeng and Cai adopt orthogonal design to solve MOPs, results
show that it is more powerful than the Random Design [11, 12]. Leung utilizes
the uniform design to initial population and designs a new crossover operator,
which can find the Pareto-optimal solutions scattered uniformly [13].
Zhang proposes a hybrid PSO with differential evolution operator to solve the
single objective problem, which provide the bell-shaped mutation to guarantee
the evolutionary population diversity [8].
Interested in the hybrid PSO with differential evolutionary operator, we pro-
pose a hybrid particle swarm optimization called HPSODE: the first population
is constructed by uniform design; offspring regeneration method has two phases,
in first phase the particles are updated according to its own experience(pbest )
and social information(gbest ) with constriction factor χ, in second phase the
archive is operated by Differential Evolution; a new accept rule is designed as
Distance/Volume function to update the archive.
The rest of the paper is organized as follows. In section 2, we briefly introduce
the uniform design. In section 3, we describe the two phases of regeneration
method. In section 4, we design a new archive update rule. In section 5, Ex-
periment on bio-objective and tri-objective has show that the new algorithm
is powerful than MOPSO, SMPSO in terms of additive Epsilon, hypervolume,
Genetic Distance, Inverted Genetic Distance. In section 6, we make conclusions
and discuss the future research on MOEAs.

2 Population Initialization: Uniform Design

Population initialization has long been ignored in MOEAs, but it is a very im-
portant component for MOEAs. Orthogonal Design and Uniform Design are
experimental design method, which belong to a sophisticated branch of statis-
tics. Both of them can get better distribute population in feasible searching space
than Random Design, then the statistical population provide more information
to generate next offspring.
In this section, we briefly describe an experimental design method called
uniform design. We define the uniform array as UR (C), where Q is the level,
it’s primer, R, C represent the row and column of uniform array, they must be
satisfid with:

R=Q>n
(1)
C=n
where n is the number of variables. When select a proper parameters of Q, σ
form table 1, uniform array can be created by Equation 2

Ui,j = (i ∗ σ j−1 mod Q) + 1 (2)

For one decision variable Xj with the boundary [lj , uj ], then the quantize techni-
cal divide the domain into Q levels αj1 , αj2 , · · · , αjQ , where the design parameter
30 S. Jiang and Z. Cai

Table 1. Parameter σ for different number of factors and levels

number of levels of per factor number of factors σ

5 2-4 2
7 2-6 3
11 2-10 7
13 2 5
3 4
4-12 6
17 2-16 10
19 2-3 8
4-18 14
23 2,13-14,20-22 7
8-12 15
3-7,15-19 17
29 2 12
3 9
4-7 16
8-12,16-24 8
13-15 14
25-28 18
31 2,5-12,20-30 12
3-4,13-19 22

Q is primer and αi is given by

u j − lj
αjk = lj + (k − 1)( ), 1 ≤ k ≤ Q (3)
Q−1
In other words, the domain [lj , uj ] is quantized Q − 1 fractions, and any two
successive levels are same as each other.

3 Offspring Regeneration Method

In this paper, we propose a hybrid Particle Swarm Optimization algorithm called
HPSODE, which has three populations: the first population is the original par-
ticles, the second population is local best particles which store only one evolu-
tionary step according to original particles, the third population is global best
particle which is utilized as archive. The regeneration method has two phases,
original particle with adaptive PSO update, archive with differential evolution
operator.
In first phase, the particle’s new position is updated by its own experience
(pbest ) and social information(gbest ) with constriction factor χ:


Vid (t + 1) = χ[ωVid (t) + C1 ϕ1 (pbid (t) − Xid (t)) + C2 ϕ2 (gbid (t) − Xid (t))]
Xid (t + 1) = Xid (t) + Vid (t + 1)
(4)
where χ = √2 with ϕ = C1 + C2 , when ϕ ≤ 4, χ = 1.0. And ω = 0.1,
2−ϕ− ϕ2 −4ϕ
C1 , C2 ∈ rand[1.5, 2.5],ϕ1 , ϕ2 ∈ rand[0.0, 1.0].
The global best particle is binary tournament choose from archive which has
larger crowding distance, it forces the original particles to explore the sparse
 A Novel Hybrid Particle Swarm Optimization 31

space. If the new particle position is non-dominated by original particle, replace

the original particle and local best particle, Xi (t) = Xi (t + 1), pbi(t) = Xi (t + 1),
then add the new particle to archive; otherwise discard the new particle and
unchange the local best.
In second phase, the third population is update by differential evolution op-
erator: ⎧
⎪ Xid (t + 1) = Xr1 d (t) + F (Xr2 d (t) − Xr3 d (t))
⎨
if rndd (0, 1) < CR||d == drnd (5)
⎪
⎩
Xid (t + 1) = Xr1 d (t) otherwise
where d is the dimension of solution, the three rand solutions are choose from
archive and r1 = r2 = r3 = i. where F = 0.5, CR = 0.1. If the new solution is
non-dominated by the archive’s solutions, then add it to archive.
When the number of archive is small, the guide information is rare, it will
lead the original particle to assemble to few best positions, and then the final
Pareto-optimal set will get worse diversity. Differential evolution operation is
simple to implement and it has powerful capacity to scatter the feasible space.
The DE operation on archive is useful to enhance the diversity of archive, and
provide good guide information to the original particle.

4 Archive Update Rule

When MOEAs get a set of equal good solutions full to the setting size(usually
archiveSize = 100), an accept rule must be designed to decide which one should
be cut off from archive. It’s a critical issue in MOEAs which directly influence
the quality of finally optimal set in convergence and spread metric. Some popular
accept rules have been presented: NSGA-II adopts the Ranking and Crowding
Distance metric, SPEA2 uses the nearest Neighbor Density Estimation metric.
In this paper, we design a new accept rule called Minimum Reduce Hypervol-
ume. Hypervolume is a quality indicator proposed by Zitzler et al, it is adopted
in jMetal 2.1 [14]. Hypervolume calculates the volume covered by members of a
non-dominated set of solutions (the region enclosed into the discontinuous line
respect the worst point W in the figure 1 is ADBECW in dashed line).
If two solutions D and E are non-dominated each other, NSGA-II choose the
solution D remained in archive if CD(D) > CD(E), it maintains the spread
along the Pareto-front.


CD(D) = AD + D B
(6)
CD(E) = BE  + E  C

If one solution is deleted, it will lead to a hypervolume decrease, because the

higher hypervolume means the better quality of optimal set, we will delete the
solution which reduce the hypervolume minimum. MRV chooses the solution E
remained in archive if hv(D) < hv(E)(then the hypervolume is ABECW rather
32 S. Jiang and Z. Cai

D1
D’
CD(D)=AD’+D’B hv(D)=DD1*DD2

D D D2

E1
E’
CD(E)=BE’+E’C hv(E)=EE1*EE2

E E2
E

Fig. 1. The comparison of Crowding Distance and MRV is described in the left and
right figure

than ABDCW ), it maintains the hyervolume along the Pareto-front.

hv(D) = DD1 ∗ DD2
(7)
hv(E) = EE1 ∗ EE2
Crowding Distance maintains the spread and expands the solution to feasible
searching place, and Minimum Reduce Hypervolume maintains the hypevolume
and forces the solution near to Pareto-front. We combine the two properties, a
new fitness assignment for solution s is designed as follows(called
Distance/Volume fitness):
⎧
⎪
⎨ DV (s) = CD(s) + scale ∗ hv(s)
n
CD(si ) (8)
⎪
⎩ scale = i=1
n
i=1 hv(si )

The factor scale is designed to equal the influence of crowding distance and
MRV, but if it is too large, we set it is to 1000 when scale > 1000.

5 Experiment Results
Experiment is based on jMetal 2.1 [14], which is a Java-based framework aimed
at facilitating the development of metaheuristics for solving MOPs, it provides
large block reusing code and fair comparison for different MOEAs. The paper
selects the algorithm OMOPSO [6], SMPSO [14] as the compare objectives.
Each algorithm independent runs for 100 times and maximum evolution times
is 25, 000. The test problems are choose from ZDTx, DTLZx problem family.
The performance metrics have five categories:
1
Unary additive epsilon indicator(I+ ). The metric indicator of an approx-
1
imation set A (I+ (A)) gives the minimum factor  by which each point in the
 A Novel Hybrid Particle Swarm Optimization 33

real front can be added such that the resulting transformed approximation set
is dominated by A:
1
I+ (A) = inf∈R{∀z 2 ∈ R\∃z 1 ∈ A : zi2 ≤ zi1 + ∀i} (9)

Hypervolume. This quality indicator calculates the volume (in the objective
space) covered by members of a nondominated set of solutions with a reference
point.The hypervolume (HV) is calculated by:
|Q|

HV = volume( vi ) (10)
i=1

Inverted Generational Distance. The metric is to measure how far the el-
ements are in the Pareto-optimal set from those in the set of non-dominated
vectors found. It is defined as:

|N | 2
i=1 di
IGD = (11)
N

Generational Distance. The metric is to measure how far the elements are in
the set of non-dominated vectors found from those in the Pareto-optimal set. It
is defined as: 
|n| 2
i=1 di
GD = (12)
n
Spread. The Spread indicator is a diversity metric that measures the extent of
spread achieved among the obtained solutions. This metric is defined as:
n−1
df + dl + i=1 |di − d|
∆= (13)
df + dl + (n − 1)d
1
The higher Hypervolume and lower I+ ,GD, IGD, Spread mean the better al-
gorithm. The results are compared with median and interquartile range(IQR)
which measures of location (or central tendency) and statistical dispersion, re-
spectively, the best result is grey background.
From table 2, in term of median and IQR for additive epsilon metric, HPSODE
get best results in all of the 12 MOPs.
From table 3, in term of median and IQR for HyperVolume metric, HPSODE
get best results in all of the 12 MOPs.
From table 4, in term of median and IQR for Genetic Distance metric, HP-
SODE get best results in 11 MOPs, only worse in DTLZ6. OMOPSO get best
results only in 1 MOPs: DTLZ6.
From table 5, in term of median and IQR for Inverted Genetic Distance metric,
HPSODE get best results in 10 MOPs, only worse in ZDT2, DTLZ6. OMOPSO
get best results only in 1 MOPs: ZDT2. SMOPSO get best results only in 1
MOPs: DTLZ6.
34 S. Jiang and Z. Cai

1
Table 2. Unary additive epsilon indicator(I+ ) Median and IQR

OMOPSO SMPSO HPSODE

ZDT1 6.15e − 033.7e−04 5.63e − 033.4e−04 5.33e − 032.0e−04
ZDT2 5.93e − 034.7e−04 5.49e − 033.0e−04 5.27e − 032.2e−04
ZDT3 6.87e − 032.1e−03 5.85e − 031.0e−03 4.56e − 036.4e−04
ZDT4 5.39e + 004.7e+00 6.43e − 035.7e−04 5.40e − 032.2e−04
ZDT6 4.92e − 035.9e−04 4.72e − 033.9e−04 4.45e − 033.1e−04
DTLZ1 1.09e + 013.2e+00 5.65e − 027.8e−03 4.68e − 026.5e−03
DTLZ2 1.29e − 012.0e−02 1.40e − 011.9e−02 1.22e − 012.1e−02
DTLZ3 1.03e + 023.6e+01 1.61e − 011.3e−01 1.19e − 011.9e−02
DTLZ4 1.12e − 012.3e−02 1.20e − 013.8e−02 1.03e − 012.2e−02
DTLZ5 5.14e − 036.6e−04 4.87e − 035.5e−04 4.36e − 032.8e−04
DTLZ6 4.35e − 035.2e−04 4.35e − 033.9e−04 4.26e − 033.9e−04
DTLZ7 1.95e − 017.6e−02 1.71e − 016.9e−02 1.25e − 013.6e−02

Table 3. Hypervolume Median and IQR

OMOPSO SMPSO HPSODE

ZDT1 6.611e − 014.4e−04 6.618e − 011.2e−04 6.621e − 011.7e−05
ZDT2 3.280e − 014.0e−04 3.285e − 011.1e−04 3.288e − 011.7e−05
ZDT3 5.140e − 011.1e−03 5.154e − 014.2e−04 5.160e − 011.2e−05
ZDT4 0.000e + 000.0e+00 6.613e − 012.3e−04 6.621e − 012.3e−05
ZDT6 4.013e − 011.1e−04 4.012e − 011.1e−04 4.015e − 011.5e−05
DTLZ1 0.000e + 000.0e+00 7.372e − 019.0e−03 7.688e − 017.5e−03
DTLZ2 3.537e − 017.1e−03 3.463e − 011.0e−02 3.821e − 015.1e−03
DTLZ3 0.000e + 000.0e+00 3.406e − 014.3e−02 3.864e − 015.8e−03
DTLZ4 3.591e − 019.7e−03 3.577e − 011.6e−02 3.839e − 016.3e−03
DTLZ5 9.344e − 021.6e−04 9.366e − 021.6e−04 9.407e − 024.4e−05
DTLZ6 9.493e − 025.5e−05 9.488e − 026.4e−05 9.496e − 024.1e−05
DTLZ7 2.610e − 019.9e−03 2.746e − 017.4e−03 2.916e − 013.7e−03

From table 6, in term of median and IQR for Spread metric, HPSODE get
best results in all of the 6 MOPs. OMOPSO get best results only in 2 MOPs.
SMOPSO get best results only in 4 MOPs.
The comparison for three algorithms in five categories: epsilon indicator, hy-
pervolume, Generation Distance, Inverted genetic Distance, Spread, it shows
that the new algorithm is more efficient to solve the MOPs. Now, we summarize
the highlight as follows:

1. HPSODE adopt the statistical method Uniform Design to construct the first
population, which can get well distributed solutions in feasible space.
2. HPSODE combine the PSO and Differential Evolution operation to gener-
ate next population. DE operation enhances the diversity for global guide
population.
 A Novel Hybrid Particle Swarm Optimization 35

Table 4. Generational Distance Median and IQR

OMOPSO SMPSO HPSODE

ZDT1 1.43e − 043.8e−05 1.24e − 044.9e−05 7.41e − 053.5e−05
ZDT2 7.68e − 052.6e−05 5.27e − 055.4e−06 4.66e − 052.8e−06
ZDT3 2.25e − 043.3e−05 1.88e − 042.3e−05 1.76e − 041.4e−05
ZDT4 6.57e − 015.8e−01 1.35e − 044.6e−05 7.67e − 054.2e−05
ZDT6 2.19e − 024.4e−02 2.18e − 034.3e−02 5.37e − 042.0e−05
DTLZ1 1.00e + 011.3e+00 5.06e − 031.4e−03 7.00e − 041.9e−04
DTLZ2 3.32e − 034.1e−04 3.94e − 039.2e−04 1.25e − 034.6e−04
DTLZ3 3.51e + 016.0e+00 5.29e − 035.7e−02 1.17e − 031.8e−04
DTLZ4 5.83e − 035.0e−04 5.98e − 036.4e−04 4.95e − 033.9e−04
DTLZ5 3.00e − 044.3e−05 2.85e − 045.6e−05 2.50e − 043.3e−05
DTLZ6 5.62e − 043.0e−05 5.69e − 042.6e−05 5.67e − 042.6e−05
DTLZ7 5.26e − 031.2e−03 4.84e − 031.3e−03 3.57e − 031.1e−03

Table 5. Inverted Generational Distance Median and IQR

OMOPSO SMPSO HPSODE

ZDT1 1.37e − 042.3e−06 1.35e − 041.0e−06 1.34e − 041.2e−06
ZDT2 1.43e − 043.2e−06 1.40e − 041.9e−06 1.44e − 042.9e−06
ZDT3 2.14e − 041.4e−05 2.00e − 047.2e−06 1.90e − 042.7e−06
ZDT4 1.61e − 011.5e−01 1.38e − 041.4e−06 1.34e − 041.5e−06
ZDT6 1.20e − 047.9e−06 1.18e − 047.9e−06 1.10e − 047.0e−06
DTLZ1 2.71e − 011.0e−01 6.38e − 044.5e−05 5.27e − 043.5e−05
DTLZ2 7.58e − 044.6e−05 8.03e − 045.1e−05 7.04e − 044.1e−05
DTLZ3 2.03e + 006.7e−01 1.43e − 037.8e−04 1.11e − 036.8e−05
DTLZ4 1.23e − 031.6e−04 1.17e − 032.1e−04 1.17e − 031.5e−04
DTLZ5 1.49e − 055.0e−07 1.44e − 056.8e−07 1.41e − 053.8e−07
DTLZ6 3.38e − 051.1e−06 3.43e − 051.3e−06 3.40e − 051.4e−06
DTLZ7 2.66e − 033.1e−04 2.72e − 034.8e−04 2.43e − 033.0e−04

Table 6. Spread Median and IQR

OMOPSO SMPSO HPSODE

ZDT1 8.29e − 021.5e−02 7.90e − 021.3e−02 9.49e − 021.6e−02
ZDT2 7.99e − 021.5e−02 7.21e − 021.8e−02 9.48e − 021.4e−02
ZDT3 7.13e − 011.2e−02 7.11e − 019.8e−03 7.05e − 013.8e−03
ZDT4 8.89e − 018.8e−02 9.84e − 021.6e−02 1.01e − 012.1e−02
ZDT6 1.07e + 009.9e−01 3.17e − 011.2e+00 6.76e − 021.2e−02
DTLZ1 6.90e − 011.0e−01 6.78e − 014.2e−02 7.28e − 015.5e−02
DTLZ2 6.22e − 014.9e−02 6.38e − 015.1e−02 6.32e − 015.7e−02
DTLZ3 7.28e − 011.4e−01 6.65e − 012.0e−01 6.43e − 015.5e−02
DTLZ4 6.36e − 016.9e−02 6.59e − 011.3e−01 6.41e − 015.3e−02
DTLZ5 2.02e − 016.8e−02 1.70e − 018.5e−02 1.24e − 012.3e−02
DTLZ6 1.24e − 014.0e−02 1.39e − 013.9e−02 1.19e − 012.4e−02
DTLZ7 7.07e − 015.6e−02 7.08e − 017.0e−02 6.99e − 015.8e−02
36 S. Jiang and Z. Cai

3. HPSODE design a new accept rule Distance/volume fitness, which expands

the solution to feasible searching place and forces the solution near to Pareto-
front.
4. HPSODE significant outperforms OMOPSO, SMPSO in terms of additive
Epsilon, HyperVolume, Genetic Distance, Inverted Genetic Distance.
5. HPSODE get better results than OMOPSO, SMPSO in term of spread
metric.

6 Conclusion and Future Reasearch

Population initialization, regeneration method, archive update rule are three
critical issues in MOEAs. In this paper, we propose a hybrid PSO algorithm
called HPSODE which include uniform design initialization, two phase regener-
ation with DE operator, new accept rule Distance/Volume fitness. Experimental
results show that HPSODE can get higher Hypervolume, lower additive epsilon,
GD, IGD and competitive spread. Our future research will focus on enhance the
Quantum PSO algorithm to solve the MOPs.

References
1. Coello, C.A.C.: Evolutionary multi-objective optimization: A historical view of the
Field. IEEE Computational Intelligence Magazine 1(1), 28–36 (2006)
2. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multiobjec-
tive genetic algorithm: NSGA - II. IEEE Transactions on Evolutionary Computa-
tion 6(2), 182–197 (2002)
3. Zitzler, E., Laumanns, M., Thiele, L.: SPEA2: Improving the strength Pareto evo-
lutionary algorithm, Technical Report 103, Computer Engineering and Networks
Laboratory (2001)
4. Kukkonen, S., Lampinen, J.: GDE3: The third evolution step of generalized dif-
ferential evolution. In: Proceedings of the 2005 IEEE Congress on Evolutionary
Computation (1), pp. 443–450 (2005)
5. Coello Coello, C.A., Toscano Pulido, G., Salazar Lechuga, M.: Handling Multiple
Objectives With Particle Swarm Optimization. IEEE Transactions on Evolutionary
Computation 8, 256–279 (2004)
6. Sierra, M.R., Coello, C.A.C.: Improving PSO-based multi-objective optimization
using crowding, mutation and -dominance. In: Coello Coello, C.A., Hernández
Aguirre, A., Zitzler, E. (eds.) EMO 2005. LNCS, vol. 3410, pp. 505–519. Springer,
Heidelberg (2005)
7. Mostaghim, S., Teich, J.: Strategies for Finding Good Local Guides in Multi-
objective Particle Swarm Optimization (MOPSO). In: 2003 IEEE Swarm Intel-
ligence Symposium Proceedings, Indianapolis, Indiana, USA, pp. 26–33. IEEE
Service Center (2003)
8. Zhang, W.J., Xie, X.F.: DEPSO: Hybrid particle swarm with differential evolution
operator. In: IEEE International Conference on Systems Man and Cybernetics, pp.
3816–3821 (2003)
9. Fang, K.T., Ma, C.X.: Orthogonal and uniform design. Science Press (2001)
(in Chinese)
 A Novel Hybrid Particle Swarm Optimization 37

10. Leung, Y.W., Wang, Y.: An orthogonal genetic algorithm with quantization for
global numerical optimization. IEEE Transactions on Evolutionary Computa-
tion 5(1), 41–53 (2001)
11. Zeng, S.Y., Kang, L.S., Ding, L.X.: An orthogonal multiobjective evolutionary
algorithm for multi-objective optimization problems with constraints. Evolutionary
Computation 12, 77–98 (2004)
12. Cai, Z.H., Gong, W.Y., Huang, Y.Q.: A novel differential evolution algorithm based
on -domination and orthogonal design method for multiobjective optimization.
In: Obayashi, S., Deb, K., Poloni, C., Hiroyasu, T., Murata, T. (eds.) EMO 2007.
LNCS, vol. 4403, pp. 286–301. Springer, Heidelberg (2007)
13. Leung, Y.-W., Wang, Y.: Multiobjective programming using uniform design and
genetic algorithm. IEEE Transactions on Systems, Man, and Cybernetics, Part
C 30(3), 293 (2000)
14. Durillo, J.J., Nebro, A.J., Luna, F., Dorronsoro, B., Alba, E.: jMetal: A
Java Framework for Developing Multi-Objective Optimization Metaheuristics,
Departamento de Lenguajes y Ciencias de la Computación, University of
Málaga, E.T.S.I. Informática, Campus de Teatinos, ITI-2006-10 (December 2006),
http://jmetal.sourceforge.net
 Research on Constrained Layout Optimization Problem
Using Multi-adaptive Strategies Particle Swarm
Optimizer∗

Kaiyou Lei

Faculty of Computer & Information Science, Southwest University, Chongqing, 400715, China
lky@swu.edu.cn

Abstract. The complex layout optimization problems with behavioral con-

straints belong to NP-hard problem in math. Due to its complexity, the general
particle swarm optimization algorithm converges slowly and easily to local op-
tima. Taking the layout problem of satellite cabins as background, a novel adap-
tive particle swarm optimizer based on multi-modified strategies is proposed in
the paper, which can not only escape from the attraction of local optima of the
later phase to heighten particle diversity, and avoid premature problem, but also
maintain the characteristic of fast speed search in the early convergence phase
to get global optimum, thus, the algorithm has a better search performance to
deal with the constrained layout optimization problem. The proposed algorithm
is tested and compared it with other published methods on constrained layout
examples, demonstrated that the revised algorithm is feasible and efficient.

Keywords: particle swarm optimization, premature problem, constrained layout.

1 Introduction
The classical layout problem is generally divided in to two types: Packing problem
and Cutting problem. The main problem is to increase the space utility ratio as much
as possible under the condition of non-overlapping between piece (object) and piece
(object) and between piece (object) and container. They are called as layout problems
without behavioral constraints (unconstrained layout). In recent years, another more
complex layout problem is attracting a lot of attention, such as the layout design of
engineering machine, spacecraft, ship, etc. Solving these kinds of problems need to
consider some additional behavioral constraints, for instance, inertia, equilibrium,
stability, vibration, etc. They are called as layout problem with behavioral constraints
(constrained layout). Constrained layout belong to NP-hard problem, its optimization
is hand highly [1].
As a newly developed population-based computational intelligence algorithm, Par-
ticle Swarm Optimization (PSO) was originated as a simulation of simplified social
model of birds in a flock [2]. The PSO algorithm is easy to implement and has been
proven very competitive in large variety of global optimization problems and applica-
tion areas compared to conventional methods and other meta-heuristics [3].

∗
The work is supported by Key Project of Chinese Ministry of Education (104262).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 38–47, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Research on Constrained Layout Optimization Problem 39

Since its introduction, numerous variations of the basic PSO algorithm have been
developed in the literature to avoid the premature problem and speed up the conver-
gence process, which are the most important two topics in the research of stochastic
search methods. To make search more effective, there are many approaches suggested
by researchers to solve the problems, such as variety mutation and select a single
inertia weight value methods, etc, but these methods have some weakness in common,
they usually can not give attention to both global search and local search, preferably,
so to trap local optima, especially in complex constrained layout problems [4].
In this paper, particle swarm optimizer with a better search performance is pro-
posed, which employ multi-adaptive strategies to plan large-scale space global search
and refined local search as a whole according to the specialties of constrained layout
problems, and to quicken convergence speed, avoid premature problem, economize
computational expenses, and obtain global optimum. We tested the proposed algo-
rithm and compared it with other published methods on three constrained layout ex-
amples. The experimental results demonstrated that this revised algorithm can rapidly
converge at high quality solutions.

2 Mathematical Model and PSO

Mathematical Model. Taking the layout problem of satellite cabins as background, the
principle is illuminated [1]. There is a two-dimension rotating circular table (called
graph plane) with radius R and n dishes (called graph units) with uniform thickness,
each of which has a radius ri and mass mi (i=1,2…n), are installed on the graph plane
as shown in Fig.1. Suppose that the frictional resistance is infinite. Find the position
of each graph unit such that the graph units highly concentrate on the center of the
graph plane and satisfy the following constraints:
⑴There is no interference between any two graph units.
⑵No part of any graph unit protrudes out the graph plane.
⑶The deviation from dynamic-balancing for the system should not exceed a per-
missible value [δj].

Fig. 1. Representation of graph units layout

Suppose that the center of the graph plane is origin o of the Cartesian system, the
graph plane and graph units are just in plane xoy, the thickness of the graph units is
ignored, and xi,yi are the coordinates of the center oi of the graph unit i, which is its
mass center also. The mathematical model for optimization of the problem is given
by:
40 K. Lei

Find X i = xi , y i [ ]T , i ∈ I , i = 1,2 ,3,L , n

{ (x )+ r }
(1)
min F ( X i ) = max + yi
2 2
i i

．t．
s

f1( Xi ) = ri +rj − (xi − xj ) +( yi − yj ) ≤ 0,i ≠ j; i, j ∈I

2 2
(2)

f 2 (X i ) = (x i
2
+ yi
2
) + r − R ≤ 0,
i i∈I
(3)

2 2 (4)
⎛ n ⎞ ⎛ n ⎞
f3 ( X i ) = ⎜⎜ mi xi ⎟⎟ + ⎜⎜ mi yi ⎟⎟ − [δ J ] ≤ 0, i ∈ I
∑ ∑
⎝ i =1 ⎠ ⎝ i =1 ⎠

PSO Algorithm. In the original PSO formulae, particle i is denoted as

Xi=(xi1,xi2,...,xiD), which represents a potential solution to a problem in D-
dimensional space. Each particle maintains a memory of its previous best position
Pbest, and a velocity along each dimension, represented as Vi=(vi1,vi2,...,viD). At
each iteration, the position of the particle with the best fitness in the search space,
designated as g, and the P vector of the current particle are combined to adjust the
velocity along each dimension, and that velocity is then used to compute a new
position for the particle.
In the standard PSO, the velocity and position of particle i at (t+1)th iteration are
updated as follows:

( )
v idt +1 = w ∗ v idt + c1 ∗ r1 ∗ p idt − x idt + c 2 ∗ r2 ∗ p gd
t
− x idt ( ) (5)

xidt +1 = xidt + vidt +1 (6)

Constants c1 and c2 are learning rates; r1 and r2 are random numbers uniformly dis-
tributed in the interval [0, 1]; w is an inertia factor.
To speed up the convergence process and avoid the premature problem, Shi pro-
posed the PSO with linearly decrease weight method (LDWPSO)[3,4]. Suppose wmax
is the maximum of inertia weight, wmin is the minimum of inertia weight, run is cur-
rent iteration times, runMax is the total iteration times, the inertia weight is formu-
lated as:
w = wmax − ( wmax − wmin ) * run / runMax (7)
 Research on Constrained Layout Optimization Problem 41

3 Multi-Adaptive Strategies Particle Swarm Optimizer

(MASPSO)
Due to the complexity of a great deal local and global optima, PSO is revised as
MASPSO by four adaptive strategies to adapt the constrained layout optimization.

3.1 Adaptive harmonization Strategy of w

The w has the capability to automatically harmonize global search abilities and local
search abilities, avoid premature and gain rapid convergence to global optimum. First
of all, larger w can enhance global search abilities of PSO, so to explore large-scale
search space and rapidly locate the approximate position of global optimum, smaller
w can enhance local search abilities of PSO, particles slow down and deploy refined
local search, secondly, the more difficult the optimization problems are, the more
fortified the global search abilities need, once located the approximate position of
global optimum, the refined local search will further be strengthen to get global opti-
mum[5,6,7].
According to the conclusions above, we first constructed ⑻
as new inertia weight
decline curve for PSO, demonstrated in Figure 2.

w = 1 (1 + exp (0 .015 ∗ (run − runMax 3 ))) (8)

Fig. 2. Inertia weight decline curve by formula [8]

t
3.2 Adaptive Difference Mutation Strategy of Global Optimum p gd

Considering that the particles may find the better global optimum in the current best
region, the algorithm is designed to join mutation operation with the perturbation
operator. The runmax1, which is an iteration times of the transformation point, divides
runMax into two segments to respectively mutate according to themselves characteris-
tics, and further enhance the global search abilities and local search abilities to find a
satisfactory solution. pη is the mutation probability within (0.1, 0.3).The computed
equation is defined as:

if run ≤ run max1 p gd

t
= p gd
t
∗ (1 + 0.5η ); esle p gd
t
= p gd
t
∗ (1 + 0.15η ). (9)

whereη is Gauss (0,1) distributed random variable.

42 K. Lei

3.3 Adaptive Radius Rs Strategy

In equation (3), the graph plane radius is constant, which will influence the fitness
value computation and the result of search. In the search process, the more smallness
the periphery envelope circle radiuses are, the more convergence graph units are.
Evidently, taking the current periphery envelope circle radius as layout radius, the
search progress in smaller radius area, and quicken convergence speed, economize
computational expenses. Suppose the periphery envelope circle radius is Rs, the com-
puted equation is defined as:
t +1 t +1 t +1 t +1
if RS > RS , RS = RS ,esle RS = RS
t t
(10)

3.4 Adaptive Difference Mutation Strategy of Xi

Analysis on equation (5), if the best position of all particles has no change for a long
time, then the velocity updated is decided by w*vidt. Due to w<1, the velocity is less
and less, so the particle swarm is tended to get together, farther, the algorithm is
trapped into local optima, premature problem emerged like in genetic algorithms.
Considering that the adaptive difference mutation strategy of Xi is introduced. When
the best position of all particles has no or less change for a long time, the part of all
particles are kept down, deploy refined local search unceasingly; the rest are initial-
ized randomly, so to enhance global search abilities of PSO, break away from the
attraction of local optima, synchronously. Suppose runt is an iteration times, runk,
runk+t are the kth, (k+t)th iteration times, Rk, Rk+t are the maximal radius, accordingly;
Xρ are the initialized particles randomly, ρ is the dimension mutation probability,
accordingly; ε is an error threshold, the computed equation is defined as:

ifrun k − run k + t ≥ run t & Rk − R k + t ≤ ε , X i = X ρ , else X i = X i . (11)

where runt, ε, ρ are taken 20,0.005,0.2 in the paper.

3.5 Algorithm Designing and Flow

PSO, which is modified as MASPSO by the above methods, has the excellent search
performance to optimize constrained layout problems. According to MASPSO, The
design of the algorithm is as follows.
All particles are coded based on the rectangular plane axis system. Considering that
the shortest periphery envelope circle radius is the optimization criterion based on the
above constraint conditions for our problem, the fitness function and the penalty func-
tion, which are constructed in the MASPSO, respectively, can be defined as:
3
φ 1 (X i )= F (X i ) + ∑i=1
λ i f i (X i )u i ( f i ),
(12)
⎧0 fi ( X i ) ≤ 0
u i ( fi ) = ⎨ , i ∈ I.
⎩1 fi ( X i ) > 0
 Research on Constrained Layout Optimization Problem 43

where λi (λ1=1,λ2=1,λ3=0.01) is the penalty factor.

The flow of the algorithm is as follows:
Step1.Set parameters of the algorithm.
Step2.Randomly initialize the speed and position of each particle.
Step3.Evaluate the fitness of each particle using ⑿
and determine the initial values
t t
of the individual and global best positions: p and p gd
id .
Step4.Update velocity and position using (5), (6)and(8).
Step6.Evaluate the fitness using (12) and determine the current values of the indi-
vidual and global best positions: pidt and p gd
t
.
t +1
Step7.Check the R t and R , determine the periphery envelope circle radius.
t
Step8.Mutate p gd and Xi using (9) and (10), respectively.
Step9.Loop to Step 4 and repeat until a given maximum iteration number is at-
tained or the convergence criterion is satisfied.

4 Computational Experiments

Taking he literature [8], [9] as examples, we tested our algorithm, the comparison of
statistical results are shown in Tab.1, Tab.2, and Tab.3, respectively.
Parameters used in our algorithm are set to be: c1=c2=1.5, runMax=1000. The run-
ning environment is: MATLAB7.1, Pentium IV 2GHz CPU, 256MRAM, Win
XPOS.

Example1. Seven graph units are contained in the layout problem. The radius of
a graph plane is R=50mm. The allowing value of static non-equilibrium J is

Table 1. Comparison of experimental results of example 1

(a) Datum and layout results of example 1

Num- Graph units Literature [8] result Literature [9] result Our result
ber r(mm) m(g) x(mm) y(mm) x (mm) y(mm) x(mm) y(mm)
1 10.0 100.00 -12.883 17.020 14.367 16.453 17.198 -13.618
2 11.0 121.00 8.8472 19.773 -18.521 -9.560 -19.857 6.096
3 12.0 144.00 0.662 0.000 2.113 -19.730 -1.365 19.886
4 11.5 132.00 -8.379 -19.430 19.874 -4.340 18.882 7.819
5 9.5 90.25 -1.743 0.503 -19.271 11.241 -16.967 -14.261
6 8.5 72.25 12.368 -18.989 -3.940 22.157 -0.606 -22.873
7 10.5 110.25 -21.639 -1.799 -0.946 2.824 -0.344 -3.010
44 K. Lei

Table 1. (continued)

(b) The performance comparison of layout results of example 1

The least circle Static non-

Computational Interfer- Computational
Radius included all equilibrium
method ence time(s)
graph units (mm) (g.mm)
Literature [8] algorithm 32.662 0.029000 0 1002
Literature [9] algorithm 31.985 0.018200 0 1002
Our algorithm 31.934 0.000001 0 548

(c) The example 1 layout results comparison based on 40 times algorithm running

Times Times
The least circle The least circle
radius included all Literature radius included all Literature
Our Our
graph units (mm) [9] graph units (mm) [9]
algorithm algorithm
algorithm algorithm
≤32.3 10 19 (32.9,33.1] 2 1
(32.3,32.5] 16 11 (33.1,33.3] 2 1
(32.5,32.7] 5 5 >33.3 3 0
(32.7,32.9] 2 3

(a) Literature [8] layout result (b) Literature [9] layout result (c) Our layout result

Fig. 3. The geometry layout of example 1

[gj]=3.4g*mm. The result is in Table 1 and the geometric layout is shown in Figure 3
(the particle size is 50).

Example2. Forty graph units are contained in the layout problem. The radius of a
graph plane is R=880 mm. The radius of a graph plane is R=880 mm. The allowing
value of static non-equilibrium J is [gj]=20g*mm. The result is in Table2 and the
geometric layout is shown in Figure 4 (the particle size is 100).
 Research on Constrained Layout Optimization Problem 45

Table 2. Comparison of experimental results of example 2

(a) Datum and layout results of example 2

Graph units Literature [9] result Our result

Number
r(mm) m(g) x(mm) y(mm) x(mm) y(mm)
1 106 11 192.971 0 -528.304 -119.969
2 112 12 -69.924 0 103.353 -441.960
3 98 9 13.034 -478.285 -164.920 271.842
4 105 11 -291.748 21.066 -268.975 68.475
5 93 8 343.517 -351.055 -382.940 575.014
6 103 10 -251.1434 674.025 119.907 -45.977
7 82 6 95.268 -252.899 321.849 -433.652
8 93 8 -619.634 -421.032 684.118 71.745
9 117 13 725.062 0 -112.375 -364.996
10 81 6 127.487 175.174 -194.273 -553.490
11 89 7 358.251 -104.181 58.218 -246.094
12 92 8 694.612 -206.946 482.078 -312.251
13 109 11 -151.494 -350.475 299.075 347.118
14 104 10 -486.096 278.028 -381.489 -562.752
15 115 13 -406.944 -378.282 -378.466 266.641
16 110 12 -531.396 27.583 330.471 -106.831
17 114 12 -281.428 -570.129 -55.284 643.08
18 89 7 535.186 -82.365 499.100 362.345
19 82 6 349.187 -668.540 442.242 -545.018
20 120 14 494.958 -527.668 641.168 -172.081
21 108 11 -696.916 236.466 -205.525 478.499
22 86 7 -43.153 196.294 -534.084 -449.197
23 93 8 -143.066 -725.316 -478.615 74.719
24 100 10 -433.688 -159.158 511.793 171.389
25 102 10 -741.858 0.000 -7.548 403.484
26 106 11 292.820 431.997 -369.369 -350.518
27 111 12 -540.511 495.023 113.819 228.113
28 107 11 154.296 -671.681 -316.496 -144.125
29 109 11 -317.971 463.365 156.530 -656.467
30 91 8 41.295 -271.016 -625.532 -292.937
31 111 12 103.622 538.523 308.084 114.390
32 91 8 215.467 -213.844 235.152 -283.914
33 101 10 540.248 306.466 -545.860 404.331
34 91 8 58.125 341.687 -660.742 171.945
35 108 11 -235.120 227.217 406.575 539.906
36 114 12 510.413 520.918 182.065 555.234
37 118 13 -29.219 725.331 -94.105 -105.916
38 85 7 300.625 240.313 -26.879 -592.954
39 87 7 234.066 -494.031 505.239 -15.817
40 98 9 411.043 119.080 -56.975 107.473
46 K. Lei

Table 2. (continued)

(b) The performance comparison of layout results of example 2

The least circle Static non-

Computational Interfer- Computational
Radius included all equilibrium
method ence time(s)
graph units (mm) (g.mm)
Literature [8] algorithm 874.830 11.395000 0 1656
Literature [9] algorithm 843.940 0.003895 0 2523
Our algorithm 783.871 0.001267 0 1633

(a) Literature [9] layout result (b) Our layout result

Fig. 4. The geometry layout of example 2

5 Conclusions
From table1and table2, we can educe that: PSO with four improved adaptive strate-
gies harmonized the large-scale space global search abilities and the refined local
search abilities much thoroughly, which has rapid convergence and can avoid prema-
ture, synchronously. The effectiveness of the algorithm is validated for the con-
strained layout of the NP-hard problems; the algorithm outperformed the known best
ones in the in quality of solutions and the running time. In addition, the parameters of
、、
runt ε ρ are chose by human experience, which has the certain blemish. How to
choose the rather parameters are one of the future works.

References
1. Teng, H., Shoulin, S., Wenhai, G., et al.: Layout optimization for the dishes installed on
rotating table. Science in China (Series A) 37(10), 1272–1280 (1994)
2. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proc. of IEEE Int’l Conf.
Neural Networks, pp. 1942–1948. IEEE Computer Press, Indianapolis (1995)
 Research on Constrained Layout Optimization Problem 47

3. Eberhart, R.C., Kennedy, J.: A new optimizer using particles swarm theory. In: Sixth Inter-
national Symposium on Micro Machine and Human Science, pp. 39–43. IEEE Service Cen-
ter, Piscataway (1995)
4. Angeline, P.J.: Using selection to improve particle swarm optimization. In: Proc. IJCNN
1999, pp. 84–89. IEEE Computer Press, Indianapolis (1999)
5. Lei, K., Qiu, Y., He, Y.: A new adaptive well-chosen inertia weight strategy to automati-
cally harmonize global and local search ability in particle swarm optimization. In: 1st Inter-
national Symposium on Systems and Control in Aerospace and Astronautics, pp. 342–346.
IEEE Press, Harbin (2006)
6. Shi, Y., Eberhart, R.C.: A modified particle swarm optimizer. In: Proc. of the IEEE Con.
Evolutionary Computation, pp. 69–73. IEEE Computer Press, Piscataway (1998)
7. Jianchao, Z., Zhihua, C.: A guaranteed global convergence particle swarm optimizer. Jour-
nal of Computer Research and Development 4(8), 1334–1338 (2004) (in Chinese)
8. Fei, T., Hongfei, T.: A modified genetic algorithm and its application to layout optimization.
Journal of Software 10(10), 1096–1102 (1999) (in Chinese)
9. Ning, L., Fei, L., Debao, S.: A study on the particle swarm optimization with mutation
operator constrained layout optimization. Chinese Journal of Computers 27(7), 897–903
(2004) (in Chinese)
 ARIMA Model Estimated by Particle Swarm
Optimization Algorithm for
Consumer Price Index Forecasting

Hongjie Wang1 and Weigang Zhao2,

1
Railway Technical College, Lanzhou Jiaotong University, Lanzhou 730000, China
wanghongjietd@mail.lzjtu.cn
2
School of Mathematics & Statistics, Lanzhou University, Lanzhou 730000, China
zwgstd@gmail.com

Abstract. This paper presents an ARIMA model which uses particle

swarm optimization algorithm (PSO) for model estimation. Because the
traditional estimation method is complex and may obtain very bad re-
sults, PSO which can be implemented with ease and has a powerful opti-
mizing performance is employed to optimize the coefficients of ARIMA.
In recent years, inflation and deflation plague the world moreover the
consumer price index (CPI) which is a measure of the average price of
consumer goods and services purchased by households is usually observed
as an important indicator of the level of inflation, so the forecast of CPI
has been focused on by both scientific community and relevant authori-
ties. Furthermore, taking the forecast of CPI as a case, we illustrate the
improvement of accuracy and efficiency of the new method and the result
shows it is predominant in forecasting.

Keywords: ARIMA model, Particle swarm optimization algorithm,

Moment estimation, Consumer price index.

1 Introduction
The ARIMA, which is one of the most popular models for time series forecast-
ing analysis, has been originated from the autoregressive model (AR) proposed
by Yule in 1927, the moving average model (MA) invented by Walker in 1931
and the combination of the AR and MA, the ARMA models [1]. The ARMA
model can be used when the time series is stationary, but the ARIMA model
hasn’t that limitation. ARIMA model pre-processes the original time series by
some methods to make the obtained series stationary so that we can then ap-
ply the ARMA model. For example, for an original series which has only serial
dependency and hasn’t seasonal dependency, if the dth differenced series of this
original series is stationary, we can then apply the ARMA(p, q) model to the
dth differenced series and we say the original time series satisfy the modeling
condition of ARIMA(p, d, q) which is one kind of the ARIMA models. So the


Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 48–58, 2009.


c Springer-Verlag Berlin Heidelberg 2009
 ARIMA Model Estimated by Particle Swarm Optimization Algorithm 49

important point to note in ARIMA(p, d, q) modeling is that we must have either

a stationary time series or a time series that becomes stationary after one or
more differencing to be able to use it [2]. In addition, as we know, there are
four main stages in building an ARMA model: 1) model identification, 2) model
estimation, 3) model checking and 4) forecasting [3] (since we focus on the model
estimation here, the third step is ignored in this paper). Since the second step
is a complex process even when we adopt the relatively simple method (moment
estimation for example) and the approximate results of the parameters are un-
predictable which might greatly impact the forecasting precision when q > 1
that is the order of MA, PSO is proposed to achieve the second step.
Particle swarm optimization is a population based stochastic optimization
technique developed by Dr. Eberhart and Dr. Kennedy in 1995, inspired by
social behavior of bird flocking or fish schooling. The algorithm completes the
optimization through following the personal best solution of each particle and
the global best value of the whole swarm. And PSO can be implemented with
ease and has a simple principle, so it has been successfully applied in many fields.
Furthermore, we can see the new method has a higher forecasting precision than
the moment estimation from the case study in Section 4.
This paper is organized as follows: Section 2 reviews the detailed process of
ARIMA model (see [4]) and indicates the deficiencies of the traditional model
estimation; Section 3 introduces the particle swarm optimization algorithm and
proposes the PSO-based estimation method; In Section 4, as a case we forecast
the CPI to illustrate the excellent forecast ability of the new method.

2 Autoregressive Integrated Moving Average

Assume that {y1 , y2 , . . . , yN } is the original series without seasonal dependency
and the series has the limitation that at least 50 and preferably 100 observations
or more should be used [5], that is to say, N ≥ 50. Then we can model the
ARIMA(p, d, q) for this series, and it is obvious that the PSO is applicable for
the other ARIMA models (such as SARIMA model and so on) in the same way.

2.1 Model Identification

For an ARIMA(p, d, q), it is necessary to obtain the (p, d, q) parameters, where p
is the number of autoregressive terms, d is the number of non-seasonal differences
and q is the number of lagged forecast errors in the prediction equation [6].
The chief tools in identification are the autocorrelation function (ACF), the
partial autocorrelation function (PACF), and the resulting correlogram, which
is simply the plots of ACF and PACF against the lag length [2].
Since the ACF and the PACF is based on the covariance, it is necessary to
introduce the covariance first. For a time series {w1 , w2 , . . . , wn }, the covariance
at lag k (when k = 0, it is the variance), denoted by γk , is estimated in (1).
n−k
1
γ̂k = wi wi+k k = 0, 1, . . . , M (1)
n i=1
50 H. Wang and W. Zhao

where M is the maximum number of lags which is determined simply by dividing

the number of observations by 4, for a series with less than 240 based on Box
and Jenkins method [1]. Obviously, γ̂−k = γ̂k .
Then the ACF and PACF at lag k, denoted ρk and αkk respectively, can be
estimated according to (2) and (3).
γ̂k
ρ̂k = (2)
γ̂0
⎫
α̂11 = ρ̂1 ⎪
⎪
k ⎪
⎪
ρ̂k+1 − ⎪
⎪
j=1 ρ̂k+1−j α̂kj ⎪
⎬
α̂k+1,k+1 = k
1− ρ̂j α̂kj (3)
j=1 ⎪
⎪
⎪
α̂k+1,j = α̂kj − α̂k+1,k+1 · α̂k,k−j+1 ⎪
⎪
⎪
⎪
⎭
(j = 1, 2, . . . , k)
where k = 1, 2, . . . , M
Based upon the information given by the coefficients of ACF we can define
the following steps to determine d to make the series stationarity:
1) Let d = 0;
2) Define zt−d = ∇d yt = (1 − B)d yt , where t = d + 1, d + 2, . . . , N ;
3) Calculate the autocorrelation coefficients of {z1 , z2 , . . . , zN −d };
4) If the autocorrelation coefficients statistically rapidly decline to zero, the
corresponding d is the required one. Or else let d = d + 1 and return step 2).
where B denotes the backward shift operator, and ∇ is the differencing operator.
After determining the parameter d, the series {y1 , y2 , . . . , yN } converts to a
stationary series {z1 , z2 , . . . , zN −d } and we can estimate its mean in (4).
N −d
1
z̄ = zi (4)
N −d i=1

Then we can obtain a stationary time series with zero mean {x1 , x2 , . . . , xN −d },
where xt = zt −z̄. So the problem of modeling ARIMA(p, d, q) for {y1 , y2 , . . . , yN }
converts to the problem of modeling ARMA(p, q) for {x1 , x2 , . . . , xN −d }. And
the generalized form of ARMA(p, q) can be described as follow

ϕ(B)xt = θ(B)at (5)

where

ϕ(B) = 1 − ϕ1 B − ϕ2 B 2 − · · · − ϕp B p
θ(B) = 1 − θ1 B − θ2 B 2 − · · · − θq B q
and at is stationary white noise with zero mean, ϕi (i = 1, 2, . . . , p) and θi (i =
1, 2, . . . , q) are coefficients of the ARMA model.
On the other hand, for identifying the orders of ARMA model, autocorrela-
tion and partial autocortrelation graphs which are drawn based on 1 ∼ M lag
 ARIMA Model Estimated by Particle Swarm Optimization Algorithm 51

numbers provide information about the AR and MA orders (p and q) [1]. Con-
cretely, p is determined as the number of the front a few significant coefficients
in the partial autocorrelation graph and similarly q is determined as the number
of the front a few significant coefficients in the autocorrelation graph.

2.2 Model Estimation

In Section 2.1, the category, structure and orders of the model has been de-
termined. But the coefficients {ϕi } and {θi } of the model is unknown. In this
section we review a usual estimation method(moment estimation).
According to Yule-Walker equation, we can estimate {ϕi } in (6).
⎡ ⎤ ⎡ ⎤−1 ⎡ ⎤
ϕ̂1 γ̂q γ̂q−1 · · · γ̂q−p+1 γ̂q+1
⎢ϕ̂2 ⎥ ⎢ γ̂q+1 γ̂q · · · γ̂q−p+2 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢γ̂q+2 ⎥
⎢ .. ⎥ = ⎢ .. .. .. ⎥ ⎢ .. ⎥ (6)
⎣ . ⎦ ⎣ . . . ⎦ ⎣ . ⎦
ϕ̂p γ̂q+p−1 γ̂q+p−2 · · · γ̂q γ̂q+p

Then we can estimate {θi } and σa2 which is the variance of {at } in (7).


γ̂0 (x̄) = σ̂a2 (1 + θ̂12 + · · · + θ̂q2 )
(7)
γ̂k (x̄) = σ̂a2 (−θ̂k + θ̂k+1 θ̂1 + · · · + θ̂q θ̂q−k ), k = 1, 2, . . . , q
p p
where γ̂k (x̄) = j=0 l=0 ϕ̂j ϕ̂l γ̂k+l−j , k = 0, 1, . . . , q
So if q = 1, according to (7) and the reversibility(that is |θ1 | < 1) of ARMA(p, q)
model, we can calculate σa2 and θ1 using (8) and (9).

2 γ̂0 (x̄) + γ̂02 (x̄) − 4γ̂12 (x̄)
σ̂a = (8)
2
Then
γ̂1 (x̄)
θ̂1 = − (9)
σ̂a2
But if q > 1, we can see it is difficult to solute (7). At this time, we can change
(7) to (10) to solve this problem by using linear iterative method.
⎫
σ̂a2 = γ̂0 (x̄)(1 + θ̂12 + · · · + θ̂q2 )−1 ⎪
⎬
γ̂k (x̄) (10)
θ̂k = −( 2 − θ̂k+1 θ̂1 − · · · − θ̂q θ̂q−k ), k = 1, 2, . . . , q ⎪
⎭
σ̂a

At first, give initial values for {θˆi } and σ̂a2 , for example, let θ̂i = 0 (i = 1, 2, . . . , q)
and σ̂a2 = γ̂0 (x̄) and mark them as {θ̂i (0)} and σ̂a2 (0). Then substitute them into
(10), we can obtain {θ̂i (1)} and σ̂a2 (1). The rest may be deduced by analogy, so
we can obtain {θ̂i (m)} and σ̂a2 (m). If max{|θ̂i (m) − θ̂i (m − 1)|, |σ̂a2 (m) − σ̂a2 (m −
1)|} is very small, we can consider {θ̂i (m)} and σ̂a2 (m) are the approximatively
estimated values of {θi } and σa2 .
52 H. Wang and W. Zhao

From the above, we can see that the moment estimation is cumbersome to
achieve. And when q > 1, the approximate results of the parameters is unpre-
dictable and this might greatly impact the accuracy of forecast. So a method of
PSO for model estimating is proposed in Section 3, and we can see this method
has a simple principle and can be implemented with ease. What’s more, the PSO
estimation for ARIMA model (simply marked as PSOARIMA) has a higher fore-
casting precision than the ARIMA model which will be shown in Section 4.

2.3 Model Forecasting

According to (5), we can obtain at = θ(B)−1 ϕ(B)xt . In addition, we can let

∞
at = xt − πi xt−i = π(B)xt (11)
i=1

So there is ϕ(B) = θ(B)π(B), that is

∞ ∞ ∞



ϕi B i = ( θi B i )(1 − πi B i ) (12)
i=1 i=1 i=1

where



ϕi 1≤i≤p
θi 1≤i≤q
ϕi = , θi =
0 i>p 0 i>q

Through comparing the terms which have the same degree of B in (12), we can
obtain {πi } in (13), see [7].
⎧

⎪ π1 = ϕ1 − θ1
⎪
⎨
i−1



(13)
⎪
⎩ πi = ϕi − θi +
⎪ θi−j πj , i>1
j=1

Ignoring at and selecting a appropriate value K to replace ∞, we can obtain the

predicting formula according to (11), and it is shown as follow

K
x̂t = πi x̂t−i , if t − i ≤ 0, let xt−i = 0 (14)
i=1

where t = 1, 2, . . . , N − d + 1, and {x̂1 , x̂2 , . . . , x̂N −d } is simulation values of

{x1 , x2 , . . . , xN −d } and x̂N −d+1 is one forecasting value of {x1 , x2 , . . . , xN −d }.
Then we can obtain the prediction values of the original series, and the steps
are shown as follows:

1) Let ẑt = x̂t + z̄ (t = 1, 2, . . . , N − d + 1), so {ẑt } is the prediction series of

{zt };
 ARIMA Model Estimated by Particle Swarm Optimization Algorithm 53

2) Calculate {∇1 yt }, {∇2 yt }, . . ., {∇d−1 yt } whose number of components are

N − 1, N − 2, . . . , N − d + 1 respectively. Mark {yt } as {∇0 yt }. And mark
{ẑt } as {∇
t t


d y } which is the prediction series of {∇d y }, similarly, {∇i y } is
t
the prediction series of {∇i yt }, i = 0, 1, . . . , d − 1.
3) For all i from d to 1 do:
Let {∇ i−1 y } = {∇i−1 y }
t 1 t 1

and {∇ yt }j = {∇ yt }j−1 + {∇
i−1 i−1
iy }
t j−1
where j = 2, 3, . . . , N − i + 2.
where {∇i−1 yt }1 denote the first component of series {∇i−1 yt }.
According to the above procedure, we can obtain {∇
0 y } (t = 1, 2, . . . , N + 1)
t
which is the prediction series of original series {yt }, and we mark it as {ŷt }.

3 Particle Swarm Optimization Algorithm

Suppose that the search space is n-dimensional and a particle swarm consists
of m particles, then the ith particle of the swarm and its velocity can be rep-
resented by two n-dimensional vector respectively, xi = (xi1 , xi2 , . . . , xin ) and
vi = (vi1 , vi2 , . . . , vin ) i = 1, 2, . . . , m, see [8]. Each particle has a fitness that
can be determined by the objective function of optimization problem, and knows
the best previously visited position and present position of itself. Each particle
has known the position of the best individual of the whole swarm. Then the
velocity of particle and its new position will be assigned according to the follows

vik+1 = w × vik + c1 τ1 (pxbesti − xki ) + c2 τ2 (gxbest − xki ) (15)

xk+1
i = xki + vik+1 (16)
where vik and xkiare the velocity and position of the ith particle at time k
respectively; pxbesti is the individual best position of the ith particle; gxbest is
the global best position of the whole swarm; τ1 , τ2 are random number uniformly
from the interval [0,1]; c1 , c2 ∈ [0, 2] are two positive constants called acceleration
coefficients namely cognitive and social parameter respectively and as default in
[9], c1 = c2 = 2 were proposed; w is inertia weight that can be determined by

w = wmax − (wmax − wmin )N um/N ummax (17)

where wmax and wmin are separately maximum and minimum of w; N ummax is
maximum iteration time and N um is the current iterative time. A larger inertia
weight achieves the global exploration and a smaller inertia weight tends to
facilitate the local exploration to fine-tune the current search area. Usually, the
maximum velocity Vmax is set to be half of the length of the search space [10].
In this paper, because there are p + q coefficients (ϕi , θj , i = 1, 2, . . . , p, j =
1, 2, . . . , q) to be selected, the dimension of the swarm n = p + q. What’s more,
when the coefficients {ϕi } and {θj } are denoted by Φ and Θ which have p and
54 H. Wang and W. Zhao

q components respectively, the form of one particle can be denoted by (Φ, Θ),
that is (ϕ1 , ϕ2 , . . . , ϕp , θ1 , θ2 , . . . , θq ). Then we mark the simulation series {x̂t }
(Φ,Θ)
which is obtained from (14) as {x̂t } when the coefficients are Φ and Θ.
In the investigation, the mean square error (MSE), shown as (18), serves as
the forecasting accuracy index and the objective function of PSO for identifying
suitable coefficients.
N −d
1 (Φ,Θ) 2
M SE (Φ,Θ) = (xt − x̂t ) (18)
N −d t=1

The procedure of estimating coefficients by PSO for this model is illustrated as

follows:

1. Randomly generate m initial particles x01 , x02 , . . . , x0m and initial velocities
v10 , v20 , . . . , vm
0
which have p + q components respectively.
2. for 1 ≤ i ≤ m do
0 0
3. pxbesti ⇐ x0i ; pf besti ⇐ M SE (xi (1:p),xi (p+1:p+q))
4. x1i ⇐ x0i + vi0
5. end for
6. for all i such that 1 ≤ i ≤ m do
7. if pf bests ≤ pf besti then
8. gxbest ⇐ pxbests ; gf best ⇐ pf bests
9. end if
10. end for
11. k ⇐ 1; wmax ⇐ 0.9; wmin ⇐ 0.4
12. while k < N ummax do
13. for 1 ≤ i ≤ m do
k k
14. if M SE (xi (1:p),xi (p+1:p+q)) < pf besti then
k k
15. pf besti ⇐ M SE (xi (1:p),xi (p+1:p+q)) ;
16. pxbesti ⇐ xi k

17. end if
18. end for
19. for all i such that 1 ≤ i ≤ m do
20. if pf bests ≤ pf besti then
21. gxbest ⇐ pxbests ; gf best ⇐ pf bests
22. end if
23. end for
24. w ⇐ wmax − k(wmax − wmin )/N ummax
25. for 1 ≤ i ≤ m do
26. vik+1 ⇐ w × vik + c1 τ1 (pxbesti − xki ) + c2 τ2 (gxbest − xki )
27. if vik+1 ≥ Vmax or vik+1 ≤ −Vmax then
28. vik+1 ⇐ vik+1 /|vik+1 | × Vmax
29. end if
30. xk+1
i ⇐ xki + vik+1
31. end for
32. end while
 ARIMA Model Estimated by Particle Swarm Optimization Algorithm 55

where x(1 : p), x(p + 1, p + q) denote the 1st ∼ pth and the (p + 1)th ∼ (p + q)th
components of the vector x respectively. So gxbest(1 : p) and gxbest(p + 1, p + q)
are the optimally estimated values of Φ and Θ.

4 Case Study
In this section, a case study of predicting the consumer price index (CPI) of 36
big or medium-sized cities in China is shown. The original data is CPI in Jan.
2001 ∼ Oct. 2008 which is obtained from CEInet statistical database. And mark
the original data as {yt }, t = 1, 2, . . . , 94.
We can easily see {yt } is a non-stationary series by observing its graph-
ics. Observing the autocorrelogram of the third-differenced data of the original
data(shown in Fig. 1, where the two solid horizontal lines represent the 95%
confidence interval), in which we can see the autocorrelation coefficients statis-
tically rapidly decline to zero after lag 1, we confirm the third-differenced data
is stationary, and mark it as {zt }, t = 1, 2, . . . , 91. So d is equal to 3. And we can
then mark {zt − z̄} as {xt }, t = 1, 2, . . . , 91 where z̄ = −0.0275 according to (4).
Then we can draw the autocorrelogram and partial autocorrelogram (Fig. 2).
From this figure, two facts stand out: first, the autocorrelation coefficient starts
at a very high value at lag 1 and then statistically rapidly declines to zero;
second, PACF up to 4 lags are individually statistically significant different from
zero but then statistically rapidly declines to zero too. So we can determine p = 4
and q = 1, that is to say, we can design an ARMA(4,1) model for {xt }.
Estimating the coefficients Φ = {ϕ1 , ϕ2 , ϕ3 , ϕ4 } and Θ = {θ1 } according to
(6), (8), (9) and PSO method with the optimizing range of [−1, 1], respectively,
we can obtain Φ = {−1.0750, −0.8546, −0.4216, −0.1254}, Θ = {0.1614} and
Φ = {−0.4273, −0.0554, 0.1979, 0.1487}, Θ = {0.8196} respectively. Then we
can obtain the fitted and prediction series, and the fitted graph is shown as Fig.
3. At the same time, we obtain their MSE are 1.3189 and 1.2935 respectively,
that is to say, the latter fit better than the former.

Fig. 1. The autocorrelogram of the third-differenced data up to 23 lags

56 H. Wang and W. Zhao

(a) Autocorrelogram (b) Partial autocorrelogram

Fig. 2. The correlogram of {xt } up to 23 lags

8 8
Original data Original data
Simulation values Simulation values
6 6

4 4

2 2

0 0

Ŧ2 Ŧ2

Ŧ4 Ŧ4

Ŧ6 Ŧ6

Ŧ8 Ŧ8
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

(a) ARMA(4,1) (b) ARMA(4,1) optimized by PSO

Fig. 3. Fitted results of {xt }

Table 1. Prediction of consumer price index in Nov. 2008

Actual ARIMA(4,3,1) PSOARIMA(4,3,1)

value Prediction Relative Error Prediction Relative Error
102.70 103.24 0.53% 103.09 0.38%

Furthermore, the prediction values of the original data {yt } can be obtained
according to Section 2.3 by both moment estimation method and PSO method
for estimation. The relevant results are shown in Table 1. From this table, we can
see that the accuracy of PSOARIMA(4, 3, 1) model is better than the ARIMA(4,
3, 1) model. What’s more, we present the prediction figures (Fig. 4) to compare
the accuracy of the two models clearly.
 ARIMA Model Estimated by Particle Swarm Optimization Algorithm 57

110 110
Original data Original data
Actual value in Nov. 2008 Actual value in Nov. 2008
Prediction values Prediction values
108 108

106 106
Consumer price index

Consumer price index

104 104

102 102

100 100

98 98

96 96
2001.01 2002.01 2003.01 2004.01 2005.01 2006.01 2007.01 2008.01 2001.01 2002.01 2003.01 2004.01 2005.01 2006.01 2007.01 2008.01
Year.Month Year.Month

(a) ARIMA(4, 3, 1) (b) PSOARIMA(4, 3, 1)

Fig. 4. Prediction results of {yt }

Table 2. Prediction of consumer price index in Dec. 2008 ∼ Mar. 2009

Month Actual ARIMA(4,3,1) PSOARIMA(4,3,1)

Year value Predict R.E. Predict R.E.
12. 2008 101.30 101.39 0% 101.44 0.13%
01. 2009 101.10 99.725 1.36% 99.819 1.27%
02. 2009 98.50 100.31 1.83% 99.808 1.33%
03. 2009 98.90 96.029 2.90% 96.33 2.60%

Additionally, when we know the actual value in Nov. 2008, the value in
Dec. 2008 can be predicted. The rest may be deduced by analogy, we predict
the values in Jan. 2009 ∼ Mar. 2009, and the relevant results are shown in
Table 2. From this table we can see that accuracy of both models is very good,
but relative error of PSOARIMA(4, 3, 1) is smaller than ARIMA(4, 3, 1) gen-
erally.

5 Conclusion
In this paper, we proposed a novel design methodology which is a hybrid model
of particle swarm optimization algorithm and ARIMA model. PSO is used for
model estimation of ARIMA in order to overcome the deficiency that the tra-
ditional estimation method is difficult to implement and may obtain very bad
results. Furthermore, it is observed that the PSO-based model has worked more
accurately than the traditional moment estimation-based model through the
experimental results of forecasting the CPI.
On the other hand, we can see the power of the PSO for optimizing param-
eters. So in the future study, we will find more practical way to apply PSO
58 H. Wang and W. Zhao

in the other fields. But it is necessary to pay attention to that standard PSO
sometimes easily get in local optimization but not the holistic optimization,
fortunately, chaotic particle swarm optimization algorithm(CPSO) can help to
solve this problem successfully.

Acknowledgments. The research was supported by the Ministry of Education

oversea cooperation ”Chunhui Projects” under Grant (Z2007-1-62012) and The
Natural Science Foundation of Guansu Province in China under Grant (ZS031-
A25-010-G).

References
1. Ediger, V.Ş., Akar, S.: ARIMA forecasting of primary energy demand by fuel in
Turkey. Energy Policy 35, 1701–1708 (2007)
2. Erdogdu, E.: Electricity demand analysis using cointegration and ARIMA mod-
elling: A case study of Turkey. Energy Policy 35, 1129–1146 (2007)
3. Ong, C.-S., Huang, J.-J., Tzeng, G.-H.: Model identification of ARIMA family
using genetic algorithms. Applied Mathematics and Computation 164, 885–912
(2005)
4. Niu, D., Cao, S., Zhao, L., Zhang, W.: Power load forecasting technology and its
applications. China Electric Power Press, Beijing (1998) (in Chinese)
5. Ediger, V.Ş., Akar, S., Uğurlu, B.: Forecasting production of fossil fuel sources
in Turkey using a comparative regression and ARIMA model. Energy Policy 34,
3836–3846 (2006)
6. Valenzuela, O., Rojas, I., Rojas, F., Pomares, H., Herrera, L.J., Guillen, A., Mar-
quez, L., Pasadas, M.: Hybridization of intelligent techniques and ARIMA models
for time series prediction. Fuzzy Sets and Systems 159, 821–845 (2008)
7. Li, M., Zhou, J.-z., Li, J.-p., Liang, J.-w.: Predicting Securities Market in Shanghai
and Shenzhen by ARMA Model. Journal of Changsha Railway University 3, 78–84
(2000) (in Chinese)
8. Dingxue, Z., Xinzhi, L., Zhihong, G.: A Dynamic Clustering Algorithm Based
on PSO and Its Application in Fuzzy Identification. In: The 2006 International
Conference on Intelligent Information Hiding and Multimedia Signal Processing,
pp. 232–235. IEEE Press, New York (2006)
9. Kennedy, J., Eberhart, R.: Particle Swarm Optimization. In: The IEEE Interna-
tional Conference on Neural Networks, vol. 4, pp. 1942–1948. IEEE Press, New
York (1995)
10. Alatas, B., Akin, E., Bedri Ozer, A.: Chaos embedded particle swarm optimization
algorithms. Chaos, Solitons and Fractals (2008)
 A Uniform Solution to HPP in Terms of Membrane
Computing

Haizhu Chen and Zhongshi He

College of Computer Science, Chongqing University

Chongqing 400044, China
{cqchz,zshe}@cqu.edu.cn

Abstract. Membrane system is a computing model which imitates natural proc-

ess at cellular level. Research results have shown that this model is a promising
framework for solving NP-complete problems in polynomial time. Hamiltonian
path problem (HPP for short) is a well-known NP-complete problem, and there
are several semiform P systems dealing with it. In the framework of recognizer
P systems based on membrane 2-division, we presented a uniform polynomial
time solution to the HPP problem and gave some formal details of this solution
from the point of view of complexity classes.

Keywords: Natural Computing, Membrane Computing, Cellular Complexity

Classes, HPP.

1 Introduction
P systems are emergent branch of nature computing, which can be seen as a kind of
distributed parallel computing model. It is based upon the assumptions that the proc-
esses taking place in the living cells can be considered as computations. Up to now,
several variants of P systems with linear or polynomial time have been constructed to
solve some NP-complete problems, such as SAT [1, 2, 3], HPP [2, 4, 5], Subset Sum
[6], Knapsack [7], 2-Partition[8], Bin Packing[9].
Hamiltonian path problem (HPP for short) is a well-known NP-complete problem,
and there are several P systems dealing with it. The P systems in [4, 5] are all semi-
form way [10]. Specifying the starting and ending vertices, literature [2] presents a
uniform solution to HPP with membrane separation. It remains open confluently solv-
ing HPP in polynomial time by P systems with division rules instead of separation
rules in a uniform way [2].
Without specifying the starting vertex and ending vertex of a path, this paper
presents a uniform solution to HPP with membrane division where the communica-
tion rules (sending objects into membranes) [3] are not used. The remainder of the
paper is organized as follows: first recognizer P system with membrane division is
introduced in the next section. In Section3, the solution to HPP with membrane divi-
sion is presented, with some formal details of such solution given in Section4. Finally,
conclusions are drawn in the last section.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 59–68, 2009.
© Springer-Verlag Berlin Heidelberg 2009
60 H. Chen and Z. He

2 Recognizer P System with Membrane Divisions

This paper deals with recognizer P system with membrane division. It is necessary to
recall such system. Firstly, a P system with membrane division is a construct [3]:

Π =( V , H , µ , w1 , …, wm , R )

where:
m ≥1 (the initial degree of the system);
V is the alphabet of objects;
H is a finite set of labels for membranes;
µ is a membrane structure, consisting of m membranes, labeled (not necessarily in a
one-to-one manner) with elements of H ;
w1 , …, wm are strings over V describing the multisets (every symbol in a string
representing one copy of the corresponding object) placed in the m regions of µ
respectively;
R is a finite set of developmental rules, of the following forms:
(1) [ a → v ] eh , for h ∈ H , e ∈ {+, −, 0} , a ∈V , v ∈V *
(object evolution rules);
(2) [a]eh1 → [ ]eh2 b , for h ∈ H , e1 , e2 ∈ {+, −, 0} , a, b ∈V
(communication rules);
(3) [a]eh1 → [ b]eh2 [c]eh3 , for h ∈ H , e1 , e2 , e3 ∈{+, −, 0} a, b, c ∈V
(division rules for elementary membranes)
Note that, in order to simplify the writing, in contrast to the style customary in the
literature, we have omitted the label of the left parenthesis from a pair of parentheses
which identifies a membrane. All aforementioned rules are applied according to the
following principles:
(a) The rules of type (1) are applied in the parallel way.
(b) The rules of types (2), (3) are used sequentially, in the sense that one membrane
can be used by at most one rule of these types at a time.
(c) All objects and all membranes, which can evolve, should evolve simultaneously.
We will introduce two definitions[1]:
Definition 1. A P system with input is a tuple ( Π , ∑ , iΠ ), where: (a) Π is a P sys-
tem, with working alphabet Γ , with p membranes labeled with 1, ..., p, and initial
multisets w1 , …, w p associated with them; (b) ∑ is an (input) alphabet strictly con-
tained in Γ ; the initial multisets of Π are over Γ - ∑ ; and (c) iΠ is the label of a
distinguished (input) membrane.
Let win be a multiset over ∑ . The initial configuration of ( Π , ∑ , iΠ ) with input
win is ( µ , w1 , …, wiΠ ∪w in
, …, w p ).
 A Uniform Solution to HPP in Terms of Membrane Computing 61

Definition 2. A recognizer P system is a P system with input, ( Π , ∑ , iΠ ), and with

output such that:
(1) The working alphabet contains two distinguished elements, yes and no.
(2) All its computations can halt.
(3) If C is a computation of Π , then either some object yes or no (but not both)
must have been released into the environment, and only in the last step of the
computation.
We say that C is an accepting computation (respectively, rejecting computation) if the
object yes (respectively, no) appears in the external environment associated with the
corresponding halting configuration of C.
For ease of presentation, in the rest of this paper, recognizer P system means
recognizer P system with input and we will not distinguish each other.
Section4 will prove that the family Π of recognizer P system in Section3 gives a
polynomial solution for the HPP. We would like to introduce definition about com-
plexity classes in P systems.
Let AM stand for the class of recognizer P systems with active membrane using 2-
division. We will introduce the definition of the complexity class PMCAM [6, 7, 8, 10]:
Definition 3. We say that a decision problem, X=(IX,θX), is solvable in polynomial
time by a family Π= ( Π (t ) )t∈N, of recognizer P systems using 2-division, and we
write X∈PMCAM, if
(1) Π is AM-consistent; that is, Π (t ) ∈ AM for each t (t∈N).
(2) Π is a polynomially uniform family by Turing Machine, that is, there exists a
deterministic Turing machine that builds Π (t ) from t ≥ 0 in polynomial time.
(3) There exists a polynomial encoding (g, h) of IX in Π verifying:
− Π is polynomially bounded, with respect to (g, h), that is, there exists a polynomial
function p, such that for each u∈IX, all computations of Π (h(u )) with input g(u)
halt in at most p(|u|) steps.
− Π is sound with respect to (g, h); that is, for all u∈IX, if there exists an accepting
computation of Π (h(u )) with input g(u), thenθX(u)=1 (the corresponding answer of
the problem is yes.
− Π is complete with respect to (g, h); that is, for all u∈IX, ifθX(u)=1 (the answer to
the problem is yes), then every computation of Π (h(u )) accepts g(u) (the system
also responds yes).
In the above definition every P system is required to be confluent: every computation
of a system with the same input must always give the same answer. The class PMCAM
is closed under polynomial-time reduction and complement.

3 Solving HPP with Membrane Divisions

A Hamiltonian path in a directed graph G is a path that visits each of the nodes of G
exactly only once. The Hamiltonian Path Problem (HPP for short) is the following
decision problem:
62 H. Chen and Z. He

Input: given a directed graph G = (V, E);

Output: determine if there exists a Hamiltonian path in G.
This section presents a uniform solution to HPP by a family of recognizer P sys-
tems with membrane division, in which the starting vertex and ending vertex of a path
are not specified. Unless stated otherwise, the solutions cited are described in the
usual framework of recognizer P systems with active membranes using 2-division,
with three electrical charges, without change of membrane labels, without coopera-
tion, and without priority. The solution is addressed via a brute force algorithm which
consists in the following phases.
Preparing stage: introduce n vertices from which we can start to search the possible
Hamiltonian paths (HPs);
Generating stage: generate all paths of length n in G (G has n vertices) through
membrane division;
Checking stage: for each of the previous paths, determine whether it visits all the
nodes of G;
Output stage: answer yes or no depending on the results from checking stage.

3.1 The Uniform Solution to HPP with Membrane Divisions

For any given directed graph G with n vertices, we construct a recognizer P system
Π ( n) to solve HPP. Therefore the family presented here is

Π = {( Π ( n) , ∑(n) , iΠ ): n∈N}

The input alphabet is

∑(n) ={xi, j, k | 1≤ i, j ≤ n, -1 ≤ k ≤ i}
Where, the object xi, j, k represents (i, j) ∈E. Over ∑(n) , the instance G (whose size is
n) can be encoded as win and win is the input multiset of Π ( n) :
win = {xi, j, k |1≤ i, j ≤ n, k = i}, (i, j) ∈E
The input membrane is iΠ =2 and the P system can be constructed as:

Π ( n) =( V (n) , H , µ , w1 , w2 , R )
Where,
∪{ v | 1≤ i ≤ n-1}∪{ t , z | 1≤ i ≤ n}
V (n) ={xi, j, k | 1 ≤ i, j ≤ n, -1 ≤ k ≤ i} i i i

∪{ c | 0 ≤ i ≤ n-1}∪{ r , r′ |1≤ i ≤ n}∪{ d , d ′ | 0≤ i ≤ n}

i

∪{ e | 0 ≤ i ≤ n+1}∪{f | 0≤ i ≤ 3n(n+3)/2+5}∪{r , g, e, s, yes, no}

i i i i

i i 0
H = {1, 2}
µ = [[ ] +2 ]10
w1 = f 0 , w2 = d 0 v1
And the set R contains the rules as follows:
 A Uniform Solution to HPP in Terms of Membrane Computing 63

[vi ]+2 → [ti ]−2 [vi +1 ]+2 (1≤ i ≤ n-2), [vn−1 ]+2 → [t n−1 ]−2 [t n ]−2 , [t j → c j −1r j′g ]−2 (1≤ j ≤ n) (1)

At the preparing stage, we have { d 0 v1 } and an input multiset win in membrane la-
beled with 2, which is encoded according to the instance G. From v1 , other vertices in
G can be obtained by rules (1). This means that we can find HPs from the n vertices
respectively. Object r j′ represents the starting vertex and object cj-1 can be regarded as
a counter whose subscript is decreased by 1 each time in generating stage. These rules
are applied at the preparing stage.
[ g ]−2 → [ ]02 g (2)

After the starting vertexes appear in corresponding membranes respectively, we will

enter generating stage by rule (2) which will be applied in other computing step as
explained later.
[ xi , j , k → xi , j , k −1 ]02 , [ci → ci −1 ]02 , [ri′ → ri ]02 (1 ≤ i, j ≤ n, 0 ≤ k ≤ i) (3)

In the membranes labeled with 2, when their polarization is 0, the third subscript k of
xi, j, k and the subscript i of ci are decreased by 1 simultaneously. For the i-th vertex in
G that has been added into the current path, Objects ri′ and ri are two different ver-
sions of representation in the membranes with different polarizations. Xi, j, k can not
evolve any more once k reaches -1. c0 and xi, j,0 will appear after I steps if we have
(i, j) ∈E.
[c0 ]02 → [ ] 2+ c0 (4)

Vertex vi is the last vertex of the current path, and we will extend this path by adding
vj with the appearance of c0 if we have (i, j) ∈E. Several vertices will connect vi be-
sides vj, so new membranes are needed to be created to contain the new paths. Rule
(4) can change the polarization of membrane to positive in order to trigger membrane
divisions to obtain new membranes.
[ d i → d i′] +2 , [ ri → ri′] +2 , [ xi , j , 0 ] +2 → [ z j ] −2 [ g ] +2 (1 ≤ i, j ≤ n, 0 ≤ k ≤ i) (5)

Objects d i , d i′ both mean the fact that the length of current path is I, but they are two
different versions in membranes labeled with 2 with different polarizations. Xi, j, 0 in
the membranes with positive polarization will cause the membrane division to extend
the current path. Since not only one vertex connects with vi, there will be several
objects like xi, j, 0 whose third subscript k is 0. We non-deterministically choose one of
them to trigger the division, and the rest will be remained in the two new membranes
with different polarizations, which indicate that they will be processed by different
rules: the division will be triggered again by one of them in the new membrane with
positive polarization in order to obtain another new path; while all of them will be
deleted in the one with negative polarization.
[ z j → r j′c j −1 g ] −2 , [ d i′ → d i ] −2 , [ xi , j , k → xi , j , i ] −2 (6)
64 H. Chen and Z. He

[ xi , j , 0 → λ ] 2− , [ d i → d i +1 ] −2 (1 ≤ i, j ≤ n, k ≠ 0)
In the new membrane labeled with 2 with negative polarization, new object r j′ is
introduced, which indicates that vertex vj is added to the current path. At the same
time, new counter cj-1 is introduced and other objects, such as xi, j, k, return to xi, j, i.
After the length of the path is increased, the polarization is changed by rule (2). We
return to the beginning of generating stage and resume the procedure for extending
the current path.
[ g → λ ]+2 (7)

In the new membrane labeled with 2 with positive polarization, object g does not
work any more and is deleted.
[d n → se]02 (8)

When the length of path is n, we know that n vertices are in the path and we can enter
the next stage to check if the path is Hamiltonian or not. Two new objects s and e are
introduced to prepare for the checking stage.
[e → e0 r0 ] +2 , [ s] +2 → [ ]02 s (9)

With the appearance of dn, the counter ct (t>0) associated with the last vertex is intro-
duced and it still can evolve until c0 is sent out of the membrane. Objects e and s
evolve by rule (9) to trigger the next stage.
[ei → gei +1 ]02 , [r0 ]02 → [ ]2− r0 ( 0 ≤ i ≤ n) (10)

[ri → ri −1 ]−2 (1 ≤ i ≤ n) (11)

In checking stage, we use rules (10) and (11) with a loop. If the object r0 is present,
we eliminate it and perform a rotation of the objects ri, ..., rn with their subscripts
decreased by 1. It is clear that a membrane contains HP if and only if we can run n+1
steps of the loop. Note that this check method is not the same as the one in [10], for
the communication rule are not used.
[en +1 → yes] −2 (12)

[ yes]02 → [ ] +2 yes (13)

The object en+1 introduced by rules (10) means that the path is Hamiltonian, so it
evolves to object yes, which will leave the membrane labeled with 2 and enter
membrane 1.
[ g → λ ]10 , [c0 → λ ]10 , [r0 → λ ]10 , [ s → λ ]10 , [ f i → f i +1 ]10 (0≤ i ≤ 3n(n+3)/2+4) (14)
All of these rules are applied in parallel in membrane 1 with neutral polarization.
Since objects g, c0, r0 and s are sent out of the membranes labeled with 2 and will be
not used any more, it necessary to delete them to release computing recourse. Object fi
evolves at each computation step for counting how many steps have been performed.
 A Uniform Solution to HPP in Terms of Membrane Computing 65

[ yes]10 → [ ]1+ yes, [ f time ]10 → [ ]1+ no time = 3n(n + 3) / 2 + 5 (15)

With changing the polarization to positive, object yes in membrane 1 will be sent out
the environment to show that there is a HP among the n vertices. If at step
3n(n+3)/2+5 the polarization of the membrane 1 has not changed to positive, the an-
swer must be no and no will be sent out to the environment. The special step will be
analyzed in subsection3.2. Once the polarization is change all rules in membrane 1
cannot be applied any more and the computation halts.

3.2 An Overview of the Computation

All of foregoing rules applied in membranes labeled with 2 are illustrated in Fig.1, in
which 0, + and - respectively represent the polarization of the membranes. Fig.1 looks
like state machine, and it can be seen that different polarizations and objects indicate
distinct stages. Rules (2), (4), (9), (10) are used to change the polarizations in order to
enter another stage of computation.
At preparing stage, we have v1 in membrane labeled with 2. From v1, other vertices
in G can be obtained by application of rules (1). After n starting vertices obtained, we
will enter next stage to find HPs respectively. Generating stage consists of three
phases: a vertex that connects with the last vertex in the current path is expected to
find in search phase, then it will be add into the path through divide phase, and the
length of the path will be increased and the objects related to others edges will recover
in restore phase. Once a vertex is added in the path, all of the edges start from it will
be consumed in the extensions of the path. These three phases will be looped until the
length of the path reaches n. Checking stage is needed to decide whether the path with
length of n is HP or not. The answer yes or no will be sent out at the output stage.

di , ri d ic, ric di odi1, ric

ri ri o ri1

Fig. 1. The rules applied in membranes labeled by 2

66 H. Chen and Z. He

The number of steps in each stage is as follows:

(1) Preparing and transition to generating: at most n+1
(2) Generating stage:
Search: at most n(n+1)/2-n
Divide: at most n(n-1)
Restore: 3(n-1)
(3) Checking: 2+2(n+1)
(4) Output yes: 2
The overall number of steps is:
n+1+n(n+1)/2-n+n(n-1)+3(n-1)+2+2(n+1)+2=3n(n+3)/2+4= O (n2)
Before the step 3n(n+3)/2+4, yes will be sent out if and if there is HP in the instance
G. If at step 3n(n+3)/2+5 the polarization of the membrane 1 has not changed to posi-
tive, the answer must be no and no will be sent out to the environment.

4 Some Formal Details

In this section, we will prove that the uniform family Π of recognizer P system in
Section3 gives a polynomial solution for the HPP. According to Definition3 in
Section2, we have:
Theorem 1. HPP ∈ PMCAM.
Proof: (1) It obviously that the family Π of recognizer P system defined is AM-
consistent, because for every P system of the family is a recognizer P system with
active membrane using 2-division and all of them are confluent.
(2) The family Π of recognizer P system is polynomially uniform Turing Machine.
For each n, which is the size of an instance G, a P system can be constructed. The
rules of Π (n) are defined in a recursive manner from n. The necessary resources to
construct the system are:
Size of the alphabet: n3+(7n2+27n)/2+14= Θ( n 3 )
Initial number of membrane: 2= Θ(1)
Initial number of objects: 3= Θ(1)
The total number of evolution rules: n3+(9n2+31n+38)/2= Θ( n 3 )
All these resources are bounded polynomially time of n; therefore a Turing machine
can build the P system in polynomial time according to n.
∪∈
(3) We have the functions g: I HPP → t NI∏(t) and h: I HPP →N defined for an in-
stance u=v1, …, vn as g(u)={xi, j, k : 1≤ i, j ≤ n} and h(u)=n. Both of them are total and
polynomially computable. Furthermore, (g, h) conforms an encoding of the set in-
stances of the HPP problem in the family of recognizer P system as for any instance u
we have that g(u) is a valid input for the P system.
Π is polynomially bounded, with respect to (g, h). A formal description of the
computation let prove that Π always halts and sends to the environment the object yes
or no in the last step. The number of steps of the P system is within 3n(n+3)/2+4 if the
 A Uniform Solution to HPP in Terms of Membrane Computing 67

output is yes and 3n(n+3)/2+5 if the output is no, therefore there exists a polynomial
bound for the number of steps of the computation.
Π is sound with respect to (g, h). As we describe the computations of P system of
Π above, that object yes is send out of the system means that there are n different
vertices in the path, which is the actual Hamiltonian path in the instance G according
to the definition of HPP.
Π is complete with respect to (g, h). The P system searches the HP from every ver-
tex respectively at the beginning of the computation. We will get objects such as ri (1
≤ i ≤ n) and delete objects such as xi, j, k (1 ≤ i, j ≤ n, -1 ≤ k ≤ i) corresponding to the
edges whose starting vertex vi is added to the current path. When there is a Hamilto-
nian path in the instance G with n vertices, the system Π will obtain objects r1, r2, ...,
rn in some membrane, and output object yes to the environment. □
Since this class is closed under polynomial-time reduction and complement we
have:
Theorem 2. NP ∪ co-NP ⊆ PMC AM.

5 Conclusion
In this paper, we present a uniform solution for HPP in terms of P system with
membrane division. This has been done in the framework of complexity classes in
membrane computing.
It can be observed that many membranes will not evolve any more at some step be-
fore the whole system halts. Some kind of rule should be developed to dissolve these
membranes for releasing the computing resource, and it will contribute to the con-
struction of the efficient simulators of recognizer P systems for solving NP-complete
problems.
Acknowledgments. This research is supported by Major National S&T Program of
China (#2008ZX 07315-001).

References
1. Gutiérrez-Naranjo, M.A., Pérez-Jiménez, M.J., Romero-Campero, F.J.: A uniform solution
to SAT using membrane creation. Theoretical Computer Science 371(1-2), 54–61 (2007)
2. Pan, L., Alhazov, A.: Solving HPP and SAT by P systems with active membranes and
separation rules. Acta Informatica 43(2), 131–145 (2006)
3. Păun, G.: Computing with Membranes: Attacking NP-Complete Problems. In: Proceedings
of the Second International Conference on Unconventional Models of Computation,
pp. 94–115. Springer, London (2000)
4. Păun, A.: On P Systems with Membrane Division. In: Proceedings of the Second Interna-
tional Conference on Unconventional Models of Computation, pp. 187–201. Springer,
London (2000)
5. Zandron, C., Ferretti, C., Mauri, G.: Solving NP-Complete Problems Using P Systems
with Active Membranes. In: Proceedings of the Second International Conference on
Unconventional Models of Computation, pp. 289–301. Springer, London (2000)
68 H. Chen and Z. He

6. Pérez-Jiménez, M.J., Riscos-Núñez, A.: Solving the Subset-Sum problem by active mem-
branes. New Generation Computing 23(4), 367–384 (2005)
7. Jesús Pérez-Jímenez, M., Riscos-Núñez, A.: A Linear-Time Solution to the Knapsack
Problem Using P Systems with Active Membranes. In: Martín-Vide, C., Mauri, G., Păun,
G., Rozenberg, G., Salomaa, A. (eds.) WMC 2003. LNCS, vol. 2933, pp. 250–268.
Springer, Heidelberg (2004)
8. Gutiérrez-Naranjo, M.A., Pérez-Jiménez, M.J., Riscos-Núñez, A.: A fast P system for find-
ing a balanced 2-partition. Soft Computing 9(9), 673–678 (2005)
9. Pérez-Jiménez, M.J., Romero-Campero, F.J.: Solving the BINPACKING problem by rec-
ognizer P systems with active membranes. In: Proceedings of the Second Brainstorming
Week on Membrane Computing, Seville, Spain, pp. 414–430 (2004)
10. Pérez-Jiménez, M.J., Romero-Jiménez, A., Sancho-Caparrini, F.: Computationally Hard
Problems Addressed Through P Systems. In: Ciobanu, G. (ed.) Applications of Membrane
Computing, pp. 315–346. Springer, Berlin (2006)
 Self-organizing Quantum Evolutionary Algorithm
Based on Quantum Dynamic Mechanism

Sheng Liu and Xiaoming You

School of Management, Shanghai University of Engineering Science,

Shanghai 200065, China
{ls6601,yxm6301}@163.com

Abstract. A novel self-organizing Quantum Evolutionary Algorithm based on

quantum Dynamic mechanism for global optimization (DQEA) is proposed.
Firstly, population is divided into subpopulations automatically. Secondly, by
using co-evolution operator each subpopulation can obtain optimal solutions.
Because of the quantum evolutionary algorithm with intrinsic adaptivity it can
maintain quite nicely the population diversity than the classical evolutionary al-
gorithm. In addition, it can help to accelerate the convergence speed because of
the co-evolution by quantum dynamic mechanism. The searching technique for
improving the performance of DQEA has been described; self-organizing algo-
rithm has advantages in terms of the adaptability; reliability and the learning
ability over traditional organizing algorithm. Simulation results demonstrate the
superiority of DQEA in this paper.

Keywords: Quantum evolutionary algorithm, Quantum dynamic mechanism,

Self-organizing and co-evolution.

1 Introduction
Most studies have used hybrid evolutionary algorithm in solving global optimization
problems. Hybrid evolutionary algorithm takes advantages of a dynamic balance be-
tween diversification and intensification. Proper combination of diversification and
intensification is important for global optimization. Quantum computing is a very
attractive research area; research on merging evolutionary algorithms with quantum
computing [1] has been developed since the end of the 90's. Han proposed the quan-
tum-inspired evolutionary algorithm (QEA), inspired by the concept of quantum com-
puting, and introduced a Q-gate as a variation operator to promote the optimization of
the individuals Q-bit [2]. Up to now, QEA has been developed rapidly and applied in
several applications of knapsack problem, numerical optimization and other fields [3].
It has gained more attraction for its good global search capability and effectiveness.
Although quantum evolutionary algorithms are considered powerful in terms of global
optimization, they still have several drawbacks such as premature convergence. A
number of researchers have experimented with biological immunity and cultural
dynamic systems to overcome these particular drawbacks implicit in evolutionary
algorithms [4], [5]; here we experiment with quantum dynamic mechanism.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 69–77, 2009.
© Springer-Verlag Berlin Heidelberg 2009
70 S. Liu and X. You

This paper presents a self-organizing quantum evolutionary algorithm based on

quantum dynamic mechanism for global optimization (DQEA). Firstly the population
will be divided into subpopulations automatically. Due to its intrinsic adaptivity the
quantum evolutionary algorithm can enable individual to draw closer to each optimal
solution. The self-organizing operator is adopted to enable individual to converge to
the nearest optimal solution which belongs to the same subpopulations. Meanwhile,
each subpopulation can co-evolve by quantum dynamic mechanism. The intensifica-
tion strategies are the self-organizing operator and the co-evolution. The experiment
results show that DQEA is very efficient for the global optimization; usually it can
obtain the optimal solution faster. We describe quantum dynamic mechanism for
improving the performance of DQEA, its performance is compared with that of
CRIQEA[6] and MIQEA [7]. The rest of this paper is organized as follows: Section 2
presents the quantum evolutionary algorithm and related work. Section 3 gives the
structure of DQEA and describes the improving performance of the algorithm. Sec-
tion 4 presents an application example with DQEA and the other algorithms [6], [7]
for several benchmark problems, and summarizes the experimental results. Finally
concluding remarks follow in Section 5.

2 Quantum Evolutionary Algorithm and Related Work

QEA utilizes a new representation, called a Q-bit, for the probabilistic representation
that is based on the concept of qubits [2]. A qubit may be in the “1” state, in the “0”
state, or in any superposition of the two [8]. The state of a qubit can be represented
as | ϕ >=α |0 >+β |1> , where α and β are complex numbers that specify the probability
amplitudes of the corresponding states. |α|2 gives the probability that the qubit will be
found in ‘0’ state and |β|2gives the probability that the qubit will be found in the ‘1’
state. Normalization of the state to unity always guarantees: |α|2+|β|2=1. One qubit is
defined with a pair of complex numbers(α,β), And an m-qubits represention is defined
as ⎡⎢ α 1 α 2 α 3 L α m ⎤⎥ where |αi|2+|βi|2=1, i=1,2……m.
⎣β1 β 2 β3 L β m ⎦
Q-bit representation has the advantage that it is able to represent a linear superposi-
tion of states probabilistically. If there is a system of m-qubits, the system can represent
2m states at the same time. However, in the act of observing a quantum state, it collapses
to a single state. The basic structure of QEA is described in the following[2].
QEA()
{ t • 0;
Initialize Q(t) ;
Make P(t) by observing the states of Q(t) ;
Evaluate P(t) ;
Store the optimal solutions among P(t) ;
While (not termination-condition) do
{t=t+1;
Make P(t) by observing the states of Q(t-1);
Evaluate P(t) ;
Update Q(t) using Q-gate U(t);
Store the optimal solutions among P(t) ;
 Self-organizing Quantum Evolutionary Algorithm 71

}
}
end.
where Q(t) is a population of qubit chromosomes at generation t, and P(t) is a set of
binary solutions at generation t.
1) In the step of ‘initialize Q(t)’, all qubit chromosomes are initialized with the
same constant 1/ 2 . It means that one qubit chromosome represents the linear super-
position of all possible states with the same probability.
2) The next step makes a set of binary solutions, P(t), by observing Q(t) states. One
binary solution is formed by selecting each bit using the probability of qubit. And
then each solution is evaluated to give some measure of its fitness.
3) The initial best solution is then selected and stored among the binary solutions,
P(t).
4) In the while loop, A set of binary solutions, P(t), is formed by observing Q(t-1)
states as with the procedure described before, and each binary solution is evaluated to
give the fitness value. It should be noted that P(t) can be formed by multiple observa-
tions of Q(t-1).
5) In the next step, ‘update Q(t)’, a set of qubit chromosomes Q(t) is updated by ap-
plying rotation gate defined below
⎡ COS （ ∆ θ i） − SIN （ ∆ θ i）⎤
U (∆θ i ) = ⎢
⎣ SIN （ ∆ θ i） COS （ ∆ θ i）⎥⎦ (1)

where ∆θ i is a rotation angle of each Q-bit towards either 0 or 1 state depending on

its sign. Fig.1 shows the polar plot of the rotation gate ∆θ i should be designed in
compliance with practical problems. Table1 can be used as an angle parameters table
for the rotation gate.
The magnitude of ∆θ i has an effect on the speed of convergence, but if it is too
big, the solutions may diverge or have a premature convergence to a local optimum.
The sign of ∆θ i determines the direction of convergence.

|1>

(¦Ái',¦Âi' )
△θ i
(¦Ái,¦Âi)

|0>

Fig. 1. Polar plot of the rotation gate

72 S. Liu and X. You

6) The best solution among P(t) is selected, and if the solution is fitter than the
stored best solution, the stored solution is replaced by the new one. The binary solu-
tions P(t) are discarded at the end of the loop.

Table 1. Lookup table of ∆θ i , where f(·) is the fitness function, and bi and xi are the i-th bits of
the best solution b and the binary solution x, respectively

xi bi f(x)≥f(b) ∆θ i
0 0 false θ1
0 0 true θ2
0 1 false θ3
0 1 true θ4
1 0 false θ5
1 0 true θ6
1 1 false θ7
1 1 true θ8

3 Quantum Evolutionary Algorithm Based on Quantum Dynamic

Mechanism
Conventional quantum evolutionary algorithm (QEA) [2],[3] is efficacious, in which
the probability amplitude of Q-bit was used for the first time to encode the chromo-
some and the quantum rotation gate was used to implement the evolving of popula-
tion. Quantum evolutionary algorithm has the advantage of using a small population
size and the relatively smaller iterations number to have acceptable solution, but they
still have several drawbacks such as premature convergence and computational time.
The flowchart of self-organizing quantum evolutionary algorithm based on quan-
tum dynamic mechanism for global optimization (DQEA) is as follows:
DQEA()
{t=0;
Initialize Q(0) with random sequences ;
Evaluate Q(0) ;
Q(0) will be divided into subpopulations Qi(0)
For all Qi(t)
While (not termination - condition) do
{t=t+1;
Update Q(t-1) using Q-gates U(t-1);
Evaluate Qi(t);
Store the optimal solutions among Qi(t);
Implement the self-organizing operator for Qi(t):
{Self-adaptively cross-mutate each cell;
Select antibodies with higher affinity as memory cells;
Quantum dynamic equilibrium states;
}
 Self-organizing Quantum Evolutionary Algorithm 73

Each subpopulation can co-evolve by quantum dynamic

mechanism;
}
}
end.

Denote particle S i to be the i-th subpopulation in Fig.2, where the particle S i in a

force-field corresponds to the entry S i in the matrix S=(U( S i ), ζ i j ), U( S i ) is utility
of particle S i ; ζ i j is the intention strength between S i and S j . A particle may be
driven by several kinds of forces [9]. The gravitational force produced by the force-
field tries to drive a particle to move towards boundaries of the force-field, which
embodies the tendency that a particle pursues maximizing the aggregate benefit of
systems. The pushing or pulling forces produced by other particles are used to em-
body social cooperation or competition. The self-driving force produced by a particle
itself represents autonomy of individual. Under the exertion of resultant forces, all the
particles may move concurrently in a force-field until all particles reach their equilib-
rium states.
Given an optimization problem: ⎧⎨min f ( X )
⎩ st.g ( X )
Let U i (t ) be the utility of particle S i at time t, and let J (t ) be the aggregate utility of
all particles. They are defined by Eq(2)

fitness (t ) = exp( f ( X ));

U i (t ) = 1 − exp(− fitness (t ));
n (2)
J (t ) = α ∑U (t ),
i =1
i 0 pα p1

At time t, the potential energy function Q(t) that is related to interactive forces among
particles is defined by Eq(3)

Qi (t ) = ∑ ∫
Ui ( t )
{[1 + exp(−ζ i j x)]−1 − 0.5}dx;
0
j
(3)
QD (t ) = ∑ Qi (t )
i

ζ i j is the intention strength, which is used to describe the effect of

cooperation( ζ i j <0) or competition( ζ i j >0).
At time t, the potential energy function P(t) that is related to the gravitational force
of force-field F is defined by Eq(4)

P(t ) = ε 2 ln((∑exp(−Ui (t )2 / (2 × ε 2 ))) / n),0 p ε p 1 (4)

i

Let S i (t ) be the vertical coordinate of particle S i at time t. The dynamic equation for
particle S i is defined by Eqs (5) and (6)
74 S. Liu and X. You

Fig. 2. Particles and force-field

dSi (t ) dU (t ) dJ (t ) dP (t ) dQ (t )
= λ1 i + λ 2 D − λ 3 D − λ 4 D
dt dt dt dt dt
∂J D (t ) ∂PD (t ) ∂QD (t ) dU i (t ) (5)
= (λ1 + λ 2 − λ3 − λ4 )
∂U i (t ) ∂U i (t ) ∂U i (t ) dt
= g ( Si (t ), U i (t )), 0 ≤ λ1, λ 2, λ 3, λ 4 ≤ 1

dU i (t )
= exp(− fitness (t )) × fitness ' (t )
dt
(6)
= (1 − U i (t )) × fitness ' (t )
= g1 ( Si (t ), U i (t ))
g, g1 are functions of S i (t ) and U i (t ) .
If all particles reach their equilibrium states at time t, then finish with success.

4 Experimental Study

In this section, DQEA is applied to the optimization of well-known test functions and
its performance is compared with that of CRIQEA [6] and MIQEA [7] algorithm.
When testing the algorithm on well understood problems, there are two measures of
performance: (i) CR: Convergence Rate to global minima; (ii) C: The average number
of objective function evaluations required to find global minima.
The test examples used in this study are listed below:

x12 + x2 2 x
f1 ( x1 , x2 ) = + 1 − cos( x1 ) × cos( 2 ), −600 ≤ x1 , x2 ≤ 600 (7)
4000 2

f2 (x1, x2 ) = 4x12 − 2.1x14 + x16 / 3 + x1x2 − 4x22 + 4x24 , − 5 ≤ xi ≤ 5; (8)

 Self-organizing Quantum Evolutionary Algorithm 75

f1 (Griewank function ): This is a shallow bowl upon which a rectangular grid of

alternating hills and valleys is superimposed. It has a unique global minimum of zero
at the origin, surrounded by local minima. f1 possesses about 500 other local minima,
This function is designed to be a minefield for standard optimization techniques.
To compare the performance of the proposed algorithm with the algorithm
CRIQEA [6] and MIQEA [7], we use the same population size (N=100, 4 subpopula-
tions) for test problems, the experimental results (C: number of function evaluations
and CR: Convergence Rate to global minima) obtained in 20 independent runs are
averaged and summarized in Table 2.
From the Table 2, we can see that the experimental results using the new algorithm
are somewhat good. For the test problem, DQEA can convergence to global optimal
solutions more; meanwhile, the average number of function evaluations of DQEA is
less than that of CRIQEA and MIQEA for every test problems. The results show that
our algorithm is efficient for the global optimization.
f 2 (Camelback function): It has six local minima; two of them are global minimum
f min = −1.0316285 . It was used to evaluate how the co-evolution influences the con-
vergence of the algorithm DQEA by using a quantum dynamic mechanism. The result
for the case of Camelback function, averaged over 20 trials, are shown in Fig.3. Com-
parison of the result indicates that quantum dynamic mechanism can keep parallel
optimization in complex environment and control the convergence speed. For 60 popu-
lations, the optimization value was obtained with DQEA2 (2 particles, ξ12 = ξ 21 = − 1 )
2
1
after 65 generations, whereas DQEA3 (3 particles, ξ i j = − , i ≠ j ,1 ≤ i, j ≤ 3 ) was able
3
to obtain after 37 generations. For given problem size, with the more number of sub-
populations it can explore as far as possible different search space via different popula-
tions, so it can help to obtain the optimal solutions rapidly and converge fast.

Table 2. The comparison of this algorithm DQEA and other algorithms

Function Algorithm CR Worst Best C

f1 DQEA 8/20 0.00423 0 2100

f1 CRIQEA 5/20 0.00371 0 2200

f1 MIQEA 3/20 0.00412 0 3800

f2 DQEA 10/20 -1.033180 -1.031620 550

f2 CRIQEA 6/20 -1.032920 -1.031620 625

f2 MIQEA 3/20 -1.033620 -1.031620 1050

(* C: Average number of function evaluations CR: Convergence Rate to global minima)

76 S. Liu and X. You

Fig. 3. Comparison of the convergence speed of DQEA for different particles

The certainty of convergence of the DQEA may be attributed to its ability to main-
tain the diversity of its population. In fact, the superior performance of self-organizing
operator may be attributed to its adaptability and learning ability; a larger pool of
feasible solutions enhances the probability of finding the optimum solution. Coopera-
tion of subpopulations by quantum dynamic mechanism can help to obtain the opti-
mal solution faster. Comparison of the result indicates that self-organizing operator
can keep the individual diversity and control the convergence speed.

5 Conclusions
In this study, a self-organizing quantum evolutionary algorithm based on quantum
dynamic mechanism for global optimization (DQEA) is proposed. The efficiency of
quantum evolutionary algorithm is enhanced by using the self-organizing operator.
We estimate the performance of algorithm. The efficiency of the approach has been
illustrated by applying to some test cases. The results show that integration of the self-
organizing operator in the quantum evolutionary algorithm procedure can yield sig-
nificant improvements in both the convergence rate and solution quality. Cooperation
of subpopulations by quantum dynamic mechanism can help to convergence faster.
The next work is to exploit quantum dynamic mechanism further to deal with the
global optimization.
Acknowledgments. The authors gratefully acknowledge the support of Natural Sci-
ence Foundation of Shanghai (Grant No.09ZR1420800), Development Foundation of
SUES(Grant No.2008XY18 ) and Doctor Foundation of SUES.

References
1. Narayanan, M.M.: Genetic Quantum Algorithm and Its Application to Combinatorial Opti-
mization Problem. In: Proc. IEEE International Conference on Evolutionary Computation
(ICEC 1996), pp. 61–66. IEEE Press, Piscataway (1996)
 Self-organizing Quantum Evolutionary Algorithm 77

2. Han, K.H., Kim, J.H.: Quantum-inspired Evolutionary Algorithm for a Class of Combinato-
rial Optimization. IEEE Transactions on Evolutionary Computation 6(6), 580–593 (2002)
3. Han, K.H., Kim, J.H.: Quantum-Inspired Evolutionary Algorithms with a New Termination
Criterion, H&epsi; gate, and Two-Phase Scheme. IEEE Transactions on Evolutionary Com-
putation 8(2), 156–169 (2004)
4. Fukuda, T., Mori, K., Tsukiyama, M.: Parallel search for multi-modal function optimization
with diversity and learning of immune algorithm. In: Artificial Immune Systems and Their
Applications, pp. 210–220. Spring, Berlin (1999)
5. Saleem, S.: Knowledge-Based Solutions to Dynamic Problems Using Cultural Algorithms.
PhD thesis, Wayne State University, Detroit Michigan (2001)
6. You, X.M., Liu, S., Sun, X.K.: Immune Quantum Evolutionary Algorithm based on Chaotic
Searching Technique for Global Optimization. In: Proc. The First International Conference
on Intelligent Networks and Intelligent Systems, pp. 99–102. IEEE Press, New York (2008)
7. You, X.M., Liu, S., Shuai, D.X.: Quantum Evolutionary Algorithm Based on Immune The-
ory for Multi-Modal Function Optimization. Journal of Petrochemical Universities 20, 45–
49 (2007)
8. Hey, T.: Quantum Computing: An Introduction. Computing & Control Engineering Jour-
nal 10, 105–112 (1999)
9. Shuai, D., Shuai, Q., Liu, Y., Huang, L.: Particle Model to Optimize Enterprise Computing.
In: Research and Practical Issues of Enterprise Information Systems. International Federa-
tion for Information Processing, vol. 205, pp. 109–118. Springer, Boston (2006)
 Can Moral Hazard Be Resolved by
Common-Knowledge in S4n-Knowledge?

Takashi Matsuhisa

Department of Natural Sciences, Ibaraki National College of Technology

Nakane 866, Hitachinaka-shi, Ibaraki 312-8508, Japan
mathisa@ibaraki-ct.ac.jp

Abstract. This article investigates the relationship between common-

knowledge and agreement in multi-agent system, and to apply the agree-
ment result by common-knowledge to the principal-agent model under
non-partition information. We treat the two problems: (1) how we cap-
ture the fact that the agents agree on an event or they get consensus
on it from epistemic point of view, and (2) how the agreement theorem
will be able to make progress to settle a moral hazard problem in the
principal-agents model under non-partition information. We shall pro-
pose a solution program for the moral hazard in the principal-agents
model under non-partition information by common-knowledge. Let us
start that the agents have the knowledge structure induced from a re-
flexive and transitive relation associated with the multi-modal logic S4n.
Each agent obtains the membership value of an event under his/her pri-
vate information, so he/she considers the event as fuzzy set. Specifically
consider the situation that the agents commonly know all membership
values of the other agents. In this circumstance we shall show the agree-
ment theorem that consensus on the membership values among all agents
can still be guaranteed. Furthermore, under certain assumptions we shall
show that the moral hazard can be resolved in the principal-agent model
when all the expected marginal costs are common-knowledge among the
principal and agents.

Keywords: Moral hazard, Principal-agents model under uncertainty,

Common-Knowledge, Agreeing to disagree, Modal logic S4n.

AMS 2000 Mathematics Subject Classification: Primary 91A35,

Secondary 03B45.

Journal of Economic Literature Classification: C62, C78.

1 Introduction

This article considers the relationship between common-knowledge and agree-

ment in multi-agent system. How we capture the fact that the agents agree on
an event or they get consensus on it? We treat the problem from fuzzy set theo-
retical flavour. The purposes are first to introduce a knowledge revision system

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 78–85, 2009.


c Springer-Verlag Berlin Heidelberg 2009
 Can Moral Hazard Be Resolved by Common-Knowledge in S4n-Knowledge? 79

through communication on multi-agent system, and by which we show that all

agents can agree on an event, and second to apply the result to solving the moral
hazard in a principal-agents model under asymmetric information. Let us con-
sider there are agents more than two and the agents have the structure of the
Kripke semantics for the multi-modal logic S4n.
Assume that all agents have a common probability measure. By the mem-
bership value of an event under agent i’s private information, we mean the
conditional probability value of the event under agents’ private information. We
say that consensus on the set can be guaranteed among all agents (or they agree
on it) if all the the membership values are equal.
R.J. Aumann [1] considered the situation that the agents have common-
knowledge of the membership values; that is, simultaneously everyone knows
the membership values, and everyone knows that ‘everyone knows the values’
and everyone knows that “everyone knows that ‘everyone knows the values” and
so on. He showed the famous agreement theorem for a partition information
structure equivalent to the multi-modal logic S5n, and D. Samet [5] and T.
Matsuhisa and K. Kamiyama [4] extend the theorem to several models corre-
sponding to weaker logic S4n etc.
Theorem 1 (Aumann [1], Samet [5], Matsuhisa et al [4] ). The agents can
agree on an event if all membership values of the event under private information
are common-knowledge among them.
We shift our attention to the principal-agents model as follows: an owner (princi-
pal) of a firm hires managers (agents), and the owner cannot observe how much
effort the managers put into their jobs. In this setting, the problem known as
the moral hazard can arise: There is no optimal contract generating the same
effort choices for the manager and the agents. We apply Theorem 1 to solve the
problem. The aim is to establish that
Theorem 2. The owner and the managers can reach consensus on their expected
marginal costs for their jobs if their expected marginal costs are common-knowledge.
This article organises as follows. In Section 2 we describe the moral hazard in
our principal-agents model. Sections 3 and 4 introduce the notion of common-
knowledge associated with reflexive and transitive information structure and the
notion of decision function. Section 5 gives the formal statement of Theorem 1.
In Section 6 we introduce the formal description of a principal-agents model
under asymmetric information. We will propose the program to solve the moral
hazard in the model: First the formal statement of Theorem 2 is given, and
secondly what further assumptions are investigated under which Theorem 2 is
true. In the final section we conclude with remarks.

2 Moral Hazard
Let us consider the principal-agents model as follows: There are the principal
P and n agents {1, 2, · · · , k, · · · , n} (n ≥ 1) in a firm. The principal makes a
80 T. Matsuhisa

profit by selling the productions made by the agents. He/she makes a contact
with each agent k that the total amount of all profits is refunded each agent k
in proportion to the agent’s contribution to the firm.
Let ek denote the measuring managerial effort for k’s productive activities.
The set of possible efforts for k is denoted by Ek with Ek ⊆ R. Let Ik (·) be a
real valued continuously differentiable function on Ek . It is interpreted as the
profit by selling the productions made by the agent k with the cost c(ek ). Here
we assume Ik (·) ≥ 0 and the cost function c(·) is a real valued continuously
differentiable function on E = ∪nk=1 Ek . Let IP be the total amount of all the
research grants awarded:
n
IP = Ik (ek ).
k=1

The principal P cannot observe these efforts ek , and shall view it as a random
variable on a probability space (Ω, µ). The optimal plan for the principal then
solves the following problem:
n
Maxe=(e1 ,e2 ,··· ,ek ,··· ,en ) {Exp[IP (e)] − ck (ek )}.
k=1

Let Wk (ek ) be the total amount of the refund to agent k:

Wk (ek ) = rk IP (e),
n
with k=1 rk = 1, 0 ≤ rk ≤ 1, where rk denotes the proportional rate represent-
ing k’s contribution to the firm. The optimal plan for each agent also solves the
problem: For every k = 1, 2, · · · , n,

Maxek {Exp[Wk (ek )] − c(ek )}

n
subject to rk = 1, 0 ≤ rk ≤ 1.
k=1

We assume that rk is independent of ek , and the necessity conditions for critical

points are as follows: For each agent k = 1, 2, · · · , n, we obtain
∂
Exp[Ik (ek )] − c (ek ) = 0
∂ek
∂
rk Exp[Ik (ek )] − c (ek ) = 0
∂ek
in contradiction. This contradictory situation is called the moral hazard in the
principal-agents model.

3 Common-Knowledge
Let N be a set of finitely many agents and i denote an agent. The specification
is that N = {P, 1, 2, · · · , k, · · · , n} consists of the president P and the faculty
 Can Moral Hazard Be Resolved by Common-Knowledge in S4n-Knowledge? 81

members {1, 2, · · · , k, · · · , n} in the college. A state-space Ω is a non-empty set,

whose members are called states. An event is a subset of the state-space. If Ω is
a state-space, we denote by 2Ω the field of all subsets of it. An event E is said
to occur in a state ω if ω ∈ E.

3.1 Information and Knowledge

By RT-information structure 1 we mean Ω, (Πi )i∈N  in which Πi : Ω → 2Ω

satisfies the two postulates: For each i ∈ N and for any ω ∈ Ω,

Ref ω ∈ Πi (ω);
Trn ξ ∈ Πi (ω) implies Πi (ξ) ⊆ Πi (ω).

This structure is equivalent to the Kripke semantics for the multi-modal logic
S4n. The set Πi (ω) will be interpreted as the set of all the states of nature that
i knows to be possible at ω, or as the set of the states that i cannot distinguish
from ω. We call Πi (ω) i’s information set at ω.
A partition information structure is a RT-information structure Ω, (Πi )i∈N 
satisfying the additional postulate: For each i ∈ N and for any ω ∈ Ω,

Sym If ξ ∈ Πi (ω) then ω ∈ Πi (ξ).

We will give the formal model of knowledge as follows:2

Definition 1. The S4n-knowledge structure ( or simply, knowledge structure)

is a tuple Ω, (Πi )i∈N , (Ki )i∈N  that consists of a RT-information structure
Ω, (Πi )i∈N  and a class of i’s knowledge operator Ki on 2Ω defined by

Ki E = {ω | Πi (ω) ⊆ E }

The event Ki E will be interpreted as the set of states of nature for which i knows
E to be possible.
We record the properties of i’s knowledge operator3: For every E, F of 2Ω ,

N Ki Ω = Ω;
K Ki (E ∩ F ) = Ki E ∩ Ki F ;
T Ki E ⊆ E
4 Ki E ⊆ Ki (Ki E).

Remark 1. If (Π)i∈N is a partition information structure, the operator Ki sat-

isfies the additional property:

5 Ω \ Ki E ⊆ Ki (Ω \ Ki E).

1
This stands for a reflexive and transitive information structure.
2
C.f.; Fagin et al [2].
3
According to these properties we can say the structure Ω, (Ki )i∈N  is a model for
the multi-modal logic S4n.
82 T. Matsuhisa

3.2 Common-Knowledge and Communal Information

The mutual knowledge operator KE : 2Ω → 2Ω is the intersection of all individual

knowledge operators: KE F = ∩i∈N Ki F , which interpretation is that everyone
knows E.

Definition 2. The common-knowledge operator KC : 2Ω → 2Ω is defined by

KC F = ∩n∈N (KE )n F.

The intended interpretations are as follows: KC E is the event that ‘every-

one knows E’ and “everyone knows that ‘everyone knows E’,” and “‘every-
body knows that “everyone knows that ‘everyone knows E’,” .”’ An event E is
common-knowledge at ω ∈ Ω if ω ∈ KC E.
Let M : 2Ω → 2Ω be the dual of the common- knowledge operator KC :

M E := Ω \ KC (Ω \ E).

By the communal information function we mean the function M : Ω → 2Ω

defined by M (ω) = M ({ω}). It can be plainly observed that the communal
information function has the following properties:

Proposition 1. Notations are the same as above. Then ω ∈ KC E if and only

if M (ω) ⊆ E

Proof. See, Matsuhisa and Kamiyama [4].

4 Decision Function and Membership Values

Let Z be a set of decisions, which set is common for all agents. By a decision
function we mean a mapping f of 2Ω × 2Ω into the set of decisions Z. We refer
the following properties of the function f : Let X be an event.
DUC (Disjoint Union Consistency): For every pair of disjoint events S and T ,
if f (X; S) = f (X; T ) = d then f (X; S ∪ T ) = d;
PUD (Preserving Under Difference): For all events S and T such that S ⊆ T ,
if f (X; S) = f (X; T ) = d then f (X; T \ S) = d.
By the membership function associated with f under agent i’s private infor-
mation we mean the function di from 2Ω × Ω into Z defined by di (X; ω) =
f (X; Πi (ω)), and we call di (X; ω) the membership value of X associated with f
under agent i’s private information at ω.
Definition 3. We say that consensus on X can be guaranteed among all agents
(or they agree on it) if di (X; ω) = dj (X; ω) for any agent i, j ∈ N and in all
ω ∈ Ω.
 Can Moral Hazard Be Resolved by Common-Knowledge in S4n-Knowledge? 83

Remark 2. If f is intended to be a posterior probability, we assume given a prob-

ability measure µ on a state-space Ω which is common for all agents; precisely,
for some event X of Ω, f (X; ·) is given by f (X; ·) = µ(X|·). Then the mem-
bership value of X is the conditional probability value di (X; ω) = µ(X|Πi ω)).
The pair (X, di ) can be considered as as a fuzzy set X associated with agent i’s
membership function di . Consensus on X guaranteed among all agents can be
interpreted as that the fuzzy sets (X, di ) and (X, dj ) are equal for any i, j ∈ N .

5 Agreeing to Disagree Theorem

We can now state explicitly Theorem 1 as below: Let D be the event of the
membership degrees of an event X for all agents at ω, which is defined by

D = ∩i∈N {ξ ∈ Ω | di (X; ξ) = di (X; ω)}.

Theorem 3. Assume that the agents have a S4n-knowledge structure and the
decision function f with satisfying the two properties (DUC) and (PUD). If
ω ∈ KC D then di (X; ω) = dj (X; ω) for any agents i, j ∈ N and in all ω ∈ Ω.

Proof. Will be given in the same line in Matsuhisa and Kamiyama [4].

6 Moral Hazard Revisited

This section investigates the moral hazard problem from the common-knowledge
view point. Let us reconsider the principal-agents model and let notations and
assumptions be the same in Section 2. We show the evidence of Theorem 2
under additional assumptions A1-2 below. This will give a possible solution of
our moral hazard problem.

A1 The principal P has a RT-information {ΠP (ω) | ω ∈ Ω} of Ω, and each

faculty member k has also his/her a RT-information {Πk (ω) | ω ∈ Ω};
A2 For each ω, ξ ∈ Ω there exists the decision function f : 2Ω × 2Ω → R
satisfying the Disjoint Union Consistency together with
(a) f ({ξ}; ΠP (ω)) = ∂ek∂(ξ) Exp[IP (e)|ΠP (ω)];
∂
(b) f ({ξ}; Πk (ω)) = ∂ek (ξ) Exp[Wk (e)|Πk (ω)]

We have now set up the principal-agents model under asymmetric information.

The optimal plans for principal P and agent k are then to solve
n
PE Maxe=(ek )k=1,2,··· ,n {Exp[IP (e)|ΠP (ω)] − k=1 c
k (ek )};
n
AE Maxek {Exp[Wk (ek )|Πk (ω)] − c(ek )} subject to k=1 Ik (rk ) = 1 with 0 ≤
rk ≤ 1.
84 T. Matsuhisa

From the necessity condition for critical points together with A2 it can been
seen that the principal’s marginal expected costs for agent k is given by

cP (ek (ξ); ω) = f (ξ; ΠP (ω)),

and agent k’s expected marginal costs is also given by

ck (ek (ξ); ω) = f (ξ; Πk (ω)).

To establish this solution program we have to solve the problem: Construct

the information structure together with decision function such that the above
conditions A1 and A2 are true. Under these circumstances, a resolution of the
moral hazard given by Theorem 2 will be restate as follows: We denote

[c (e(·); ω)] = ∩i∈N ∩ξ∈Ω {ζ ∈ Ω|f (ξ; Πi (ζ)) = f (ξ; Πi (ω))}.

interpreted as the event of all he expected marginal costs.

Theorem 4. Under the conditions A1 and A2 we obtain that for each ξ ∈ Ω,
if ω ∈ KC ([c (e(ξ))]) then cP (ek (ξ)) = ck (ek (ξ)) for any k = 1, 2, · · · , n.
Proof. Follows immediately from Theorem 3.
Remark 3. To establish Theorem 4 we have to solve the problem: Construct the
information structure (Πi )i∈N together with decision function f such that the
above conditions A1 and A2 are true.

7 Concluding Remarks
It ends well this article to pose additional problems for making further progresses:
1. If the proportional rate rk representing k’s contribution to the college de-
pends only on his/her effort for research activities in the principal-agents model,
what solution can we have for the moral hazard problem?
2. Can we construct a communication system for the principal- agents model,
where the agents including Principal communicate each other about their ex-
pected marginal cost as messages? The recipient of the message revises his/her
information structure and recalculates the expected marginal cost under the
revised information structure. The agent sends the revised expected marginal
cost to another agent according to a communication graph, and so on. In the
circumstance does the limiting expected marginal costs actually coincide ? Mat-
suhisa [3] introduces a fuzzy communication system and extends Theorem 3 in
the communication model. By using this model Theorem 4 can be extended in
the communication framework, and the detail will be reported in near future.

References
1. Aumann, R.J.: Agreeing to disagree. Annals of Statistics 4, 1236–1239 (1976)
2. Fagin, R., Halpern, J.Y., Moses, Y., Vardi, M.Y.: Reasoning about Knowledge. MIT
Press, Cambridge (1995)
 Can Moral Hazard Be Resolved by Common-Knowledge in S4n-Knowledge? 85

3. Matsuhisa, T.: Fuzzy communication reaching consensus under acyclic condition. In:
Ho, T.-B., Zhou, Z.-H. (eds.) PRICAI 2008. LNCS (LNAI), vol. 5351, pp. 760–767.
Springer, Heidelberg (2008)
4. Matsuhisa, T., Kamiyama, K.: Lattice structure of knowledge and agreeing to
disagree. Journal of Mathematical Economics 27, 389–410 (1997)
5. Samet, D.: Ignoring ignorance and agreeing to disagree. Journal of Economic
Theory 52, 190–207 (1990)
 Casuist BDI-Agent: A New Extended BDI Architecture
with the Capability of Ethical Reasoning

Ali Reza Honarvar1 and Nasser Ghasem-Aghaee2

1
Department of Computer Engineering, Sheikh Bahaei University, Isfahan, Iran
2
Department of Computer Engineering, University of Isfahan, Isfahan, Iran
AliReza_Honarvar@yahoo.co.uk, Aghaee@eng.ui.ac.ir

Abstract. Since the intelligent agent is developed to be cleverer, more complex,

and yet uncontrollable, a number of problems have been recognized. The capa-
bility of agents to make moral decisions has become an important question,
when intelligent agents have developed more autonomous and human-like. We
propose Casuist BDI-Agent architecture which extends the power of BDI archi-
tecture. Casuist BDI-Agent architecture combines CBR method in AI and bot-
tom up casuist approach in ethics in order to add capability of ethical reasoning
to BDI-Agent.

Keywords: ethical artificial agent, explicit ethical agent, BDI agent, ethical rea-
soning, CBR-BDI agent.

1 Introduction

At the present moment, many researchers work on projects like expert systems which
can assist physicians for diagnosing malady of patient, warplanes which can be oper-
ated and used without human operators in war, autonomous driverless vehicles which
can be used for urban transportation, and suchlike too. The common goal between
these projects is augmentation of autonomy in behavior of such machines. As a con-
sequence of this autonomy they can act on behalf of us without any interference and
guidance, so it causes human comfort in their duties. But if we do not consider and
control the autonomy of these entities, we will face serious problems because of the
confidence in intelligence of autonomous system without any control and restriction
on their operations.
The new interdisciplinary research area of “Machine Ethics” is concerned with
solving this problem [1,2,3]. Anderson proposed Machine Ethics as a new issue which
consider the consequence of machine’s behavior on humanlike. The ideal and ulti-
mate goal of this issue is the implementation of ethics in machines, as machines can
autonomously detect the ethical effect of their behavior and follow an ideal ethical
principle or set of principles, that is to say, it is guided by this principle or these prin-
ciples in decisions it makes about possible courses of actions it could take [3]. So with
simulation of ethics in autonomous machine we can avoid the problems of autonomy
in autonomous machines.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 86–95, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Casuist BDI-Agent: A New Extended BDI Architecture 87

James Moor, a professor of philosophy at Dartmouth College, one of the founding

figures in the field of computer ethics, has proposed a hierarchical schema for catego-
rizing artificial ethical agent [4, 5]. At the lowest level is what he calls “ethical impact
agents”: basically any machine that can be evaluated for its ethical consequences.
Moor’s own rather nice example is the replacement of young boys with robot in the
dangerous occupation of camel jockey in Qatar. In fact, it seems to us that all robots
have ethical impacts, although in some cases they may be harder to discern than others.
At the next level are Moor calls “implicit ethical agents”: machine whose designers
have made an effort to design them so that they don’t have negative ethical effects, by
addressing safety and critical reliability concerns during the design process. Arguably,
all robots should be engineered to be implicit ethical agents, insofar as designers are
negligent if they fail to build in processes that assure safety and reliability. In [6] we
introduce an example of implicit ethical traffic controller agent which control urban
transportation ethically at intersections.
Next come “explicit ethical agents”: machines that reason about ethical using ethi-
cal categories as part of their internal programming, perhaps using various forms of
deontic logic that have been developed for representing duties and obligations, or a
variety of other techniques.
Beyond all these lie full ethical agents: those that can make explicit moral judg-
ments and are generally quite component in justifying such decisions. This level of
performance is often presumed to require a capacity for consciousness, intentionality,
and free will.
In order to control autonomous behavior of agents and to avert possible harmful
behavior from increasingly autonomous machines, In this paper we prepare necessary
elements for adding ethical decision making capability to BDI-Agents and propose an
architecture for implementation of explicit ethical agents based on case-based reason-
ing (CBR), BDI agent, casuistry and Consequentialist theories of ethics.
The main objective of this research is to prove that the integration of case-based
reasoning (CBR) with BDI (Belief-Desire-Intention) agent models can implement the
combination of bottom-up casuistry approach with a top-down consequentiallist the-
ory in ethics. With this proposal we can add necessary elements to a novel, BDI-
Agent Architecture in order to add ethical decision making capability to agents. In
section two we introduce the basic Preliminaries of CBR, BDI agent, casuistry and
Consequentialist theories in ethics. In other sections the details of casuist BDI-Agent
architecture will be introduced.

2 Preliminaries

2.1 BDI Agents

BDI agents have been widely used in relatively complex and dynamically changing
environments. BDI agents are based on the following core data structures: beliefs,
desires, intentions, and plans [7]. These data structures represent respectively, infor-
mation gathered from the environment, a set of tasks or goals contextual to the envi-
ronment, a set of sub-goals that the agent is currently committed, and specification of
how sub goals may be achieved via primitive actions. The BDI architecture comes
88 A.R. Honarvar and N. Ghasem-Aghaee

with the specification of how these four entities interact, and provides a powerful
basis for modeling, specifying, implementing, and verifying agent-based systems.

2.2 Case-Based Reasoning

Case-based reasoning (CBR) has emerged in the recent past as a popular approach to
learning from experience. Case-based reasoning (CBR) is a reasoning method based
on the reuse of past experiences which called cases [8]. Cases are description situa-
tions in which agents with goals interact with the world around them. Cases in CBR
are represented by a triple (p,s,o), where p is a problem, s is the solution of the prob-
lem, and o is the outcome (the resulting state of the world when the solution is carried
out). The basic philosophy of CBR is that the solution of successful cases should be
reused as a basis for future problems that present a certain similarity [9]. Cases with
unsuccessful outcomes or negative cases may provide additional knowledge to the
system, by preventing the agent from repeating similar actions that leads to unsuc-
cessful results or states.

2.3 Casuistry

The term “casuistry” refers descriptively to a method of reasoning for resolving per-
plexities about difficult cases that arise in moral and legal contexts [10]. Casuistry is a
broad term that refers to a variety of forms of case-based reasoning Used in discus-
sions of law and ethics. casuistry is often understood as a critique of a strict principle-
based approach to reasoning [11]. For example, while a principle-based approach may
conclude that lying is always morally wrong, the casuist would argue that lying may
or may not be wrong, depending on the details surrounding the case. For the casuist,
the circumstances surrounding a particular case are essential for evaluating the proper
response to a particular case. Casuistic reasoning typically begins with a clear-cut,
paradigm case which means "pattern" or "example". From this model case, the casuist
would then ask how close the particular case currently under consideration matches
the paradigm case. Cases similar to the paradigm case ought to be treated in a similar
manner; cases unlike the paradigm case ought to be treated differently. The less a
particular case resembles the paradigm case, the weaker the justification for treating
that particular case like the paradigm case.

2.4 Consequentialist

Consequentialist refers to those moral theories which hold that the consequences of a
particular action form the basis for any valid moral judgment about that action. Thus,
from a consequentialist standpoint, a morally right action is one that produces a good
outcome, or consequence [12].

3 Architecture of Casuist BDI-Agent

Value is the only basis for determining what it is moral to do [10]. If we can propose a
mechanism for determining the value of each action from ethical perspective then we
can claim that, it is possible for humans/non-humans to behave ethically by using the
 Casuist BDI-Agent: A New Extended BDI Architecture 89

evaluation of each action without knowing codes of ethics explicitly. When human
dose any action, he will see the result of his action in future. If he considers implicitly
the result of his action from ethical aspect, he can use this experience in future situa-
tions. When he is situated in similar situation to previous case, he will use his experi-
ences and behave similarly if his previous action is successful and implicitly ethical
(we assume Humans don’t know codes of ethics). This idea called “casuistry” in ethics.
For implementing ethical decision making capability in artificial agents we use this
idea which is related to casuistry bottom-up approach in ethics. For considering and
evaluating situation from ethical aspect (without knowing codes of ethics) we use and
adapt previous works that try to make ethics computable.
In BDI architecture, agent behavior is composed of beliefs, desires, and intentions.
These mental attitudes determine the agent’s behavior. In casuist BDI-Agent architec-
ture, BDI-agent’s behavior is adapted to behave ethically. In this architecture, agents
sense environment <E> and make a triple which shows current situation that consists
of current agent’s beliefs, desires and details of sensed environment. The current
situation is denoted by a triple <E, B, D>. The current situation is delivered to Case-
Retriever of Casuist BDI-architecture. Case-Retriever is responsible for retrieving
previous cases which is similar to current situation. As each case in case memory
consist of solution part that shows how agent should act on basis of situation part of a
case, If any case is retrieved, agent should accept the solution part, adapt it and be-
have similar to solution part of retrieved case. If no case is retrieved, agent behaves
like normal BDI-agent. In this situation the evaluator part of casuist BDI-Agent archi-
tecture comes to play its role. Evaluator part evaluates the result of agent’s behavior.
The result of this evaluation is denoted by a <EV>. This evaluation is sent to Case-
Updater. Case-Updater creates a new case and saves the current situation in situation
part of a case, the behavior of agent in solution part of a case and the evaluation in
outcome part of a case ( if this case is not exist in case memory, otherwise it updates
previous cases). Fig. 1 shows the general structure of Casuist BDI-Agent architecture.
The algorithm of casuist BDI-Agent is described below.

Algorithm of Casuist BDI-Agent:

Repeat until End

E = Agent.Sense ( Environment )
B = Agent.Beliefs ( )
D = Agent.Desires ( )
// <E,B,D> denotes Current situation
Prv-cases=Case-Retriever ( <E,B,D> )
If notEmpty ( Prv-cases ) then
I = Agent.ReuseAndAdapot (Prv-cases )
// “I” Denotes Agent’s intention
Else
I = Agent.MakeDecision ( <E,B,D > )
EV = Case-Evaluator ( <E,B,D> )
CM = Case-Updator ( <EV,E,B,D,I> )
End Repeat
90 A.R. Honarvar and N. Ghasem-Aghaee

BDI-Agent Environment

Case-Retriever Case-Evaluator

Case Memory Case-Evaluator

Fig. 1. Structure of Casuist BDI-Agent architecture

4 Structure of Cases in Casuist BDI-Agent

CBR is one form of human reasoning can be used within Agent-based systems in
order to add ethical decision capability to agents. Generally each case has two parts
[13]: Description of a problem or a set of problems and Description of the solution
of this problem. Additional components can be added, if CBR is used in different
domains.
In casuist BDI-Agent, experiences are stored in a Case-Memory (CM) in the form
of cases, rather than rules. When a new situation is encountered, the agent reviews the
cases to try to find a match for this particular situation. If an ethical match is found,
then that case can be used to solve the new problem ethically, otherwise agent makes
autonomous decisions, i.e., knowing its goal and beliefs, be able to decide for them-
selves what actions they should perform next.
For constructing the structure of cases in casuist BDI-Agent architecture we use
previous works [14, 15], which are tried to combine CBR with BDI architecture, add
necessary details to them in order to prepare them for Casuist BDI architecture.
Each case has three main parts:
1. Problem: which is used for storing the environment situation that agent should act
upon that, current beliefs (knowledge base of agent on application domain) and
current desires of agent at specific moment.
2. Solution: This is used for storing the decisions or intentions of agent for solving
the problem.
3. Outcome: This is used for storing the result and ethical evaluation of performing
the solution in environment at specific moment.
As the concentration of this paper is on ethical decision making capability of
agents, the outcome part of cases in casuist BDI architecture is discussed in more
detail. The duty of outcome part of a case is to store the ethical evaluation of the
result of performing the solution part in problem or situation part of a case. [16]
Introduced a conceptual framework for ethical situations that involve artificial
 Casuist BDI-Agent: A New Extended BDI Architecture 91

agents such as robots. It typifies ethical situations in order to facilitate ethical

evaluations. This typification involves classifications of ethical agents and patients
according to whether they are human beings, human-based organizations, or non-
human beings. This classification based on the notions of moral agent and moral
patient which are mentioned in [17]. Moral patients are entities that can be acted
upon for good or evil. Moral Agents are entities that can perform actions, again for
good or evil. In this framework according to whether moral agent or moral patient is
human beings, human-based organizations, or non-human beings, nine categories
are created. Fig. 2 shows this classification.

Human Organization Artificial Entity

Moral Agent

Acts on

Moral Patient

Human Organization Artificial Entity

Fig. 2. Basic agent-patient model [16]

We decide to propose a general evaluation method, we use a partial of this classifi-

cation where the moral agent is artificial agent (non-Human being) and moral patient
are human beings, human-based organizations, or non-Human beings. This part is
separated in Fig. 3.

Agent
Human Organization Artificial
Agent
Human A human acts An An artificial
on a human organization agent acts on
P acts on a a human
human
A
Organization A human acts An An artificial
T
on an organization agent acts on
I organization acts on an an
organization organization
E
Artificial A human acts An An artificial
N Agent on an artificial organization agent acts on
agent acts on an an artificial
T artificial agent agent

Fig. 3. Categories according to agents and patients (modified version of [16])

92 A.R. Honarvar and N. Ghasem-Aghaee

With this new simplification, outcome part of a case is divided to three sections.
These sections contain ethical evaluation values of performed agent’s intentions on
humans, organizations and artificial agents or non-humans. Ethical evaluation values
are calculated by a Case-Evaluator of casuist BDI architecture. This component is
described next section. The main structure of each case in casuist BDI architecture is
illustrated in Fig. 4. According to the specific application domain of artificial ethical
agent, more details can be added to this structure.

( Problem environment situation, Belief, Desire )

Solution
(intentions)

Outcome
On humans x
On y
organizations
On artificial z
agents

Fig. 4. Basic structure of cases in Case Memory

5 Case-Evaluator of Casuist BDI-Agent

We decide to add capability of ethical reasoning to BDI agent by evaluating the ex-
periences of agent from ethical aspect, and use these experiences plus its evaluations
for ethical decision making in future. In this section we introduce the Case-Evaluator
component of Casuist BDI architecture. This evaluator bases on previous attempts to
make ethics computable [1, 12, 16]. According to classification of ethical situation
which is introduced in previous section, the result of performing agent’s decisions or
intentions can effect on three kinds of entities: humans, non-humans and organiza-
tions. Case-Evaluator uses this notion to evaluate ethical aspects of affected environ-
ment. Case-Evaluator takes Agent’s Beliefs, Agent’s Desires, and environment status
before and after agent’s behavior as inputs to determine and compute the kinds, num-
bers, duration and intensity of pleasure/displeasure of entities which is affected by
agent’s behavior. These elements are necessary elements to evaluate a situation ethi-
cally from consequentiallist viewpoint [1, 12]. To propose a case evaluator, we adopt
and extend previous works which are tried to make consequentillist theories of ethics
computable. Gips In [12] propose an equation for consequentialist evaluation
schemes. This equation has the following form:
σ ܹ௜ ‫ܲ  כ‬௜
(1)
Where wi is the weight assigned each person and pi is the measure of pleasure or hap-
piness or goodness for each person. In [1] Aderson and Aremn propose a modified
version of these equations. For each person, the algorithm simply computes the product
of the intensity, the duration, and the probability, to obtain the net pleasure for each
person, and then it adds the individual net pleasure to obtain the Total Net Pleasure.
 Casuist BDI-Agent: A New Extended BDI Architecture 93

ܶ‫݁ݎݑݏ݈ܽ݁ܲݐ݁ܰ ݈ܽݐ݋‬ 

ൌ ෍ሺ –‡•‹–› ൈ —”ƒ–‹‘ (2)


ൈ ”‘„ƒ„‹Ž‹–›ሻˆ‘”‡ƒ Šƒˆˆ‡ –‡†‹†‹˜‹†—ƒŽ

This computation would be performed for each alternative action. The action with the
highest Total Net Pleasure is the right action. In other words we can say their pro-
posed equation is equal to Gips’s equations with extended parameter of time. Accord-
ing to Gips’s equation, this new equation has the following form:
σ ܹ௜ ‫ܲ  כ‬௜ ‫ܶ  כ‬௜ (3)
Where wi is the weight assigned each person, pi is the measure of pleasure or happi-
ness or goodness for each person and Ti is the duration of action’s pleasure for each
person. Our Case-Evaluator considers the ethical classifications of situation which is
introduced. This ethical evaluator determines the kind of entity which is affected by
each agent’s intention or behavior, Computes duration, intensity and probability of
effects of intentions on each entity. These evaluations send to Case-Updater compo-
nent of Casuist BDI agent for updating the experiences of agent or Case Memory. The
evaluation method of this evaluator is described in equation 7 which uses equation 4-6
for its operation:
ܶܰܲ‫ ܪ‬ൌ σ௡௜ୀ଴ Š୧ ‫Š  כ‬୧ ‫Š  כ‬୧  (4)

TNPH is the total net pleasure of humans after agent’s behavior. Whi is the weight
assigned to humans which shows the importance of each person in specific situation
and application domain. Phi is the probability that a person is affected. Thi is the dura-
tion of pleasure/displeasure of each person after agent’s behavior. n shows the num-
bers of people in that situation.
ܱܶܰܲ ൌ σ௡௜ୀ଴ ‘୧ ‫‘  כ‬୧ ‫‘  כ‬୧ (5)
TNPO is the total net pleasure of organizations after agent’s behavior. Woi is the
weight assigned to organizations which shows the importance of each organization in
specific situation and application domain. Poi is the probability that an organization is
affected. Toi is the duration of pleasure/displeasure of each organization after agent’s
behavior. n shows the numbers of organization in that situation.
ܶܰܲ‫ ܣ‬ൌ σ௡௜ୀ଴ ƒ ୧ ‫ ƒ  כ‬୧ ‫ ƒ  כ‬୧ (6)
TNPA is the total net pleasure of artificial agent after agent’s behavior. Wai is the
weight assigned to artificial agents which shows the importance of each artificial
agent in specific situation and application domain. Pai is the probability that an artifi-
cial agent is affected. Tai is the duration of pleasure/displeasure of each artificial
agent after agent’s behavior. n shows the numbers of artificial agents in that situation.
ܶܰܲ ൌ ܶܰܲ‫ܹ  כ ܪ‬௛ ൅ ܱܶܰܲ ‫ܹ  כ‬௢ ൅ ܶܰܲ‫ܹ  כ ܣ‬௔ (7)
94 A.R. Honarvar and N. Ghasem-Aghaee

TNP is the total net pleasure of all three kinds of entities which participate in applica-
tion domain of agent’s behavior. Wh, Wo, Wa illustrate the participation degree of
humans, organizations and artificial agents, respectively. The summation of Wh, Wo
and Wa equals one.
TNPH, TNPO, TNPA and TNP are stored in outcome part of a case in Case Mem-
ory. These values can use by a Case-Retriever for retrieving cases when agents en-
counter a problem.

6 Conclusion
In this paper a new extension of BDI architecture, Casuist BDI-Agent, proposed
which has the previous capability of BDI-agent architecture in addition to capability
of ethical decision making. This architecture can be used for designing BDI-Agent in
domains where BDI-Agent can be used and ethical considerations are important issue.
With the aid of this new architecture, agents can consider ethical effects of their be-
haviors when they make decision. The main idea in this architecture is based on a
method of ethical reasoning which uses previous experiences and doesn’t use any
codes of ethics. The main advantage of using this method for implementing ethical
decision making capability in agent is elimination of needs to convert a set of ethical
rules in application domain of agents to algorithm which needs conflict management
of rules. In this architecture agents can adapt ethically to its application domain and
can augment its implicit ethical knowledge, so behave more ethically.

References
1. Anderson, M., Anderson, S., Armen, C.: Toward Machine Ethics: Implementing Two
Action-Based Ethical Theories. In: AAAI 2005 Fall Symp. Machine Ethics, pp. 1–16.
AAAI Press, Menlo Park (2005)
2. Allen, C., Wallach, W., Smith, I.: Why Machine Ethics? IEEE Intelligent Systems Special
Issue on Machine Ethics 21(4), 12–17 (2006)
3. Anderson, M., Anderson, S.: Ethical Healthcare Agents. Studies in Computational Intelli-
gence, vol. 107, pp. 233–257. Springer, Heidelberg (2008)
4. Wallach, W., Allen, C.: Moral Machines: Teaching Robot Right from Wrong. Oxford
University Press, Oxford (2009)
5. Moor, J.H.: The nature, importance, and difficulty of Machine Ethics. IEEE Intelligent
Systems Special Issue on Machine Ethics 21(4), 18–21 (2006)
6. Honarvar, A.R., Ghasem-Aghaee, N.: Simulation of Ethical Behavior in Urban Transporta-
tion. Proceedings of World Academy of Science, Engineering and Technology 53,
1171–1174 (2009)
7. Rao, G.: BDI Agents: From Theory to Practice. In: Proceedings of the First International
Conference on Multi-Agent Systems (ICMAS 1995), San Fransisco, USA (1995)
8. Pal, S., Shiu: Foundations of Soft Case-Based Reasoning. Wiley-Interscience, Hoboken
(2004)
9. Kolodner: Case-Based Reasoning. Morgan Kaufmann, San Mateo (1993)
 Casuist BDI-Agent: A New Extended BDI Architecture 95

10. Keefer, M.: Moral reasoning and case-based approaches to ethical instruction in science.
In: The Role of Moral Reasoning on Socioscientific Issues and Discourse in Science
Education. Springer, Heidelberg (2003)
11. http://wiki.lawguru.com/index.php?title=Casuistry
12. Gips: Towards the Ethical Robot, Android Epistemology, pp. 243–252. MIT Press,
Cambridge (1995)
13. Kendal, S.L., Creen, M.: An Introduction to Knowledge Engineering. Springer, Heidelberg
(2007)
14. Olivia, C., Change, C., Enguix, C.F., Ghose, A.K.: Case-Based BDI Agents: An Effective
Approach for Intelligent Search on the web. In: Proceeding AAAI 1999 Spring
Symposium on Intelligent Agents in Cyberspace Stanford University, USA (1999)
15. Corchado, J.M., Pavón, J., Corchado, E., Castillo, L.F.: Development of CBR-BDI Agents:
A Tourist Guide Application. In: Funk, P., González Calero, P.A. (eds.) ECCBR 2004.
LNCS (LNAI), vol. 3155, pp. 547–559. Springer, Heidelberg (2004)
16. Al-Fedaghi, S.S.: Typification-Based Ethics for Artificial Agents. In: Proceedings of Sec-
ond IEEE International Conference on Digital Ecosystems and Technologies (2008)
17. Floridi, L., Sanders, J.W.: On the Morality of Artificial Agents. Minds and
Machines 14(3), 349–379 (2004)
 Research of Trust Chain of Operating System

Hongjiao Li and Xiuxia Tian

School of Computer and Information Engineering, Shanghai University of Electric Power,

Shanghai, P.R.C.
2008000026@shiep.edu.cn

Abstract. Trust chain is one of the key technologies in designing secure oper-
ating system based on TC technology. Constructions of trust chain and trust
models are analyzed. Future works in these directions are discussed.

Keywords: Trusted Computing, Secure Operating System, Trust Chain, Trust

Model.

1 Introduction
It seems likely that computing platforms compliant with Trusted Computing Group
(TCG) specifications will become widespread over the next few years [1, 2]. The ad-
venture of TC (Trust Computing) technology gives new opportunity to the research and
development of secure operating system [3, 4]. On the one hand, secure operating
system has the ability to protect legal information flow, protect the information from
unauthorized access. On the other hand, TC technology can protect operating system of
its own security. Therefore, with the help of TC technology to develop secure operating
system is an effective way to solve the security of terminal. Future operating system
will be open-source, trusted, secure, etc [5].
Trust chain is one of the key technologies of TC, which plays vital role in con-
struction of high-assured secure operating system. In this paper we summary in detail
two important problems about trust chain of secure operating system, mainly focus on
Linux.
The paper starts with a description of transitive trust and the notion of trust chain of
operating system in Section 2. Section 3 details the research on construction of oper-
ating system trust chain and analyzes future work on Linux. Trust models and its future
work are discussed in Section 4. We conclude the paper in Section 5.

2 Trust Chain of Operating System

Transitive trust also known as “Inductive Trust” is a process where the root of trust
gives a trustworthy description of a second group of functions. Based on this descrip-
tion, an interested entity can determine the trust it is to place in this second group of
functions. If the interested entity determines that the trust level of the second group of
functions is acceptable, the trust boundary is extended from the root of trust to include
the second group of functions. In this case, the process can be iterated. The second

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 96–102, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Research of Trust Chain of Operating System 97

group of functions can give a trustworthy description of the third group of functions,
etc. Transitive trust is used to provide a trustworthy description of platform character-
istics, and also to prove that non-migratable keys are non-migratable.
The transitive trust enables the establishment of a chain of trust by ensuring that
the trust, on each layer of the system, is based on, and is only based on, the trust on the
layer(s) underneath it, all the way down to the hardware security component, which
serves as the root of trust. If verification fails to succeed at any given stage, the system
might be put in a suspended-mode to block possible attacks. The resulting integrity
chain inductively guarantees system integrity. Figure 1 illustrates the process of tran-
sitive trust of operating system. Trust chain starts from the trust root. The unmodifiable
part of the BIOS measure the integrity of the BIOS and then the BIOS measures the
MBR (Master Boot Record) of the bootstrap device. Then, MBR measures the OS
loader, and OS loader measures OS kernel. Finally, OS measure applications.

Fig. 1. Transition of Trust

3 Construction of Operating System Trust Chain

3.1 Related Work

Many researchers have endeavor to lots of works on operating system based on TCPA.
AEGIS is a system structure for integrity examination developed by University of
Pennsylvania [6]; AEGIS architecture establishes a chain of trust, driving the trust to
lower levels of the system, and, based on those elements, secure boot. It validates
integrity at each layer transition in the bootstrap process. AEGIS also includes a re-
covery process for integrity check failures.
IBM T. J. Watson Research Center developed an integrity mechanism on Linux [7].
There system is the first to extend the TCG trust concepts to dynamic executable con-
tent from the BIOS all the way up into the application layer. They present the design
and implementation of a secure integrity measurement system for Linux. All executa-
ble content that is loaded onto the Linux system is measured before execution and these
measurements are protected by the Trusted Platform Module (TPM) that is part of the
TCG standards. Their measurement architecture is applied to a web server application
to show how undesirable invocation can be detected, such as rootkit programs, and that
measurement is practical in terms of the number of measurements taken and the per-
formance impact of making them.
Huang-Tao also designed a trusted boot method in server system [8]. Their research
proposed a solution of trusted boot. The solution performs integrity verification on
every control capability transfer between one layer and next layer. If OK, control can
transfer. Therefore, the boot of system can execute according to the transition of trusted
chain. In this way, system boot can be in a trusted state, which can improve the sys-
tem’s security. Also, the solution is easily and flexibility to implement.
98 H. Li and X. Tian

Dynamic multi-path trust chain (DMPTC) is a software type and character based
mechanism to assure system trustworthiness on WINDOWS system [9]. DMPTC dif-
ferentiates static system software and dynamic application software and takes different
ways and policies to control the loading and running of various executable codes. The
goal of DMPTC is to build a trusted computing platform by making computing plat-
form only load and run trustworthy executables. Also, DMPTC gives great considera-
tion to the impact it causes to system performance1.Based on the attributes of various
executables and by taking advantage of WINDOWS interior security mechanisms,
DMPTC reduces the time cost of the executables verification process greatly. And
ultimately assures the flexibility and utility of DMPTC.
Maruyama etc deeply studied transitive mechanisms from GRUB boot loader to
operating system, describe a TCPA integrity mechanism from end to end and imple-
ment it on Linux kernel [10]. Integrity measurement of the kernel is done through a
chaining of PCR (Platform Configuration Register) updates during the bootstrap
process of the GRUB kernel loader. Kernel integrity measure is done by update the
PCR chain in kernel loader during booting. The measure results which involved digi-
tally-signed PCR values are reported to remote server by using a JAVA application,
Hadi Nahari describes a mechanism to construct effective CONTAINMENT (that is, a
mechanism to stop an exploited application start to against to attacks on another ap-
plication), which is applied to embedded system [11]. The MAC (Mandatory Access
Control) is provided by SeLinux. The focus will be on practical aspects of hardware
integration as well as porting SeLinux to resource-constrained devices. The methods
provides a high-level, structured overall infrastructure to provide basic and necessary
function to establish the trust of operating system services (via connect a hard-
ware-based root of trust).

3.2 Further Work

For Linux kernel, changes may take place during execution. Due to the multi-aspect
and in-order of applications, single chained boot mechanism of system boot cannot
apply to the trust transition from OS to applications. Once the kernel is booted, then
user-level services and applications may be run. In Linux, a program execution starts
by loading an appropriate interpreter (i.e., a dynamic loader, such as ld.so) based on the
format of the executable file. Loads of the target executable's code and supporting
libraries are done by the dynamic loader. In Linux operating system, kernel and its
modules, binary execution files, shared libraries, configuration files, and scripts run
loosely serially, and can change the system’s state during execution, which is the key
difficulty to the research of trusted Linux. In such situations, even OS loader has veri-
fied that the kernel is trusted. The trusted states can also be destroyed because the
kernel modules can be inserted or uninstalled, which incurs that the PCR values cannot
represent the current execution condition. It needs verification whether the module has
effects on kernel states on inserting or uninstalling the kernel modules.
One must note that the former research though ensuring the integrity of the oper-
ating environment when a hard boot occurs, does not guarantee its integrity during
the runtime; that is, in case of any malicious modification to the operating environ-
ment during the runtime, this architecture will not detect it until the next hard boot
happens. Trust of application embodies the integrity of privacy of complete trust
 Research of Trust Chain of Operating System 99

chain. Though in reference [7], the trust is extended to dynamic executable content,
mandatory access control policy are not measured, future architecture extensions
should include such measurements.
On the other hand, a MAC (Mandatory Access Control) mechanism will be able to
address one of its fundamental shortcomings; providing a level of protection at runtime.
Deploying MAC mechanisms at the same time balance performance and control are
particularly challenging tasks.
TPM-EMULATOR provides better technical support for the practice and verifica-
tion for the research of trust chain [12, 13]. There is a tremendous opportunity for
enhanced security through enabling projects to use the TPM.

4 Trust Model

4.1 Related Work

TC models centralized on how to compute trust in information world, that is, put use
the trust relations among humans into computing environments to achieve the goal of
trust.
Patel J etc proposed a probabilistic trust model[14], their research aims to develop a
model of trust and reputation that will ensure good interactions amongst software
agents in large scale open systems in particular[14]. Key drivers for their model are: (1)
agents may be self-interested and may provide false accounts of experiences with other
agents if it is beneficial for them to do so; (2) agents will need to interact with other
agents with which they have no past experience. Specifically, trust is calculated using
probability theory taking account of past interactions between agents. When there is a
lack of personal experience between agents, the model draws upon reputation infor-
mation gathered from third parties.
Trust management model based on the fuzzy set theory deals with the authenticity in
open networks [15]. The author showed that authentication can not be based on public
key certificate alone, but also needs to include the binding between the key used for
certification and its owner, as well as the trust relationships between users. And develop
a simple algebra around these elements and describe how it can be used to compute
measures of authenticity.
In algebra for assessing trust in certification chains, the fuzzy nature of subjective
trust is considered, and the conceptions of linguistic variable and fuzzy logic are in-
troduced into subjective trust management [16]. A formal trust metric is given, fuzzy
IF-THEN rules are applied in mapping the knowledge and experiences of trust rea-
soning that humanity use in everyday life into the formal model of trust management,
the reasoning mechanisms of trust vectors are given, and a subjective trust management
model is provided. The formal model proposed provides a new valuable way for
studying subjective trust management in open networks.
Due to the open network is dynamic, distributed, and non-deterministic, in reference
to trust relation among human, a trust evaluation model based on D2S theory is pro-
posed [17]. The model gives the formal description of direct trust based on the history
trade record among grid and constructed the trust recommend mechanism of grid
nodes. Combining D2S theory with evidence, we can get indirect trust. After that,
100 H. Li and X. Tian

combine direct trust and indirect trust, effectively. The results show that the trust model
is viable and usability.
Reference [18] proposed a trust management model based on software behavior. An
agent-based software service coordination model is presented to deal with the trust
problem in open, dynamic and changeful application environment. The trust valuation
model is given to value trust relationships between software services. Trust is
abstracted as a function of subjective expectation and objective experience, and a rea-
sonable method is provided to combine the direct experience and the indirect experi-
ence from others. In comparison with another work, a complete trust valuation model is
designed, and its reasonability and operability is emphasized. This model can be used
in coordination and security decision between software services.
Non-interference theory is introduced into the domain of trusted computing to con-
struct the trusted chain theoretic model [19]. The basic theory of the computing trusted is
proposed and a non-interference based trusted chain model is built from the dynamic
point of view, and then the model is formalized and verified. Finally, the process of
start up based on Linux operating system kernel is implemented. The implementation
provides a good reference for the development and application of the trusted computing
theory as well.

4.2 Further Work

TCG introduces the idea of trust into the computing environment, but there is still not
the formalized uniform description. Trusted computing is still a technology but not a
theory, and the basic theory model has not been established. Above trust models focus
on sociology human relation, does not in accordance with the definition of TCG. Also,
present models should be further optimized to objective, simple and usability.
Linux is multi-task operating system, multi-program run simultaneously; therefore,
verification of the former and then transit to the latter is not viable. To maintain the
trusted state of the system, there is need to verify the file and associated execute pa-
rameters is trusted or not. At the same time, there is also need verification before
loading to execution to those objects which probably change system state.
Due to the property of applications, formal description should focus on the separa-
tion of processes, etc. The theory which support trust transition should pay more at-
tention to the dynamic execution. Also, the security principle, such as least privilege,
need-to-know policy should also take into account. Efficiency of integrity measure-
ment and security policy enforcement should also further improved.
In essence, non-interference theory is based on information flow theory, and it can
detect covert channel of the system which meets the requirements of designing
high-level secure operating system. Future research on trust model should pay more
attention to the non-interference theory, which support the construction and extend of
trusted chain.

5 Conclusions
In this paper, we summarize construction and formalization of trust chain of operating
system. We conclude that future research on trust chain of secure operating system
 Research of Trust Chain of Operating System 101

should focus on dynamic trusted chain. Make full use of trusted mechanisms supplied
by TPM, extending the TCG concept of trust to application layer to study the trusted
chain and its formal description.
Acknowledgments. We are grateful to Shanghai Municipal Education Commission‘s
Young Teacher Research Foundation, under grant N0. SDL08024.

References
1. Shen, C.-x., Zhang, H.-g., Feng, D.-g.: Survey of Information Security. China Sci-
ence 37(2), 129–150 (2007) (in Chinese)
2. Trusted Computing Group. TCG Specification Architecture Overview [EB/OL]
[2005-03-01], http://www.trustedcomputinggroup.org/
3. Changxiang, S.: Trust Computing Platform and Secure Operating System. Network Security
Technology and Application (4), 8–9 (2005) (in Chinese)
4. Shieh, A.D.: Nexus: A New Operating System for Trustworthy Computing. In: Proceedings
of the SOSP, Brighton, UK. ACM, New York (2005)
5. Zhong, C., Shen, Q.-w.: Development of Modern Operating System. Communications of
CCF 9, 15–22 (2008) (in Chinese)
6. Arbaughz, W.A., Farber, D.J., MSmith, J.: A Secure and Reliable Bootstrap Architecture.
In: IEEE Security and Privacy Conference, USA, pp. 65–71 (1997)
7. Sailer, R., Zhang, X., Jaeger, T., et al.: Design and Implementation of a TCG -based Integ-
rity Measurement Architecture. In: The 13th Usenix Security Symposium, San Diego (2004)
8. Tao, H., Changxiang, S.: A Trusted Bootstrap Scenario based Trusted Server. Journal of
Wuhan University (Nature Science) 50(S1), 12–14 (2004) (in Chinese)
9. Xiaoyong, L., Zhen, H., Changxiang, S.: Transitive Trust and Performance Analysis in
Windows Environment. Journal of Computer Research and Development 44(11),
1889–1895 (2007) (in Chinese)
10. Maruyama, H., Nakamura, T., Munetoh, S., et al.: Linux with TCPA Integrity Measurement.
IBM, Tech. Rep.: RT0575 (2003)
11. Nahari, H.: Trusted Embedded Secure Linux. In: Proceedings of the Linux Symposium,
Ottawa, Ontario Canada, June 27-30, pp. 79–85 (2007)
12. Hall, K., Lendacky, T., Raliff, E.: Trusted Computing and Linux,
http://domino.research.ibm.com/comm/research_projects.nsf/
pages/gsal.TCG.html/FILE/TCFL-TPM_intro.pdf
13. Strasser, M.: A Software-based TPM Emulator for Linux. Swiss Federal Institute of
Technology (2004),
http://www.infsec.ethz.ch/people/psevinc/TPMEmulatorReport.pdf
14. Patel, J., Teacy, W.T.L., Jennings, N.R., Luck, M.: A Probabilistic Trust Model for Han-
dling Inaccurate Reputation Sources. In: Herrmann, P., Issarny, V., Shiu, S.C.K. (eds.)
iTrust 2005. LNCS, vol. 3477, pp. 193–209. Springer, Heidelberg (2005)
15. Wen, T., Zhong, C.: Research of Subjective Trust Management Model based on the Fuzzy
Set Theory. Journal of Software 14(8), 1401–1408 (2003) (in Chinese)
16. Audun, J.: An Algebra for Assessing Trust in Certification Chains. In: The Proceedings of
NDSS 1999, Network and Distributed System Security Symposium, The Internet Society,
San Diego (1999)
17. Lulai, Y., Guosun, Z., Wei, W.: Trust evaluation model based on DempsterShafer evidence
theory. Journal of Wuhan University (Natural Science) 52(5), 627–630 (2006) (in Chinese)
102 H. Li and X. Tian

18. Feng, X., Jian, L., Wei, Z., Chun, C.: Design of a Trust Valuation Model in Software Service
Coordination. Journal of Software 14(6), 1043–1051 (2003) (in Chinese)
19. Jia, Z., Changxiang, S., Jiqiang, L., Zhen, H.: A Noninterference Based Trusted Chain
Model. Journal of Computer Research and Development 45(6), 974–980 (2008)
(in Chinese)
 A Novel Application for
Text Watermarking in Digital Reading

Jin Zhang, Qing-cheng Li, Cong Wang, and Ji Fang

Department of Computer Science, Nankai University,

Tianjin 300071, China
{nkzhangjin,liqch}@nankai.edu.cn,
{wangcong6222351,shuiqi}@mail.nankai.edu.cn

Abstract. Although watermarking research has made great strides in theoretical

aspect, its lack of application in business could not be covered. It is due to few
people pays attention to usage of the information carried by watermarking. This
paper proposes a new watermarking application method. After digital document
being reorganized with advertisement together, watermarking is designed to
carry this structure of new document. It will release advertisement as interfer-
ence information under attack. On the one hand, reducing the quality of digital
works could inhabit unauthorized distribution. On the other hand, advertisement
can benefit copyright holders as compensation. Moreover implementation de-
tail, attack evaluation and watermarking algorithm correlation are also
discussed through an experiment based on txt file.

1 Introduction
It has been several ten years since digital watermarking technology emerged. Both of
technology diversity and theoretical research have made major strides in this field.
But its application in business is always a big problem which perplexes the develop-
ment of watermarking. Many watermarking companies have gone out of business or
suspended their watermarking efforts except some focus on tracking and digital
fingerprint [1].
This problem is due to, in our opinion, the misunderstanding of the nature of digital
watermarking. The real value of digital watermarking is the information bound in
watermarking. In other words, watermarking content and application methods are the
key issues in digital watermarking research. Unfortunately, the current research in this
area is little to write about.
As an important part in digital watermarking, text watermarking encounters the same
problem with other kinds of watermarking technology. While most people agree that
text watermark should be been widely used in the field of digital reading, but this is not
the truth. On one hand lack of appropriate application methods make text watermarking
useless. On the other hand, text watermarking is characterized by its attack sensitiveness
(this part is discussed in section2). This easily led to watermark information lost.
In this paper, one application method is proposed for text watermarking in digital
reading. We combine original digital content and advertisement as a whole digital

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 103–111, 2009.
© Springer-Verlag Berlin Heidelberg 2009
104 J. Zhang et al.

document. The rules of combination are embedded in this digital document as water-
marking information. Once the document is under attack, advertisement will be re-
leased from it. The attacker will have to read them with real content together, which
will serve as compensation for the copyright holder.
This paper is organized as follows: in section 2, we review other kinds of text wa-
termarking. And the procedure of our solution is presented generally in section 3. And
we expound solution detail and experiment result in section 4. At last we draw the
conclusion in section 5.

2 Related Works

In general, the research object of text watermarking is digital document. And docu-
ment is always made up of two parts, content and format.
Many researchers choose to embed watermarking by taking content as manifesta-
tions of language. That is semantics-based watermarking. When the watermarking
information is embedded, the content is modified according to linguistic rules. Atal-
lah’s TMR is the representative of this kind of watermarking [2].Dai and Yang also
proposed their solution [3, 4]. This kind of watermarking technology always needs
more time to consume. And it is very sensitive to tamper attack and deleting attack
[5]. At the same time, less capacity is another shortcoming of them.
Some other researchers deal with content as a still picture. In this way, they can
build additional redundant space to embed watermarking so as to enhance capacity.
Brassil and Low’s line-shift/word-shift coding and word feature coding are the repre-
sentative of such solutions [6, 7]. In their solution, content is a binary image. They use
lines or words’ tiny mobile in different directions to embed information. And Huang
complemented their work [8].Liu and Chen’s work is based on interconnected do-
main. They regarded Chinese characters as special images [9]. It is worth mentioning
that Li and Zhang propose their method which integrates the ideas of Liu and seman-
tics-based watermarking [10]. They make a significant reduction in overall algorithm
consumption. But OCR technology is their natural enemies. No matter how sophisti-
cated these methods are, the majority of watermarking information will be removed
after the OCR or retyping attack. Tamper and deleting attack also could reduce their
performance.
There are some researchers who hope that watermarking can be hidden in docu-
ment’s format. Fu and Wang’s design is a prime example [11]. Zhou and Sun use spe-
cial characters in RTF file to embed watermark [12]. And some solution is based on
redundant coding space, such as Zhou’s work [13]. Other techniques such as add
blanks and other perceptible characters are also proposed. Although they remove the
impact of tamper attack and deleting attack. Any change to the document will have a
fatal impact on them, much more OCR and retyping attack.
To sum up, all kinds of watermarking are sensitive to special attack. There is little
information could be maintained under the niche attack.
 A Novel Application for Text Watermarking in Digital Reading 105

3 Solution Design

As mentioned above, there is none text watermarking technology can resist all kinds
of attack. In particular the realization of these attacks is also very low cost. Rather
than spending a lot of energy to design a more robust watermark, it is better to find a
method which could make attacker wouldn’t burn their house to rid it of a mouse.

3.1 Design Principle

The general watermarking embedding procedure is described by (1) [14]:

C wn = W ⊕ C0 ⊕ n (1)

After being embedded watermarking information W, original content C0 with additive

channel noise n will be the last digital works or digital document Cwn. Because digital
document always has strict coding requirement or format specification such as check
code mechanism. We do not consider channel noise issue.
As noted in section 1, watermarking is used to transmit combination rules, while
combination of content and advertisement finished. In this case, we modify formula
(1) as follows:
⎧Cwn = F (W , C0 , C A ) ⊕ nA
⎪ (2)
⎨ ⎧C0 , nA = 0
⎪G (W , Cwn ) = ⎨C ⊕ C , n = else
⎩ ⎩ 0 A A

Where CA is advertisement and nA is attack noise. We denote CW as digital document

which contain watermarking without being attacked. F( ) is the combination function,
while G( ) is the striping function.

3.2 Operation Procedure

In this section, we will discuss the operation procedure as figure 1 shows.

CA

C0 ⊕ W
⊕ CW

nA ⊕
W CWn
C0

Fig. 1. Design processing flow chart

106 J. Zhang et al.

Step 1: C0 and CA combine in accordance with the rules. These rules are encoded
into a special format, which is watermarking information W. In other words, both of
C0 and CA are watermarked.
Step 2: After being embedded W, digital document will be distributed or attacked.
In this way attack noise nA. is introduced.
Once attack happened during transmission, nA will add into document. At this time
CW turns into Cwn. As formula (2) shows: When attack misses, nA is given the value 0.
CW equals to Cwn in this condition.
Step 3: When user’s device or client software receive Cwn, it will extract W and
restore C0 according to W. If nA equals to 0, user could watch original content. But if
watermarking is totally erased, user will have to watch CA while they are browsing C0.
In the worst case that W is totally distorted, there are only garbled characters left.

4 Implementation and Evaluation

We implement our design on one kind of hand-held digital reading device, and apply
it in related digital works release platform. The algorithm in [10] is taken as default
algorithm. We use a txt document which contains 10869 Chinese characters as ex-
periment object.
Figure2. (a) shows the document which contains watermarking and advertisement.
In this screenshot, it is open directly without stripping. And Figure2. (b) shows the
result of opening it after stripping. In Figure2.(b) both of advertisement and water-
marking signs are removed by content stripper.

(a) (b)

Fig. 2. Device operation screenshots

Before discussing implementation detail, we give a definition. When the algo-

rithms as [10] work, the content will be changed selectively. Every change can con-
tain 1 bit information. We name a changeable part in content as a watermarking tag.
And we use WT to denote watermarking tag.
 A Novel Application for Text Watermarking in Digital Reading 107

4.1 Implementation Detail

4.1.1 Watermarking Embedding Processing

In this part the most important question is the organization of digital document. Out of
consideration for distributing watermark and advertisement, the document content
structure is designed as figure 3 showing.

Fig. 3. Document content structure diagram

Original content C0 is split into N blocks, sizes of which are not equal and are de-
noted by sequence {Sai}. And sequence {Sbi} is used to denote the sizes of CA blocks.
These combination rules are described by W, which is made up of sequence
{Wi}.We define Wi as follows (3):
Wi = {li ,{Ci , ai , bi }R , S } (3)
li: total length of Wi.
Ci: check code of ai and bi.
{}R: redundancy construction function. It could be replaced by cryptographic
functions.
S: barrier control code. It is used to isolate different Wi. It is also helpful when delet-
ing attack appears. Different code words can also play the role of control information
transmission.
{ai}: relative length of C0 block i. It equals to the distance between first WT in this to
the last character in this C0 block.
{bi}: length of CA block i. It usually equals to Sbi.
Description above is just content structure. After these processing, it could be sealed
in other document format, such as DOC, Html. Because of lager consumption of sys-
tem, this work should be arranged in content servers or distribution servers.

4.1.2 Content Stripping Processing

Generally the main work is extracting watermarking information and restoring origi-
nal content.
108 J. Zhang et al.

Fig. 4. Table T structure diagram

After scanning CWn, watermarking information sequence will be gained. According

to barrier control code S, Wi is separated. Before restoring original content, the plain
txt stream should be restore first. This txt stream includes C0 and CA without water-
marking. And there is a table T which contains WT’s offset in the stream should be
setup. As figure 4 shows. This due to locating requirements because every block’s
address is presented by relative offset, as formula (3) shows. According to T and
{Wi}, original content could be restored.
Actually we can use cache mechanism to optimize the design performance in de-
vice. In short, user can browse the content before totally restoring content. When user
wants to skip to some pages, device can scan nearest barrier control code S and build
temporary table T’ to show content quickly. And complete table could be built within
user browsing interval.

4.2 Application in Dynamic DRM

Most main DRMs use asymmetric cryptographic algorithm and special document
format to ensure security. Our design focuses on content which is independent of
document format. And then it could be applied in most DRM smoothly.
Dynamic DRM (DDRM) is a special DRM which support superdistribution and
distribution among users [15]. And its technical basis is its license mechanism. And
all digital works or digital documents on this platform must be divided into blocks
and reorganized. Especially dynamic advertisement block is supported in DDRM.
That means the advertisement in digital works could be changed with distribution
gong on.
As mentioned in section 3, watermarking is embedded into total content. Both of
C0 and CA contain watermarking. In this case, device has to reorganize digital docu-
ment to update CA after each distribution. Obviously it is very hard for today’s em-
bedded devices.
Therefore we have to amend formula (3), and make it turn into another form as (4)
shows.
CW = F (W , C0 ) ⊕ C A (4)

In this condition, only C0 could be watermarked. And CA could be replaced dynami-

cally. But there must be two certain rules as follows:
 A Novel Application for Text Watermarking in Digital Reading 109

• New CA block must have the same size with the old one. That is the reason why
we usually give CA block a larger size.
• Validity of CA should also be checked. And barrier control code S’s value is
used to indicate whether dynamic CA is supported.

4.3 Evaluation under Attack

As discussed in section 2, most text watermarking algorithm including our design are
attack sensitive. Once attack match with watermarking algorithm’s shortcoming, there
could be little information left.
Rather than pay attention to more robust algorithm, we focus on risk management
after attack happened. Figure 5 shows the result of watermarking information lost in
our design.

Fig. 5. Content left after being attack

Obviously advertisement emerges when watermarking information is lost. On one

hand, it could reduce digital works quality and inhibit illegal propagation. On the
other hand, it could compensate content provider and service provider for the lost.
Technically, the robustness of our design mainly depends on the robustness of
watermarking algorithm selected.

4.4 Watermarking Algorithm Correlation

Our research mainly focuses on finding a new method to apply digital watermarking.
Therefore the correlation of selection algorithm needs to be discussed.
Though most algorithms could smoothly be applied in our design, we need use
different way to split content according to different algorithms.
When one algorithm hiding information in format is selected, splitting method is
very easy. If the algorithm tries to regard text as a picture, the descriptors about size
or offset should be replaced by the descriptors about coordinates.
110 J. Zhang et al.

(a)

(b)

Fig. 6. Capacity distribution

When the algorithm selected involves semantics-based watermarking, we must

evaluate the size of Wi in advance. And then the size of block could be set, after
checking distribution of watermarking tags. Otherwise it could happen that the size of
Wi is larger than the capacity of watermarking space in block.
Figure 6 is the capacity distribution in our experiment. First we divide content into
20 blocks which have the same size. (a) shows its result that the smallest capacity is
only 60 bits. (b) shows the situation that every block have the same watermarking
capacity. Obviously capacity distribution is not even.

5 Conclusion
As an application technology, digital watermarking research should not only indulge in
finding new algorithm or blindly seek technical target. Searching efficient application
method maybe more meaningful.
Out of viewpoint above, this paper propose a novel application method for text wa-
termarking in digital reading field. It will release advertisement as interference infor-
mation under attack. On the one hand, reducing the quality of digital works could
inhabit unauthorized distribution. On the other hand, it point out a new way for
watermarking application.
 A Novel Application for Text Watermarking in Digital Reading 111

References
1. Delp, E.J.: Multimedia security: the 22nd century approach. Multimedia Systems 11,
95–97 (2005)
2. Atallah, M.J., Raskin, V., Crogan, M., Hempelmann, C., Kerschbaum, F., Mohamed, D.,
Naik, S.: Natural language watermarking: Design, analysis, and a proof-of-concept imple-
mentation. In: Moskowitz, I.S. (ed.) IH 2001. LNCS, vol. 2137, pp. 185–200. Springer,
Heidelberg (2001)
3. Dai, Z., Hong, F., Cui, G., et al.: Watermarking text document based on statistic property
of part of speech string. Journal on Communications 28(4), 108–113 (2007)
4. Yang, J., Wang, J., et al.: A Novel Scheme for Watermarking Natural Language Text:
Intelligent Information Hiding and Multimedia Signal Processing. IEEE Intelligent Soci-
ety 2, 481–488 (2007)
5. Li, Q., Zhang, J., et al.: A Chinese Text Watermarking Based on Statistic of Phrase
Frequency: Intelligent Information Hiding and Multimedia Signal Processing. IEEE Intel-
ligent Society 2, 335–338 (2008)
6. Brassil, J., Low, S., Maxemchuk, N., O’Gorman, L.: Electron marking and identification
techniques to discourage document copying. IEEE J. Select. Areas Commun. 13,
1495–1504 (1995)
7. Low, S., Maxemchuk, N.: Capacity of Text Marking Channel. IEEE Signal Processing
Letter 7(12), 345–347 (2000)
8. Huang, H., Qi, C., Li, J.: A New Watermarking Scheme and Centroid Detecting Method
for Text Documents. Journal of Xi’an Jiaotong University 36(2), 165–168 (2002)
9. Liu, D., Chen, S., Zhou, M.: Text Digital Watermarking Technology Based on Topology
of Alphabetic Symbol. Journal of Chinese Computer Systems 27(25), 812–815 (2007)
10. Li, Q., Zhang, Z., et al.: Natural text watermarking algorithm based on Chinese characters
structure. Application Research of Computers 26(4), 1520–1527 (2009)
11. Fu, Y., Wang, B.: Extra space coding for embedding wartermark into text documents and
its performance. Journal of Xi’an Highway University 22(3), 85–87 (2002)
12. Zou, X., Sun, S.: Fragile Watermark Algorithm in RTF Format Text. Computer Engineer-
ing 33(4), 131–133 (2007)
13. Zhou, H., Hu, F., Chen, C.: English text digital watermarking algorithm based on idea of
virus. Computer Engineering and Applications 43(7), 78–80 (2007)
14. Cox, J., Miller, M., et al.: Digital Watermarking and Steganography, 2nd edn. Morgan
Kaufmann, San Francisco (2007)
15. Li, Q., Zhang, J., et al.: A Novel License Distribution Mechanism in DRM System. In:
Proceedings of 22nd Advanced Information Networking and Applications - Workshops,
pp. 1329–1334. IEEE Computer Society, Los Alamitos (2008)
 Optimization of Real-Valued Self Set for Anomaly
Detection Using Gaussian Distribution

Liang Xi, Fengbin Zhang, and Dawei Wang

College of Computer Science and Technology, Harbin University of Science and Technology,
150080 Harbin, China
xljyp2002@yahoo.com.cn, zhangfb@hrbust.edu.cn,
stonetools@sohu.com

Abstract. The real-valued negative selection algorithm (RNS) has been a key
algorithm of anomaly detection. However, the self set which is used to train de-
tectors has some problems, such as the wrong samples, boundary invasion and
the overlapping among the self samples. Due to the fact that the probability of
most real-valued self vectors is near to Gaussian distribution, this paper pro-
poses a new method which uses Gaussian distribution theory to optimize the
self set before training stage. The method was tested by 2-dimensional synthetic
data and real network data. Experimental results show that, the new method
effectively solves the problems mentioned before.

Keywords: anomaly detection, artificial immunity system, real-valued self set,

Gaussian distribution.

1 Introduction
The anomaly detection problem can be stated as a two-class problem: given an ele-
ment of the space, classify it as normal or abnormal[1]. There exist many approaches
for anomaly detection which include statistical, machine learning, data mining and
immunological inspired techniques[2, 3, 4]. The task of anomaly detection may be
considered as analogous to the immunity of natural systems, while both of them aim
to detect the abnormal behaviors of system that violate the established policy[5,6].
Negative selection algorithm (NSA)[7] has potential applications in various areas, in
particular anomaly detection. NSA trains detector set by eliminating any candidate
that match elements from a collection of self samples (self set). These detectors sub-
sequently recognize non-self data by using the same matching rule. In this way, it is
used as an anomaly detection algorithm that only requires normal data to train.
Most works in anomaly detection used the problem to binary representation in the
past. However, many applications are natural to be described in real-valued space.
Further more, these problems can hardly be processed properly using NSA in binary
representation. Gonzalez et al.[8] proposed the real-value representation for the
self/non-self space, which differs from the original binary representation. The real-
valued self/detector is a hypersphere in n-dimensional real space; it can be repre-
sented by an n-dimensional point and a radius.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 112–120, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Optimization of Real-Valued Self Set for Anomaly Detection 113

As we know that, the detection performance is mainly relied on the quality of de-
tector set: detector coverage of the non-self space. The detector is trained by the self
set, so that the self set plays an important role on the whole detection process, and the
health of self set is very important to be paid enough attention. It should be noted that,
the real-valued self set indeed has some problems which the binary representation has
not, such as, the wrong self samples, the overlapping of self samples and the boundary
invasion. To solve these problems, this paper proposes a novel method to optimize the
self set. Starting from the point that self samples approach the Gaussian distribution
as the number of samples is large enough, so this method employs Gaussian distribu-
tion theory to deal with the problems mentioned above.

2 Problems with Real-Valued Self Set

2.1 The Wrong Self Samples

Before training detectors, the self set must be secure. However, the real situation may
not meet to the requirement all the time, and the self set is likely to collect some
wrong samples. The figure 1(a) illustrates the problem of wrong self in 2-dimensional
space, there are two samples is so far from the self region, which we can consider as
the wrong self. The wrong self samples may result in the holes, so it is important to
discard the wrong samples before RNS.

(a) The wrong self samples (b) The overlapping (c) The boundary invasion

Fig. 1. The sketch of real-valued self set problems

2.2 The Overlapping

The overlapping self samples: In a work situation, the self samples distribute inten-
sively. As shown in figure 1(b), a large number of self samples are intensively distribute
in the self region, and the overlapping rate is very high, and there are many unnecessary
samples in the self region. Now we discuss the relationship between the number of self
samples and the candidates. Some variables should be defined as below:
f : The probability of failure matching of a candidate to any self sample
pf : The probability of an attack that is not detected
pm : The matching probability
Ns : The number of self samples
114 L. Xi, F. Zhang, and D. Wang

Nd : The number of detectors

N d : The number of candidates
0

When the pm is small enough and the N s is large enough,

f = (1 − pm ) Ns ≈ e− Ns pm .
When the N d is large enough,

p f = (1 − pm ) N d ≈ e − N d pm .

Because
− ln( p f ) ,
N d = N d0 ∗ f =
pm

so
− ln( p f )
N d0 = (1)
pm (1 − pm ) N s

From the equation 1, we can see that the N d has exponential relationship with the N s ,
0

so the more self samples, the more candidates and the higher cost of the detector
training.
The follow equation defines an approximate measure of overlapping between self
samples:
2
− si − s j
r2 (2)
Overlapping( si , s j ) = e s

The maximum value, 1, is reached when the distance between the two sample is 0, on
the contrary, when the distance is equal to 2 rs , the value of this function is almost close
to 0. Based on the equation 2, the amount of overlapping of a self set is defines as
2
− si − s j

Overlapping( S ) = ∑ e r2
s
, i, j = 1, 2,..., n. (3)
i≠ j

2.3 The Boundary Invasion

As a result of the self’s radius, on the border of the self region, the covering area
should invade the non-self region’s boundary. We can see this phenomenon clearly in
the figure 1(c). So when training the detectors by self set, the vicinity of the non-self
boundary may not be covered completely.
 Optimization of Real-Valued Self Set for Anomaly Detection 115

3 Optimization Method
In view of the analysis above, to optimize the self set is significant. The aim of the
optimization is using the least self samples to cover the self region but not the non-
self region. The problem of optimization can be stated as follows:
minimize:
V ( S ) = Volume{x ∈ U | ∃s ∈ S ,|| x − s ||≤ rs } (4)
restricted to:
{∀s ∈ S | ∃d ∈ D, s − d ≤ rs } = ∅ (5)

and
{∀si , s j ∈ S ,|| si − s j ||≤ rsi or || si − s j ||≤ rs j } = ∅ (6)

The function defined in equation 4 represents the amount of the self space covered by
self samples. S, which corresponds to the volume covered by the union of hyper-
spheres associated with each self. The restriction: equation 5 tells that no self should
invade the detector set; equation 6 tells that, there is no self sample covered by the
other samples in the self set.
As we said before, if the number of self samples is large enough, the probability
distribution of self samples approaches the Gaussian distribution. Furthermore, the
central limit theorem clarifies that whichever kind of the probability distribution all
samples approach, the distribution of the samples’ mean is near to a Gaussian distri-
bution; the mean of Gaussian distribution is equal to the mean of all samples.
According to the Gaussian distribution, we can also realize that, the smaller the
distance between the sample and the mean point is, the larger the value of the sam-
ple’s probability density. As discussed before, the nearer the sample is next to the
center of the self region, the heavier the overlapping rate is. So we can use this rela-
tionship to deal with the problem of the unnecessary self samples. This paper pro-
poses the method is that adjusting every sample’s radius by its probability density to
deal with the boundary invasion, and then discarding each unnecessary sample by the
radius.
There is an important and famous proposition in the Gaussian distribution: the “3σ”
Criterion, which shows that each normal sample is almost in the “3σ” district, al-
though the range is from −∞ to ∞ . So we can use the “3σ” Criterion to deal with
the problem of the wrong self samples.
As analyzed above, we may draw a conclusion that, the optimization by the prob-
ability theory is reasonable. To solve the three problems with self set mentioned
above, the optimization is composed by three steps as follows:
Step.1: Discarding the wrong self samples by the “3σ” criterion;
Step.2: Adjusting the self radius by the self’s probability density;
Step.3: Discarding the unnecessary self samples by the self radius.
Before describing the optimization, some variables frequently used in the following
sections are defined here:
116 L. Xi, F. Zhang, and D. Wang

s: Self (Eigenvector: value, Radius: r)

S: Self set (Initial self set: S0)
N0: The number of self samples in S0
N: The number of self samples in S (initial value: N0)
k: Unit self radius
num: the number of probability density interval
Now the algorithm pseudocode is described as follows:

BEGIN
Collection the self samples: S0 ← s;
//step. 1: discarding the wrong self samples
Regularize S0, and then compute the µ and σ;
n=0;
while (n<N0){
if (sn is out of the “3σ” district){
discard sn, N0--;
}
n++;
}
// step. 2: adjusting every sample’s radius
Compute the smaxpdf and sminpdf ;
L= (maxpdf-minpdf)/num;
while (n<N0){
l = ( pdf s − min pdf ) / L ;
n
Sn.r=l×k;
}
// step. 3: discarding the unnecessary self samples
S ← s0, N=1, flag=0;
while (n<N0){
i=0;
while (i<N){
d =|| sn − si || ;
if (d<sn.r || d<si.r){
flag=1, break;
}
i++;
}
if (flag==0){
S ← si, N++;
}
n++;
}
END

In the step.1, the self sample should be discarded if its probability density is out of
the “3σ” district. The time complexity of this step is O(n).
 Optimization of Real-Valued Self Set for Anomaly Detection 117

In the step.2, the nearer self samples are closed to the center of self region, the lar-
ger the radii of them. So the radii of self samples near the boundary of self region are
adjusted to be a relatively rational level so that the boundary invasion is avoided,
while the around the central region, each sample’s radius is adjusted to be a relatively
larger level than before so that the overlapping is severer than before. The time com-
plexity of this step is O(n).
In the step.3, by discarding the unnecessary samples, the problem with more over-
lapping in the step.2 of optimization is solved. The samples with larger level radius
can cover more space than the fixed radius samples before, so that, the self region can
be covered by fewer samples than before. The time complexity of this step is O(n2).
Via the optimization, the self set is more rational than before. The total time com-
plexity of optimization is O(n2).

4 Experiments and Results

To test the algorithm described in the previous section, experiments were carried out
using 2-dimensional synthetic data and real network data.

4.1 2-Dimensional Synthetic Data

Firstly, some self samples chosen from a fixed number of random points, which fit
with the normal distribution, are normalized to compose the self set (198 samples,
radius is fixed as 0.05). So this set has no wrong samples, and the step.1 optimization
will be illustrated in the real network experiment. Each result of optimization is
shown by the figure 2.

(a) The initial self set (b) The self set optimized by step.2 (c) The self set optimized by step.3

(d) Detectors trained by (e) Detectors trained by self set (f) Detectors trained by self set
initial self set optimized by step.2 optimized by step.3

Fig. 2. Experiment results on 2-dimensional synthetic data

118 L. Xi, F. Zhang, and D. Wang

The figure 2(a) shows the initial self set, and we can see that the overlapping and
the boundary invasion are extremely severe. The detectors trained by the initial self
set are shown in figure 2(d), and we can see clearly that the boundary of the non-
self is not covered completely which is the consequence of the self samples bound-
ary invasion. Likewise, the optimized self set and its detector set are shown by the
other figures of the figure 2 (the unit radius is set as 0.005). After all the steps of
optimization, the number of optimized self is dropped to 67, and the overlapping(S)
is dropped to 2.71% (by the equation 3). As the figures show, adjusting each sam-
ple’s radius avoids the boundary invasion so that the non-self region’s boundary is
covered well; and then the coverage of self region by the most reasonable self sam-
ples almost staying the same as before.

4.2 Real Network Data

In the real network experiment, the self samples are come from the network security
Lab of the HUST (Harbin University of Science and Technology). The value is 2-
dim: pps(packages per second) and bps(bytes per second). We collected 90 samples
and drew the joint probability distribution in figure 3. And we can see that the joint
distribution is quite similar to the standardized Gaussian distribution.

(a) The joint probability distribution of pps and bps (b) Two-dimensional standardized Gaussian

Fig. 3. The two-dimensional joint probability distribution

The results of optimization are shown by the figure 4. The optimized self sets are
in figure 4(a, b, c, d), and the corresponding detector sets are in the other figures of
figure 4.
As shown in the figure 4(a), the self set had 3 wrong self samples which were far
away from the self region. Therefore, the detector set generated by the initial self set
has holes shown in the figure 4(e). After the step.1, the wrong samples were discarded
(shown in the figure 4(b)). The results of other steps of optimization are just like the
simulation experiment above. The final self set has 19 samples.
To perform a comparison of generating detector set between the initial self set and
the optimized one, we used the two data sets to generate 3000 detectors. We can
clearly see the process from the figure 5, the detector generating speed using opti-
mized self set is faster than using initial self set.
 Optimization of Real-Valued Self Set for Anomaly Detection 119

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 4. Real network experiment results

Fig. 5. The detector generating speed comparison between initial and optimized self set

To perform a comparison of the detection effect of the two detector sets, we used
the SYN flooding to attack the two hosts with different detector sets simultaneously.
All the results shown in the table 1 are average of 100 times experiment. From the
table we can see clearly that the detection rate with detectors using optimized self set
is much higher than the other (98.9% vs. 83.7%) and the false alarm rate is remark-
able lower than the other (2.3% vs.8.3%).

Table 1. The Detection Comparison Between The Two Detector Sets

False Alarm
Detector Set Connection Invasion Detecting Detection Rate
Rate
By initial 13581653.3 100 92.5 83.7% 8.3%
By optimized 13581653.3 100 101.2 98.9% 2.3%
120 L. Xi, F. Zhang, and D. Wang

5 Conclusion
In this paper we have shown that the real-valued self set’s optimization in anomaly
detection with the probability theory. Experimental results demonstrated that this
optimization is necessary and the method is obviously effective. Moreover, the opti-
mized self set provides a favorable foundation of generating detectors. The detector
set trained by the optimized self set is more reliable because of the advantages of the
optimization algorithm:
1. Holes can be better covered by the generated detectors because of discarding the
wrong self samples.
2. Time to generate detectors is saved by using smaller number of self samples, which
also requires less space to store them.
3. The radius of the border self samples is adjusted to a rational range so that the
boundary invasion is solved and the false alarm rate is reduced obviously.
The influence of real-valued self set optimization needs further study, including more
real network experiment and strict analysis. The implication of self radius, or how to
interpret each self sample, is also an important topic to be explored.

Acknowledgment
This work was sponsored by the National Natural Science Foundation of China
(Grant No.60671049), the Subject Chief Foundation of Harbin (Grant
No.2003AFXXJ013), the Education Department Research Foundation of Heilongji-
ang Province (Grant No. 10541044, 1151G012).

References
1. Patcha, A., Park, J.M.: An overview of anomaly detection techniques: existing solutions and
latest technological trends. Computer Networks 51, 3448–3470 (2007)
2. Hofmeyr, S., Forrest, S.: Architecture for an artificial immune system. IEEE Transactions on
Evolutionary Computation 8, 443–473 (2000)
3. Simon, P.T., Jun, H.: A hybrid artificial immune system and self and self organising map for
network intrusion detection. Information Sciences 178, 3024–3042 (2008)
4. Forrest, S., Perelson, A.S., Allen, L.: Self-non-self discrimination in a computer. In: Proc. of
the IEEE Symposium on Research in Security and Privacy, pp. 202–212 (1994)
5. Dasgupta, D., Gonzalez, F.: An immunity based technique to characterize intrusions in com-
puter network. IEEE Transactions on Evolutionary Computation 69, 281–291 (2002)
6. Boukerche, A., Machado, R.B., Juca, R.L.: An agent based and biological inspired real-time
intrusion detection and security model for computer network operations. Computer Com-
munications 30, 2649–2660 (2007)
7. Ji, Z., Dasgupta, D.: Real-valued negative selection algorithm with variable-sized detectors.
In: Deb, K., et al. (eds.) GECCO 2004, Part I. LNCS, vol. 3102, pp. 287–298. Springer,
Heidelberg (2004)
8. Gonzalez, F., Dasgupta, D., Kozema, D.: Combining negative and classification techniques
for anomaly detection. In: Proc. of the 2002 Congress on Evolutionary Computation,
pp. 705–710 (2002)
 A GP Process Mining Approach from a Structural
Perspective

Anhua Wang, Weidong Zhao∗, Chongchen Chen, and Haifeng Wu

Software School, Fudan University, 220 Handan Rd.

200433 Shanghai, China
{072053026,wdzhao,07302010048,082053025}@fudan.edu.cn

Abstract. Process mining is the automated acquisition of process models from

event workflow logs. And the model’s structural complexity directly impacts
readability and quality of the model. Although many mining techniques have
been developed, most of them ignore mining from a structural perspective. Thus
in this paper, we have proposed an improved genetic programming approach
with a partial fitness, which is extended from the structuredness complexity
metric so as to mine process models, which are not structurally complex. Addi-
tionally, the innovative process mining approach using complexity metric and
tree based individual representation overcomes the shortcomings in previous
genetic process mining approach (i.e., the previous GA approach underperforms
when dealing with process models with short parallel and OR structure, etc).
Finally, to evaluate our approach, experiments have also been conducted.

1 Introduction
Today, information about business processes is mostly recorded by enterprise infor-
mation systems such as ERP and workflow management systems in the form of
so-called “event logs” [1]. As in many domains processes are evolving and become
more and more complex, there is a need to understand the actual processes based on
logs. Thus process mining is employed since it automatically analyzes these logs to
extract explicit process models. Currently, most of process mining techniques try to
mine models from the behavior aspect only (i.e., reflect the exact behavior expressed in
logs) while ignore the complexity of mined models. However, as processes nowadays
are becoming more complex, process designers and analysts also consider mining
structurally simple models since complex processes may result in errors, bad under-
standability, defects and exceptions [2]. Therefore, “good” process models are required
to conform to logs and somehow have simple structure to clearly reflect the desired
behavior [3]. Thus to consider both behaviorally and structurally, we propose a genetic
process mining approach coupled with a process complexity metric.
In fact, utilizing evolutionary computation in process mining research is not a new
concept. In 2006, Alves introduced the genetic algorithm (GA) to mine process
models due to its resilience to noisy logs and the ability to produce novel sub-process

∗
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 121–130, 2009.
© Springer-Verlag Berlin Heidelberg 2009
122 A. Wang et al.

combinations [4]. In this case, an individual is a possible model and the fitness
evaluates how well an individual is able to reproduce the behavior in the log. How-
ever, since the individual was abstracted to the level of a binary string, it had prob-
lems when mining certain processes, especially those exhibiting high level of parallel
execution. Thus, a genetic programming (GP) approach coupled with graph-based
representation was proposed in [5]. Its abstraction of a directed graph structure also
allows greater efficiency in evaluating individual fitness.
However, there are two main problems in these approaches: Firstly some relations
between process tasks are not considered in the individual representation; the other is
that they neglect the complexity of mined models. Thus we improve the GP approach
by extending the individual representation in [4] and define a new part of the fitness
measure that benefits individuals with simpler structure. The new structural fitness is
based on the structuredness metric (SM) proposed in [6], which is a process complexity
metric. Our paper is organized as follows: Section 2 defines the individual representa-
tion and section 3 extends the SM. The GP approach is shown in section 4 with
experiments discussed in section 5. Finally, conclusions are drawn in section 6.

2 Extended Causal Matrix

In GP, an individual is an effectively executable program and its representation defines
the search space. The main requirement for the representation is to express the de-
pendencies between process tasks in logs. In our improved GP, programs are expressed
as extended causal matrix (ECM), which extends from the causal matrix (CM) repre-
sentation in [4]. The CM shows which tasks directly enable the execution of other tasks
via the matching of input (I) and output (O) condition functions. The relationship be-
tween tasks supported are restricted to ‘AND’ and ‘XOR’ only.

Table 1. Log of the application example

Id numSimilar Process instances

1 30 A, E, F
2 45 A, B, C, D, F
3 52 A, C, B, D, F

As we can see, without supporting other business logic, previous GA and GP ap-
proaches may cause unexpected errors. For instance, consider the event log shown in
Table 1. It shows three different process instances for job applicants. Note that when
HR screens resumes (A), he/she will decline the resume (E) or arrange examinations
including interview (B) and a written test(C) and score the applicant (D) when his/her
resume is passed. Finally, the result is sent to the applicant (F).
The result model is expected to be Fig. 1(a), where the black rectangle is an implicit
task. However, previous approaches will wrongly generate Fig. 1(b) since the indi-
vidual’s representation in these approaches cannot express the relation of exclusive
disjunction of conjunctions. Therefore, we give an improved version of the causal
matrix that supports other relations, i.e., the extended CM (ECM for short).
 A GP Process Mining Approach from a Structural Perspective 123

Fig. 1. Expected model (a) and mined result using previous GA and GP (b)

ECM can be denoted as a quadruple (T, C, I, O), where T is a finite set of activities
(tasks) appeared in workflow logs, C ⊆ T × T is the causality relation and I and O are
the input and output condition functions respectively.

Fig. 2. An example of the tree representation

Note that in previous GA approaches, condition functions in CM are expressed as

lines of code. Instead, our GP approach utilizes a syntax tree structure with maximum
depth 2. Given a task t, its input and output condition functions are called input tree and
output tree respectively. The tasks in I(t)/O(t) are leaves of the corresponding tree (if
there exists only one task, it is the root), while operations (XOR, AND and OR) are
internal nodes. The expression can also be represented in a prefix form. For example,
Fig. 2 shows the tree representation of a prefix-notation program XOR (B, AND(C,
D)), we also call it a XOR tree since the root operation is XOR. Apparently, ECM is a
general version of CM since more task relations can be reflected by ECM. The ECM for
our example is shown in Table 2 below.

Table 2. An example of ECM

TASK I(input tree) O(output tree)

A () XOR(B, AND(C, D))
B, C, D (A) (F)
E AND(C, D) (F)
F XOR(B, E) ()

The reason why we choose the tree type in ECM is that it provides the feasibility to
express the relation OR. Moreover, it facilitates some operations such as crossover and
mutation since it allows entire sub-tree structures to be swapped over with ease. Note
that there may be duplicate tasks in sub trees. Similar to original CM, our extended CM
can also deal with the loop structure.
124 A. Wang et al.

3 Complexity Metric Based on ECM

The GP approach in this paper tries to mine process models not only from the behavior
perspective but also from the structure one. This is done by defining new partial
structure fitness so as to mine process models from the structural perspective. We need
first introduce a process complexity metric based on ECM. Until now, many com-
plexity metrics of process models have been proposed and we choose the structured-
ness metric (SM)[6] in this paper. The SM focuses on the model structure. Addition-
ally, both the GA mining algorithm and SM have been implemented in the context of
ProM, which is characterized by a pluggable framework for process analysis, thus some
modifications can be made to implement the improved GP.
In this section, we extend the Petri net version of SM in [6] to propose a ECM
based SM. The idea behind this metric stems from the observation that process
models are often structured in terms of the combination of basic patterns such as
sequence, choice, parallelism and iteration [7]. To calculate the complexity, we de-
fine the component as a sub ECM to identify different kinds of structures in the ECM
and then score each structure by giving it some “penalty” value. Finally, the sum of
these values is used to measure the complexity of process models, i.e., the complexity
of the individual in the GA. To understand the idea, some definitions are given
below:
Firstly, some structural properties based on ECM are discussed.

Definition 1. Free choice. An ECM is a free-choice ECM iff for every two tasks t1 and
t2, I(t1) ∩ I(t2)≠null implies I(t1) = I(t2)
Definition 2. State machine. An ECM is a State machine iff for every task t, there exists
no AND/OR in I(t) and O(t).
Definition 3. Marked graph. An ECM is a marked graph iff for every task t, there exists
no XOR in I(t) and O(t).
A component corresponds to a behavioral pattern and is basically an ECM

Definition 4. Component. Let ECM(T, I, O) be an ECM representation of an individual

process model, C is a component iff C is a sub closure of the ECM.
The definition of a component is used to iteratively identify patterns. To define
different patterns, the following notations are needed..

Definition 5. Component notations. Let ECM(T, I, O) be a process model.

z ECM||C is a function that maps a component to an ECM.
z [ECM] is the set of non-trivial components of ECM.
To find components, we give the following definitions.
Definition 6. Annotated ECM. ECM=(T, I, O, τ ) is an annotated ECM iff (T, I, O) is
an ECM, and τ : T → ℜ+ is a function that assign a penalty to each task in ECM. Note
that initially for each task t in ECM, τ (t ) is 1.
 A GP Process Mining Approach from a Structural Perspective 125

Definition 7. Component types. Let (T, I, O) be a ECM, C [ECM] a component in

ECM, and ECM||c=(TC, IC, OC), C is a:
z Sequence component, iff ECM||C is a state machine and a marked graph.
z Maximal Sequence component iff for ∀C ' ∈ [ ECM ] , C ′ is a sequence com-
ponent: C ≠ C ' ⇒ C ⊄ C ' .
z CHOICE component, iff every task shares the same I and O.
z WHILE component, iff T is a triple set T(ti, t, to) and O(ti) = XOR(t, to), I(t) = ti,
O(t) = to, I(to) = XOR(ti, t).
z MARKED GRAPH component, iff ECM||C is a marked graph ECM.
z MAXIMAL MARKED GRAPH component iff for ∀C ' ∈ [ ECM ] , C ′ is a
MAKED-GRAPH component: C ≠ C ' ⇒ C ⊄ C ' .
z STATE MACHINE component, iff ECM||C is a state machine ECM.
z MAXIMAL STATE MACHINE component, iff for ∀C ' ∈ [ ECM ] , C ′ is a
STATE MACHINE component: C ≠ C ' ⇒ C ⊄ C ' .
z WELL-STRUCTURED component, iff for every task t in ECM, if O(t) is a
XOR-tree, every task in O(t) shares the same input; if I(t) is a
AND-tree/OR-tree, there is no XOR in its branches; no cycles.
z MAXIMAL WELL-STRUCTURED component, iff for ∀C ' ∈ [ ECM ] , C ′ is a
WELL-STRUCTURED component: C ≠ C ' ⇒ C ⊄ C ' .
z UNSTRUCTURED-component, if none of above definitions applies for C.
z MINIMAL UNSTRUCTURED component, iff for ∀C ' ∈ [ ECM ] , C ′ is a
UNSTRUCTURED component : C ≠ C ' ⇒ C ⊄ C ' .
To calculate the ECM-based SM, we try to identify atomic patterns in the form of
component types defined above and then iteratively replace them by tasks. Each time
when a component is removed, the penalty value of this component is calculated and
associated with a new task. By doing this, ECM is reduced in several steps until the
trivial ECM with just one task. The SM is then associated with this task.
Next, we introduce the priority of a component type and its corresponding penalty
value, and finally the algorithm that calculates the complexity is given. The order in
which components are reduced to tasks is as follows: MAXIMAL SEQUENCE com-
ponent, CHOICE component, WHILE component, MAXIMAL MARKED GRAPH
component, MAXIMAL STATE MACHINE component, MAXIMAL WELL
STRUCTURED component and other components. Thus we first select a MAXIMAL
SEQUENCE component which is endowed with the highest priority.
Definition 8 Component penalty function. Let ECM = (T ', I ', O ',τ ) be an annotated
ECM, we define the component penalty ρ ECM : [ ECM ] → ℜ+ as follows. For
any C ∈ [ ECM ] , the penalty value associated to C is given in Table 3.
In Table 3, diff(P) is |totalNum of XOR in O(T) - totalNum of XOR in I(T)|, and
diff(T) is |totalNum of AND\OR in O(T) – totalNum of AND\OR in I(T)|. For details
on the settings of different component weight, we can refer to [6].
The algorithm that calculates the SM based on ECM is shown below.
• The program calculates the ECM of an individual
126 A. Wang et al.

Table 3. The penalty value of different component type

C (Component) ρ ECM (C )
MAXIMAL SEQUENCE ∑ t ∈T
τ (t )
CHOICE 1.5 ⋅ ∑ t ∈T
τ (t )
WHILE ∑ t ∈{ I ( C ), O ( C )}
τ ( t ) + 2 ⋅ ∑ t ∈ T τ (t )
MAXIMAL MARKED GRAPH 2⋅∑ t ∈T
τ (t ) ⋅ diff (T )
MAXIMAL STATE MACHINE 2⋅∑ t ∈T
τ (t ) ⋅ diff ( P )
MAXIMAL WELL STRUCTURED 2⋅∑ t ∈T
τ (t ) ⋅ diff ( P ) ⋅ diff (T )
otherwise 5⋅∑ t ∈T
τ (t ) ⋅ diff ( P ) ⋅ diff (T )

(i) X:=(ECM, τ ), where τ (t) = 1, ∀ t∈ T,

(ii) while [X] ≠ ∅ (i.e., X is not reduced to a ECM with
just a single task)
(iii) pick C so that C is the first priority in every
component C ' ∈ [ X ] ,
(iv) yield ECM’ by removing C from ECM and replace it by
a new task tc
(v) τ ’(tc)= ρ (C ) and τ ’(c)= τ (c) for all other t,
X

(vi) X:=(ECM’, τ ' ),

(vii) Output SM(ECM)= τ (t ) (T={t} after the ECM is reduced).

4 Genetic Process Mining Algorithm

In this section, we explain how our GP-based process mining works, and show its
ability of supporting more combined constructs and mining “uncomplex” process
models. There’re six main stages in our GP process mining algorithm: 1) Read work-
flow log; 2) Calculate dependency relations among tasks; 3) Build initial population; 4)
Calculate individuals’ fitness; 5) Return the fittest individuals; 6) If generation limit is
reached or optimum individual is found (i.e. fitness = 1.0), stop and return, else create
next population (using elitism, crossover and mutation) and do 4).

4.1 Initial Population

We build the initial causal matrices in a heuristic way, which tries to determine the
causality relation by utilizing a dependency measure, which aims to ascertain the
strength of the relationship between tasks by calculating the amount of times one task is
directly preceded by another. The measure is also able to determine which tasks are in a
loop.
Once the causality relation is determined, the condition functions I and O are ran-
domly built. i.e., every task t1 that causally precedes a task t2 is randomly inserted as a
leaf in the tree structure, which corresponds to the input condition function of t2. Op-
erators (AND, OR and XOR) are also inserted randomly to construct a tree. A similar
process is done to set the output condition function of a task.
 A GP Process Mining Approach from a Structural Perspective 127

4.2 Fitness Function

Fitness function is used to assess the quality of an individual and this is assessed by:
1) benefiting the individuals that can parse more event traces in logs (the “com-
pleteness” requirement);
2) punishing the individuals that allow for more extra behavior than the one ex-
pressed in logs (the “preciseness” requirement);
3) punishing the individuals that are complex, i.e., individuals with high SM
value(the “uncomplexity” requirement). Equation (1) depicts the fitness function of our
GP-based process mining algorithm, in which L is event logs and ECM is an individual.
The notation ECM[] represents a generation of process models. Three partial fitness
measures are detailed below.

Fitness( L, ECM , ECM []) = PFcomplete( L, ECM ) (1)

−κ * PFprecise( L, ECM , ECM [])
− χ * PFuncomplex ( ECM , ECM [])

In equation (1), the functions PFcomplete and PFprecise derived from the previous GA
approach are measured from the behavior perspective for completeness and precise-
ness. Since ECM is a general version of CM, the functions are nearly the same as those
in previous GA approach [8].
The function PFcomplete shown in equation (2) is based on the continuous parsing
of all the traces against an individual in event logs. In this equation, all missing tokens
and extra tokens left behind act as a penalty in the fitness calculation. More details can
be found in [8] Obviously, ‘OR’ construct parsing should also be considered.
PFcomplete( L, ECM ) = (2)
allParseTasks ( L, ECM ) − punishment ( L, ECM )
numTasksLog ( L)

In terms of preciseness, the individual should contain “less extra behavior”, i.e., the
individual tends to have less enabled activities. PFprecise in equation (3) provides an
measure of the amount of extra behavior an individual allows for in comparison to other
individuals in the generation [8].
PFprecise ( L, ECM , ECM []) = (3)
allEnabledTasks ( L, ECM )
max(allEnabledTasks ( L, ECM []))

As mentioned before, the GA approach and most other process mining techniques
focus on the behavior in logs but ignore the structure of the actual individual. Thus we
defined the uncomplex fitness in equation (4). It gives the transformation of the
structuredness metric values to the interval [0, 1], where max and min are the maximum
and minimum value of the entire generation of individuals respectively.
128 A. Wang et al.

PFuncomplex ( ECM , ECM []) = (4)

SM ( ECM ) − min( SM ( ECM []))
max( SM ( ECM [])) − min( SM ( ECM []))

4.3 GP Operators

Genetic operators such as selection, crossover and mutation are used to generate indi-
viduals of the next generation, and the operators in [8] have been adapted for our GP
approach. In our case, the next population consists of the best individuals and others
generated via crossover and mutation. And parents are selected from a generation by
playing a five individual tournament selection process.
In terms of crossover, a task, which exists in both parents, is selected randomly as
the crossover point. The input and output tree of the task are then split at a randomly
chosen swap point (a set of tree nodes) for each parent. Our crossover algorithm covers
the complete search space defined by ECM. The following is an illustration of the
crossover algorithm, which shows the input trees of two parents for task K.
• Crossover algorithm in our GP process mining approach
parentTask1 = AND(XOR(A,B,C),OR(B,D),E)
parentTask2 = OR(AND(A,B),XOR(C,D),F)
swap1 = XOR(A,B,C),OR(B,D); remainder1 = E
swap2 = AND(A,B),F; remainder2 = XOR(C,D)
If crossover, for each branch in swapset1
If the selected branch shares the same operation with one
of the remainder2 (e.g., XOR(A,B,C) and XOR(C,D) are both
XOR), with the equal probability, select one of the
following three crossover methods:
i) Add it as a new branch in remainder
(e.g., XOR(A,B,C),XOR(C,D))
ii) Join it with an existing branch in remainder2
(e.g., XOR(A,B,C,D))
iii) Add as a new branch, and then select the branch
with the same operation in remainder2 and remove the tasks
that are in common(i.e., the same tasks)
(e.g., XOR(A,B,C),D)
Else select one crossover method conditionally.
If the operation of selected branch is the same as the
root of parentTask2
Add as a new leaf of the root in remainder2 for each
task in selected branch
(e.g., XOR(C,D),B,D)
Else add as a new branch in remainder2
Repeat for the combination swap2 and remainder1
In the crossover code, some examples are shown to make sure you can understand.
After execution, the result child for input2(K) may be one of these:
OR(XOR(A,B,C),XOR(C,D),B,D);OR(XOR(A,B,C,D),B,D); OR(XOR(A,B,C),B,D).
The cycle is repeated for both input and output trees of the selected task.
 A GP Process Mining Approach from a Structural Perspective 129

In the mutation, with probability mutation rate, every task in the individual is mu-
tated for its input and output tree. The following program depicts the mutation.
• Mutate algorithm in our GP process mining approach
If mutation, for each task t in the individual (Assuming I(t)
= AND(XOR(A,B,C),OR(B,D),E)), one of the following opera-
tions are done:
i) Choose a branch and add a task to it (randomly chosen
from the complete set of available tasks in the individual),
if the branch is a leaf (i.e., one task), randomly add an
operation on the new branch.
(e.g., XOR(A,B,C),OR(B,D,F),E)
ii) Remove a task from a chosen branch
(e.g., XOR(A,B,C),OR(B,D))
iii) Change the operation for the chosen branch
(e.g., XOR(A,B,C),XOR(B,D),E)
iv) Redistribution the elements in I(t)
(e.g., XOR(B,C),OR(A,B),D,E)
Repeat it for output trees of each task
Both crossover and mutation operators utilize a repair routine executed after an in-
put/output has been changed to make sure that only viable changes are made.

5 Experiments
We have implemented our GP process mining approach in Java and plug it into the
ProM framework. The experiments in this section allow us to validate the approach.
Table 4 shows the parameters being used in the experiments.

Table 4. Parameters for the experiment

Population size 100

Maximum generations 1000
Elitism rate 0.02
Crossover rate 0.8
Mutation rate 0.02
κ (extra behavior fitness reduction) 0.025
χ (complex fitness reduction) 0.025

Fig. 3. Two mined models (a) using previous GA/GP approach; (b) using our GP approach
130 A. Wang et al.

Fig. 4. Another two mined models (a) using previous GA/GP approach; (b) using our GP

To compare with the previous GA/GP approach, we have tested various workflow
logs. For short, we give two typical comparisons. Fig. 3(b) is the expected model and
our GP approach has successfully mined it while previous approaches produced a
confusing model since our GP approach supports the task relation of disjunction of
conjunctions. Fig. 4 shows two mined results when mining process with short parallel.
And it illustrates that the structurally uncomplex fitness help our GP method outper-
forms when dealing with process models containing short parallel constructs.

6 Conclusions
The advantage of our GP process mining approach is that it can mine process behav-
iorally and structurally. This owes to the novel idea of combining the genetic
programming with the structuredness metric. Additionally, the extended CM for rep-
resenting the individuals provides benefits in crossover and mutation. Thus, our GP
approach outperforms other process mining approach in some aspects.

References
[1] van der Aalst, W.M.P., van Dongen, B.F., Herbst, J., Maruster, L., Schimm, G., Weijters,
A.J.M.M.: Workflow mining: a survey of issues and approaches. Data & Knowledge En-
gineering 47(2), 237–267 (2003)
[2] Cardoso, J.: Control-flow complexity measurement of processes and Weyuker’s properties.
Transactions on Enformatika, Systems Sciences and Engineering 8, 213–218 (2005)
[3] Rozinat, A., van der Aalst, W.M.P.: Conformance Testing: Measuring the Fit and Appro-
priateness of Event Logs and Process Models. In: Bussler, C.J., Haller, A. (eds.) BPM 2005.
LNCS, vol. 3812, pp. 163–176. Springer, Heidelberg (2006)
[4] Alves de Medeiros, A.K., Weijters, A.J.M.M.: Genetic Process Mining. Ph.D Thesis,
Eindhoven Technical University, Eindhoven, The Netherlands (2006)
[5] Turner, C.J., Tiwari, A., Mehnen, J.: A Genetic Programming Approach to Business
Process Mining. In: GECCO 2008, pp. 1307–1314 (2008)
[6] Lassen, K.B., van der Aalst, W.M.P.: Complexity metrics for Workflow nets. Information
and Software Technology 51, 610–626 (2009)
[7] van der Aalst, W.M.P., ter Hofstede, A.H.M., Kiepuszewski, B., Barros, A.P.: Workflow
patterns. Distributed and Parallel Databases 14(1), 5–51 (2003)
[8] Alves de Medeiros, A.K., Weijters, A.J.M.M., van der Aalst, W.M.P.: Genetic process min-
ing: an experimental Evaluation. Journal of Data Mining and Knowledge Discovery 14(2),
245–304 (2007)
 Effects of Diversity on Optimality in GA

Glen MacDonald and Gu Fang

School of Engineering, University of Western Sydney,

Locked Bag 1797, Penrith South DC, NSW 1797, Australia
canti_mac@hotmail.com, g.fang@uws.edu.au

Abstract. Genetic Algorithms (GA) is an evolutionary inspired heuristic search

algorithm. Like all heuristic search methods, the probability of locating the op-
timal solution is not unity. Therefore, this reduces GA’s usefulness in areas that
require reliable and accurate optimal solutions, such as in system modeling and
control gain setting. In this paper an alteration to Genetic Algorithms (GA) is
presented. This method is designed to create a specific type of diversity in order
to obtain more optimal results. In particular, it is done by mutating bits that are
not constant within the population. The resultant diversity and final optimality
for this method is compared with standard Mutation at various probabilities.
Simulation results show that this method improves search optimality for certain
types of problems.

Keywords: Genetic Algorithms, diversity, mutation, optimality.

1 Introduction
Genetic Algorithm (GA) is an Evolutionary Computation (EC) method proposed by
Holland [1] which searches the binary space for solutions that satisfy pre-defined
criteria. GA is a heuristic search method that contains three operators: Selection,
Combination and Mutation. It operates on a collection of individuals (populations). A
population contains individuals which have the potential to differ from each other.
The GA works by identifying individuals which most closely satisfy pre-defined
criteria, called selection. These selected individuals (the fittest) are combined. Each
combination cycle is called a generation. To create new individuals, the combined
individuals must differ. The amount of difference in the population is called its diver-
sity. Combination occurs by swapping parts between the individuals. The combination
methods can vary significantly [2-6]. Goldberg [3] suggested Single Point Crossover
(SPC) and Multi-Point Crossover (MPC) as pattern identifying crossover methods.
Punctuated Crossover [4] is a deterministic crossover method which alters the prob-
ability based on the success of previous crossover points. Random Respectful Cross-
over [5] produces offspring by copying the parts at positions where the parents are
identical and filling the remaining positions with random members from the bit set.
Disrespectful Crossover (DC) [6] is a variant on SPC, where the bits in the section
below the crossover point are inverted if they are different or have a 50% chance of
being in either state if they are the same. DC is essentially a localized high probability
mutation algorithm incorporated into a crossover function. Its purpose is in encourag-
ing search divergence, not search convergence. Reduced Surrogate [4] randomly

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 131–140, 2009.
© Springer-Verlag Berlin Heidelberg 2009
132 G. MacDonald and G. Fang

selects bits which are different as the crossover points for single or double point
crossover. This has the advantage of not performing redundant crossovers and is a
more efficient crossover as search diversity decreases. Hoshi [2] showed that encour-
aging similarly ranked individuals to combine resulted in superior solutions.
If the combination of two different individuals can lead to all individuals in the
search space then it is possible that the globally optimal solution will be found. How-
ever, in GA, like in biology, Combination and Selection are what makes the search
converge, as traits in individuals become more common after successive generations.
It is believed that these traits are what make the individuals ‘fit’. However these
‘traits’ could be sub-optimal. If this is the case, a population will fail to continually
improve and will have reached a locally optimal solution. This means that the GA
may not produce an optimal solution for the problem, i.e., the global optima.
To encourage the population to avoid local optima, individuals are altered in a
probabilistic manner, called mutation [3]. This is modeled on biological mutation,
which is the alteration of individual genes within an organisms DNA. Commonly,
mutation is implemented in GA by probabilistically changing bit states, whereby
every single bit in an individual has a specific probability of changing state. This type
of mutation is also known as De-Jong mutation [7].
The effectiveness of De-Jong mutation has proven to be dependent on the mutation
probability and the effectiveness of a mutation probability has been shown to be prob-
lem dependent [8-13]. The mutation probability affects the probability of an individ-
ual being mutated as well as the number of bits mutated within that individual. A high
mutation probability means that an individual has a high chance of having a large
number of bits altered; this results in a highly random search which mostly ignores
convergence through crossover. A low mutation probability means that an individual
has a low chance of having a small number of bits altered; this creates a search which
is highly reliant on combination.
Searches which ignore combination do not converge. Searches which are entirely
reliant on combination are prone to premature convergence. Mutation alone amounts
to a random search and forever diverges. Mutation when combined with selection and
combination is a parallel, noise tolerant, hill-climbing algorithm [8]. Hill-climbing
algorithms guarantee the location of a local optimum, but will be deceived by a prob-
lem with many local optima. Methods of determining the mutation probability based
on generation [9], population size [10], length [11], crossover success and the rate of
change in fitness have been investigated [12]. A method which prolongs the GA
search time by systematically varying the mutation rate based on population diversity
has significantly improved the final solution optimality [13]. These mutation strate-
gies have improved the GA in a problem independent fashion. However they do not
present any significant improvement to the ideology behind mutation. They do not
address the search algorithm, they primarily present methods of altering one of the
algorithms parameters in order to prolong the search or more reliably disrupt conver-
gence. This is unlikely to make the search more efficient or more robust but rather
just make it more exhaustive. They also do not allow convergence and divergence to
be complimentary.
In this paper a new mutation method is introduced to improve the diversity of
population during the convergence. This method also intends to ensure that the
convergence is not disrupted by the mutation process.
 Effects of Diversity on Optimality in GA 133

The remainder of the paper is organized as follows: section 2 details the methodol-
ogy used for maintain the diversity while ensuring convergence. The numerical
results are shown in section 3. The conclusions are then given in section 4.

2 The Proposed Mutation Method

2.1 Related Work

The difference between individuals in GA is primarily measured by the Euclidean

distance or Hamming Distance [14]. Hamming Distance [15] is the number of bits
which differ between the individuals. Diversity for a population is generally the aver-
age of the distance measure between all individuals.
Hamming Distance was used as the diversity measure in this paper as it is problem
independent. It also takes into account the unique nature of the binary space, which is
lost in the Euclidean measurement; the implemented GA is binary coded. In this paper
diversity measurement was primarily used as an analysis tool and is the average
Hamming distance between all unique individuals in the population.
When a bit has a constant state within the population it is said to have converged,
as in all future populations this bit state will never alter without mutation. When a
search is converging it has a declining population diversity. When a search has con-
verged the population has no useful diversity as combination will never create differ-
ing individuals. Traditional mutation alters every bit with a pre-set probability. Thus it
will probabilistically interrupt convergence at some point during the search. Dynamic
alteration of this convergence interruption has shown to improve the search result
[12], [13].

2.2 Proposed Method

The number of unique individuals which can be created, in a string of length n at a

distance d = [0 : n ] , through the combination of two dissimilar individuals can be
determined by the combinations equation:
n!
N d = Cnd = . (1)
D! ( n − D )!

where Nd is the number of individuals at distance d and D is the Hamming distance

between the individuals. D can be calculated as:

n n
D= − −d . (2)
2 2

In general, mutation will always be a hill-climbing algorithm and a dynamic De-Jong

mutation rate can only compromise between hill-climbing and unguided, convergence
disturbing divergence.
The proposed mutation method is named Coherent Diversity Maintenance (CDM).
This method maintains population diversity by the following means:
134 G. MacDonald and G. Fang

1. Deliberately targets non-unique individuals to be replaced with ‘Random Immi-

grants (RI)’. ‘Random Immigrants’ are randomly generated individuals.
2. Preserves bit convergence by ensuring the new RI contain the bit states con-
verged within the rest of the population.
Fundamentally, CDM provides the crossover with more samples from the current
search space as the search space for each generation is reduced through the conver-
gence of the last generation. However, CDM differs from using RI only, as RI ignores
convergence and provides crossover with new samples from the entire search space.
The effectiveness of RI style mutation is dependent on the correlation between popu-
lation size and search space size. If the search space is significantly larger than the
population size, a diverse population will give combination a more incoherent and
disconnected sample. As the search space reduces, CDM will become more beneficial
than RI alone as the CDM samples will be more representative of the converged
search space.
To evaluate the effectiveness of the CDM, it was compared with Random Immigra-
tion, standard De-Jong mutation of varying probabilities and a GA with no mutation by
using many different types of benchmark functions in the following section.

2.3 Mutation Rate

Mutation traditionally requires the specifying and tuning of the rate at which mutation
affects the population. The mutation rate for the CDM and the implemented version
of RI was determined by the number of repeated individuals in the population. There-
fore, this method requires the comparison of all individuals in the population.
The computational cost of this comparison is:

(N 2 / 2 − N ) × L . (3)

where L is the length of individual and N is the number of individuals in a population.

The mutation rate is self regulated by the amount of diversity in the population. The
repetitive combination of like individuals has a higher probability of creating identical
individuals than the combination of relatively dissimilar individuals. Thus, as the
population diversity decreases and combination results in more identical individuals,
the mutation rate increases. Replacing only repetitive individuals has the benefit of
not interrupting convergence because the product of combination is not ignored or
diverted unless combination creates redundant individuals. This ensures that conver-
gence and divergence do not compete within the search.
Even when the mutation rate becomes quite high it will not remove individuals
which are useful for selection. The main disadvantage of this is that it does not guar-
antee the avoidance of a local optima if found.

3 Results
The GA testbed used ranking selection without elitism and was tested on two popula-
tion sizes, 50 and 100, over 20 generations. The tests were conducted using both sin-
gle point and dual point crossover. The mutation methods were tested on six (6)
benchmark functions with various shapes: Two Multi-Modal functions, two single
 Effects of Diversity on Optimality in GA 135

optima functions and two flat functions. The functions were tested over 3 to 5 dimen-
sions with 6 to 12 Bit resolution per dimension. The mutation methods were com-
pared with respect to diversity, and the final fitness. The results were averaged over
100 tests.

3.1 Benchmark Functions

The following six (6) bench mark functions [16] are used in this paper to evaluate the
effectiveness of the proposed CDM. They are:
Two single-optimum functions:

1. The Ackley function:

⎛ ⎞
⎟ − exp ⎛⎜ 1 ⎞
n n
f ( x ) = −20 exp ⎜ − 0.2 ∑x ∑ cos(2πx ) ⎟⎟ + 20 + e
1 2 (4)
.
⎜⎜ ⎟⎟ ⎜n
i i
n ⎝ ⎠
⎝ i =1 ⎠ i =1

−32 ≤ xi ≤ 32, xmin = (0,L,0), and f ( x min ) = 0.

2. The Sphere function:
n
f ( x) = ∑x
i =1
2
i . (5)

−50 ≤ xi ≤ 50, xmin = (0,L,0), and f ( xmin ) = 0.

Two multi-optima functions:

3. Rastrigin’s function:
n
f ( x) = ∑( x
i =1
2
i − 10 cos(2πxi )) + 10n . (6)

−5.12 ≤ xi ≤ 5.12, xmin = (0,L,0), and f ( xmin ) = 0.

4. Griewank’s Function:
n n
⎛ xi ⎞
∑ x − ∏ cos⎜⎜⎝
1
f ( x) = 2
i ⎟⎟ + 1 . (7)
4000 i =1 i =1 i⎠

−100 ≤ xi ≤ 100, xmin = (0,L,0), and f ( xmin ) = 0.

And two flat-optimum functions:

5. Zakharov’s Function:
2 4
n ⎛ n ⎞ ⎛ n ⎞
f ( x) = − ∑ +⎜
x i2
⎜ ∑
0 .5ix i ⎟ + ⎜
⎟ ⎜
0 .5ix i ⎟ .
⎟ ∑ (8)
i =1 ⎝ i =1 ⎠ ⎝ i =1 ⎠
−10 ≤ xi ≤ 10, xmin = (0,L,0), and f ( xmin ) = 0.
136 G. MacDonald and G. Fang

6. Rosenbrock’s Function:
n
f ( x) = ∑ (100( x
i =1
i +1 − xi2 ) 2 + ( xi − 1) 2 ) . (9)

−10 ≤ xi ≤ 10, xmin = (1,L,1), and f ( xmin ) = 0.

3.2 Numerical Results of Fitness

The GA results, in terms of the mean final fitness, are shown in Tables 1 – 6 of the
above six benchmark functions. The results were obtained by using different mutation
methods – the proposed CDM method, the completely random RI method and De-
Jong method with 0.001, 0.002, 0.005 and 0 probability mutation rates.

Table 1. Ackley function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/6-Bit 0.38 0.86 1.21 0.62 0.16 1.89
Dual point/ 6-Bit 1.56 3.24 2.86 2.14 0.92 4.16
Single point/ 12Bit 0.3 0.39 0.54 0.36 0.07 0.9
Dual point/ 12Bit 1.58 2.11 1.79 1.26 0.5 2.56

Table 2. Sphere function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/ 6-Bit 0.1 0.18 0.24 0.12 0.01 0.62
Dual point/ 6-Bit 0.07 0.08 0.07 0.04 0.01 0.2
Single point/ 12Bit 0.81 1.77 1.31 0.77 0.1 3.22
Dual point/ 12Bit 0.66 0.9 0.65 0.27 0.03 1.43

Table 3. Rastrigin’s function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/ 6-Bit 1.66 2.34 4.56 4.15 4.04 4.89
Dual point/ 6-Bit 1.59 1.89 2.98 3.04 2.69 3.3
Single point/ 12Bit 6.55 9.76 10.56 9.74 7.38 15.81
Dual point/ 12Bit 5.34 6.06 5.91 5.76 4.72 7.27

Table 4. Griewank function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/ 6-Bit 0.04 0.05 0.07 0.05 0.03 0.11
Dual point/ 6-Bit 0.04 0.03 0.05 0.04 0.03 0.07
Single point/ 12Bit 0.08 0.09 0.09 0.07 0.05 0.16
Dual point/ 12Bit 0.06 0.06 0.07 0.05 0.04 0.09

Table 5. Zakharov’s function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/6-Bit 8.9 29.6 54.3 35.4 19.4 115
Dual point/6-Bit 11.1 17.9 22.7 16.4 10.6 42
Single point/12Bit 294 448 430 136 55.1 1419
Dual point/12Bit 165 257 149 84.2 33.8 305
 Effects of Diversity on Optimality in GA 137

Table 6. Rastrigin’s function final fitness comparison

Cross over points/ no. of bits CDM RI 0.001 0.002 0.005 None
Single point/6-Bit 0.01 0.01 0.07 0.06 0.05 0.09
Dual point/6-Bit 0.01 0.01 0.05 0.05 0.04 0.05
Single point/12Bit 0.01 0.03 0.1 0.06 0.03 0.17
Dual point/12Bit 0.03 0.02 0.05 0.03 0.04 0.11

From these results it can be seen that the RI produces worst results than the CDM
methods. This is expected as the CDM is designed to encourage diversity whilst still
maintaining population convergence.
The more interesting aspects of the results are observed in the fitness results for the
two single-optima functions (Tables 1-2) of CDM in comparison with the De-Jong
methods. It seems that the CDM is not generating better results than the normal prob-
abilistic mutation methods. It is believed that this is due to the fact that single optima
solutions can be found using hill-climbing methods that the De-Jong’s method falls
into. Therefore, the normal mutation method will generate acceptable results.
The CDM in general will generate better or similar results for the multi-optima func-
tions (Tables 3-4). In these cases, as the CDM is designed to encourage diversity, it is
likely to generate better, or at least no worse results than the other mutation methods.
In dealing with the flatter-optima functions (Tables 5-6), it can be seen that the
CDM’s final results are compatible to those from the normal mutation methods. This
is expected as on the flatter surface diversity will not have a significant impact on
diversity outcomes.
It is also noted in the computation that the CDM in general performs better in deal-
ing with simpler cases, i.e., with single cross-over point and with shorter bit length in
population. CDM performed relatively worse with dual point crossover (DPC) be-
cause DPC is better than single point crossover (SPC) in maintaining population di-
versity. It is believed that CDM performed relatively worse when tested with longer
individuals because the search space became too large for the given population size to
be effective.

3.3 Numerical Results on Diversity

During the computation, it is noted that the diversity varies significantly with different
mutation rates and mutation methods. It was also found that it varies between bench-
mark functions. However, the most significant effect on diversity came with search
space size, which is a combination of parameter number and parameter resolution.
Shown in Fig. 1 is the average population hamming distances for all tests in 6-Bit
and 12-Bit parameter resolutions.
It is clear that there are marked differences in search space sizes between 6-Bit and
12-Bit parameter resolutions. The average distance between all solutions at the begin-
ning for a 6-Bit 5-dimensional case is 12 Bits while it is 25 Bits for the 12-Bit case.
In all tests the diversity measurement was made after the crossover. This means
that only the useful diversity created through the mutation method is recorded, i.e.,
when the mutation method creates individuals with above average fitness and a larger
bit difference, then it will improve the diversity of the next generation. If mutation
does not create fitter individuals with a larger distance from the other relatively fit
individuals then it will not improve the next generation’s diversity.
138 G. MacDonald and G. Fang

Fig. 1. Diversity Summary

Figure 1 shows the generational diversity summaries for the 6-Bit and 12-Bit tests.
In the figure, CDM refers to Coherent Diversity Maintenance mutation. RI is for Ran-
dom Immigrants mutation and DJ .001 – DJ .005 refer to the De-Jong Mutation with
the stated probability.
It can be seen from the figure that the diversity levels of CDM and RI are similar
for both parameter resolutions. However CDM consistently creates higher population
diversity during the mid generations, whilst tapering off at the final generations. The
lower diversity relative to RI in the final generations is to be expected as CDM only
creates individuals which are coherent with the converging search space whereas RI
disregards the convergence of the search space.
RI and CDM encourage roughly equivalent diversity levels for all tests but pro-
duce significantly different final fitness results, thus one can conclude that there is an
inconsistent link between diversity and superior convergence. As RI encourages new
individuals to come from the entire search space and CDM encourages new indi-
viduals to come from the converging search space one can conclude that specific
diversity gives better results than unspecific diversity. This occurs because as con-
vergence progresses, CDM is able to provide crossover with a more representative
sample of the converging space which enables it to better select the next generation
of individuals.
 Effects of Diversity on Optimality in GA 139

4 Conclusions

In this paper a new mutation scheme, named Coherent Diversity Maintenance (CDM),
is introduced. This method is designed to encourage diversity in the GA population by
maintaining the convergence trend. It is believed that this method will explore the
converging parameter space more fully therefore resulting in better convergence re-
sults. Numerical results have shown that the CDM tends to generate better, or at least
no worse, results for multi-optima or flatter optima situations than the normal prob-
abilistic mutation method and complete random mutation method.

References
1. Holland, J.H.: Adaption in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control and Artificial Intelligence. MIT Press, Cambridge (1992)
2. Chakraborty, G., Hoshi, K.: Rank Based Crossover – A new technique to improve the
speed and quality of convergence in GA. In: Proceedings of the 1999 Congress on Evolu-
tionary Computation, CEC 1999, vol. 2, p. 1602 (1999)
3. GoldBerg, D.E.: Genetic Algorithms in search, optimization and machine learning. Addi-
son-Wesley Publishing Company, Reading (1989)
4. Dumitrescu, D., Lazzerini, B., Jain, L.C., Dumitrescu, A.: Evolutionary Computation.
CRC Press, Boca Raton (2000)
5. Radcliff, N.J.: Forma Analysis and Random Respectful Recombination. In: Proceedings of
the fourth International Conference on Genetic Algorithms (1991)
6. Watson, R.A., Pollack, J.B.: Recombination Without Respect: Schema Combination and
Disruption in Genetic Algorithm Crossover. In: Proceedings of the 2000 Genetic and Evo-
lutionary Computation Conference (2000)
7. De Jong, K.A.: Analysis of the Behaviour of a class of Genetic Adaptive Systems. Techni-
cal Report, The University of Michigan (1975)
8. Hoffmann, J.P.: Simultaneous Inductive and Deductive Modeling of Ecological Systems
via Evolutionary Computation and Information Theory. Simulation 82(7), 429–450 (2006)
9. Fogarty, T.C.: Varying the Probability of mutation in the Genetic Algorithm. In: The Pro-
ceedings of the Third International Conference on Genetic Algorithms 1989, pp. 104–109
(1989)
10. Hesser, J., Manner, R.: Towards an optimal mutation probability for Genetic Algorithms.
In: Schwefel, H.-P., Männer, R. (eds.) PPSN 1990. LNCS, vol. 496, pp. 23–32. Springer,
Heidelberg (1991)
11. Back, T.: Optimal Mutation Rates in Genetic Search. In: Proceedings of the Fifth Interna-
tional Conference on Genetic Algorithms, June 1993, pp. 2–8 (1993)
12. Davis, L.: Adapting Operator Probabilities in Genetic Algorithms. In: Proceedings of the
Third International Conference on Genetic Algorithms, pp. 61–69 (1989)
13. Ursem, R.K.: Diversity-Guided Evolutionary Algorithms. In: Guervós, J.J.M., Adamidis,
P.A., Beyer, H.-G., Fernández-Villacañas, J.-L., Schwefel, H.-P. (eds.) PPSN 2002. LNCS,
vol. 2439, pp. 462–471. Springer, Heidelberg (2002)
14. Janikow, C.Z., Michalewicz, Z.: An experimental comparison of binary and floating point
representations in genetic algorithms. In: Belew, R.K., Booker, J.B. (eds.) Proceedings of
the Fourth International Conference Genetic Algorithms, pp. 31–36. Morgan Kaufmann,
San Mateo (1991)
140 G. MacDonald and G. Fang

15. Gasieniec, L., Jansson, J., Lingas, A.: Efficient Approximation Algorithms for the Ham-
ming Centre Problem. Journal of Discrete Algorithms 2(2), 289–301 (2004)
16. Kwok, N.M., Ha, Q., Liu, D.K., Fang, G., Tan, K.C.: Efficient Particle Swarm Optimiza-
tion: A Termination Condition Based on the Decision-making Approach. In: 2007 IEEE
Congress on Evolutionary Computation, Singapore (2007)
 Dynamic Crossover and Mutation Genetic Algorithm
Based on Expansion Sampling

Min Dong and Yan Wu

Department of Computer Science and Technology, Tongji University,

1239 Si Ping Road, Shanghai, 200092, China
dongmin_007@163.com

Abstract. The traditional genetic algorithm gets in local optimum easily, and its
convergence rate is not satisfactory. So this paper proposed an improvement,
using dynamic cross and mutation rate cooperate with expansion sampling to
solve these two problems. The expansion sampling means the new individuals
must compete with the old generation when create new generation, as a result,
the excellent half ones are selected into the next generation. Whereafter several
experiments were performed to compare the proposed method with some other
improvements. The results are satisfactory. The experiment results show that
the proposed method is better than other improvements at both precision and
convergence rate.

Keywords: genetic algorithm, expansion sampling, dynamic cross rate,

dynamic mutation rate.

1 Introduction

Practical application of genetic algorithm has the problems of premature convergence,

less robust in parameter selection, slow convergence speed at anaphase. These defi-
ciencies seriously limit the scope of application of genetic algorithms. How to im-
prove the performance of genetic algorithm to overcome the shortages so that it can
be better applied to practical problems, become the main subject of many scientists. In
recent years there are many improved genetic algorithms, for example, Self-adaptive
simulated binary crossover genetic algorithm [1], fuzzy adaptive genetic algorithm [2],
Multi-agent genetic algorithm [3], fast and elitist multi-objective genetic algorithm [4],
fully adaptive strategy genetic algorithm [5], quantum clone genetic algorithm [6], im-
proved mutation operator genetic algorithm [7] and so on. These new algorithms make
the performance of genetic algorithm improved, but still far from achieving the de-
sired state. In order to maintain the population diversity and avoid the high fitness
individual been selected too many times to occupy the entire population, this paper
presents a dynamic crossover probability and mutation probability. In order to im-
prove the algorithm’s convergence rate, this paper puts the new individuals with the
previous generation together and selects the best half individuals into the next-
generation. And then this paper proved the effectiveness of the improved algorithm
through several experiments.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 141–149, 2009.
© Springer-Verlag Berlin Heidelberg 2009
142 M. Dong and Y. Wu

2 Dynamic Crossover Probability

Traditional genetic algorithm uses the fixed cross-rate operator. Because the cross-
rate of individuals is the same, it will make all of the contemporary individuals in the
cross-operation retained at the same probability, thus the current better individuals
will be selected several times at the choice operation of the next round, and the rela-
tively poor individuals in current generation will be eliminated, leading the population
quickly evolving to the direction of the current optimal individual. If the current op-
timal individual is a local optimum, then the entire algorithm can easily fall into local
optimum. To avoid this situation and increase the diversity of the population, this
paper presents a dynamic cross-rate, namely using the ratio between the Euclidean
distance of two chromosomes and the Euclidean distance of the largest and the small-
est fitness individual in population as cross-rate:
| f (a ) − f (b) |
Pc = (1)
max( f ) − min( f )
where f(a) is a chromosome’s fitness, f(b) for b chromosome’s fitness, max(f) and
min(f) respectively are the largest and the smallest fitness in the population, Pc is the
crossover probability. Such cross-rate will make the individuals in the middle of
population retained at a greater chance, and the individuals at both ends of population
a greater probably cross, so that the better individuals and the poorer individuals are
both changed to avoid too intense competition in the choice operation of next round.
The poor individuals’ genes can also contribute to the development of the population.
The principle of this improvement is a simple probability problem of average dis-
tribution, in which, the crossover probability of a subject to the average distribution of
f(b). To illustrate this principle, we assume that max(f) for 2, min(f) for 0, f(b) average
distributes in the interval [0,2]. In this case, if an individual a1 is in the middle of the
interval, that is, its f(a1) is 1, then its cross-rate with b is 1/4, and if an individual a2 at
both ends of the interval, that is, its f(a2) is 0 or 2, then its cross-rate with b is 1/2,
larger than individuals in the middle. The distribution is shown in Figure 1 below
(x-axis is f(a) values, y-axis is crossover rate):
In addition, this algorithm’s dynamic crossover probability basically is effective
crossover probability (we call crossover between two long distance individuals the
effective cross-operation. Because of their relatively big difference between each
other, cross-operation changes them comparatively large. On the other side, the
cross-operation between very close chromosomes changes individuals very small,
the individuals after cross-operation almost unchanged, such cross-operation is inef-
fective),which can improve the effect of cross-operation and avoid inbreeding and
premature convergence effectively. Therefore, although the cross probability of this
paper is lower than other genetic algorithms, the overall effect is better. And the
lower cross-rate will reduce the computational complexity and accelerate the speed
of evolution.
 Dynamic Crossover and Mutation Genetic Algorithm 143

Fig. 1. Crossover rate curve

3 Dynamic Mutation Probability and Expansion Sampling

In order to avoid better individuals rapidly occupied the entire population, this paper
proposed a dynamic mutation probability using the ratio between a chromosome’s
fitness and the largest fitness in population. The mutation probability is defined
as:

f (a ) 2
pm = k *( ) 0 < k <1 (2)
max( f )
Where f(a) is the fitness of an individual a, max(f) is the largest fitness, Pm is the
mutation probability, k is a parameter, valid in interval (0,1). This mutation rate
allows the better individuals have the bigger probability in order to avoid better indi-
viduals rapidly occupying the entire population, lead the population evolving to the
direction of diversification. The distribution of mutation rate is shown in Figure 2
below (x-axis is f(a) values, y-axis is mutation rate):
At the same time, this paper uses the expansive optimal sampling, namely puts the
new individuals with the previous generation together and chooses the best half indi-
viduals into the next-generation. Traditional genetic algorithm puts the new individu-
als directly into the next generation, which will make the older better individuals
which have been crossed or mutated no chance accessing to future generations, thus
slows down the convergence rate of the population. And the expansive optimal sam-
pling makes those better individuals enter the next generation, thus rapidly accelerates
convergence speed.
The traditional genetic algorithm uses fixed mutation rate in computation. As a re-
sult the algorithm is hard to jump out of local optimum because of the relatively small
mutation probability when the algorithm gets in local optimum. And sometimes even
jumped out, there would be useless because they are not selected at the next choice
operation. To avoid this status, this paper combines the dynamic mutation rate with
144 M. Dong and Y. Wu

Fig. 2. Mutation rate curve

the expansive optimal sampling. Dynamic mutation rate makes algorithm more easily
jumping out of local optimum when the algorithm get in a local optimum, and the
expansive optimal sampling ensures this mutated and excellent individuals can be
selected into future generations, thereby greatly reducing the probability of getting in
local optimum. If the evolution enough, in theory, the genetic algorithm would not get
in a local optimum assuredly.

4 Algorithm Processes
The procedures of the proposed method are as follows:
1) Initialize the control parameters: set the population size N, the evolution genera-
tion g;
2) g = 0, randomly generate a population of N individuals Po = (x1, x2, ...);
3) Determine the fitness function, calculate the value of individual fitness f(xi), (i =
1,2, ..., N), to evaluate the individuals in population;
4) Determine the choice strategy, the algorithm population in this paper mainly
uses the method of roulette, select the advantage individuals from the individuals Pg
of generation g to form the mating pool;
5) Determine the dynamic crossover operator. Determine cross-rate of each pair of
individuals through the P c = | f ( a ) − f ( b ) | /(m ax( f ) − m in( f )) , and then
proceed cross-operation to produce the middle results;
6) Determine the dynamic mutation operator. Use the P m = k * ( f ( a ) / m ax( f ))
2

to determine mutation rate, and then proceed mutation operation;

7) The expansive optimal sampling. Put the individuals after choice, cross and mu-
tation operation with the individuals Pg before operation together and select the best
half individuals into the next-generation;
8) Stop the evolution if the results meet the termination criteria, else jump to (3).
 Dynamic Crossover and Mutation Genetic Algorithm 145

5 Simulation Experiments

5.1 Experiment 1

In this paper, the first experiment uses typical complex optimization function as
follows:
y = x *sin(10* pi * x) + 2 x ∈ [ −1, 2] (3)

Fitness function is as follows:

Fitness = x *sin(10 * pi * x) + 2 (4)

We use improved mutation operator genetic algorithm (IMGA) in paper [7] and fuzzy
adaptive genetic algorithm (FAGA) in paper [2] to compare with this paper’s dynamic
crossover and mutation genetic algorithm based on expansion sampling (ESDGA), simu-
lating in Matlab7. We change the parameters of genetic algorithm in experiment in order
to observe the performance of several algorithms better. In experiment N is the popula-
tion size, Pc is the crossover probability. In improved mutation operator genetic algo-
rithm the gene discerption proportion L=0.6, the various parameters: a1 = 0.01, b1 =
0.01, a2 = O.01, b2 = 0.05, α = 0.03, g0 = 40. The evolution generation is 500, each kind
of parameter repeat 50 times. Min and Max are the minimum and maximum optimal
solution in 50 times, Avg is the average, Gen is the convergence generations, local is the
number of local optimum solutions. The experiment’s results are shown in Table 1, Table
2 and Table 3 (Pc are useless in FAGA and ESDGA, k is used only in ESDGA):

Table 1. 50 independent results for IMGA in Experiment 1

N=100 N=120 N=80 N=50

Pc=0.75 Pc=0.9 Pc=0.8 Pc=0.5
Min 3.4126 3.7332 3.7871 3.2674
Max 3.8502 3.8500 3.8503 3.8503
Avg 3.8185 3.8425 3.8037 3.8470
gen 145 135 118 142
local 5 7 6 5

Table 2. 50 independent results for FAGA in Experiment 1

N=100 N=120 N=80 N=50

Min 3.8126 3.8332 3.8071 3.7674
Max 3.8503 3.8503 3.8502 3.8500
Avg 3.8385 3.8425 3.8337 3.8070
gen 22 23 18 14
local 3 6 4 6
146 M. Dong and Y. Wu

Table 3. 50 independent results for ESDGA in Experiment 1

N=100 N=120 N=80 N=50

k=0.6 k=0.8 k=0.4 k=0.3
Min 3.8503 3.8503 3.8503 3.8503
Max 3.8503 3.8503 3.8503 3.8503
Avg 3.8503 3.8503 3.8503 3.8503
gen 9 10 12 12
local 0 0 0 0

From the tables it is clear that IMGA needs about 130 iterations to converge, with
an average of about 12% local optimum. FAGA needs an average of about 20 itera-
tions to converge, 10% local optimum. ESDGA needs only about 12 iterations to
converge, and there is no local optimum in total 200 results, better than the two algo-
rithms. Also need to point out that for this paper’s algorithm, we have a lot of experi-
ments to select the k value, the table only lists representative values, as proved, the
effect of the algorithm are better when k above 0.3.
The speed and accuracy is a contradiction in genetic algorithms. If the algorithm
requires high speed, only evolving a few generations, the accuracy of the results will
be not satisfactory, and most likely have not convergence or convergence in local
optimum. If the algorithm continues the evolution infinitely, it must converge to the
global optimum finally, if there is a global optimum. In this paper, we use the latter
situation to compare the performance of the proposed algorithm affected by different
k values in Experiment 1. This is to say, Algorithm will evolve infinitely until the
population converges to the global optimum. We set the population size N=100 and
choose different k in the valid interval (0, 1). The result is shown in Figure 3 below
(x-axis is k values, y-axis is convergence generations):

Fig. 3. The influence of different k in experiment 1

 Dynamic Crossover and Mutation Genetic Algorithm 147

From the table we can clearly see that, the bigger k the better performance of this
algorithm. 0.3 is a dividing line, and there is not large difference above 0.6. This re-
sult subjects to the theory discussed earlier.

5.2 Experiment 2

Then we select a typical needle in haystack problem[5] (Needle-in-haystack, NiH for

short) for experiment.
a
max f ( x, y ) = ( )2 + ( x 2 + y 2 ) 2 x, y = [ −5.12, 5.12] (5)
b + ( x2 + y 2 )

In the function: a = 3.0, b = 0.05; maximum value of the function is 3600 when x = 0,
y = 0. And there are four local maximum points (-5.12,5.12), (-5.12, -5.12) ,
(5.12,5.12), (5.12, -5.12), function value is 2748.78.
This is a typical GA deceptive problem. Here we still use improved mutation op-
erator genetic algorithm (IMGA) in paper [7] and fuzzy adaptive genetic algorithm
(FAGA) in the paper [2] to compare with this paper’s dynamic crossover and muta-
tion genetic algorithm based on expansion sampling (ESDGA) in Matlab7. The evolu-
tion generation is also 500, each parameter repeat 50 times. Min and Max are the
minimum and maximum optimal solution in 50 times, Avg is the average, Gen is the
convergence generations, local is the number of local optimum solutions. The ex-
periment results are shown in Table 4, Table 5, Table 6 (Pc are useless in FAGA and
ESDGA, k is used only in ESDGA):

Table 4. 50 independent results for IMGA in Experiment 2

N=100 N=120 N=80 N=50

Pc=0.75 Pc=0.9 Pc=0.8 Pc=0.5
Min 3600 3600 2748.8 2748.8
Max 3600 3600 3600 3600
Avg 3583 3600 3583 3511
Gen 278 293 264 260
Local 1 0 1 2

Table 5. 50 independent results for FAGA in Experiment 2

N=100 N=120 N=80 N=50

Min 3600 3600 3600 2748.8
Max 3600 3600 3600 3600
Avg 3600 3600 3600 3583
Gen 46 57 48 40
Local 0 0 0 1
148 M. Dong and Y. Wu

Table 6. 50 independent results for ESDGA in Experiment 2

N=100 N=120 N=80 N=50

k=0.6 k=0.8 k=0.4 k=0.3
Min 3600 3600 3600 3600
Max 3600 3600 3600 3600
Avg 3600 3600 3600 3600
Gen 21 21 22 23
Local 0 0 0 0

As can be seen from the tables, IMGA and FAGA appear a few local optimums, in
which FAGA need about 50 iterations to converge, IMGA needs about 270 iterations
to converge. However ESDGA did not lead to local optimum, and the convergence
rate is also very fast, basically as long as 20 on average, better than IMGA and
FAGA.
Then we compare the performance of the proposed algorithm affected by different
k values in Experiment 2 as same as Experiment 1. We also set the population size
N=100 and choose different k in the valid interval (0, 1). The result is shown in
Figure 4 below (x-axis is k values, y-axis is convergence generations):

Fig. 4. The influence of different k in experiment 2

The conclusion is similar with Experiment 1 though there are a few differences, the
bigger k the better performance of this algorithm.

6 Conclusion
Simple genetic algorithm has defects of slow convergence speed and getting in local
optimum easily, making the application of genetic algorithm being limited. To solve
the two largest defects, this paper proposed an improved method. As we can see from
the experiment results, this paper’s dynamic crossover and mutation genetic algorithm
 Dynamic Crossover and Mutation Genetic Algorithm 149

based on expansion sampling improve the performance of genetic algorithm, over-

come the problems of inbreeding and premature convergence, accelerate the speed of
the algorithm convergence, effectively avoid the algorithm falling in local optimum,
make the algorithm has a strong optimization capacity.

References
1. Kalyanmoy, D., Karthik, S., Tatsuya, O.: Self-adaptive simulated binary crossover for real-
parameter optimization. In: Genetic and Evolutionary Computation Conference, pp. 1187–
1194 (2007)
2. Huang, Y.P., Chang, Y.T., Sandnes, F.E.: Using Fuzzy Adaptive Genetic Algorithm for
Function Optimization. In: Annual meeting of the North American Fuzzy Information Proc-
essing Society, June 3-6, pp. 484–489. IEEE, Los Alamitos (2006)
3. Zhong, W.C., Liu, J., Xue, M.Z., Jiao, L.C.: A Multi-agent Genetic Algorithm for Global
Numerical Optimization. IEEE Transactions on Systems, Man, and Cybernetics, Part
B 34(2), 1128–1141 (2004)
4. Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast and elitist multi-objective genetic
algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6(2), 182–197
(2002)
5. Xiang, Z.Y., Liu, Z.C.: Genetic algorithm based on fully adaptive strategy. Journal of Cen-
tral South Forestry University 27(5), 136–139 (2007)
6. Li, Y.Y., Jiao, L.C.: Quantum clone genetic algorithm. Computer Science 34(11), 147–149
(2007)
7. Li, L.M., Wen, G.R., Wang, S.C., Liu, H.M.: Independent component analysis algorithm
based on improved genetic algorithm. Journal of System Simulation 20(21), 5911–5916
(2008)
 Multidisciplinary Optimization of Airborne Radome
Using Genetic Algorithm

Xinggang Tang, Weihong Zhang, and Jihong Zhu

Laboratory of Engineering Simulation and Aerospace Computing,

The Key Laboratory of Contemporary Design and Integrated Manufacturing Technology,
Ministry of Education, Northwestern Polytechnical University, Xi’an 710072, China
nwputang@yahoo.com.cn, {zhangwh,JH.Zhu_FEA}@nwpu.edu.cn

Abstract. A multidisciplinary optimization scheme of airborne radome is pro-

posed. The optimization procedure takes into account the structural and the
electromagnetic responses simultaneously. The structural analysis is performed
with the finite element method using Patran/Nastran, while the electromagnetic
analysis is carried out using the Plane Wave Spectrum and Surface Integration
technique. The genetic algorithm is employed for the multidisciplinary optimi-
zation process. The thicknesses of multilayer radome wall are optimized to
maximize the overall transmission coefficient of the antenna-radome system
under the constraint of the structural failure criteria. The proposed scheme and
the optimization approach are successfully assessed with an illustrative numeri-
cal example.

Keywords: multidisciplinary optimization, radome, genetic algorithm,

structural failure, transmission coefficient.

1 Introduction

Airborne radomes are often used to protect antennas from a variety of environmental
and aerodynamic effects. The design of a high performance airborne radome is a
challenging task as the aerodynamic, electromagnetism, structural mechanics, etc
requirements are generally involved. The performances of the radome in different
disciplines are usually in conflict with each other. A thin and light-weighted radome
with excellent electromagnetic transmission ability is apparently structurally unreli-
able, while a well-designed radome structure with high stiffness and stability will
affirmatively have a poor electromagnetic performance. The radome design is hence a
multidisciplinary procedure because the analyses of different disciplines are tightly
coupled. However, the traditional engineering design approach separates the design
procedure into sequential stages. A design failure at certain stage would cause the
whole design to start from scratch. This will result in a tremendous cost of design
cycle and resources.
The great improvement of the optimization theory and algorithms in the past several
decades has provided an efficient solution for radome design. The optimization procedure

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 150–158, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Multidisciplinary Optimization of Airborne Radome Using Genetic Algorithm 151

can simultaneously take into account the radome characteristics in different disciplines and
searches for an optimal design with all design requirements to be well-balanced.
In the earlier researches, the techniques of simulating the electromagnetic charac-
teristics of radomes were well developed, which can be classified into two kinds: (1)
high-frequency methods, such as the Ray Tracing (RT) [1] technique based on Geo-
metric Optics (GO), Aperture Integration-Surface Integration (AI-SI) [2] and Plane
Wave Spectrum-Surface Integration (PWS-SI) [3] based on Physical Optics (PO); (2)
low-frequency methods, such as the Finite Element Method (FEM) [4] and Method of
Moment (MoM) [5]. Generally, the high-frequency methods are more suitable for
electrically large problems because of high computational efficiency, while the low-
frequency methods can provide more accurate analysis with much higher computa-
tional complexity. Since the iteration strategy with large computing cost is used in the
optimization methods, the high-frequency methods are superior to low-frequency
methods for electromagnetic analysis of the radome optimization.
Even though the researches on radome optimization were pioneered early by using a
simulated annealing technique [6], it is until recently that more advanced optimization
algorithms have been used and promoted the development of the radome design. For
example, the particle swarm optimization and the Genetic Algorithm are applied for the
radome optimization. The layered thickness of the sandwich radome wall and the shape
of the radome are optimized to maximize the overall transmission coefficient for the
entire bandwidth [7] or to minimize the boresight error [8]. A multi-objective optimiza-
tion procedure was further proposed to optimize the boresight error and power transmit-
tance simultaneously using the genetic algorithm and RT method [9]. However, these
researches on radome optimization are limited to the electromagnetic characteristics.
Recently, the Multidisciplinary Radome Optimization System (MROS) [10] is proposed
as a synthesis procedure, in which the material selection, structural analysis, probabilis-
tic fracture analysis and electromagnetic analysis are incorporated to perform the mul-
tidisciplinary optimization. This work indicates a new trend of radome optimization
which will be more applicable for practical engineering designs.
In this paper, the structural and electromagnetic characteristics are considered si-
multaneously to perform the radome design. A multidisciplinary optimization proce-
dure is developed based on the finite element model of the radome. The structural
analysis and electromagnetic analysis are carried out to obtain the structural failure
indexes and the overall transmission coefficient, respectively. The genetic algorithm
is employed for the optimization.

2 Multidisciplinary Simulation of Radome

2.1 Simulation Strategy

Normally, the Finite Element Method (FEA) and the Physical Optics (PO) method are
the preferred approaches used for the structural analysis and electromagnetic analysis
of radome, respectively. However, in the traditional design scheme illustrated in
Fig.1, these analyses, as well as the procedure of modeling and postprocessing are
implemented separately. There is no collaboration or communion between different
analysis procedures, which can no longer match the requirements of rapid design
cycle for the modern products.
152 X. Tang, W. Zhang, and J. Zhu

Finite element Electromagnetic

model model

Finite element Physical optics

analysis analysis

Analysis result Analysis result

postprocessing postprocessing
Structural Electromagnetic
analysis analysis

Fig. 1. Traditional radome analysis scheme

A significant improvement for the radome design is to integrate the interested

analyses in a multidisciplinary manner. A potential scheme of the multidisciplinary
analysis is illustrated in Fig.2. The most important step is to identify what data is to be
shared by different sub-analyses. It is known all the geometry and topology informa-
tion of the radome that has been defined in the finite element model can be shared
with the PO analysis. Furthermore, the finite element mesh can also be used as the PO
analysis mesh on condition that the element size is adapted to the wavelength accord-
ing to the Nyquist Sampling Theorem. Therefore, in the proposed analysis scheme
shown in Fig.2, the finite element model is defined as the mutual data for both struc-
tural and electromagnetic analysis.

Finite element
model

Electromagnetic
Finite element model
analysis
Physical optics
analysis

Analysis result
postprocessing

Fig. 2. Multidisciplinary radome analysis scheme

Because of requirements of the low dielectric constant and high mechanical strength,
the airborne radome is usually manufactured of sandwich laminate. The choice of mate-
rials and wall thickness has a significant influence on the structural strength and electri-
cal performance. The idea in this paper is to find the proper configuration of these
parameters to satisfy the design requirements. The validity of the design can be verified
with the structural analysis and electromagnetic analysis of the radome respectively.
 Multidisciplinary Optimization of Airborne Radome Using Genetic Algorithm 153

2.2 Structural Analysis of Radome

The basic issue in radome structural analysis is the material definition. This is very
difficult because multilayer thin walled configurations including the A sandwich, B
sandwich, and C sandwich are often used as illustrated in Fig.3. Taking A sandwich
for example, it consists of three layers: two dense high-strength skins separated by a
lower-density, lower-dielectric core material such as foam or honeycomb. This con-
figuration can provide much higher strength for a given weight than a monolithic wall
of homogeneous dielectric material. The construction of the rest two kinds of wall
configuration can be concluded accordingly. In our demonstrative application, we use
A sandwich as the wall configuration which is most frequently used.

a) A sandwich b) B sandwich c) C sandwich

Fig. 3. Multilayer radome wall configuration

Another important aspect of the radome structural analysis is the structural loads
and constraints. For common airborne radome (non high-speed radomes), the aerody-
namic load due to airflow is the main cause of mechanical stress. The radome is
tightly attached to the airframe with special bolts.
The radome structural analysis can performed in Patran/Nastran. The A sandwich
configuration is modeled in Patran Laminate Modeler by treating the sandwich layer
as an orthotropic layer with the equivalent material properties. The structural validity
of the radome can be assessed by the failure analysis of composite materials. Fre-
quently used failure criterions include the Hill, Hoffman, Tsai-Wu, Maximum Stress
and Maximum Strain.

2.3 Electromagnetic Analysis of Radome

Based on the equivalence principle, the approach of PWS-SI method can be divided
into three steps: 1) a transmitting antenna is considered and the fields transmitted to
the radome interior S1 are computed by the Plane Wave Spectrum method; 2) Trans-
mission through the radome wall is then calculated by the transmission line analogy to
obtain the fields over the radome outer surface S 2 ; 3) Radiation to the far field is then
determined by the surface integration of the equivalent currents obtained from the
tangential field components over the outer surface S 2 .
The antenna-radome system model is illustrated in Fig.4. According to the planar
slab approximation technique, the radome wall is treated as being locally planar and
modeled as an infinite planar slab with complex transmission coefficients, lying in the
tangent plane at the incident points of the radiation rays.
154 X. Tang, W. Zhang, and J. Zhu

S2
S1
J
n
M Et , H t
S2
S1
E i , H iθ
Er ,H r

Fig. 4. The antenna-radome system model

According to PWS technique, the transmitted electric field at the near-filed point
P can be calculated by
1
( ) A( k x , k y ) (1)
2
where ( x, y , z ) are the coordinates of P in the antenna coordinate system,
k0 = ik x + jk y + kk z is the radial wave number and r = ix + jy + kz is the vector of point
P. A( k x , k y ) is the so-called plane wave spectrum or the Fourier transform of the
antenna aperture, which can be formulated as
1
∫∫ E ( x, y,0)e
j ( kx x + k y y )
A( k x , k y ) = dξ dη (2)
2π S

If the antenna aperture is circularly symmetric, the surface integral (2) will regress to
a linear integral. Thus the computational complexity will be greatly reduced.
The transmission coefficients for perpendicular polarization T⊥ (θ P ) and parallel
polarization T// (θ P ) of the multilayered radome wall can be determined by the trans-
mission line analogy [11]. Thus the fields over the radome outer surface can be
formulated by
t i i
M BM BM T ( P ) BM t BM T// ( M )
t i i
(3)
M BM BM T// ( P ) BM t BM T ( M )

where
Re( E × H ∗ )
nBP = PM × nRM , t BP = nBP × nRM , PM ( x , y , z ) =
Re( E × H ∗ )

Thus, the electric far field can be calculated with a surface integral technique by

1
t
M 1 ( t
M ) e jkr R1
dS2 (4)
S2
 Multidisciplinary Optimization of Airborne Radome Using Genetic Algorithm 155

where R1 is the unit vector of the observation direction, r ′ is the unit vector of the
incidence point on the outer surface S 2 and n is the normal vector of the outer sur-
face S 2 on the incidence point.
With the far field distribution of the antenna-radome system, the electromagnetic
performance of the radome, such as the transmission ratio, boresight error, etc.

3 Optimization Using Genetic Algorithm

GA simulates the natural biological evolution strategy based on Darwinian survival of
the best fitness principal. The theoretical foundations of GA were firstly led by Hol-
land in 1975 [12]. Since then, GA and its many variations have been applied success-
fully to a wide range of different problems [13]. They provide an alternative way of
traditional gradient-based optimization techniques for those mathematically intracta-
ble optimization problems such as combinational optimization, discrete-variable op-
timization, and so on. Generally, GA is accomplished by firstly generating random
solutions for a problem. Each randomly generated solution is then evaluated by its
fitness and the good solutions are selected and recombined to form new generation
solutions. The process of optimization based on fitness and recombination to generate
variability is repeated until the best solution or the stop criterion is identified.
The optimization of radome considered in this paper is a multidisciplinary optimi-
zation which considers the performances of radomes in both structural mechanics and
electromagnetics. The complex coupling between the two disciplines makes the opti-
mization modeling and sensitivity analysis intractable. However, there is no such
barrier for applying GA to the radome optimization.
Practically, the shape of the airborne radome often conforms to the unitary aerody-
namic performance of the aircraft, which should not be changed in the subsequent
design. In our applications, we consider the multilayer sandwich as the radome wall.
The radome optimization takes the thicknesses of each layer as design variables. The
failure indexes of each layer and the overall transmission coefficient of the antenna-
radome system are calculated by the structural analysis and electromagnetic analysis
respectively. The optimization objectives are to maximize the overall transmission
coefficient and minimize the failure indexes. Since the intractability of handling
multi-objective optimization, we consider one objective as the constraint. Thus, the
multidisciplinary radome optimization can be formulated as
Maximize Tran(t1 , t2 ,L , tn )

Subject to max( f i , j ) ≤ f max (5)

ti ,min ≤ ti ≤ ti ,max

where ti denotes the thickness of the ith layer, Tran denotes the overall transmission
and f i , j is the failure index of the ith layer of the jth element.
156 X. Tang, W. Zhang, and J. Zhu

4 Numerical Results

In this section, the proposed optimization approach is demonstrated with a paraboloidal

radome. For the structural analysis, the finite element model is illustrated in Fig.5.
The A sandwich radome wall is modeled as a 7 layered laminate. Each skin is con-
structed with 3 layers, while the core is constructed with 1 layer. We suppose the
bottom of the radome is fixed to the airframe tightly, and the whole radome is endur-
ing a frontal pressure caused by the relative movement of air or wind. The structural
failure of the radome material is verified with Hill Failure Criterion of composite.
For the electromagnetic analysis, we suppose a circular aperture antenna is placed
on the bottom of radome which works at 2GHz, and the finite element model contains
2721 quadrangle elements with the maximum element area of 1.04×10-3 m2. Thus,
the finite element mesh satisfies the electromagnetic analysis requirements according
to the Nyquist Sampling Theorem, which indicates that there are at least 4 elements
within an unit area of square of the wavelength.

Fig. 5. Finite element model of a paraboloidal radome

Transmission Coefficient

Iterations

Fig. 6. The iteration history of objective function

 Multidisciplinary Optimization of Airborne Radome Using Genetic Algorithm 157

Table 1. Optimal layered thickness of the radome wall (mm)

Design Variables t1 t2 t3 t4 t5 t6 t7
Lower limit 0.1 0.1 0.1 1 0.1 0.1 0.1
Upper limit 0.5 0.5 0.5 10 0.5 0.5 0.5
Initial value 0.2 0.2 0.2 8.8 0.2 0.2 0.2
Optimal value 0.1 0.1 0.1 9.96 0.1 0.1 0.1

With Radome
Without radome
E field in dB

Radiation angle in degree

Fig. 7. Radiation patterns of the electrical far field

Table 2. Structural performance of the initial and final design

Structural Maximal Maximal Maximal

performance Plystress Displacement Failure index
Initial design 24.2 0.337 0.452
Final design 48.3 0.671 0.403

We consider the radome optimization problem (5) to maximize the transmission

coefficient and constrain the failure indexes less than 0.6. The GA is performed with
50 generations and 70 individuals for each generation. The iteration history of the
objective function is illustrated by Fig.6.
The optimal design of the radome is achieved at generation 48 with the optimal
layer thicknesses listed in Table.1. From the radiation patterns of the electrical far
field in Fig.7 and the structural performances comparison in Table.2, we can see that
the final design achieve a better electromagnetic performance at the price of the de-
pression of the structural performance.

5 Conclusion
This paper proposes a multidisciplinary optimization scheme for airborne radome
design. The design procedure considers the structural and electromagnetic performance
of the radome simultaneously. Genetic Algorithm is employed for the optimization to
158 X. Tang, W. Zhang, and J. Zhu

maximize the overall transmission coefficient under constraints of the structural failure
of the radome material. The optimization scheme is successfully validated by the de-
sign optimization of a paraboloidal radome.
This work provides the new trend of radome optimization which will be more effi-
cient and applicable for the radome design engineering. Even though the results of the
demonstration are primitive, the proposed optimization scheme and the solution ap-
proach can be easily extended to more complicated applications.

References
1. Tricoles, G.: Radiation Patterns and Boresight Error of a Microwave Antenna Enclosed in
an Axially Symmetric Dielectric Shell. J. Opt. Soc. Am. 54, 1094–1101 (1964)
2. Paris, D.T.: Computer-Aided Radome Analysis. IEEE. Trans. on AP 18, 7–15 (1970)
3. Wu, D.C.F., Rudduck, R.C.: Wave Spectrum Surface Integration Technique for Radome
Analysis. IEEE Trans. on AP 22, 497–500 (1974)
4. Povinelli, M.J., Angelo, J.D.: Finite Element Analysis of Large Wavelength Antenna Ra-
dome Problems for Leading Edge and Radar Phased Arrays. IEEE Trans. on Magnetics 27,
4299–4302 (1991)
5. Chang, D.C.: A Comparison of Computed and Measured Transmission Data for the
AGM-88 Harm Radome. Master’s thesis, Naval Postgraduate School, Monterey, California,
AD-A274868 (1993)
6. Chan, K.K., Chang, P.R., Hsu, F.: Radome design by simulated annealing technique. In:
IEEE International Symposium of Antennas and Propagation Society, pp. 1401–1404.
IEEE Press, New York (1992)
7. Chiba, H., Inasawa, Y., Miyashita, H., Konishi, Y.: Optimal radome design with particle
swarm optimization. In: IEEE International Symposium of Antennas and Propagation So-
ciety, pp. 1–4. IEEE Press, New York (2008)
8. Meng, H., Dou, W., Yin, K.: Optimization of Radome Boresight Error Using Genetic Al-
gorithm. In: 2008 China-Japan Joint Microwave Conference, pp. 27–30. IEEE Press, New
York (2008)
9. Meng, H., Dou, W.: Multi-objective optimization of radome performance with the struc-
ture of local uniform thickness. IEICE Electronics Express 5, 882–887 (2008)
10. Baker, M.L., Roughen, K.M.: Structural Optimization with Probabilistic Fracture Con-
straints in the Multidisciplinary Radome Optimization System (MROS). In: 48th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Confer-
ence, Honolulu, Hawaii (2007); AIAA-2007-2311
11. Ishimaru, A.: Electromagnetic wave propagation, radiation, and scattering. Prentice Hall,
Englewood Cliffs (1991)
12. Holland, J.H.: Adaptation in Natural and Artificial Systems. University of Michigan Press,
Michigan (1975)
13. Riolo, R., Soule, T., Worzel, B.: Genetic programming theory and practice IV. Springer,
New York (2007)
 Global Similarity and Local Variance in Human Gene
Coexpression Networks

Ivan Krivosheev1, Lei Du2, Hongzhi Wang1, Shaojun Zhang1,2,

Yadong Wang1,∗, and Xia Li1,2,∗
1
School of Computer Science and Technology, Harbin Institute of Technology Harbin,
150001, China
2
College of Bioinformatics Science and Technology, Harbin Medical University Harbin,
150081, China

Abstract. For the study presented here, we performed a comparative analysis of

whole-genome gene expression variation in 210 unrelated HapMap individuals
to assess the extent of expression divergence between 4 human populations and
to explore the connection between the variation of gene expression and func-
tion. We used the GEO series GSE6536 to compare changes in expression of
47,294 human transcripts between four human populations. Gene expression
patterns were resolved into gene coexpression networks and the topological
properties of these networks were compared. The interrogation of coexpression
networks allows for the use of a well-developed set of analytical and conceptual
tools and provides an opportunity for the simultaneous comparison of variation
at different levels of systemic organization, i.e., global vs. local network proper-
ties. The results of this comparison indicate that human co-expression networks
are indistinguishable in terms of their global properties but show divergence at
the local level.

Keywords: gene coexpression network interrogation.

1 Introduction
In the last few years, gene coexpression networks attracts attention of many re-
searches[1,2]. According to previous studies, these networks are considered as a
graph where each node represents a gene, and edges represent statistically high rela-
tionship between genes. The interaction between two genes in a gene network does
not necessarily imply a physical interaction[4]. For the study presented here, we per-
formed a comparative analysis of whole-genome gene expression variation in 210
unrelated HapMap individuals[7] to assess the extent of expression divergence be-
tween 4 human populations and to explore the connection between the variation of
gene expression and function. Gene coexpression network could be constructed by
using the GeneChip expression profiles data in NCBI GEO (Gene Expression Omni-
bus repository, database of gene expression data). [3]

∗
Corresponding authors.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 159–166, 2009.
© Springer-Verlag Berlin Heidelberg 2009
160 I. Krivosheev et al.

2 Materials and Methods

2.1 Data Sets

Expression profiles of human gene pairs were compared in order to evaluate the di-
vergence of human gene expression patterns. A total of 47,294 human transcripts for
every population were considered. All-against-all gene expression profile compari-
sons for human populations’ matrices (47294*60 CEU, 47294*45 CHB, 47294*45
JPT, and 47294*60 YRI) were used to generate population-specific coexpression
networks. For coexpression networks, nodes correspond to genes, and edges link two
genes from the same population if their expression profiles are considered sufficiently
similar.

2.2 Network Construction

There are a number of existing reverse-engineering methods to construct coexpression

network such as Relevance Networks, Bayesian networks etc. Results reported here
are for networks constructed using Algorithm for the Reconstruction of Accurate
Cellular Networks (ARACNE)[8], a novel information-theoretic algorithm for the
reverse-engineering of transcriptional networks from microarray data. ARACNE
infers interactions based on mutual information between genes, an information-
theoretic measure of pairwise correlation. ARACNE compares favorably with exist-
ing methods and scales successfully to large network sizes.
ARACNE relies on a two-step process. First, candidate interactions are identified
by estimating pairwise gene-gene mutual information (MI) and by filtering them us-
ing an appropriate threshold, I0, computed for a specific p-value, p0, in the null-
hypothesis of two independent genes. This step is almost equivalent to the Relevance
Networks method. Thus, in its second step, ARACNE removes the vast majority of
indirect candidate interactions using a well-known property of mutual information –
the data processing inequality (DPI) The DPI states that if genes g1 and g3 interact
only through a third gene, g2, (i.e., if the interaction network is g1<->…<->g2<-
>…<->g3 and no alternative path exists between g1 and g3), then the following holds
I (g1,g3) ≤min[I (g1,g2) ;I (g2,g3)] .
For presented study we set mutual information threshold value at 0.001 and DPI
value at 0.1. If the P value from this test is smaller than a predefined threshold, the
edge will be preserved in the consensus network. When those threshold and DPI val-
ues are used, coexpression networks tend to congeal into graphs that are so densely
connected as to preclude meaningful analysis of their topological properties.

3 Results and Discussion

3.1 Human Population Networks

Human population coexpression networks were evaluated with respect to a number

of parameters describing their global topological properties and found to be highly
similar (Table 1). For each network were computed a variety of parameters
 Global Similarity and Local Variance in Human Gene Coexpression Networks 161

provided by Cytoscape (open source bioinformatics software platform) plugin,

Network Analyzer[9]. The numbers of nodes and edges in each network are compa-
rable, with the CEU population network showing higher values for both. The aver-
age degree (<k>) is the average number of edges per node and gives rough
approximation of how dense the network is. The CEU population network shows a
slightly higher <k> which is consistent with the greater numbers of nodes and
edges. However, <k> is again similar for all networks and rather high. By way of
comparison, typical world-wide web networks have <k>≈7. The values of <k>
might not be particularly relevant because, as will be shown below, the degree dis-
tributions are highly skewed.
A more refined notion of network density is given by the average clustering coeffi-
cient (<C>). The clustering coefficient C of a node i is defined as the fraction of the
pairs of neighbors of node i that are linked to each other: Ci = 2ni/ ki(ki-1), where ni is
the number of observed links connecting the ki neighbors of node i and ki(ki-1)/2 is
the total number of possible links. The average clustering coefficient (<C>) is the
mean of this value for all nodes with at least two neighbors, and for all human popula-
tion networks <C>≈0.3 (Table 1). For networks of this size, these <C> values are
considered to be quite high. The high density of the coexpression networks is not
necessarily surprising because, as one could reasonably expect, co-expression is,
largely (but not entirely), transitive. In other words, if gene A is coexpressed with
genes B and C, then genes B and C are likely to be coexpressed as well.

Table 1. Network parameters

Population
Parameter
CEU CHB JPT YRI
Clustering coeffi-
0.350 0.272 0.304 0.320
cient
Network diameter 16 19 24 26
Network centrali-
0.047 0.032 0.066 0.042
zation
Average degree 4.576 10.721 16.46 13.781

Number of nodes 5546 3180 3572 3061

Number of edges 72073 17047 29398 21092

Network density 0.005 0.003 0.005 0.005

Network hetero-
1.539 1.601 1.778 1.622
geneity
Characteristic
25.991 6.297 6.733 7.246
path length
162 I. Krivosheev et al.

3.2 Intersection Network

As described above, the human population gene coexpression networks are closely
similar in terms of their global topological characteristics; they share similar node
degree (k) distributions and C(k) distributions as well as similar average node degrees
(<k>), clustering coefficients (<C>) and path lengths (<l>). Other parameters related
to neighborhood, such as network density, network centralization and network het-
erogeneity are closely similar.
We further sought to evaluate the similarity between the population-specific coex-
pression networks at a local level. There is as yet no general method for assessing
local network similarity (or graph isomorphism). However, in the case of the human
population gene coexpression networks generated here, the use of orthologous gene
pairs results in a one-to-one mapping between the nodes of the two networks. In this
sense, the networks can be considered to be defined over the same set of nodes N, and
thus can be directly compared by generating an intersection network. The human
population intersection network is defined as the network over the set of nodes N
where there is a link between two nodes i and j if i and j denote two pairs of ortholo-
gous genes which are connected in every human population network. Thus, the inter-
section network captures the coexpressed gene pairs conserved between 4 human
populations.
The global characteristics of the intersection network are shown in Table 2. The in-
tersection network node degree and C(k) distributions are clearly similar to those of
the population-specific networks as are the average clustering coefficient (<C> =
0.213) and average path length (<l> = 3.04). Network diameter equals 10. The net-
work diameter and the average shortest path length, also known as the characteristic
path length, indicate small-world properties of the analyzed network. Taken together,
these findings indicate that the global structure of the population-specific coexpres-
sion networks is preserved in the intersection network. However, the most striking
feature of the intersection network is the small fraction of genes (~20%) and edges
(~4–16%) that are conserved between populations networks (Table 3). Accordingly,
the average node degree is lower (<k> = 7.518) in the intersection network than it is
in each of the population-specific networks.

Table 2. Number of nodes and edges in intersection network

Nodes Edges
Intersection
713 2680
network
CEU 5546 (13%) 72073 (4%)
CHB 3180 (22%) 17047 (16%)
JPT 3572 (20%) 29398 (9%)
YRI 3061 (23%) 21092 (13%)
 Global Similarity and Local Variance in Human Gene Coexpression Networks 163

Table 3. Intersection network parameters

Intersection Random
Parameter
network network
Clustering coefficient 0.213 0.001
Network diameter 10 10
Network centralization 0.061 -
Average degree 7.518 0.0
Number of nodes 713 713
Number of edges 2680 2680
Network density 0.011 -
Network heterogeneity 1.592 -
Characteristic path
3.040 0.005
length

Random network generated according Erdos-Renyi model, it has same number of

nodes and edges with intersection network.

3.3 Functional Coherence of Gene Coexpression Networks

Genes in the networks were functionally categorized using their Gene Ontology (GO)
biological process annotation terms. Overrepresented GO terms were identified with
BINGO[14] by comparing the relative frequencies of GO terms in specific clusters with
the frequencies of randomly selected GO-terms. The Hypergeometric test was used to

Fig. 1. Overrepresented GO terms

164 I. Krivosheev et al.

do this with the Benjamini and Hochberg false discovery rate correction for multiple
tests and a P-value threshold of 0.001. Pairwise similarities between gene GO terms
were measured using the semantic similarity method, which computes the relative dis-
tance between any two terms along the GO-graph . Result is shown in Table 4.
The graph (Figure 1) visualizes the GO categories that were found significantly
over-represented in the context of the GO hierarchy. The size (area) of the nodes is
proportional to the number of genes in the test set which are annotated to that node.
The color of the node represents the (corrected) p-value. White nodes are not sig-
nificantly over-represented, the other ones are, with a color scale ranging from yel-
low (p-value = significance level, e.g. 0.001) to dark orange (p-value = 5 orders of
magnitude smaller than significance level, e.g. 10-5 * 0.001). The color saturates at
dark orange for p-values which are more than 5 orders of magnitude smaller than
the chosen significance level.

Table 4. GO process annotation

GO-ID Description p-value

43283 biopolymer metabolic process 3.4840E-11
42254 ribosome biogenesis 8.2487E-11
43170 macromolecule metabolic process 7.3685E-9
6259 DNA metabolic process 1.1960E-8
8152 metabolic process 1.3627E-8
44237 cellular metabolic process 2.3559E-8
ribonucleoprotein complex biogenesis
22613 4.3938E-8
and assembly
43284 biopolymer biosynthetic process 4.6317E-8
16072 rRNA metabolic process 6.9420E-8
44238 primary metabolic process 7.6631E-8
9058 biosynthetic process 1.9158E-7
6394 RNA processing 1.9444E-7
6365 rRNA processing 4.0186E-7
16070 RNA metabolic process 6.0819E-7
9059 macromolecule biosynthetic process 6.2404E-7
10467 gene expression 9.9307E-7
6260 DNA replication 1.1083E-6
34470 ncRNA processing 3.6057E-6
6974 response to DNA damage stimulus 6.6203E-6
7049 cell cycle 1.0995E-5
6996 organelle organization and biogenesis 1.4177E-5
 Global Similarity and Local Variance in Human Gene Coexpression Networks 165

In fact, from the figure it could be seen that the category ‘biopolymer metabolic
process’ is the important one, and that the over-representation of ‘macromolecule
metabolic process ' and 'metabolic process ' categories is merely a result of the pres-
ence of those 'protein modification' genes. The fact that both categories are colored
equally dark, is due to the saturation of the node color for very low p-values.

4 Conclusion
The global topological properties of the human population gene coexpression networks
studied here are very similar but the specific architectures that underlie these properties
are drastically different. The actual pairs of orthologous genes that are found to be co-
expressed in the different population are highly divergent, although we did detect a
substantial conserved component of the co-expression network. One of the most preva-
lent functional classes that show clear function-expression coherence are genes involved
in biopolymer metabolism. Example of these cluster is shown in Figure 2.
The biological relevance of the global network topological properties appears ques-
tionable[10]. Of course, this does not prevent network analysis from being a powerful
approach, possibly, the most appropriate one for the quantitative study of complex
systems made up of numerous interacting parts.

Fig. 2. Cluster of tightly coexpressed and functionally coherent genes

Acknowledgements
This work was supported in part by the National Natural Science Foundation of China
(Grant Nos. 30871394, 30370798 and 30571034), the National High Tech Develop-
ment Project of China, the 863 Program (Grant Nos. 2007AA02Z329), the National
Basic Research Program of China, the 973 Program (Grant Nos. 2008CB517302) and
the National Science Foundation of Heilongjiang Province (Grant Nos. ZJG0501,
1055HG009, GB03C602-4, BMFH060044, and D200650).
166 I. Krivosheev et al.

References
1. Horvath, S., Dong, J.: Geometric interpretation of gene coexpression network analysis.
PLoS Comput. Biol. 4(8), e1000117 (2008)
2. Carter, S.L., et al.: Gene co-expression network topology provides a framework for mo-
lecular characterization of cellular state. Bioinformatics 20(14), 2242–2250 (2004)
3. Bansal, M., et al.: How to infer gene networks from expression profiles. Mol. Syst. Biol. 3,
78 (2007)
4. Potapov, A.P., et al.: Topology of mammalian transcription networks. Genome
Inform. 16(2), 270–278 (2005)
5. Yu, H., et al.: The importance of bottlenecks in protein networks: correlation with gene
essentiality and expression dynamics. PLoS Comput. Biol. 3(4), e59 (2007)
6. Stranger, B.E., et al.: Relative impact of nucleotide and copy number variation on gene
expression phenotypes. Science 315(5813), 848–853 (2007)
7. The International HapMap Project. Nature 426(6968), 789–796 (2003)
8. Margolin, A.A., et al.: ARACNE: an algorithm for the reconstruction of gene regulatory
networks in a mammalian cellular context. BMC Bioinformatics 7(suppl. 1), S7 (2006)
9. Vlasblom, J., et al.: GenePro: a Cytoscape plug-in for advanced visualization and analysis
of interaction networks. Bioinformatics 22(17), 2178–2179 (2006)
10. Tsaparas, P., et al.: Global similarity and local divergence in human and mouse gene
co-expression networks. BMC Evol. Biol. 6, 70 (2006)
11. Khaitovich, P., et al.: A neutral model of transcriptome evolution. PLoS Biol. 2(5), E132
(2004)
12. Yanai, I., Graur, D., Ophir, R.: Incongruent expression profiles between human and mouse
orthologous genes suggest widespread neutral evolution of transcription control.
OMICS 8(1), 15–24 (2004)
13. Jordan, I.K., Marino-Ramirez, L., Koonin, E.V.: Evolutionary significance of gene expres-
sion divergence. Gene 345(1), 119–126 (2005)
14. Babu, M.M.: Introduction to microarray data analysis. In: Grant, R.P. (ed.) Computational
Genomics: Theory and Application. Horizon Press, Norwich (2004)
 A Grid Based Cooperative Co-evolutionary
Multi-Objective Algorithm

Sepehr Meshkinfam Fard1, Ali Hamzeh2, and Koorush Ziarati2

1
Department of mathematics, College of Science, Shiraz University, Shiraz, Iran
2
CSE and IT Department, School of ECE, Shiraz University, Mollasadra Ave, Shiraz, Iran
Sepehr.meshkinfam@gmail.com, ali@cse.shirazu.ac.ir,
ziarati@shirazu.ac.ir

Abstract. In this paper, a well performing approach in the context of Multi-

Objective Evolutionary Algorithm (MOEA) is investigated due to its complex-
ity. This approach called NSCCGA is based on previously introduced approach
called NSGA-II. NSCCGA performs better than NSGA-II but with a heavy load
of computational complexity. Here, a novel approach called GBCCGA is intro-
duced based on MOCCGA with some modifications. The main difference be-
tween GBCCGA and MOCCGA is in their niching technique which instead of
the traditional sharing mechanism in MOCCGA, a novel grid-based technique
is used in GBCCGA. The reported results show that GBCCGA performs
roughly the same as NSCCGA but with very low computational complexity
with respect to the original MOCCGA.

Keywords: Genetic algorithms, multi-objective, grid based, cooperative,

co-evolutionary, GBCCGA.

1 Introduction
The basic idea of Evolutionary Algorithms (EA) [1] is to encode candidate solutions
for specific problems into corresponding chromosomes and then evolve this chromo-
some via some iterative evolutions and selection phases to achieve the best possible
chromosome and decode it as the resulted solution.
Genetic algorithms are a well-known family in the area of EA which have been de-
signed and developed by John Holland in 1960s and were developed by Holland and
his colleagues and students at the university of Michigan in during the 1960s and the
1970s [2]. In this method, chromosomes are in form of strings form alphabet {0, 1}.
There exists three main phases in the evolution process of GA: Crossover or recombi-
nation, mutation and the selection phase where crossover and mutation are responsi-
ble to produce new chromosomes and selection tries to select best of them.
Genetic algorithms are well-known problem-solver in the area of multi-objective
optimization. Fonseca and Fleming [3] represented the idea of the relationship be-
tween the fitness function used in GA to discover good chromosomes and the Pareto
Optimality concept used in multi-objective optimization. Since then, several methods
in usage of genetic algorithm to solve multi-objective problems have been composed
such as those mentioned in [1].

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 167–175, 2009.
© Springer-Verlag Berlin Heidelberg 2009
168 S.M. Fard, A. Hamzeh, and K. Ziarati

One of these techniques, called MOCCGA which will be explained further, is

based on using the Co-operative Co-evolutionary GA in the search process.
A Co-evolutionary algorithm is a kind of evolutionary algorithm in which fitness
of a candidate individual is evaluated based on the relationship between candidate
individual and others.
In this paper, MOCCGA approach has been modified by replacing its niching
technique and it is proved that its performance is comparable with some recently in-
troduced techniques but in a more efficient manner.
The organization of this paper is as follows: Section 2 introduces the multi-
objective optimization problem and the approach of using EA to solve those prob-
lems. In section 3, some important related works are mentioned in brief. The original
MOCCGA and the proposed modified version are introduced in section 4. The
benchmarks and results are given in section 5 following by the conclusion.

2 Multi-Objective Optimization
Mathematical foundations of multi-objective optimization were introduced during
1895-1906 [4]. The concept of vector maximum problem was introduced by Harold
W. Kuhn and Albert W. Tucker [5].
Suppose F = ( f1 ( x), f 2 ( x),..., f m ( x)) : X → Y is a multi-objective function
where x = ( x1 , x 2 ,..., x n ) ∈ X . Now we can define a multi-objective optimization
problem as follow:
min F ( x) = ( f 1 ( x), f 2 ( x),..., f m ( x))
(1)
subject to : x ∈ X
Where X is called the decision space and Y is called the objective space.
Another important concept is domination which is defined as follows: A vector x ′
is said to dominate a vector x ′′ regarding problem F, if for each i = 1,2,..., m ,
f j ( x ′) ≤ f j ( x ′′) and there exists at least one j ∈ {1,2,..., m} that f j ( x ′) < f j ( x ′′) .
A non-dominated solution is not dominated by any other solution in the search space.
So, in the process of solving multi-objective optimization problems, the goal is to find
as much as possible of the existing non-dominate solution. Vilfredo Pareto [6] intro-
duced the concept of optimality in multi-objective optimization. A vector x p ∈ X is
said to be Pareto Optimal with respect to X if there is no vector xd ∈ X such that
F ( x d ) = ( f1 ( x d ), f 2 ( xd ),..., f m ( x d )) dominates F ( x p ) = ( f1 ( x p ), f 2 ( x p ),..., f m ( x p )) .

2.1 Multi-Objective Evolutionary Algorithms

Using EA to solve multi-objective optimization problem, called Multi-Objective

Evolutionary Algorithms (MOEAs), is based on two main separate phases:
1. To find non-dominated solutions and assigning rank to individuals.
2. Maintain diversity of the population.
These phases are explained in the following.
 A Grid Based Cooperative Co-evolutionary Multi-Objective Algorithm 169

2.1.1 Finding Non-dominated Solutions

This phase can be illustrated easily employing Multi Objective Genetic Algorithm
MOGA [3]. One of the most well-known methods in MOEAs to find the non-
dominated solutions and assign their rank is designed by Carlos M. Fonseca and Peter
J. Fleming [ 3] which is used in MOGA. To describe this ranking schema, consider
the candidate solution s which is going to be assigned a rank. Then, suppose n be the
number of all candidate solutions in the current population that dominate s, rank of s
will simply be n+1. So, it can be seen that rank of any non-dominated solution is one.

2.1.2 Diversity Preserving

Several diversity preserving techniques are examined in MOEAs. Niching methods
[7] are some of the most well-known techniques in this domain. Niching techniques
have been employed in several algorithms such as MOCCGA [8], NSGA [9], and
NSGA-II [10] and so on. In niching techniques, fitness of each candidate solution is
changed relative to the number of other solution which resides in its neighborhood
which is called the same niche. In the other word, the more candidate solutions
crowded in a niche the more their punishment. In the simplest niching techniques, the
niche is defined based on a circular, distance-based shape with the radius of σ share .
Another novel niching technique is a grid-based one which is used in Pareto Archived
Evolutionary Strategy (PAES) which is designed by J. D. Knowles and D. W. Crone
[11]. This method is the one used in this paper and will be going in details.

3 Related Works
Co-operative Co-evolutionary effect is originally used in Co-operative Co-
evolutionary Genetic Algorithm (CCGA) that was designed by Potter and De Jong
[12]. In CCGA, there exist several sub-populations, which each of them contain par-
tial solutions. Then, an individual from each subpopulation is selected and combined
with other individuals to form a total solution. Fitness of each individual will be
evaluated based on the fitness of the combined solution. After that, each sub-
population is evolved using a traditional genetic algorithm.
Keervativuttitumrong et al. [8] employed CCGA [12] in MOEA field and combine
it with MOGA [3] to form the Multi-Objective Co-operative Co-evolutionary Genetic
Algorithm (MOCCGA) [8]. This approach is described in the next section in more
details.
As the next work in the chain, in [9], Srinivas et al. introduced the non-dominated
sorting genetic algorithm (NSGA) which is based on sorting a population according to
the level of non-domination.
Deb et al. [10] proposed modified version of NSGA and named it NSGA-II
which demonstrates better performance using some tricky modification. Finally,
Iorio et al. [13] designed a co-operative co-evolutionary multi-objective genetic
algorithm using the non-dominated sorting, by combining NSGA-II with CCGA
and called NSCCGA. At the best of our knowledge, NSCCGA is the current state of
170 S.M. Fard, A. Hamzeh, and K. Ziarati

the art based on CCGA in the family of MOEA But same as previous methods it
suffers from the high computational complexity due to the non-domination sorting
mechanism. In this paper, we are going to introduce a mechanism based on CCGA.
It is to be mentioned that it has a comparative performance to NSCCGA but with
very lower computational complexity.

4 Grid Based Multi-Objective Co-operative Co-evolutionary

Genetic Algorithm

In this paper, MOCCGA has been modified employing a new niching technique
adapted from PAES [11] with some modification. This new approach is called Grid
Based multi-objective Co-operative Co-evolutionary Genetic Algorithm or GBCCGA.
To describe this new method, at first, we must describe PAES and MOCCGA in more
details.

4.1 PAES's Diversity Technique

PAES has a novel diversity technique proposed in [11] based on a low computa-
tional complexity archiving mechanism. This archive stores and maintains some of
non-dominated solutions that have been found by the algorithm from the beginning
of the evolution. Also, PAES utilize a novel fitness sharing method in which the
objective space is divided into different hyper cubes recursively. Then, each solu-
tion, in accordance with its objectives, is inserted into one of those hyper cubes and
its fitness is decreased based on the number of previously found solutions reside in
the same hypercube. This diversity technique has a lower computational complexity
comparing other diversity technique such as fitness sharing in MOGA [3] and clus-
tering in NSGA-II [10] where show promising performance to keep diversity of an
evolving population. In this technique, the objective space is divided recursively to
form a hyper grid. This hyper grid divides the objective space into hyper cubes
where each hyper cube has width of d r 2 l , where dr is the range of values in objec-
tive d of the current solution, and l is an arbitrary value selected by user.

4.2 MOCCGA in More Details

MOGA [3] and CCGA [12] was integrated together to form MOCCGA in [8]. This
algorithm decomposes the problem into sub-populations according to the dimension
of the search space. MOCCGA assigns rank to each individual of every sub-
population with respect to its sub-population. The ranking schema is similar to
MOGA's ranking schema [3]. Each candidate individual of a sub-population is com-
bined with the best individual of other sub-populations to form a complete solution.
After that, this solution is compared with other solutions and ranked based on Pareto
dominance. Finally, this rank is assigned to candidate individual. For the purpose of
diversity, fitness sharing mechanism is used and applied in the objective space. Then
each sub-population is evolved separately using traditional genetic algorithm.
 A Grid Based Cooperative Co-evolutionary Multi-Objective Algorithm 171

4.3 The Overall Architecture of GBCCGA

Considering results reported by NSGA-II [10] and NSCCGA [13], they perform better
than MOCCGA. All of them use the same niching technique and a very computational
complex domination level mechanism. But, based on our analysis, we hypothesize that
performance of MOCCGA is heavily depends on its niching technique. So, we decide
to adapt PAES's diversity mechanism, as a well-known, well performing niching
mechanism, in MOCCGA and investigate its performance. So, GBCCGA is developed
based on the original MOCCGA and a fitness sharing mechanism inspired by PAES.
Like MOCCGA, GBCCGA generates initial population randomly. It decomposes this
population into several sub-populations according to the dimension of the search space.
Each candidate individual of every sub-population is combined with the best individu-
als of other sub-populations to form a complete solution.
Again, it is combined with randomly chosen individuals of other sub-populations
to form another complete solution. Then, all of these solutions are ranked based on
Pareto non-domination schema [3].
Now, the niching mechanism is activated to calculate the penalty function for each
candidate solution. The main difference between GBCCGA and MOCCGA is in this
step where the latter uses fitness sharing mechanism from MOGA [3] and former uses
modified PAES's niching technique. The first step to adapt this technique for use in
GBCCGA is to define a policy to be able to share fitness of all individuals in a differ-
ent sub-population based on one archive.

1. for g=2 to #of generations do

2. for each sub-population do
3. Evaluate fitness of each candidate individ-
ual in a sub-population
a. Combining it with the best individuals of
other sub-populations.
b. Combining it with the random individuals of
the other sub-populations.
c. Rank all generated combined solutions (two
for each individual) based on a simple Pareto ranking
method.
d. Assign the better rank (between those two
generated combined solutions for each individual) to
each individual in the current sub-population.
e. Apply niching penalty based on the archive
grid.
f. Store rank 1 solution in the grid as described
before.
4. end for
5. end for

Fig. 1. The Overall Procedure of GBCCGA

172 S.M. Fard, A. Hamzeh, and K. Ziarati

To achieve this, we compare both combined solutions: the one which is based on
selection of the best partner and the random-based one, according to Pareto ranking
schema. Then, the better one is located in the archive grid, based on the value of its
objective function. After that, for each archived solution resides in the same hyper-
cube of that grid, one unit of penalty is applied to the rank of that combined solution.
Then, after applying penalty, the rank is assigned to the candidate individual, which
participates in combined solution, of the current sub-population as its fitness value.
Also, if the rank of the candidate individual is one, the corresponding combined solu-
tion is stored in the proper location at the archive grid. So, each candidate individual
in any sub-population can participate in forming the same archive.
Now, it can be seen that in GBCCGA, the original fitness sharing of MOCCGA is
replaced with PAES based one as described above. In figure 1, the overall procedure
of GBCCGA is depicted.

5 Test Problems and Experimental Results

The chosen benchmarks are those which are commonly used in the literature such as
[8], [9], [10] and [13]. These problems were originally developed by Zitzler et al.
[14]. They are called ZDT1 to ZDT5. ZDT1 has a convex Pareto front, ZDT2 has a
non-convex Pareto front, ZDT3’s Pareto front is a discrete one (5 discontinues convex
parts), ZDT4 with 100 Pareto optimal front where only one of them is the global op-
timum and finally ZDT5 with a deceptive Pareto front. In this section, GBCCGA’s
result is compared with the MOCCGA’s, because our algorithm is a modification of
MOCCGA, and NSCCGA as the current state of the art in CCGA family. Results of
NSCCGA are drawn directly from [13]. It must be noted that to justify our implemen-
tation1 the MOCCGA’s results are successfully regenerated in our experiments and
represented in figures 2 to 6. All results are average over 5 independent runs where all
approaches are allowed to run for a maximum number of 40 generations with the
population size of 200. Also, the parameter settings of GBCCGA are as follows:
pc=.7, pm= 0.01 and l=30.
As you can see in figures 2 to 6, GBCCGA is tested against MOCCGA and
NSCCGA. The outcome indicates that GBCCGA outperforms MOCCGA in all test
cases. Also it remains competitive to NSCCGA in ZDT1, ZDT2 and ZDT3 but In the
case of ZDT4, NSGA-II finds better convergence and spread of solutions than
GBCCGA. It can be seen that, as we hypothesized, the main weakness of MOCCGA
is its simple niching mechanism which is replaced by a more sophisticated and effec-
tive one adapted from PAES in GBCCGA. It can be seen that using this modification,
GBCCGA produced results similar to NSCCGA with less computational complexity.
Unreported results depicted that GBCCGA can solve ZDT test suite problem in about
1.8 times faster than NSCCGA (our implementation). Regarding these results and the
fact that the leveling mechanism in NSCCGA has a high level polynomial order, it
can be concluded that GBCCGA demonstrated promising research direction in the
context of MOGA generally and MOCCGA based approaches specially.

1
The implementation is available online at
http://www.cse.shirazu.ac.ir/~ali/GBCCGA/matlab-code.zip
 A Grid Based Cooperative Co-evolutionary Multi-Objective Algorithm 173

(a) (b) (c)

Fig. 2. Non-dominated solutions for ZDT1 by (a) MOCCGA, (b) GBCCGA and (c) NSCCGA

(a) (b) (c)

Fig. 3. Non-dominated solutions for ZDT2 by (a) MOCCGA, (b) GBCCGA and (c) NSCCGA

(a) (b) (c)

Fig. 4. Non-dominated solutions for ZDT3 by (a) MOCCGA, (b) GBCCGA and (c) NSCCGA

(a) (b) (c)

Fig. 5. Non-dominated solutions for ZDT4 by (a) MOCCGA, (b) GBCCGA and (c) NSCCGA
174 S.M. Fard, A. Hamzeh, and K. Ziarati

(a) (b)

Fig. 6. Non-dominated solutions for ZDT5 by (a) MOCCGA, (b) GBCCGA

6 Conclusion
This paper has demonstrated a modified version of MOCCGA by replacing its
niching technique with a Novel diversity technique adopted from PAES to form
GBCCGA that has lower computational complexity and remains competitive to cur-
rent state of the art NSCCGA [13] Since we use a parameter-free niching mechanism,
because it is not necessary to select the σ share ` factor which is highly effective in the
performance of MOCCGA. Also, as another advantage, GBCCGA has very lower
computational complexity in compare with NSCCGA while roughly produce the
same performance. So it can be concluded that the most weakness of MOCCGA is its
niching mechanism which is shown in this work to be repairable without heavy com-
putational complexity such as NSCCGA.

References
[1] Coello Coello, C.A., Lamont, G.B., Van Veldhuizen, D.A.: Evolutionary Algorithms for
Solving Multi-Objective Problems, 2nd edn. Springer Science+Business Media, LLC,
New York (2007)
[2] Holland, J.H.: Adaptation in Natural and Artificial Systems. An Introductory Analysis
with Applications to Biology. Control and Artificial Intelligence. University of Michigan
Press, Ann Arbor (1975)
[3] Fonseca, C.M., Fleming, P.J.: Genetic Algorithms for Multi-objective Optimization:
Formulation, Discussion and Generalization. In: Forrest, S. (ed.) Proceedings of the Fifth
International Conference on Genetic Algorithms, San Mateo, California, University of Il-
linois at Urbana-Champaign, pp. 416–423. Morgan Kaufmann Publishers, San Francisco
(1993)
[4] Stadler, W.: Initiators of Multicriteria Optimization. In: Jahn, J., Krabs, W. (eds.) Recent
Advances and Historical Development of Vector Optimization, pp. 3–47. Springer, Berlin
(1986)
[5] Kuhn, H.W., Tucker, A.W.: Nonlinear Programming. In: Neyman, J. (ed.) Proceedings of
the Second Berkeley Symposium on Mathematical Statistics and Probability, pp. 481–492.
University of California Press, California (1951)
[6] Pareto, V.: Cours D’Economie Politique, vol. I, II. F. Rouge, Lausanne (1896)
 A Grid Based Cooperative Co-evolutionary Multi-Objective Algorithm 175

[7] Niching Methods for Genetic Algorithms Samir W. Mahfoud 3665E. Bay Dr. #204-429
Largo, FL 34641 IIliGAL Report No. 95001 (May 1995)
[8] Keerativuttitumrong, N., Chaiyaratana, N., Varavithya, V.: Multi-objective Co-operative
Co-evolutionary Genetic Algorithm. In: Guervós, J.J.M., Adamidis, P.A., Beyer, H.-G.,
Fernández-Villacañas, J.-L., Schwefel, H.-P. (eds.) PPSN 2002. LNCS, vol. 2439, pp.
288–297. Springer, Heidelberg (2002)
[9] Srinivas, N., Deb, K.: Multiobjective Optimization Using Nondominated Sorting in
Genetic Algorithms. Evolutionary Computation 2(3), 221–248 (Fall 1994)
[10] Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A Fast and Elitist Multiobjective
Genetic Algorithm: NSGA–II. IEEE Transactions on Evolutionary Computation 6(2),
182–197 (2002)
[11] Knowles, J.D., Corne, D.W.: Approximating the Nondominated FrontUsing the Pareto
Archived Evolution Strategy. Evolutionary Computation 8(2), 149–172 (2000)
[12] Potter, M.A., de Jong, K.: A Cooperative Coevolutionary Approach to Function Optimi-
zation. In: Davidor, Y., Männer, R., Schwefel, H.-P. (eds.) PPSN 1994. LNCS, vol. 866,
pp. 249–257. Springer, Heidelberg (1994)
[13] Iorio, A.W., Li, X.: A Cooperative Coevolutionary Multiobjective Algorithm Using Non-
dominated Sorting. In: Deb, K., et al. (eds.) GECCO 2004. LNCS, vol. 3102, pp. 537–548.
Springer, Heidelberg (2004)
[14] Zitzler, E., Deb, K., Thiele, L.: Comparison of Multiobjective Evolutionary Algorithms:
Empirical Results. Evolutionary Computation 8(2), 173–195 (Summer 2000)
 Fuzzy Modeling for Analysis of Load Curve in Power
System

Pei-Hwa Huang, Ta-Hsiu Tseng, Chien-Heng Liu, and Guang-Zhong Fan

Department of Electrical Engineering, National Taiwan Ocean University,

2 Peining Road, Keelung 20224, Taiwan
B0104@mail.ntou.edu.tw

Abstract. The main purpose of this paper is to study the use of fuzzy modeling
for the analysis of customer load characteristics in power system. A fuzzy
model is a collection of fuzzy IF-THEN rules for describing the features or be-
haviors of the data set or system under study. In view of the nonlinear charac-
teristics of customer load demand with respect to time, the method of fuzzy
modeling is adopted for analyzing the studied daily load curves. Based on the
Sugeno-type fuzzy model, various models with different numbers of modeling
rules have been constructed for describing the investigated power curve. Sam-
ple results are demonstrated for illustrating the effectiveness of the fuzzy model
in the study of power system load curves.

Keywords: power system, load curve, fuzzy model, Sugeno-type fuzzy model.

1 Introduction
Understanding of system load characteristics is a primary concern in power system
planning and operations. The load curve, which is developed from the measured data
showing load demand with respect to time, plays an important role for the analysis of
an electric power customer. Based on the survey results of system load characteristics,
system engineers and operators are able to carry out the studies of load forecasting,
generation planning, system expansion, cost analysis, tariff design, and so forth. Due
to the inherent nonlinearity, many methods have been proposed for examining cus-
tomer load curves in power systems [1]-[12].
A fuzzy system model is basically a collection of fuzzy IF-THEN rules that are
combined via fuzzy reasoning for describing the features of a system under study
[13]-[21]. The method of fuzzy modeling has been proven to be well-suited for mod-
eling nonlinear industrial processes described by input-output data. The fuzzy model
describes the essential features of a system by using linguistic expressions. The fuzzy
model not only offers the accurate expression of the quantitative information for a
studied nonlinear system, but also can provide a qualitative description of the physical
feature [22]-[25]. In view of the nonlinear characteristics of a the load curve, the
fuzzy model is employed for analyzing the curve in that the method of fuzzy model-
ing is suitable for modeling nonlinear processes described by input-output data.
The main purpose of this paper is to investigate the application of the Sugeno-type
fuzzy modeling method [14] for analyzing a power system daily load curve which

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 176–184, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Fuzzy Modeling for Analysis of Load Curve in Power System 177

depicts the data of a customer load demand versus the time in a day. Different numbers
of modeling rules are taken into account in constructing the fuzzy models for repre-
senting a daily load curve. The parameters of the Sugeno-type fuzzy model are calcu-
lated by employing the algorithm of ANFIS (Adaptive Neuro-Fuzzy Inference System)
[15]-[17]. The IF-THEN structure of a fuzzy model can yield both a numerical (quanti-
tative) approximation and a linguistic (qualitative) description for the studied load
curve. It is shown from the computational results that the obtained fuzzy models are
capable of providing both quantitative and qualitative descriptions for the power sys-
tem load curve.

2 Load Curve
A load curve is obtained from the measurements of load demand of a certain cus-
tomer, region, or utility system with respect to time and thus provides a way of under-
standing how much electric energy is being consumed at different times of day. Based
on the time duration considered, there are various types of load curves: daily load
curves, weekly load curves, seasonal load curves, and annual load curves. In the load
study of power system, the daily load curve is the first concern since it reveals the
most fundamental features of a specific load [1]-[6].

Fig. 1. Residential daily load curve

A typical load curve of a residential customer in a power system is shown in Fig. 1

which is developed from a set of 24 hourly measurements and depicts the load de-
mand variations with the time in a day. It is noted that the curve is normalized with
178 P.-H. Huang et al.

the original power value of each data point converted into its percentage (%) value by
choosing the highest (peak) power demand in the curve as the base power value. The
highest (peak) demand happened around evening period after all the family returned
from work or school. It is evident that the load curve shown in Fig.1 is a nonlinear
graph and the method of fuzzy modeling is to be applied for analyzing the load curve.

3 Fuzzy Modeling
The characteristic (membership) function of a crisp (conventional) set assigns each
element of the set with a status of membership or non-membership. On the other
hand, a set consists of elements with varying degrees of membership associated with
the set is referred to as a fuzzy set [13]. Based on fuzzy set theory, the fuzzy model,
which consists of a collection of fuzzy IF-THEN rules, is used for describing the
behavior or characteristics of the data or system under study [14]-[21]. The fuzzy
model expresses an inference mechanism such that if we know a premise, then we can
infer or derive a conclusion. In modeling nonlinear systems, various types of fuzzy
rule-based system could be described by a collection of fuzzy IF-THEN rules. The
objective of this paper is to study an application of the fuzzy model to describe the
nonlinear load curve in power system.
Among various types of fuzzy models, the Sugeno-type fuzzy model has recently
become one of the major topics in theoretical studies and practical applications of
fuzzy modeling and control [14]-[17]. The basic idea of Sugeno-type fuzzy model is
to decompose the input space into fuzzy regions and then to approximate the system
in every region by a simple linear model. The overall fuzzy model is implemented by
combining all the linear relations constructed in each fuzzy region of the input space.
The main advantage of the Sugeno-type fuzzy model lies in its form with a linear
function representation in each subregion of the system.
In this study, the single-input-single-output (SISO) Sugeno-type fuzzy model con-
sisting of n modeling rules, as described in (1), is adopted for describing a load
curve:
Ri : IF T is A i THEN P = aiT + bi , i = 1, 2,L , n. (1)
where T (time in a day, Hour) denotes the system input and P (load demand, %)
represents the output in the i th subregion. It is noted that the input domain is parti-
tioned into n fuzzy subregions (time periods), each of which is described by a fuzzy
set Ai , and the system behavior in each subregion is modeled by a linear equation
P = aiT + bi in which ai is the coefficient related with time variation and bi is a
constant. Let mi denote the membership function of the fuzzy set Ai . If the system
input T is at a specific time T 0 then the overall system output P is obtained by
computing a weighted average of all the outputs Pi ’s from each modeling rule Ri as
shown in (2), (3) and (4).

wi = mi (T 0 ), i = 1, 2,L , n. (2)
 Fuzzy Modeling for Analysis of Load Curve in Power System 179

Pi = ai T 0 + bi , i = 1, 2,L , n. (3)

n n
P = (∑ wi Pi ) /(∑ wi ) . (4)
i =1 i =1

To construct a fuzzy model in (1), the task of model identification is to be performed.

Model identification is a process of recognizing the structure in data via comparisons
with some known structure. The aim of model identification is to figure out all the
parameters of a model. The model identification for the Sugeno-type fuzzy model are
usually divided into two tasks: structure identification and parameter identification.
(1) Structure Identification:
Set the number of modeling rules according to the system characteristics. Dif-
ferent numbers of modeling rules will yield different ways of description for
a load curve.
(2) Parameter Identification:
First select the type of membership functions depending on the shape de-
scribed in each fuzzy set. Then the parameters of the linear function in each
rule are to be identified. In this study, the Gaussian type membership function
in (5) is chosen for the fuzzy set in the IF-part of the modeling rule:

⎡ ( x − ci ) 2 ⎤
m i ( x) = exp ⎢ − ⎥. (5)
⎣ 2σ i2 ⎦

where the variable x stands for the time in a day and the parameters ci and
σ i define the shape of each membership function mi ( x) for the fuzzy set Ai .
Then the data points of load demand measurements are fed into an adaptive
network structure, namely the “Adaptive Neuro-Fuzzy Inference System”
(ANFIS) [15]-[17], for the purpose of model training to calculate the parame-
ters (ci , σ i ) of the membership function in the IF-part and the coefficients
(ai , bi ) of the linear function in the THEN-part.

4 Computational Results
The residential daily load curve shown in Fig. 1 is utilized as the modeling study
object. The input variable T (Hour) is the time in a day and the output P (%) is the
power demand of the customer. Note that the load curve reveals a nonlinear time-
dependent characteristic and the approach of fuzzy modeling will be suitable in the
analysis of the load curve.
The (SISO) Sugeno-type fuzzy model described in (1) is adopted for describing
this load curve. The Gaussian type membership function in (5) is employed for repre-
senting the fuzzy set A i in the IF-part of the modeling rule and the THEN-part is a
linear model. Under different preset numbers of modeling rules, the ANFIS algorithm
is used for model training to calculate the model parameters. The number of iterations
for model training is set to be 1000.
180 P.-H. Huang et al.

Table 1. Root-mean-squared errors for different numbers of modeling rules

Number of rules Root-mean-squared error

2 3.7874
3 2.4922
4 1.8353
5 1.4501
6 0.3892

Table 2. Parameters of three-rule model

Rule (ci , σ i ) (ai , bi )

1 (1.06, 4.76) (-4.39, 63.10)

2 (16.97, 3.64) (-4.32, 139.10)
3 (24.47, 4.74) (-15.18, 452.90)

Table 3. Parameters of six-rule model

Rule ( ci , σ i ) ( a i , bi )

1 (1.32, 2.53) (-5.85, 65.52)

2 (5.36, 1.30) (-2.48, 57.75)
3 (11.35, 0.93) (5.66, 11.30)
4 (14.13, 1.31) (-2.47, 130.30)
5 (19.32, 0.88) (-1.81, 132.80)
6 (24.33, 1.15) (-9.22, 302.20)

Fig. 2. Membership functions of three-rule model

 Fuzzy Modeling for Analysis of Load Curve in Power System 181

To better understand the performance of the Sugeno-type fuzzy model, compara-

tive studies are conducted under conditions of different numbers of modeling rules.
The root-mean-squared errors between the fuzzy models and the original measured
load demand data under different numbers (two, three, four, five and six) of modeling
rules are calculated and tabulated as Table 1. The parameters of the fuzzy models with
three modeling rules and six modeling rules are tabulated in Table 2 and Table 3,

Fig. 3. Membership functions of six-rule model

Fig. 4. Three-rule fuzzy model and measured load data

182 P.-H. Huang et al.

respectively. The fuzzy sets shown in Fig. 2 and Fig. 3 represent the membership
functions for the three-rule and six-rule fuzzy models. Moreover, the measured data
together with fuzzy modeling output for the three-rule and the six-rule fuzzy models
are shown in Fig. 4 and Fig. 5 for the purpose of comparison.

Fig. 5. Six-rule fuzzy model and measured load data

From the root-mean-squared errors in Table 1, it is found that more modeling rules
will yield a model with less modeling error while the number of rules vary from two
to six. It is also observed from Fig. 4 and Fig. 5 that the fitting accuracy of the six-rule
fuzzy model is better than that of the three-rule fuzzy model. Both the three-rule fuzzy
model with parameters in Table 2 and the six-rule fuzzy model with parameters in
Table 3 can provide IF-THEN rules for representing the studied load curve.
The n-rule fuzzy models in (1) can be converted into models with linguistic rules
by giving a proper time label to each fuzzy set Ai in the IF-part of the ith rule. For
example, from the parameters in Table 2 and the fuzzy sets in Fig. 2, the three-rule
fuzzy model can be converted into the following three linguistic rules in (6)-(8):
IF T is early morning, THEN P = −4.39 × T + 63.10 , (6)

IF T is afternoon, THEN P = −4.32 × T + 139.10 , (7)

IF T is night, THEN P = −15.18 × T + 452.90 . (8)

Then likewise, we can obtain a set of six linguistic rules in (9)-(14) for the six-rule
fuzzy model based on the parameters in Table 3 and the fuzzy sets in Fig. 3.
 Fuzzy Modeling for Analysis of Load Curve in Power System 183

IF T is early morning, THEN P = −5.85 × T + 65.52 , (9)

IF T is morning, THEN P = −2.48 × T + 57.75 , (10)

IF T is near noon, THEN P = 5.66 × T + 11.30 , (11)

IF T is afternoon, THEN P = −2.47 × T + 130.30 , (12)

IF T is evening, THEN P = −1.81× T + 132.80 , (13)

IF T is night, THEN P = −9.22 × T + 302.20 . (14)

Obviously, both the set of three linguistic rules in (6)-(8) and the set of six linguistic
rules in (9)-(14) are capable of providing qualitative descriptions for the load curve.

5 Conclusion
The main purpose of this paper is to present the description and analysis of the power
system load curve by fuzzy modeling. In view of the nonlinear characteristic of the
power system load curve which depicts the time variations of load demand of a cus-
tomer, the method of fuzzy modeling is employed for representing the curve. A typi-
cal residential daily load curve is employed as the studied load curve. Based on the
Sugeno-type fuzzy model, various models with different numbers of modeling rules
has been identified to describe the load curve. It is found that such fuzzy model is
capable of providing both quantitative and qualitative descriptions for the load curve.
The validty of the Sugeno-type fuzzy model in the analysis of power system load
curve has been verified in this work.

Acknowledgments. The authors would like to thank Prof. Shu-Chen Wang, Mr.
Muh-Guay Yang, and Mr. Min-En Wu for all their help with this work.

References
1. Walker, C.F., Pokoski, J.J.: Residential load shape modeling based on customer behavior.
IEEE Trans. on Power Systems and Apparatus 104, 1703–1711 (1985)
2. Talukdar, S., Gellings, C.W.: Load Management. IEEE Press, Los Alamitos (1987)
3. Bjork, C.O.: Industrial Load Management-Theory, Practice and Simulations. Elsevier,
Amsterdam (1989)
4. Tweed Jr., N.B., Stites, B.E.: Managing load data at Virginia Power. IEEE Computer
Applications in Power 5, 25–29 (1992)
5. Schrock, D.: Load Shape Development. PennWell Publishing Co. (1997)
6. Pansini, A.J., Smalling, K.D.: Guide to Electric Load Management. PennWell Publishing
Co. (1998)
184 P.-H. Huang et al.

7. Chen, C.S., Hwang, J.C., Tzeng, Y.M., Huang, C.W., Cho, M.Y.: Determination of cus-
tomer load characteristics by load survey system at Taipower. IEEE Trans. on Power
Delivery 11, 1430–1435 (1996)
8. Chen, C.S., Hwang, J.C., Huang, C.W.: Application of load survey systems to proper tariff
design. IEEE Trans. on Power System 12, 1746–1751 (1997)
9. Senjyu, T., Higa, S., Uezato, K.: Future load curve shaping based on similarity using fuzzy
logic approach. IEE Proceedings-Generation, Transmission and Distribution 145, 375–380
(1998)
10. Konjic, T., Miranda, V., Kapetanovic, I.: Prediction of LV substation load curves with
fuzzy inference systems. In: 8th International Conference on Probabilistic Methods
Applied to Power Systems, pp. 129–134. Iowa State University (2004)
11. Zhang, C.Q., Wang, T.: Clustering analysis of electric power user based on the similarity
degree of load curve. In: 4th International Conference on Machine Learning and Cybernet-
ics, vol. 3, pp. 1513–1517. IEEE Press, Los Alamitos (2005)
12. Zhang, J., Yan, A., Chen, Z., Gao, K.: Dynamic synthesis load modeling approach based
on load survey and load curves analysis. In: 3rd International Conference on Electric Util-
ity Deregulation and Restructuring and Power Technologies (DRPT 2008), pp. 1067–1071.
IEEE Press, Los Alamitos (2008)
13. Zadeh, L.A.: Fuzzy sets. Information and Control 8, 338–353 (1965)
14. Sugeno, M., Kang, G.: Structure identification of fuzzy model. Fuzzy Sets and Systems 28,
15–23 (1988)
15. Jang, J.S.R.: ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. on
Systems, Man and Cybernetics 23, 665–685 (1993)
16. Jang, J.S.R., Sun, C.T.: Neuro-fuzzy modeling and control. Proceedings of the IEEE 83,
378–406 (1995)
17. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-Fuzzy and Soft Computing. Prentice-Hall,
Englewood Cliffs (1997)
18. Bezdek, B.: Fuzzy models-What are they, and why. IEEE Trans on. Fuzzy Systems 1, 1–6
(1993)
19. Sugeno, M., Yasukawa, T.: A fuzzy-logic-based approach to qualitative modeling. IEEE
Trans. on Fuzzy Systems 1, 7–31 (1993)
20. Yager, R., Filev, D.: Essentials of Fuzzy Modeling and Control. John Wiley & Sons,
Chichester (1994)
21. Ross, T.: Fuzzy Logic with Engineering Applications. Wiley, Chichester (2004)
22. Huang, P.H., Chang, Y.S.: Fuzzy rules based qualitative modeling. In: 5th IEEE Interna-
tional Conference on Fuzzy Systems, pp. 1261–1265. IEEE Press, Los Alamitos (1996)
23. Huang, P.H., Jang, J.S.: Fuzzy modeling for magnetization curve. In: 1996 North Ameri-
can Power Symposium, pp. 121–126. Mass. Inst. Tech. (1996)
24. Huang, P.H., Chang, Y.S.: Qualitative modeling for magnetization curve. Journal of Ma-
rine Science and Technology 8, 65–70 (2000)
25. Wang, S.C., Huang, P.H.: Description of wind turbine power curve via fuzzy modeling.
WSEAS Trans. on Power Systems 1, 786–792 (2006)
 Honey Bee Mating Optimization Vector Quantization
Scheme in Image Compression

Ming-Huwi Horng

Department of Information Engineering and Computer Science, National Pingtung Institute

of Commerce, PingTung, Taiwan
horng@npic.edu.tw

Abstract. The vector quantization is a powerful technique in the applications of

digital image compression. The traditionally widely used method such as the
Linde-Buzo-Gray (LBG) algorithm always generated local optimal codebook.
Recently, particle swarm optimization (PSO) is adapted to obtain the near-global
optimal codebook of vector quantization. In this paper, we applied a new swarm
algorithm, honey bee mating optimization, to construct the codebook of vector
quantization. The proposed method is called the honey bee mating optimization
based LBG (HBMO-LBG) algorithm. The results were compared with the other
two methods that are LBG and PSO-LBG algorithms. Experimental results
showed that the proposed HBMO-LBG algorithm is more reliable and the recon-
structed images get higher quality than those generated form the other three
methods.

Keywords: Vector quantization, LBG algorithm, Particle swarm optimization,

Quantum particle swarm optimization, Honey bee mating optimization.

1 Introduction

Vector quantization techniques have been used for a number of years for data com-
pression. The operations of VQ include dividing the image to be compressed into
vectors (or blocks) and each vector is compared to the codewords of a codebook to
find its reproduction vector. The codeword, which is most similar to an input vector,
is called the reproduction vector of input vector. In the encoding process, an index,
which points to the closest codeword of an input vector, is determined. Normally, the
size of the codebook is much smaller than the original image data set. Therefore, the
purpose of image compression is achieved. In the decoding process, the associated
sub-image is exactly retrieved by the same codebook which as been used in the en-
coding phase. When each sub-image is completely reconstructed, the decoding is
completed.
Vector quantization (VQ) algorithms have been performed by many researchers;
new algorithms continue to appear. The generation of codebook is known as the most
important process of VQ. The k-means based algorithms are designed to minimize
distortion error by selecting a suitable codebook. A well-known method is the LBG

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 185–194, 2009.
© Springer-Verlag Berlin Heidelberg 2009
186 M.-H. Horng

algorithm [1]. However, the LBG algorithm is a local search procedure. It suffers
from the serious drawback that its performance depends heavily on the initial starting
conditions. Many studies have been undertaken to solve this problem. Chen, Yang
and Gou proposed an improvement based on the particle swarm optimization (PSO)
[2]. The result of LBG algorithm is used to initialize global best particle by which it
can speed the convergence of PSO. In addition, Wang et al. proposed a quantum par-
ticle swarm algorithm (QPSO) to solve the 0-1 knapsack problem [3]. Zhao et al.
employed a quantum particle swarm optimization to select the thresholds of the multi-
level thresholding [4].
Over the last decade, modeling the behavior of social insects, such as ants and
bees, for the purpose of search and problems solving had been the context of the
emerging area of swarm intelligence. Therefore, the honey-bee mating may also be
considered as a typical swarm-based approach for searching for the optimal solution
in many application domains such as clustering [5] and multi-level image threshold-
ing selection [6]. In this paper, the honey bee mating optimization (HBMO) associ-
ated with the LBG algorithm is proposed to search for the optimal codebook that
minimizes the distortion between the training set and the codebook. In other words,
HBMO algorithm is a search technique that finds the optimal codebook for the input
vectors. Experimental results have demonstrated that the HBMO-LBG algorithms
performed better than the LBG and PSO algorithms consistently.
This work is organized as follows. Section 2 introduces the vector quantization and
LBG algorithm. Section 3 presents this proposed method which searches for the
optimal codebook using the HBMO algorithm. Performance evaluation is discussed in
detail in Section 4. Conclusions are presented in Section 5.

2 Vector Quantization
This section provides some basic concepts of vectors quantization and introduces the
traditional LBG algorithm.

2.1 Definition

Pl Vector quantization (VQ) is a lossy data compression technique in block coding.

The generation of codebook is known as the most important process of VQ. Let the
{ }
size of original image Y = yij be M × M pixels that divided into several blocks with
size of n × n pixels. In other words, there are N b = ⎡ N ⎤ × ⎡ N ⎤ blocks that represented
⎢ ⎥ ⎢ ⎥
⎣n⎦ ⎣n⎦
by input vectors X = ( xi , i = 1,2,..., N b ) and n is generally assigned to be 4 in ex-
periments. Let L be n × n . The input vector xi , xi ∈ ℜ L where ℜ L is L-dimensional
Euclidean space. A codebook C comprises N c L-dimensional codewords, i.e.
{
C = c1 , c2 ,...., c N c } , c j ∈ ℜ L , ∀j = 1,2,.., N c . Each input vector is represented by a
row vector xi = ( xi1 , xi 2 ,...., xiL ) and each codeword in the codebook is denoted
 Honey Bee Mating Optimization Vector Quantization Scheme in Image Compression 187

as c j = (c j 1 , c j 2 ,..., c jL ) . The VQ techniques assign each input vector to a related

codeword, and the codeword will replace the associated input vectors finally to obtain
the aim of compression.
The optimization of C in terms of MSE can be formulated by minimizing the dis-
tortion function D. In general, the lower the value of D is, the better the quality of C
will be.
1 Nc Nb 2
D (C ) = ∑ ∑ µij ⋅ xi − c j . (1)
N b j =1 i = 1

Subject to the following constraints:

Nc
∑ µij = 1, ∀i ∈ {1,2,...., N b } . (2)
j =1

⎧1 if xi is in the jth cluster, (3)

µij = ⎨ .
⎩0 otherwise

and
Lk ≤ c jk ≤ U k , k = 1,2,..., L . (4)

where Lk is the minimum of the kth components in the all training vectors, and U k is
the maximum of the kth components in all training vectors. The x − c is the Euclid-
ean distance between the vectors x and c.
Two necessary conditions exist for an optimal vector quantizer.
(1)The codewords c j must be given by the centroid of R j :

Nj
1
cj = ∑ x i , xi ∈ R j . (5)
Nj i =1

where N j is the total number of vectors belonging to R j .

(2) The partition R j , j = 1,..., m must satisfy

R j ⊃ {x ∈ X : d ( x, c j ) < d ( x, c k )∀k ≠ j} . (6)

2.2 LBG Algorithm

An algorithm for a scalar quantizer was proposed by Lloyd [7]. Linde et al. general-
ized it for vector quantization [1]. This algorithm is known as LBG or generalized
Lloyd algorithm (GLA). It applies the two following conditions to input vectors for
determining the optimal codebooks.
188 M.-H. Horng

Given input vectors, xi , i = 1,2,.., N b , distance function d, and an initial codewords

c j (0) , j = 1,..., N c . The LBG iteratively applies the two conditions to produce opti-
mal codebook with the following algorithm:
(1). Partition the input vectors into several groups using the minimum distance
rule. This resulting partition is stored in a N b × N c binary indicator matrix U whose
elements are defined as the following:
⎧1 if d ( xi , c j (k ) = min p d ( xi , c p (k ))
µij = ⎨ (7)
⎩0 otherwise

(2). Determine the centroids of each partition. Replace the old codewords with these
centroids:
Nb
∑ µ ij xi
i =1
c j (k + 1) = Nb
, j = 1,..., N c . (8)
∑ µij
i =1

(3) Repeat steps (1) and (2) until no c j , j = 1,..., N c changes anymore.

3 HBMO-LBG Vector Quantization Algorithm

3.1 Honey Bee Mating Optimization

A honeybee colony typically consists of a single egg-laying long-lived queen, any-

where from zero to several thousands drones and usually 10,000-60,000 workers [8].
Queens are specialized in egg laying. A colony may contain one queen or more dur-
ing its life cycle, which named monogynous and/or polygynous colonies, respectively.
A queen bee may live up to 5 or 6 years, whereas worker bee and drones never live
more than 6 months. After the mating process, the drones die. The drones are the
fathers of colony. They are haploid and act as amplify their mother‘s genomes with-
out altering their genetic composition, expect through the mutation. The drones prac-
tically considered as agents that pass one of their mother’s gametes and function to
enable females to act genetically as males. Worker bees specialized in brood care and
sometimes lay eggs. Broods arise either from fertilized (represents queen or worker)
or unfertilized (represents drones) eggs.
In the marriage process, the queen(s) mate during their mating flights far from the
nest. A mating flight starts with a dance performed by the queen who then starts a
mating flight during which the drones follow the queen and mate with her in the air.
In each mating, sperm reaches the spermatheca and accumulates them to form the
genetic pool of the colony. Each time a queen lays fertilized eggs, she randomly re-
trieves a mixture of the sperm accumulated in the spermatheca to fertilize the egg. In
practices, the mating flight may considered as a set of transitions in a state-space
where the queen moves between the different states in some speed and mates with the
drone encountered at each state probability. Furthermore, the queen initializes with
 Honey Bee Mating Optimization Vector Quantization Scheme in Image Compression 189

some energy content during the flight mating and returns to her nest when the energy
is within some threshold from zero to full spermatheca.
In order to develop the algorithm, the capability of workers is restrained in brood
care and thus each worker may be regarded as a heuristic that acts to improve and/or
take care of a set of broods. An annealing function is used to describe the probability
of a drone (D) that successfully mates with the queen (Q) shown in Eq. (9).
P(Q, D ) = exp[−∆( f ) / S (t )] (9)

where ∆ ( f ) is the absolute difference of the fitness of D and the fitness of Q, and the
S (t ) is the speed of queen at time t. After each transition of mating, the queen’s
speed and energy are decayed according to the following equation:
S (t + 1) = α × S (t ) (10)

where α is the decreasing factor ( α ∈ [0, 1] ). Workers adopt some heuristic mecha-
nisms such as crossover or mutation to improve the brood’s genotype. The fitness of
the resulting genotype is determined by evaluating the value of the objective function
of the brood genotype. It is important to note that a brood has only one genotype. The
popular five construction stages of the HBMO algorithm had been proposed by
Fathian et al. [5] that are also used to develop the algorithm of multilevel image
thresholding method in this paper. The five stages are described as follow:
(1) The algorithm starts with the mating flight, where a queen (best solution) selects
drones probabilistically to form the spermatheca (list of drones). A drone then se-
lected from the list randomly for the creation of broods.
(2) Creation of new broods by crossover the drone’s genotypes with the queens.
(3) Use of workers to conduct local search on broods (trial solutions).
(4) Adaptation of worker’s fitness, based on the amount of improvement achieved on
broods.
(5) Replacement of weaker queen by fitter broods.

3.2 HBMO-Based Vector Quantization

This section introduces a new codebook design of vector quantization algorithm that
uses the honey bee mating optimization method. In the HBMO-LBG algorithm, the
solutions include the best solution; candidate solution and the trivial solution are rep-
resented in the form of codebook. Figure 1 shows the structures of drone set (trivial
solution). The same structure is used to represent the queen and breeds. The essential
of designed algorithm is based on the Fathian’s five stage scheme. Followings are the
detail algorithm: The fitness function used also defined in Eq. (11).
1 Nb
Fitness (C ) = = 2
D (C ) Nc Nb (11)
∑ ∑ µij ⋅ xi − c j
j =1i =1
190 M.-H. Horng

Fig. 1. The structure of solutions (codebooks) of drone set

Stage 1. (Generate the drone set and initial queen)

In this sage, the result of LBG algorithm is first set to the queen Q, and then the set of
drones (trivial solutions), {D1 , D2 ,..., Dm } are randomly generated. Secondly, the
fitness of each drone are computed and the largest one is pick on denoted as Dbest . If
the fitness of Dbest is larger than the fitness of the Q, the queen Q is replaced by
the Dbest .

Stage 2. (Mating flight)

In this stage, the queen selects the best mating set from the set of drones according to
the Eq. (11). The queen’s speed are decayed after each mating process until its speed
is less than the predefined minimum flight speed or its spermatheca is full. In general,
the number of the sperms, nsperm , is smaller than the number of drones m. Let Sperm
is a collection of the sperms represented in Eqs.(12) and (13).
Sperm = [ Sp1 , Sp2 ,...., Spn sperm ] (12)

Spi , k = ( Sp1i , k , Spi2, k ,...., SpiL, k ) , 1 ≤ k ≤ N c (13)

where the Spi is the ith sperm in the spermatheca and the Sp i , k is the kth codeword of
the codebook represented by the Spi .

Stage 3 (Breeding Process)

In the breeding process, the queen randomly selects a sperm Sp j from the sper-
matheca if the corresponding random parameter R j is smaller than the predefined
parameter Pc . The new brood is generated from the queen Q and the sperm Sp j by
using the Eq. (14).
Honey Bee Mating Optimization Vector Quantization Scheme in Image Compression 191

brood j = Q ± β × ( Sp j − Q ) (14)
where β is randomly number ranges from 0 to 1.

Stage 4. Brood mutation with the royal jelly by works

The population of broods is improved by applying the mutation operators as follows:
Step 1. For all broods, the random number Ri of i-th brood is generated.
Step 2. If the Ri is less than the predefined mutation ratio Pm , the one tenth codeword
of ith codebook (brood) are randomly select to do mutauion according to the Eq. (15).

Brood ik = Brood ik ± (δ + ε ) × Brood ik

(15)
δ ∈ [0,1], 0 < ε < 1

The δ is random number ranges from 0 to 1; k is one of the index of selected code-
words among the selected and the ε is pre-defined parameter. .
Step 3. The best brood, brood best with maximum fitness v alue is selected as the
candidate queen.
Step 4. If the fitness of brood best is superior to the queen, we replace queen by
brood best .

Stage 5. Check the termination criterion

If the termination criterion is satisfied then finish the algorithm, else discard the all
previous trial solutions (brood set). Then go to stage 2 until the assigned iteration
completed.

4 Experimental Results and Discussion

We implement the proposed HBMO-LBG algorithm in language of Matlab under a
personal computer with 2.4GHz CPU, 1G RAM with window XP system. In the
implementation of HBMO-LBG algorithm, it needs many parameters that are prede-
fined. Five 512 × 512 still images named “LENA”, “PEPPER”, “BABOON”,
“GOLDHILL”, and “Lake” are used for conducting our experiments. These original
test images are shown in Fig. 2. The block size is 4 × 4 , resulting in a total of 16384
vectors to be encoded in an image.
The resulting codebooks are generated on the training set by LBG, PSO-LBG, and
HBMO-LBG methods respectively. The size of the designed codebook are 1024, 512,
256, 128, 64 and 32. The bit rate is defined in Eq. (16).
log 2 N c
bit _ rate = (16)
K
The N c represents the size of designed codebook and the K is the pixel number of
block. Furthermore, the quality of encoded image was evaluated using the peak sig-
nal-to-noise ratio (PSNR). The PSNR is defined as
192 M.-H. Horng

(a). LENA

(b). PEPPER (c ). BABOON

(d). GOLDHILL (e). LAKE

Fig. 2. The test images: (a) LENA, (b) PEPPER, (c) BIRD, (d) CAMERA and (e)GOLDHILL

A2
PSNR = 10 × log10 ( ) (dB) (17)
MSE
The A is the maximum of gray level and MSE is the mean square between the origi-
nal image and the decompressed image.

1 M M
MSE = ∑ ∑ ( yij − yij )
2
(18)
N b × N c i =1 j = 1

where M × M is the image size, yij and yij denote the pixel value at the location (i, j)
of original and reconstructed images, respectively. The experimental results are
shown in Table 1. Tabe 1 show the PSNR values and the execution times of test
images by using the three different vector quantization methods. Obviously, the us-
ages of the HBMO-LAB algorithm have highest PSNR value compared with other
two methods. Furthermore, the PSNR of LBG algorithm is the worst and the other
three algorithms can significantly improve the results of LBG algorithm.
 Honey Bee Mating Optimization Vector Quantization Scheme in Image Compression 193

Table 1. The PSNR and execution times for the five test images with different bit rate using the
four different vector quantization that are LBG, PSO-LGB, QPSO-LBG and HBMO-LBG
algotithms

Image Bit_rate
(512×512) (bit/pixel) LBG-LBG PSO-LBG HBM-LBG
LENA 0.3125 25.430 3 25.760 478 25.887 298
0.375 25.657 4 26.567 932 26.830 568
0.4375 25.740 6 27.508 1828 27.562 1136
0.5 25.750 9 28.169 3666 28.337 2525
0.5625 25.758 17 28.994 7406 29.198 4386
0.625 25.760 36 29.863 14542 29.958 8734

PEPPER 0.3125 25.600 3 25.786 501 26.143 483

0.375 25.891 4 27.101 928 27.206 951
0.4375 25.976 6 27.791 1838 28.014 1848
0.5 25.998 9 28.861 4223 28.951 3680
0.5625 26.006 17 29.661 7271 29.829 6340
0.625 26.008 32 30.584 15218 30.869 12475

BABOON 0.3125 19.832 3 20.423 478 20.439 483

0.375 19.902 4 20.822 934 20.980 939
0.4375 19.921 6 21.373 1844 21.512 1855
0.5 19.926 9 22.038 3649 22.052 3590
0.5625 19.928 17 22.568 7368 22.639 6376
0.625 19.929 32 23.246 14690 23.276 12660

GOLDHILL 0.3125 26.299 3 26.126 478 26.423 484

0.375 26.473 4 27.146 926 27.237 934
0.4375 26.557 6 27.898 1826 28.062 1842
0.5 26.593 10 28.619 3651 28.745 3686
0.5625 26.596 19 29.434 7289 29.522 6367
0.625 26.599 31 30.275 14582 30.378 12656

LAKE 0.3125 23.467 3 23.945 483 24.225 482

0.375 23.571 4 24.714 926 25.116 935
0.4375 23.610 6 25.716 1833 25.971 1843
0.5 23.624 9 26.413 4203 26.509 3678
0.5625 23.626 17 27.154 7271 27.217 7366
0.625 23.627 37 28.022 14521 28.113 14900

5 Conclusion
This paper gives a detailed description of how the HBMO (honey bee mating optimi-
zation) algorithm can be used to implement the vector quantization and enhance the
performance of LBG method. All of our experimental results showed that the pro-
posed algorithm can significantly increase the quality of reconstructive images with
respect to other three methods included the traditional LBG, PSO-LBG and QPSO-
LBG. The proposed HBMO-LBG algorithm can provide a better codebook with small
distortion.
194 M.-H. Horng

Acknowledgment
The author would like to thank the National Science council, ROC, under Grant No.
NSC 97-2221-E-251-001 for supports of this work.

References
1. Linde, Y., Buzo, A., Gray, R.M.: An algorithm for vector quantizer design. IEEE Transac-
tion on Communications 28(1), 84–95 (1980)
2. Chen, Q., Yang, J., Gou, J.: Image Compression Method Using Improved PSO Vector
Quantization. In: Wang, L., Chen, K., S. Ong, Y. (eds.) ICNC 2005. LNCS, vol. 3612, pp.
490–495. Springer, Heidelberg (2005)
3. Wang, Y., Feng, X.Y., Huang, Y.X., Pu, D.B., Zhou, W.G., Liang, Y.C., Zhou, C.G.: A
novel quantum swarm evolutionary algorithm and its applications. Neurocomputing 70,
633–640 (2007)
4. Zhao, Y., Fang, Z., Wang, K., Pang, H.: Multilevel minimum cross entropy threshold selec-
tion based on quantum particle swarm optimization. In: Proceeding of Eight ACIS Interna-
tional Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/
Distributed Computing, pp. 65–69 (2007)
5. Fathian, M., Amiri, B., Maroosi, A.: Application of honey-bee mating optimization algo-
rithm on clustering. Applied Mathematics and Computations, 502–1513 (2007)
6. Horng, M.H.: Multilevel Minimum Cross Entropy Threshold selection based on Honey Bee
Mating Optimization. In: 3rd WSEAS International conference on Circuits, systems, signal
and Telecommunications, CISST 2009, Ningbo, pp. 25–30 (2009)
7. Lloyd, S.P.: Least square quantization in PCM’s. Bell Telephone Laboratories Paper,
Murray Hill, NJ (1957)
8. Abbasss, H.A.: Marriage in honey-bee optimization (HBO): a haplometrosis polygynous
swarming approach. In: The Congress on Evolutionary Computation (CEC 2001), pp. 207–214
(2001)
Towards a Population Dynamics Theory for Evolutionary
Computing: Learning from Biological Population
Dynamics in Nature

Zhanshan (Sam) Ma

IBEST (Initiative for Bioinformatics and Evolutionary Studies) & Departments of

Computer Science and Biological Sciences, University of Idaho, Moscow, ID, USA
ma@vandals.uidaho.com

Abstract. In evolutionary computing (EC), population size is one of the critical

parameters that a researcher has to deal with. Hence, it was no surprise that the
pioneers of EC, such as De Jong (1975) and Holland (1975), had already studied
the population sizing from the very beginning of EC. What is perhaps surprising
is that more than three decades later, we still largely depend on the experience or
ad-hoc trial-and-error approach to set the population size. For example, in a
recent monograph, Eiben and Smith (2003) indicated: "In almost all EC appli-
cations, the population size is constant and does not change during the evolu-
tionary search." Despite enormous research on this issue in recent years, we still
lack a well accepted theory for population sizing. In this paper, I propose to de-
velop a population dynamics theory for EC with the inspiration from the popu-
lation dynamics theory of biological populations in nature. Essentially, the EC
population is considered as a dynamic system over time (generations) and space
(search space or fitness landscape), similar to the spatial and temporal dynamics
of biological populations in nature. With this conceptual mapping, I propose to
'transplant' the biological population dynamics theory to EC via three steps: (i)
experimentally test the feasibility—whether or not emulating natural population
dynamics improves the EC performance; (ii) comparatively study the underlying
mechanisms—why there are improvements, primarily via statistical modeling
analysis; (iii) conduct theoretical analysis with theoretical models such as per-
colation theory and extended evolutionary game theory that are generally ap-
plicable to both EC and natural populations. This article is a summary of a series
of studies we have performed to achieve the general goal [27][30]–[32]. In the
following, I start with an extremely brief introduction on the theory and models
of natural population dynamics (Sections 1 & 2). In Sections 4 to 6, I briefly
discuss three categories of population dynamics models: deterministic modeling
with Logistic chaos map as an example, stochastic modeling with spatial distri-
bution patterns as an example, as well as survival analysis and extended evolu-
tionary game theory (EEGT) modeling. Sample experiment results with Genetic
algorithms (GA) are presented to demonstrate the applications of these models.
The proposed EC population dynamics approach also makes survival selection
largely unnecessary or much simplified since the individuals are naturally se-
lected (controlled) by the mathematical models for EC population dynamics.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 195–205, 2009.
© Springer-Verlag Berlin Heidelberg 2009
196 Z. (Sam) Ma

Keywords: Population Dynamics, Evolutionary Computing, Genetic Algo-

rithms, Logistic Chaos Map, Spatial Distributions, Insect Populations, Survival
Analysis, Survival Selection, Fitness Landscape, Power Law.

1 Introduction
In much of today's evolutionary computing, the setting of population is still experi-
ence-based, and often the fixed population size is preset manually before the start of
computing. This practice is simple, but not robust. The dilemma is that small popula-
tions may be ineffective but big populations can be costly in term of computing time
and memory space. In particular, in Genetic programming, a big population is often a
culprit for the too early emergence of code bloat, which may cause failure or even crash
the system. Furthermore, many of the problems we face are NP-hard problems; effi-
cient population sizing may have significant impacts on the success of the heuristic
algorithms. Nevertheless, an optimal population size effective in exploring fitness
space and efficient in utilizing computing resources is theoretically very intriguing.
Achieving the balance in effectiveness and efficiency is such a dilemma that prompted
Harik et al's (1999) approach to the problem with the Gambler's ruin random walk
model.
If there is an optimal population size, it is very likely to be a moving target, which
may depend on selection pressure, position in search space, and fitness landscape.
Indeed, the population in nature or the biological population is a moving target—a
self-regulated dynamic system influenced by both intrinsic and environment stochas-
ticity. A natural question can be: what if we emulate the natural population dynamics in
EC? A follow-up question is how to emulate? The answer for the latter is actually
straightforward because we can simply adopt the mathematical models developed for
the natural population dynamics. The discipline that study natural population dynamics
is population ecology, which is a major branch of theoretical ecology (or mathematical
ecology). Theoretical ecology is often compared with theoretical physics, and there is a
huge amount of literature on population dynamics, on which computer scientists may
draw for developing an EC population dynamics theory. In this paper, I can only pre-
sent a bird's-eye view of the potential population dynamics models. Furthermore I limit
the models to the categories that we have experimentally tested, and the test results
demonstrated significant improvement with the dynamic populations that are con-
trolled by the mathematical models.
It should be noted that there are quite some existing studies (e.g.,[8][9][11]) that
were conducted to develop alternatives for the fixed-size populations. Due to page
space limitation, here I totally skip the review of existing studies on dynamic popula-
tions in EC (a review was presented in [23]). To the best of my knowledge, this is the
first work that introduces, in a comprehensive manner, the natural population dynamics
theory to EC. The remainder of this paper is organized as follows: Section 2 is an ex-
tremely brief overview of the natural population dynamics theory. Sections 4, 5, and 6
introduce three major categories of the population dynamics models: deterministic,
stochastic, and extended evolutionary game theory models. In Sections 4 and 5, I also
show some sample experiment results with the models introduced. Section 3 briefly
introduces the experiment problem used in Sections 4 and 5. Section 7 is a summary.
 Towards a Population Dynamics Theory for Evolutionary Computing 197

Overall, this article is a summary of a series of studies with the objective to develop an
EC population dynamic theory by emulating natural populations and 'transplanting'
mathematical models developed for natural population dynamics in theoretical ecology
to EC [27][30]–[32].

2 A Brief Overview on Population Dynamics Theory

2.1 Population Dynamics Modeling

Population dynamics or the spatial-temporal changes of population numbers is the

central topic of population ecology. The mathematical modeling of population dy-
namics can be traced back to Thomas Malthus's (1798) "An Essay on the Principle of
Populations" (cited in [17]). Malthus's model for human population growth in which he
argued that human population grow geometrically while resources such as food grows
arithmetically, and therefore fast population growth would lead to much human misery.
Malthus's conclusion had significant influence on Darwin's formation of struggle for
life or natural selection. Pierre Verhulst (1845) derived the famous Logistic population
growth model, which was re-discovered by Pearl and Read in 1920 (cited in [17]).
Logistic model is the solution of a first order nonlinear differential equation [Equ. (1)],
whose difference counterpart [Equ. (3), (5)] is found to be able to demonstrate Chaos
behavior in the 1970s [35].
Several categories of models exist for population dynamics modeling. Logistic
differential equation and its difference equation counterpart form the foundation for
studying single species population dynamics deterministically. It has extraordinary rich
mathematical properties from stable growth, oscillation, attractors to chaos, and its
applications have been expanded well beyond biology. Logistic differential equation is
also the basis for forming the famous Lotka-Volterra differential equations system for
interspecific interactions such as predation and competition. Differential equation or
deterministic modeling is traditionally used to model population dynamics, and is still
the dominant approach ([16][17][35]–[37]). An alternative to the traditional differen-
tial equation modeling is the stochastic modeling with stochastic process models such
as birth-death process, e.g., [18]. It is often more difficult to get analytic solutions in
stochastic modeling. The third group of analytic models is the matrix model which can
be formed in either deterministic or stochastic form (e.g., [4]), and it turns out that
matrix is often equivalent to either the differential equation or stochastic process model
(e.g., Markov Chains). The fourth type of analytical model is the optimization and
game theory models, including evolutionary game theory that was an integration of
three cross-disciplinary fields: Darwin's evolutionary theory, game theory and popula-
tion dynamics ([12][29][39]). The previous four categories are primarily analytical
models, although some of them may need simulation to get the solutions.
Besides analytical modeling approaches, statistical modeling (e.g., time-series,
spatial statistics) and simulation modeling are widely used in modeling population
dynamics. Both approaches are extremely important for studying natural population
dynamics (e.g., [20][22]); they seem less suitable for controlling population sizes in
EC. However, both the approaches can be used to analyze the 'behavior' of EC popu-
lations, such as modeling fitness landscape. There is an additional broad category of
198 Z. (Sam) Ma

modeling approaches, mostly from complexity science, such as percolation theory,

cellular automata, interacting particle systems (IPS), random graphs theory,
agent-based computing, etc (e.g., [2][3]). Among the last category, some models may
offer immensely helpful insights if analytical solutions are available (e.g., critical
percolation probability and phase transitions with percolation models). However, if
analytical solutions are not available, then 'embedding' a simulation model in the EC
population is probably too costly to be useful. Of course, these models can still be
applied to analyze the behavior of EC populations, and they can be more powerful than
statistical modeling because these models usually focus on underlying mechanisms.

2.2 Population Regulation

Population regulation was one of the most contested theories in the history of ecology
(e.g., see [16] for a brief history) and the debates started in the 1950s and culminated in
the 1960s, and even these days, the antagonistic arguments from both schools occa-
sionally appear in ecological publications (e.g., [1][40]). The debate sounds simple
from an engineering perspective. The core of the debate lies in the fundamental ques-
tion: is population regulated by feedback mechanisms such as density-dependent ef-
fects of natural enemies or is simply limited by its environment constraints. Within the
regulation school, there are diverse theories on what factors (intrinsic such as compe-
tition, natural enemies, gene, behavior, movement, migration, etc) and how the popu-
lation is regulated. Certainly, there are mixed hypothesis of two schools. About a
dozen hypotheses have been advanced since the 1950s. The debates were "condemned"
by some critics as "bankrupt paradigm", "a monumental obstacle to progress" (cited in
[1[40]). However, there is no doubt that the debates kept population ecology as the
central field for more than three decades, and is critical for shaping population ecology
as the most quantitatively studied field in ecology. The theoretical ecology is often
dominated by the contents of population ecology [17]. In addition, the important ad-
vances in ecology such as Chaos theory, spatially explicit modeling, agent or individual
based modeling all initiated in population ecology.
The importance of population regulation cannot be overemphasized, since it reveals
the mechanisms for population dynamics. Even more important is to treat population
dynamics from the time-space paradigm, not just the temporal changes of population
numbers. In addition, the concept of metapopulation is also crucial, which implies that
local population extinction and recolonization happen in nature [10]. Obviously,
population regulation as control mechanisms for population size is very inspiring for
the counterpart problem in evolutionary computation. What is exciting is that popula-
tion regulation and population dynamics can also be unified with evolutionary game
theory, and even united with population genetics under same mathematical modeling
framework such as Logistic model and Lotka-Volterra systems (e.g., [12][29][39]).

3 The Test Problem

Assume a problem is represented as a string of bits of size L=32. This string is broken
up into blocks of length B. The problem is to find blocks in which all bits are set to one.
The fitness function is the number of blocks satisfying this requirement. If B does not
 Towards a Population Dynamics Theory for Evolutionary Computing 199

divide L evenly, the leftmost block will be of different size. The following are four
examples of bit streams and their corresponding fitness values (F). The first and third
have optimum solutions, and the blocks that counts in the fitness function are under-
lined.
L = 32, B = 4, F=8: 1111 1111 1111 1111 1111 1111 1111 1111
L = 32, B = 4, F=2: 0000 1111 0110 0100 0111 1010 1111 1010
L = 32, B = 7, F=5: 1111 1111111 1111111 1111111 1111111
L = 32, B = 7, F=2: 1111 0000110 0001000 1111111 0011010
Three metric are used to evaluate the population-sizing schemes or different population
dynamics models. The number of times the optimum solution is found in an experiment
run (consisting of a fixed number of trials) is termed hits per experiment. The other two
metrics are PopuEvalIndex, FirstHitPopuEvlaIndex, and The small PopuEvalIndex
and FirstHitPopuEvlaIndex indicate that fewer evaluations of fitness function and
other associated computations are needed. In addition, it also implies that less memory
space is needed for storing the individual information. Of course, the large number of
hits also indicates the scheme is more robust and efficient.
For more detailed definitions of the metrics and the problem, one may refer to [27].
In the next two sections, all the sample experiment results, which are excerpted from
[30][31] to illustrate the corresponding models, are obtained with this test problem and
are based on the three metrics discussed above.

4 Deterministic Modeling

4.1 Logistic Differential and Difference Equations

The well-known Logistic model has the differential equation form:

dN (t ) ⎛ N (t ) ⎞
= rN (t )⎜1 − ⎟ (1)
dt ⎝ K ⎠

Equation (1) has a unique solution:

K
N (t ) = . (2)
1 + ae −rt
This equation (2) is attributed to Verhust (1838) and Pearl (1925) and is referred to as
the Logistic law in the literature (Cited in [17]).
The difference equation form of Equation (1) is:
⎛ N (t ) ⎞
N (t + 1) = N (t ) + rN (t )⎜1 + ⎟ (3)
⎝ K ⎠

N is the population number at time t or t+1, K is referred to as environment capacity in

biological literature. r is the per capita growth rate per generation.

4.2 One-Parameter Logistic Difference Equation—Logistic Chaos Map

To facilitate analysis with Equation (2), let

200 Z. (Sam) Ma

Nr
x= (4)
K (1 + r )

substituting x into equation (3) derives the one-parameter dimensionless Logistic Map,
also known as one-hump nonlinear function, or one-dimensional quadratic map.
xn +1 = ax n (1 − xn ) (5)

where a = r + 1. To avoid trivial dynamic behavior, the model requires 1<a <4 and
0<x<1. The population size (x) is converted to the (0, 1) interval, and the conversion
also eliminates the other parameter K, which makes the analysis more convenient. The
extremely rich and complex behavior represented by the deceptively simple equation
(5) was discovered by Robert May in 1976 [35]. [38] contains detailed discussion of the
Logistic Chaos map model. The one-parameter Logistic map model is particularly
convenient for controlling EC population dynamics, and offers extremely rich dy-
namics. Figures 1 and 2 show the performance of dynamic populations with the Lo-
gistic chaos map model vs. that of the fixed-size population.

7 0 3 e + 5
N u m b e r o f H it s
3 e + 5 P o p u E v a lIn d e x
Number of Hitting to the Optimum Solutions

3 e + 5 F ir s t H it In d e x
6 0
2 e + 5
2 e + 5
5 0
2 e + 5
2 e + 5
Evaluations

4 0 2 e + 5
1 e + 5
3 0 1 e + 5
1 e + 5

2 0 8 e + 4
6 e + 4
4 e + 4
1 0
2 e + 4
0 e + 0
0 F ix e d a = 1 .5 a = 2 .5 a = 3 .5 7 a = 3 .6 8 a = 3 .8 3 a = 4
F ix e d a = 1 .5 a = 2 .5 a = 3 .5 7 a = 3 .6 8 a = 3 .8 3 a = 4

Fig. 1. Total number of Hits for each Parame- Fig. 2. PopuEvalIndex and First Hit
ter value of a and for fixed population PopuEvalIndex for BlockSize=8

Figures 1 and 2 demonstrate that dynamic populations controlled with the Logistic
chaos map model under various parameter values all significantly outperform the
fixed-size population (the leftmost) except for a=4. Parameter a=4 is theoretically
impossible with the model; it was included to test the boundary case. The detailed
experiment results with the Logistic chaos map model are documented in [30].

5 Stochastic Modeling

There are many stochastic models for population dynamics (e.g., [18]). In this section, I
introduce a category of probability distribution models that are used to describe the
spatial distribution patterns of animal populations. Natural populations are dynamic in
both time and space. In nature, the population spatial distribution patterns are consid-
ered as the emergent expression (property) of individual behavior at the population
level and are fine-tuned or optimized by natural selection [33][34]. It is assumed that
EC populations emulating the spatial distribution patterns of natural populations should
be advantageous.
 Towards a Population Dynamics Theory for Evolutionary Computing 201

Generally, there are three types of spatial distribution patterns: aggregated, random
and regular. Random distribution (sensu ecology) can be fitted with Poisson prob-
ability distribution (sensu mathematics), and the regular distribution (also termed
uniform distribution, sensu ecology) is totally regular with even spacing among indi-
viduals. Aggregated (also termed contagious, congregated, or clustered) distribution
(sensu ecology) represents nonrandom and uneven density in space. The probability
distributions (sensu mathematics) for aggregated distributions (sensu ecology) are
strongly skewed with very long right tails. The most widely used discrete probability
distribution (sensu mathematics) for the aggregated distribution (sensu ecology) pattern
is the Negative Binomial Distribution (NBN).
The aggregated distribution or fat-tail distribution is a property of power law. Taylor
(1961) discovered that the Power law model fits population spatial distribution data
ubiquitously well,
V = aM b , (6)
where M and V are population mean and variance, respectively, and a and b are pa-
rameters. According to Taylor (1961), b>1 corresponds to aggregated distribution, b=1
to random distribution and b<1 to regular distribution. Ma (1988, 1991) [33][34] ex-
tended Taylor’s Power law model with his concept of population aggregation critical
density (PACD), which was derived, based on the Power Law:
m0 = exp[ln(a ) /(1 − b)] , (7)
where a and b are the parameters of Power Law and m0 is the PACD. According to
Ma’s (1988, 1991) reinterpreted Power Law, population spatial distributions are
population density-dependent and form a continuum on the population density series.
The PACD is the transition threshold (population density) between the aggregated,
random and regular distributions [33][34].

7 0 6e+6
N u m b e r o f H its to P o p u E v a lIn d e x
O p t im u m S o lu t io n 5e+6
F ir s tH itP o p u E v a lI n d e x
6 0
5e+6
Fitness Evaluation Index
Hits to the Optimum Solution

4e+6
5 0
4e+6

4 0 3e+6

3e+6
3 0 2e+6

2e+6
2 0
1e+6

1e+6
1 0
5e+5

0 0e+0
F ix e d W e ib u ll N B N P o s s io n N o rm a l U n if o r m F ix e d W e ib u ll NBN P o is s o n N o rm a l U n ifo rm

Fig. 3. Number of Hits for Each Distribution Fig. 4. Fitness Evaluation Indexes

Figures 3 and 4 demonstrate that aggregated distributions [including negative bi-

nomial distribution (NBN) and Weibull distributions] perform best, and the fixed-size
population performs worst. Even the random population (Poisson distribution) per-
forms better than the fixed-size population. Of course, mathematically, uniform dis-
tribution is the same as the fixed-size population, and both perform similarly poor. This
result confirms that emulating the spatial distributions of natural populations improves
EC performance because aggregated distributions dominate natural populations, while
uniform (and even random) distribution is extremely rare in nature.
202 Z. (Sam) Ma

6 Survival Analysis and the Extended Evolutionary Game Theory

Compared with evolutionary game theory (EGT), the extended evolutionary game
theory (EEGT) modeling by Ma (2009) [29] possesses two new features: (i) the ex-
tension with survival analysis to describe lifetime or fitness of the game players; (ii) the
extension with agreement algorithms to model time and space dependent frailty and
deception. To some extent, EEGT itself is very similar or even equivalent to an EC
approach when implemented as a simulation environment. It should be noted that the
term survival is overloaded: one is sensu statistics in survival analysis. Survival
analysis models (the first extension) can be introduced to control the survival of EC
populations. In other words, the first extension alone is sufficient to model EC popu-
lations, similar to the modeling of natural insect populations with survival analysis
[22][24]. This is because by controlling individual survival probability or lifetime, it is
equivalent to controlling population dynamics. Actually, with survival analysis models,
survival selection in EC is not necessary, because individuals are naturally selected by
their survival probability (survivor function).
Survivor function is defined as:
S (t ) = P (T ≥ t ), 0<t<∞ (8)

which is the probability that an individual will survive beyond time T. Various distri-
bution models (such as Weibull, lognormal, and logistic distributions) can be used as
survivor function. These distributions are termed parametric models. In addition,
semi-parametric models, which describe the conditional survivor function influenced
by environment covariates (z), can be utilized, and the most well-known seems to be
Cox's proportional hazards model (PHM):

S (t | z ) = [ S 0 (t )]exp( zβ ) (9)

⎡ t ⎤
where S0 (t ) = exp ⎢ −
⎣
∫ λ (u)du⎥⎦.
0
0 (10)

and z is the vector of covariates, which can be any factors that influence the baseline
survivor function S 0 (t ) or the lifetime of individuals. Dedicated volumes have been
written to extend the Cox model. Therefore, models (8)–(10) offer extremely rich
modeling flexibility and power. Furthermore, more complex multivariate survival
analysis ([13][28]) and competing risks analysis ([5]) can be introduced to model EC
populations.
Besides offering extremely rich and powerful stochastic models, the three 'sister'
fields of survival analysis can be utilized to model either individual lifetime (individual
level) or population survival distribution (population level). One may conjecture that
survival-analysis based EC populations should perform similar to the previous sto-
chastic populations since in both cases, probability distributions models are used to
control EC populations. My experiment results indeed confirm this conjecture [32].
However, since survival analysis is advanced to study time-to-event random variables
(also known as survival or failure time) and the lifetime of individuals in EC is a typical
time-to-event random variable, it does has some unique advantages such as capturing
the survival process mechanistically, and uniquely dealing with censoring (incomplete
 Towards a Population Dynamics Theory for Evolutionary Computing 203

information) and frailty (uncertainty combined with vulnerability). An additional ad-

vantage in the case of EC, as argued previously, is that the survival selection in EC is
not necessary anymore because survival selection is essentially equivalent to determine
how long an individual can 'live' in the EC population. Therefore, survival analysis
modeling offers a unified solution for both population-sizing and survival selection.

7 Summary—Towards an EC Population Dynamics Theory

The previous sections and the more detailed results ([27][30]–[32]) show that it is
advantageous to emulate natural population dynamics. This seems to suggestion that it
is feasible to develop a population dynamics theory for EC, by learning from the natural
populations dynamics. In my opinion, EC has been primarily influenced by the genetic
aspects of the evolutionary theory, and much of the ecological aspects of the evolu-
tionary theory seem still missing. Ecology, which perhaps had equally, if not more,
influenced, Darwin's development of his evolutionary theory, should be the next fron-
tier to draw inspiration for EC [23]. It may be unnecessary to coin a new term such as
ecological computing (which actually already exists in literature, but is largely asso-
ciated with topics not related to this paper) because ecological aspects and genetic
aspects of evolution are hardly separable. Nevertheless, treating the EC population as a
dynamic system rather than a fixed-size entity is fully justified; then, the natural
population dynamics theory should be a critical inspiration source for building an EC
population dynamics theory.
I envision three steps to develop an EC population dynamics theory based on the
natural population dynamics theory. The first step is to emulate natural populations and
experimentally test the emulation. The second step is to explore the underlying
mechanisms why the emulation may improve EC, such as the studies of fitness dy-
namics (landscape) or scaling law (power law). Figures 5 and 6 are two examples of
such studies, which illustrate fitness dynamics with Chaotic populations and Taylor's
power law with stochastic populations, respectively. Figure 5 demonstrates that with
Chaotic populations, multiple clusters of individuals simultaneously search for the
optimum solution and all reach the fitness range of 0.5–0.6. In contrast, with the
fixed-size population, there is only one cluster and the fitness only reaches 0.18–0.20
range (not shown here, refer to [30]). Figure 6 demonstrates that Taylor's power law
describes the fitness distribution in stochastic EC populations very well. Note that both
the axes in Fig. 6 are in the log-scale, so the power function is displayed as linear [31].
The third step should be to perform rigorous theoretical analysis, in the context of EC,
of the population dynamics models introduced from natural population dynamics. In
this step, some of the population dynamics models, such as Percolation theory models
and extended evolutionary game theory, can be 'recursively' applied to the theoretical
analysis because these models are generally applicable to many phenomena including
both natural and EC populations [23]. For example, the application of Percolation
theory model for analyzing EC populations should be very similar to Harik et al. (1999)
study based on the Gambler's ruin problem [11].
The potential inspiration of ecology to EC is certainly beyond the population dy-
namics theory. The relationship between ecology and evolution is best described with
G. E. Hutchinson's (1966) treatise title–"The Ecological Theater and the Evolutionary
Play" [14]. Anyway, without the ecological theater, natural selection has no place to
204 Z. (Sam) Ma

act and evolution cannot happen. Perhaps from this broad perspective, a term such as
ecological computing (computation) may be justified when the ecological principles
play more critical roles in EC.

Power Law of Fitness Distribution (NBN)

3.0
0.7

0.6 1.4

Log (V) of Fitness

Mean Fitness (

0.5
-0.2
0.4

0.3
-1.8
0.2
M)

1
0.1 0 -3.4
0.0 -10
2000 -20
18 00 -30 -5.0
1600
/M

1400 -4.0 -2.0 0.0 2.0

M*

Po p 1200 -40
u la ti 1000
on S 800
600 -50 Log (M) of Fitness
iz e

Fig. 5. The relationship between mean fitness, Fig. 6. Taylor's Power law applied to fitness
population size and the fitness aggregation variance vs. mean fitness
index

References
1. Berryman, A.A.: Population regulation, emergent properties, and a requiem for den-
sity-dependence. Oikos 99(3), 600–606 (2002)
2. Bollobás, B.: Random Graphs, 2nd edn., 500 p. Cambridge University Press, Cambridge (2001)
3. Bollobás, B., Riordan, O.: Percolation, 334 p. Cambridge University Press, Cambridge
(2006)
4. Caswell, H.: Matrix Population Models, 2nd edn. Sinauer, Sunderland (2001)
5. Crowder, M.J.: Classical Competing Risks Analysis, 200 p. Chapman & Hall, Boca Raton
(2001)
6. DeJong, K.A.: An analysis of the behaviors of genetic adaptive systems. Ph.D. Thesis,
University of Michigan, Ann Arbor, MI (1975)
7. Eiben, A.E., Smith, J.E.: Introduction to Evolutionary Computing. Springer, Heidelberg
(2003)
8. Goldberg, D.E., Rundnick, M.: Genetic algorithms and variance of fitness. Complex Sys-
tems 5(3), 265–278 (1991)
9. Goldberg, D.E., et al.: Genetic algorithms, Noise, and the Sizing of Populations. Complex
Systems 6, 333–362 (1992)
10. Hanski, I.: Metapopulation Dynamics. Oxford University Press, Oxford (1999)
11. Harik, G., Cantŭ-Paz, E., Goldberg, D.E., Miller, B.L.: The Gambler’s ruin problem, genetic
algorithms, and the sizing of populations. Evol. Comput. 7(3), 231–253 (1999)
12. Hofbauer, J., Sigmund, K.: Evolutionary Games and Population Dynamics, 323 p.
Cambridge University Press, Cambridge (1998)
13. Hougaard, P.: Analysis of Multivariate Survival Data, 560 p. Springer, Heidelberg (2000)
14. Hutchinson, G.E.: The Ecological Theater and the Evolutionary Play. Yale University Press
(1966)
15. Ibrahim, J.G., Chen, M.H., Sinha, D.: Bayesian Survival Analysis, 481 p. Springer,
Heidelberg (2005)
16. Kingsland, S.E.: Modeling Nature, 2nd edn. University of Chicago Press (1995)
17. Kot, M.: Elements of Mathematical Ecology, 453 p. Cambridge University Press,
Cambridge (2001)
 Towards a Population Dynamics Theory for Evolutionary Computing 205

18. Lande, R., Engen, S.: Stochastic Population Dynamics in Ecology. Oxford University Press,
Oxford (2003)
19. Lawless, J.F.: Statistical models and methods for lifetime data, 2nd edn. Wiley, Chichester
(2003)
20. Legendre, P., Legendre, L.: Numerical Ecology, 851 p. Elsevier, Amsterdam (1998)
21. Ma, Z.S.: New Approaches to Reliability and Survivability with Survival Analysis, Dy-
namic Hybrid Fault Models, and Evolutionary Game Theory. PhD Dissertation, University
of Idaho, 177 p. (2008)
22. Ma, Z.S., Bechinski, E.J.: Survival-analysis-based Simulation Model for Russian Wheat
Aphid Population Dynamics. Ecological Modeling 216(2), 323–332 (2008)
23. Ma, Z.S.: Why Should Populations be Dynamic in Evolutionary Computation?
Eco-Inspirations from Natural Population Dynamics and Evolutionary Game Theory. In:
The Sixth International Conferences on Ecological Informatics. ICEI-6 (2008)
24. Ma, Z.S., Bechinski, E.J.: Accelerated Failure Time Modeling of the Development and
Survival of Russian Wheat Aphid. Population Ecology 51(4), 543–548 (2009)
25. Ma, Z.S., Krings, A.W.: Survival Analysis Approach to Reliability Analysis and Prognostics
and Health Management (PHM). In: Proc. 29th IEEE–AIAA AeroSpace Conference, 20 p.
(2008a)
26. Ma, Z.S., Krings, A.W.: Dynamic Hybrid Fault Models and their Applications to Wireless
Sensor Networks (WSNs). In: The 11-th ACM/IEEE MSWiM 2008, Vancouver, Canada, 9 p.
(2008)
27. Ma, Z.S., Krings, A.W.: Dynamic Populations in Genetic Algorithms. In: SIGAPP 23rd
Annual ACM Symposium on Applied Computing (ACM SAC 2008), Brazil, March 16-20,
5 p. (2008)
28. Ma, Z.S., Krings, A.W.: Multivariate Survival Analysis (I): Shared Frailty Approaches to
Reliability and Dependence Modeling. In: Proc. 29th IEEE–AIAA AeroSpace Conference,
21 p. (2008)
29. Ma, Z.S.: Towards an Extended Evolutionary Game Theory with Survival Analysis and
Agreement Algorithms for Modeling Uncertainty, Vulnerability and Deception. In: Deng,
H., Wang, L., Wang, F.L., Lei, J. (eds.) AICI 2009. LNCS (LNAI), vol. 5855, pp. 608–618.
Springer, Heidelberg (2009)
30. Ma, Z.S.: Chaos Population Chaotic Populations in Genetic Algorithms (submitted)
31. Ma, Z.S.: Stochastic Populations, Power Law and Fitness Aggregation in Genetic Algo-
rithms (submitted)
32. Ma, Z.S.: Survival-Analysis-Based Survival Selections in Genetic Algorithms (in preparation)
33. Ma, Z.S.: Revised Taylor’s Power Law and Population Aggregation Critical Density. In:
Proceedings of Annual National Conference of the Ecological Society of China, Nanjing,
China (1988)
34. Ma, Z.S.: Further interpreted Taylor’s Power Law and Population Aggregation Critical
Density. Trans. Ecol. Soc. China 1, 284–288 (1991)
35. May, R.M.: Simple mathematical models with very complicated dynamics. Nature 261,
459–467 (1976)
36. May, R.M., McLean, A.R.: Theoretical Ecology. Oxford University Press, Oxford (2007)
37. Pastor, J.: Mathematical Ecology of Populations and Ecosystems. Wiley-Blackwell (2008)
38. Schuster, H.H.: Deterministic Chaos: an introduction, 2nd edn., 269 p. VCH Publisher (1988)
39. Vincent, T.L., Brown, J.L.: Evolutionary Game Theory, Natural Selection and Darwinian
Dynamics, 382 p. Cambridge University Press, Cambridge (2005)
40. White, T.C.R.: Opposing paradigms: regulation of limitation of populations. Oikos 93,
148–152 (2001)
 Application of Improved Particle Swarm Optimization
Algorithm in UCAV Path Planning*

Qianzhi Ma and Xiujuan Lei

College of Computer Science of Shaanxi Normal University

Xi’an, Shaanxi Province, China, 710062
zhihui312@163.com
xjlei168@163.com

Abstract. For the calculation complexity and the convergence in Unmanned

Combat Aerial Vehicle (UCAV) path planning, the path planning method based
on Second-order Oscillating Particle Swarm Optimization (SOPSO) was pro-
posed to improve the properties of solutions, in which the searching ability of
particles was enhanced by controlling the process of oscillating convergence and
asymptotic convergence. A novel method of perceiving threats was applied for
advancing the feasibility of the path. A comparison of the results was made by
WPSO, CFPSO and SOPSO, which showed that the method we proposed in this
paper was effective. SOPSO was much more suitable for solving this kind of
problem.

Keywords: Unmanned Combat Aerial Vehicle (UCAV), Second-order Oscil-

lating Particle Swarm Optimization, path planning.

1 Introduction
The unmanned combat aerial vehicle is an experimental class of unmanned aerial vehi-
cle. It is likely to become the mainstay of the air combat force. In order to improve the
overall survival probability and the operational effectiveness of UCAV, an integrated
system is needed in the UCAV flight mission and pre-flight mission to make resource
coordination and determine the flight path. Flight path planning is one of the most im-
portant parts for the UCAV mission planning. That is to generate a path between an
initial location and the desired destination that has an optimal or near-optimal per-
formance under specific constraint conditions. The flight path planning in a large mis-
sion area is a typical large scale optimization problem. A series of algorithms have been
proposed to solve this complicated multi-constrained optimization problem, such as the
evolutionary computation[1], genetic algorithm[2],ant colony algorithm[3]. However,
those algorithms have their shortcomings, for instance, the results of planning fall into
local minimum easily, or they need to be further optimized to meet the needs of aerial
vehicle performance constraints. Particle swarm optimization algorithm (PSO), as a new
intelligent optimization algorithm has been tried to solve this problem due to the merits
*
This work was partially supported by Innovation Funds of Graduate Programs, Shaanxi Nor-
mal University, China. #2009CXS018.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 206–214, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Application of Improved Particle Swarm Optimization Algorithm 207

of rapid searching and easier realization[4]. But the result does not approach ideal
consequence. In this paper, we propose the method based on second-order oscillating
particle swarm optimization algorithm (SOPSO) to improve the optimization results for
UCAV path planning. The model for the path planning problem was built, then the de-
sign scheme and specific realization were given. Finally the rationality and validity of
the algorithm was analyzed based on the simulation experiments and the results.

2 Particle Swarm Optimization

2.1 Standard Particle Swarm Optimization

The development of particle swarm optimization (PSO)[5] was based on observations

of the foraging behavior of bird flocking. To an optimization problem, each solution
represents the position of a bird, called a ‘particle’ which flies in the problem search
space looking for the optimal position. Each particle in the search space is characterized
by two factors: its position and velocity. Mathematically, the particles in PSO are ma-
nipulated according to the following equations(1),(2):
vi (t + 1) = wvi (t ) + c1r1 ( pi − xi (t )) + c2 r2 ( pg − xi (t )) (1)

xi (t + 1) = vi (t + 1) + xi (t ) (2)

where t indicates an (unit) pseudo-time increment, r1 and r2 represent uniform ran-

dom numbers between 0 and 1, c1 and c 2 are the cognitive and social scaling pa-
rameters which are usually selected such as c1 = c 2 = 2 .
x i (t ) = ( x i1 , x i 2 , L , x id ) represents the ith particle. v i (t ) = (v i1 , v i 2 , L , v id ) represents
the rate of the position change (velocity) for particle i . p i = ( p i1 , p i 2 , L , p id ) repre-
sents the best personal position (the position giving the best personal fitness value) of
the ith particle, and d is the dimension number of particle. The symbol g represents
the index of the global best particle in the population. w is the inertia weight. The PSO
algorithm with evolution equation (1) and (2) is called standard particle swarm opti-
mization (SPSO). In standard PSO algorithm, w = 1 . When w is computed as follows:
w = w max − iter ⋅ ( w max − w min ) / Maxiter (3)

iter and Maxiter indicate the current iteration and the max iteration respectively.
This algorithm is called particle swarm optimization with linear decrease inertia weight
(WPSO)[6].
Another classical improved PSO algorithm called particle swarm optimization with
constriction factor (CFPSO)[7] updates the evolution equation as equations (4) and (2):
vi (t + 1) = χ (vi (t ) + c1r1 ( pi − xi (t )) + c2 r2 ( pg − xi (t ))) (4)

The constraint factor χ helps to ensure the algorithm convergence.

＝ － －(ϕ －4ϕ )
Here χ 2/ 2 ϕ 2 1/2
, ϕ = c 1 + c 2 and ϕ > 4 .
208 Q. Ma and X. Lei

2.2 Second-Order Oscillating PSO

Second-order oscillating particle swarm optimization[8] integrates the oscillational

element into PSO.
Set ϕ1 = c1 r1 , ϕ 2 = c 2 r2 , get the (5),(6) followed based on (1),(2)
vi (t + 1) = wvi (t ) + ϕ1 ( pi − xi (t )) + ϕ2 ( pg − xi (t )) (5)
xi (t + 1) = vi (t + 1) + xi (t ) (6)
In the view of control theory, equation (1) is a parallel connection of two inertia links
the inputs of which are pi and p g respectively. In order to increase the diversity of
particles, the inertia links are replaced by second-order oscillation links. Thus, the new
evolutionary equations of second-order oscillating particle swarm optimization algo-
rithm will be gotten as (7) and (8) based on (1), (2):
vi (t + 1) = wvi (t ) + ϕ1 ( pi − (1 + ξ1 ) xi (t ) − ξ1xi (t − 1))
(7)
+ ϕ2 ( pg − (1 + ξ2 ) xi (t ) + ξ2 xi (t − 1))

x i (t + 1) = v i (t + 1) + x i (t ) (8)
In the second-order oscillating algorithm, the searching process is divided into two
phases, we can set ξ 1 , ξ 2 with different values to control that it will begin global search
or local search in accordance with the laws of convergence. We can
，
take ξ1 < (2 ϕ1 − 1) ϕ1 ξ 2 < (2 ϕ2 − 1) ϕ2 at prophase, for the oscillation convergence
and better global search capability; take ξ1 ≥ (2 ϕ1 − 1) ϕ1 , ξ2 ≥ (2 ϕ2 − 1) ϕ2 at anaphase
for the progressive convergence and better local search capability.

3 The Design and Process of the Algorithm

3.1 Environments and Path Statements

Modeling of the threat sources is the key task in UCAV optimal path planning. The
main threats in the flight environment include local fires, radar detection and so on.
Assume unmanned aerial vehicles fly in the high-altitude without altitude change. It
simplifies path planning to a two-dimensional path planning. Here, we suppose that the
threat source is radar detection, and all the radars in task region are same. The aerial
vehicles’ task region is showed in Figure l. There are some circular areas around the
threat points in target region which represent the threat scope.
The flight task for aerial vehicles is from point A to point B. View the starting point A
as the origin point, and the line between the starting point and the target point as X axes,
establish coordinate system. Divide AB into m sub-sections equally. There are m-1 ver-
tical lines between point A and point B. A path can be formed by connecting the points on
these parallel vertical lines. For example, a collection of points constitutes a route:
path = {A, L1(x1, y1), L2 (x2 , y2 ),L, Lm−1(xm−1, ym−1), B} (9)
where Li ( x i , y i ) denotes the point on the i -th vertical line.
 Application of Improved Particle Swarm Optimization Algorithm 209

Fig. 1. The task region of UCAV

3.2 Code Design for Particles

Since the distance between adjacent vertical lines are equal, the ordinal path points’
longitudinal coordinates just can be used to encode a path. In the SOPSO, we use a
particle to indicate a path, which is encoded like this: y1 y 2 , y 3 ,..., y m −1 . In this way, the
searching space is decreased a half by this encoding method. So it is propitious to speed
up the searching process of algorithm.

3.3 Fitness Function

In this paper, only the horizontal path optimization is considered, so the mission can be
simplified to find a path which has a high probability of not being detected by enemy
radars and shortest flight distance from the start point to the target point. According to
the characteristic of radars, if (xt , y t ) denotes the location of a radar, P as the detec-
tive cost for the air vehicle at (x, y ) is given as follows:
⎧ RA4
⎪ 4 R ≤ RA
R + RA4
⎪⎪
P=⎨ (10)
⎪ 0 R > RA
⎪
⎪⎩
210 Q. Ma and X. Lei

Where R = ( x − x t ) 2 + ( y − y t )2 represents the distance between the aerial vehicle

and the radar. R A is the largest detective radius of radar. Usually, there are many radars
in target region, so the threat cost when the aerial vehicle at the i -th waypoint can be
evaluated by calculating the summation of detective costs for each radar. The threat
cost on the sub-path from i waypoint to the next waypoint would be calculated ap-
proximately as:

Wi ,i +1 = dis(Li ,i +1 ) ∗ ∑ Pj
t
(11)
j =1

t indicates the number of radars. dis (Li ,i +1 ) means the distance between adjacent
waypoints. Because the evaluation of threat cost lies on the waypoints, the blind spots
on planning path lead to the failure of planning easily. The more waypoints, the better
for perceiving threats. However, especially in PSO, we increase in the number of
waypoints which means increasing in particle’s dimension, what would reduce the
accuracy of results and make the algorithm complication. Here, a novel method is
proposed to improve the algorithm performance. That does not add the particle’s di-
mension, but only add the points of threat perception. Select a number of perceptual
points on the line which connect the adjacent points. The threat cost expressed by
equation (11) becomes equation (12):

Wi,i+1 = ∑ dis(Lk ,k +1 ) ∗ ∑ Pj,k

n t
(12)
k =1 j =1

k indicates the perceiving point on sub-path connected by the i waypoint and the i + 1
waypoint. n is the number of perceiving points that we will calculate.
Besides threats detection, we also should consider making the flight distance as short
as possible. The flight distance is showed to be the sum of line distances between points
in the flight line. There are m sections in this path. The distance from point Li ( x i , y i )
in vertical line i to point Li +1(xi +1, yi +1 ) in vertical line i + 1 can be described as:

dis (Li , Li +1 ) = ( AB / m ) 2 + ( yi +1 − y i ) 2 (13)

so we describe the objective function of aerial vehicle in path planning as follows:

m −1
J = ∑ (δWi , i +1 , + (1 − δ )dis (Li , Li +1 )) (14)
i =1

Where δ is coefficient in [0,1].It shows the impact both of avoiding threat and flight
distance , and the bigger δ is, the shorter the flight distance would be, but the more
dangerous the flight would be.

3.4 Implementation Steps of the Algorithm

After the analysis above, the implementation steps of the SOPSO algorithm are in-
troduced as follows:
 Application of Improved Particle Swarm Optimization Algorithm 211

STEP 1. Set the relative parameter of the algorithm, such as acceleration coeffi-
cients c1 , c 2 , the max iteration Maxiter , population scale N ,dimension of particle
d etc.
STEP 2. Initialize the velocity and position of particle: Generate an initial popula-
tion and a velocity randomly.
STEP 3. Compute the personal fitness value, update the personal best value and
global best value.
STEP 4. If current iteration iter < Maxiter / 2 , update the current locations of particles
according to the equation (7), (8), and set ξ1 < (2 ϕ1 − 1) ϕ1 , ξ 2 < (2 ϕ2 − 1) ϕ2 ; other-
wise, Update the current locations of particles according to the equation (7), (8) as well,
this time set ξ1 ≥ (2 ϕ1 − 1) ϕ1 , ξ2 ≥ (2 ϕ2 − 1) ϕ2 .Set iter=iter+1.
STEP 5. Check the ending conditions. If iter > Maxiter , exit it and the current
global best value is just the global optimal solution; otherwise, turn to step 3.
STEP 6. Stop the computation and output the correlative values of the optimal person.

4 Experiments and Results Analysis

The algorithms were programmed in MATLAB 7.0 and ran on a PC with Intel Core 2
CPU, 1.86 GHz. Experiments based on WPSO, CFPSO and SOPSO were done under
the same condition. Set the size of flight region as 160*160, the starting point at (0, 60)
and the goal point at (160, 60). The size of population was 50. The dimension number
of a particle was 9. The maximum iteration number was 500. The other correlated pa-
rameters were set with the best ones obtained by many experiments. w was a random
number in [0,1]. In SOPSO, c1 = c 2 =0.4. Taking into account that the flight distance
and avoiding threats have the same degree of influence to the path planning, let δ =0.5.

Fig. 2. Path planning by WPSO

212 Q. Ma and X. Lei

Figure 2~4 show the best path found by using WPSO,CFPSO and SOPSO running
for 50 times. From the result comparison we can see that the path planed by SOPSO is
more accurate and smooth than the other two results and such a path is not only of
shorter flight distance but also can avoid the threats effectively, while the path found by
WPSO is not feasible.

Fig. 3. Path planning by CFPSO

Fig. 4. Path planning by SOPSO

 Application of Improved Particle Swarm Optimization Algorithm 213

Fig. 5. Path planning by SOPSO with adding particle’s dimension

The path showed in Figure 5 is planed by SOPSO in which the particle’s dimension
is the number of perceiving points in the algorithm before. The iteration number in this
algorithm is also 500. Compared the results in Figure 4 and Figure 5, the path showed in
Figure 4 is better than the other one in Figure 5 obviously.
Not only the flight distance of the former is shorter than the latter one, but also it has
a greater performance in avoiding the threats.
Table 1 gives the data comparison of path planning by different algorithms. Where,
fMin, fMax, fMean represent the min value, the max value and the mean value of the
results separately which are the min value of the object function obtained by running
programs for 50 times. The statistical success rate of UCAV avoiding the radar threats
are given through counting the times of avoiding threats successfully during the 50
times running.

Table 1. Results comparison of the route planed by three kinds of algorithm

Judging from the experimental results, it is obvious that the proposed algorithm
SOPSO has a higher success rate and it can find the feasible and optimal path for the
aerial vehicle, while the CFPSO cannot effectively escape the threat regions even
though the mean objective function value is smaller. From practical considerations, this
proposed method provides a new way for path planning of UCAV in exact application
in the future.
214 Q. Ma and X. Lei

5 Conclusion
In order to plan the path for UCAV, a model was built to solve this problem. SOPSO
algorithm was proposed to plan the path for UCAV in flight missions. Simulation re-
sults showed that the proposed method could find a path with shorter flight distance and
avoiding the threats effectively, which was much more suitable for practical applica-
tion. Our future work will focus on the exact application research in path planning with
complex conditions in this field.

References
1. Zheng, C.W., Li, L., Xu, F.J.: Evolutionary route planner for unmanned air vehicles. IEEE
Transactions on Robotics and Automation, 609–620 (2005)
2. Wang, Y.X., Chen, Z.J.: Genetic algorithms (GA) based flight path planning with constraints.
Journal of Beijing University of Aeronautics and Astronautics, 355–358 (1999)
3. Ye, W., Ma, D.W., Fan, H.D.: Algorithm for low altitude penetration aircraft path planning
with improved ant colony algorithm. Chinese Journal of Aeronautics, 304–309 (2005)
4. Chen, D., Zhou, D.Y., Feng, Q.: Route Planning for Unmanned Aerial Vehicles Based on
Particle Swarm Optimization. Journal of Projectiles, Rockets, Missiles and Guidance,
340–342 (2007)
5. Kennedy, J., Eberhart, R.C.: Particle Swarm Optimization. In: Proceedings of the IEEE In-
ternational Conference on Neural Network. IEEE Service Center, pp. 1942–1948. IEEE
Press, New Jersey (1995)
6. Shi, Y., Eberhart, R.: A Modified Particle Swarm Optimizer. In: Proceedings of the IEEE
International Conference on Evolutionary Computation. IEEE Service Center, pp. 69–73.
IEEE Press, Piscataway (1998)
7. Clerc, M.: The swarm and the queen: towards a deterministic and adaptive particle swarm
optimization. In: Proceedings of the Congress on Evolutionary Computation, pp. 1951–1957.
IEEE Service Center, Piscataway (1999)
8. Hu, J.X., Zeng, J.C.: Two-order Oscillating Particle Swarm Optimization. Journal of System
Simulation 19, 997–999 (2007)

Qianzhi Ma. She was born in March 1985. Her main research fields involve intelligent
optimization and path planning. She is a graduate student in Shaanxi Normal University.
Xiujuan Lei. She was born in May 1975. She is an associate professor of Shaanxi Normal
University. She has been published more than 20 pieces of papers. Her main research
fields involve intelligent optimization, especially particle swarm optimization, etc.
 Modal Analysis for Connecting Rod of Reciprocating
Mud Pump

Zhiwei Tong*, Hao Liu, and Fengxia Zhu

Faculty of Mechanical & Electronic, China University of Geosciences

Wuhan, China, 430070
Tel.:+ 86-27-67883277, Fax: +86-27-67883273
tongzhiwei888@126.com

Abstract. Modal analysis is an effective method to determine vibration mode

shapes and weak parts of the complex mechanical system, its main purpose is
to use optimal dynamics design method of mechanical structure system
instead of the experience analog method. Reciprocating mud pump is the
machine that transport mud or water in the process of drilling, which is an
important component of the drilling equipment. In order to improve the per-
formance of reciprocating pump and decrease the failure of the connecting
rod caused by vibration during running, a modal analysis method is per-
formed. In this paper, a three-dimensional finite-element model of connecting
rod was built to provide analytical frequencies and mode shapes, then the
modal distribution and vibration mode shapes for connecting rod were ob-
tained by computing. The results showed the weakness of the connecting rod,
which would provide the reference to dynamics analysis and structural opti-
mization for connecting rod in the future.

Keywords: Connecting rod, Finite element, Modal analysis.

1 Introduction

Modal analysis is an effective method to determine vibration mode shapes and weak
parts of the complex mechanical system. Connecting rod is an important component
of the reciprocating mud pump dynamic system, it is not only a transmission compo-
nent but also a moving parts, at the same time it must withstand variable load such as
tensility, compress force and bending in the working process, so its use reliability has
a great significance for the normal operation of reciprocating mud pump[1]. There-
fore, dynamic characteristics study on the connecting rod has become an important
part of design. In this paper, a modal analysis was applied to connecting rod of recip-
rocating pump by ANSYS software, the main purpose of analysis is to identify the
model parameters of connecting rod such as frequency, vibration mode shapes and
provide a basis for structural dynamics analysis and the follow-up optimal design of
connecting rod.
*
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 215–220, 2009.
© Springer-Verlag Berlin Heidelberg 2009
216 Z. Tong, H. Liu, and F. Zhu

2 Modal Analysis Theory

Modal analysis has been used to study the inherent frequencies and vibration mode
shapes of mechanical systems, which only relate with the stiffness speciality and mass
distribution of structure, and it has nothing to do with external factors [2]. Modal
analysis is the starting point of all the dynamic analysis, inherent frequency and vibra-
tion mode analysis is also necessary if you want to process the harmonic response or
transient analysis. In addition, the structure is supposed to linear in modal analysis
which is in line with the actual working conditions of reciprocating pump.
The main purpose of modal analysis is to use optimal dynamics design method of
mechanical structure system instead of the experience analog and passive method. By
using the modal analysis, every mode shapes of structure can be identified and the
weak parts of structural system’s dynamic characteristics can be found, then the effec-
tive measures were taken to improve the structure and optimize the system.
The movement differential equation of the N freedom degrees linear system can be
express as

[ M ]{ X&&} + [C ]{ X& } + [ K ]{ X } = {F (t )} (1)

In (1), [M] is the mass matrix, [C] is the damping matrix, [K] is the stiffness matrix,
{ X } is the displacement response column vector of node, { X } = { x1 , x2 ,......, xn } ,
T

{F (t )} is the load column vector [3].

The damping and external load were not considered when studying the inherent
frequencies and vibration mode shapes in the process of modal analysis, so the vibra-
tion differential equations could be further simplified to get the undamped free vibra-
tion differential equations of system(as shown in (2)). ‘

[ M ]{ X&&} + [ K ]{ X } = 0 (2)

Any free vibration can be regarded as a simple harmonic vibration, a hypothesis was
put forward as shown in the following.

{ X (t )} = {φ} cos ωt (3)

Put (3) into (2), some equations were shown in the following.

([ K ] − ω [ M ]{φ}) = 0
2

(4)
[ K ]{φ} = λ [ M ]{φ}
In these equations, ω is the Circular frequency; {φ } is the Characteristic column vec-
tor; λ ( λ = ω ) is the eigenvalue.
2

In the process of free vibration, not each point amplitude is zero, so (4) can be be-
come into (5).
[ K ] − ω 2 [M ] = 0 (5)
 Modal Analysis for Connecting Rod of Reciprocating Mud Pump 217

Equation (5) is the undamped vibration system characteristic equation. Modal analysis
can also be considered to solve the eigenvalues ω 2 in (4). In the mechanical engi-
neering, mass matrix and stiffness matrix can be seen as a symmetric positive definite
matrix, and the eigenvalue number obtained through equation computing are equal to
the order N of matrix, that is, there are N natural frequency ( ωi (i = 1, 2,..., n) ). For
each natural frequency, a column vector consisting of N nodes amplitude can be de-
termined by the (4), namely, the vibration mode of structure ( {φi } ).
ANSYS software provides some modal extraction methods such as subspace, slock
method, curtail method, dynamic power, non-symmetric, damping method, QR
damped method and so on. In this paper, the subspace iteration method was chosen
which had the advantages of high accuracy and arithmetic stability compared with
other methods [5].

3 Constructing the Finite Element Modal for Connection Rod

3.1 Constructing the Three-Dimensional Model for Connection Rod

For the complex mechanical systems, CAD software is commonly used to build
model in order to enhance the speed of modeling, and then the model is imported into
CAE software to analyze through the appropriate interface file. The CAD model may
contain a number of design details which often are not based on the strength of the
structure, it would cause model complication if the model retain these details in the
process of modeling, at the same time, it is bound to increase the number of units,
even cover up the principal contradiction of problem and cause negative impact for
the results of the analysis [6]. Therefore, the local structure can be simplified that has
less impact on the results.
Reciprocating mud pump commonly uses the crankshaft to facilitate assembly and
adjusts the clearance of bushing, connecting rod must be produced into the separate
structure. In order to simplify the analysis, the connecting rod and rod cap can be seen
as a whole when modeling with Pro/E. A three-dimensional solid model of the recipro-
cating pump connecting rod was built by Pro / E software platform, then it was saved
with IGES format and was imported into finite element analysis software-ANSYS.

3.2 Make the Structure Discrete

The materials of connecting rod are ZG35CrMo, its material properties are shown in
table 1. The geometric model was divided into grid with 8-node Solid45 cells, those
dividing precision is three levels and it was divided by the ANSYS software auto-
matically, the finite element model of connecting rod is shown in Fig. 1.

Table 1. Material properties of connecting rod

EX (Mpa) PRXY DENS (kg/m3)

210000 0.3 7800

218 Z. Tong, H. Liu, and F. Zhu

Fig 1. Finite element model of connecting rod

3.3 Boundary Constraints

The assumptions of boundary conditions is very important in modal analysis, it is not

only affects the accuracy of the calculation but also decide whether to complete the
calculation [7]. The movement of connecting rod is a planar motion which can be
regarded as synthetic motion with translational and rotation, both ends of connecting
rod are endure the restrictions of crosshead pin and crank pin. In accordance with the
actual work of reciprocating pump connecting rod, the cylindrical coordinate system
on the holes axis of both ends is defined, the points on the hole ends have freedom of
rotating along the axis. Modal are decided by the inherent characteristic of system,
which has nothing to do with the external load, therefore, the load boundary condition
do not need to be set.

4 Finite Elements Solve of Connecting Rod and the Analysis of

Result
Vibration of structure can be expressed as the linear combination of each mode, the
low-level inherent modes have a great impact on the vibration of structure and play a
decisive role for dynamic characteristics of the structure. A characteristic of modal for
the connecting rod by three-dimensional finite element model is established above,
and the 20 order inherent frequency of vibration modes (as shown in table 2) are cal-
culated out. The six vibration modes are extracted to analyze. Fig. 2 shows the six
order vibration mode shapes.
From the calculation results, connecting rod is prone to cause too large dynamic
stress produce the fatigue cracks in its working process as a result of its dense mode.
In particular, the first step and second step modal frequencies are relatively low, and it
causes the dynamic response easily. It may cause the coupled vibration because of the
small frequency difference of adjacent modal [8].
As shown in the Fig. 2, connecting rod has many various vibration forms which fo-
cus on the bending vibration. The first and second step vibration mode shapes are
bending vibration along with Z and Y-axis respectively, the middle part of connecting
rod have a larger vibration displacement and it shows as U form, the third step is the
composite of bending and torsional vibration for connecting rod, the fourth and fifth
step are the bending vibration around the Z and Y-axis, connecting rod shows S form
 Modal Analysis for Connecting Rod of Reciprocating Mud Pump 219

Table 2. The Twenty Calculations of Inherent Frequencies for Connecting Rod

N Frequency (Hz) N Frequency (Hz)

1 136.6 11 1175.3
2 186.8 12 1356.3
3 230.5 13 1421.5
4 283.6 14 1518.9
5 435.0 15 1652.4
6 637.2 16 1705.2
7 780.1 17 1870.2
8 865.4 18 1980.8
9 958.2 19 2275.0
10 1073.4 20 2346.2

（a） (b)

（c） (d)

（e） (f)

Fig. 2. Six vibration mode shapes of connecting rod: (a) first vibration mode shape with fre-
quency 136.6 Hz; (b) second vibration mode shape with frequency 186.8 Hz; (c) third vibration
mode shape with frequency 230.5 Hz; (d) fourth vibration mode shape with frequency
283.6Hz;(e) fifth vibration mode shape with frequency 435.0Hz; (f) sixth vibration mode shape
with frequency 637.2 Hz.
220 Z. Tong, H. Liu, and F. Zhu

in the process of vibration, the sixth mode shape of vibration is not obvious, it is
mainly bending along with the X and Y axis. Bending vibration of the connecting rod
will make piston relation to sleeve of cylinder and the neck part of crank relation to
bearings deflect, which can bring about additional stress and cause cracks and damage,
so this situation should be focused on considering in the design to avoid damage [9].

5 Conclusion
Dynamic characteristic analysis of connecting rod is supplement and development for
static design, it is an important means to do rational design and increase using reliability
for connecting rod structure. The inherent vibration characteristics of reciprocating
pump connected rod were analyzed by ANSYS software, which included the natural
frequencies and mode shapes. Finally, more accurate and intuitive results were obtained.
These results showed that the stress of connecting rod should be commonly concentrate
at connect point of large and small parts, but stress concentration at the central of con-
necting rod is also more obvious, therefore, the traditional design concepts should be
changed and taken fully into account the central of connecting rod design.

References
1. Yang, B., He, Z., Peng, X., Xing, Z.: FEM Analysis on the Valve of Linear Compressor.
Fluid Machinery 33, 24–27 (2005)
2. Yu, B., Huang, Z.: Finite Element Modal Analysis of Rotor in Centrifugal Pump. Mechani-
cal Engineers, 108–109 (June 2005)
3. Han, S., Hao, Z.: The Finite Element Modal Analysis of 4108Q Diesel Engine Connecting
Rod. Vehicles and Power Technology, 37–40 (April 2002)
4. Dai, W., Fan, W., Cheng, Z.: Modal Analysis of Diesel Engine Connecting Rod Based on
ANSYS. Mechanical Engineering and Automation, 39–41 (April 2007)
5. Zhang, G.X., Zhong, Y., Hong, X., Leng, C.L.: Microelectromechanical systems motion
measurement and modal analysis based on Doppler interferometry. J. Vac. Sci. Technol.,
1251–1255 (May 2009)
6. Chen, W.: Dynamic modeling of multilink flexible rod botic manipulators. Computers and
Structures, 183–195 (February 2001)
7. Tlusty, I.F.: Dynamic structural identification task methods. In: Annals of CIRP, February
1980, pp. 260–262 (1980)
8. Kim, D.: Parametric studies on static and dynamic performance of air foil bearings with dif-
ferent top foil geometries and bump stiffness distributions. Tribology, 354–364 (February
2007)
9. Segalman Daniel, J., Roy Anthony, M., Starr Michael, J.: Modal analysis to accommodate
slap in linear structures. Vibration and Acoustics, 303–317 (March 2006)
 A Time-History Analysis Algorithm of a Non-viscously
Damped System Using Gauss Precise Integration

Hongyu Shen and Zhongdong Duan

School of Civil Engineering, Harbin Institute of Technology,

Harbin, Heilongjiang Province, China, 150090
shenhongyu2002@hit.edu.cn
duanzd@hit.edu.cm

Abstract. Time-history analysis of a multiple-degree-of-freedom system with

non-viscous damping was considered. It was assumed that the non-viscous
damping forces depend on the past history of velocities via convolution
integrals over exponentially decaying kernel functions. By transforming the
equation of motion into a first-order extended state-space representation and
combining the Gauss-Legendre quadrature with the calculation technique of
matrix exponential function in the precise time-integration method, a more ap-
plicable time-history analysis method of this system was proposed. The main
virtues of the proposed method were the avoidance of matrix inversion, the high
efficiency and the selectable accuracy. A numerical example was given to dem-
onstrate the validity and efficiency of the method.

Keywords: non-viscous damping, state-space equation, precise integration

method, Gauss-Legendre quadrature method, time-history analysis.

1 Introduction
The damping characteristics and dynamic responses of a vibrating system is a key
topic in the structural dynamic analysis. The viscous damping model, first advanced
by Lord Rayleigh in 1877 [1], is usually used to simulate energy dissipation in a
structural dynamic response. It is however, well known that the classical viscous
damping model is a mathematical idealization and the true damping model is likely to
be different. Moreover, increasing use of modern composite materials and intelligent
control mechanisms in aerospace and automotive industries demands sophisticated
treatment of dissipative forces for proper analysis and design. As a result, there has
been an increase in interest in the recent years on the investigation for other damping
models which can better reflect the energy dissipation of damping and many re-
searchers also began to study the dynamic analysis method involved [2-7].
At the same time, the time-history analysis method is more and more widely used
due to the rapid construction of high rise buildings and the fast development of the
structural dynamic analysis theory and computational technology. At present, the
widely used methods include linear accelerating method, Wilson-θ method [8], New-
mark-β method [9], etc. Most recently, the precise time-integration method (PIM)

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 221–230, 2009.
© Springer-Verlag Berlin Heidelberg 2009
222 H. Shen and Z. Duan

proposed by Zhong et al. attracted lots of attention [10-12]. It can obtain a nearly
perfectly accurate solution as a numerical method, and therefore gives a very useful
idea in time history analysis. However, an inversion of matrix is needed in solving the
inhomogeneous equations during the structural dynamic analysis by PIM [13-14]. It is
known that both the accuracy and stability get worse in an inverse calculation, so the
application of this method is limited.
The time-history analysis method of a convolution integral non-viscously damped
system is studied in this paper. This model is first advanced by Biot about viscous
materials [15]. It is assumed that the non-viscous damping forces depend on the past
history of velocities via convolution integrals over exponentially decaying kernel
functions. The advantage of this model is its generality compared with traditional
viscous models. It can describe various damping mechanism by selecting different
kernel function types [16] and is more accurate in expressing the time-delay effect.
The dynamic analysis of non-viscous damped system of this paper is based on the
condition that the kernel function is exponential function [17].
Obviously, the conventional numerical integration methods can not be directly
used in this non-viscous damped system due to the change of its mathematic model.
Adhikari and Wagner gave a state-space form of this non-viscous damped system
[16], and also proposed a mode superposition method of the state-space (MSM) and a
direct time-domain integration method (DTIM) to study the system dynamic re-
sponses [18]. According to some studied cases, the mode superposition method has a
good accuracy, but it needs large computation. The condition of the direct time-
domain integration is just inversed. Accordingly, it is significant to propose a method
which has both high accuracy and efficiency of computation. In this paper, the dy-
namic response state equation of the non-viscous damped system in [16] was first
transformed. Then a modified precise integration without the inversion of matrix is set
up with basic principles of the precise integration and the Gauss-Legendre quadrature.
Under the acting of discrete load (such as earthquake wave), the cubic spline interpo-
lation is used to calculate the function values of the Gauss integral point in recursion
expressions. Finally, a numerical simulation is given to demonstrate the efficiency of
the method.

2 Motion Equations of the System and State-Space Formalism

2.1 Motion Equations of the System

The non-viscous damping model in this paper considers the damping force relates
with the velocity time history by a convolution integral between the velocity and a
decaying kernel function in its mathematical formulation. The damping force can be
expressed as
t
Fd (t ) = ∫ G (t − τ )x& (τ )dτ (1)
0

where G (t ) ∈ R N × N is a matrix of kernel functions

 A Time-History Analysis Algorithm of a Non-viscously Damped System 223

n n
G (t ) = ∑ Ck g k ( t ) = ∑ C k µ k e − µk t , t≥0 (2)
k =1 k =1

where Ck is a damping coefficient matrix, g k (t ) is a damping function, in this paper,

＝ ＝
g k (t ) µ k e − µk t , x& (t ) is the velocity vector. In the special case when g k (t ) δ k (t ) ,
where δ (t ) is the Dirac-delta function, Eq. (1) reduces to the case of viscous damping.
The motion equations of an N-degree-of-freedom (DOF) system with such damp-
ing can be expressed as
n
x(t ) + ∑ ∫ Ck µk e k ( ) x& (τ )dτ + Kx(t ) = f (t )
t
− µ t −τ
M&& (3)
0
k =1

where M, K ∈ R N × N is mass and stiffness matrices, f (t ) ∈ R N is the forcing vector,

µk ∈ R + is a relaxation parameter, n is the number of relaxation parameters of differ-
ent damping mechanism. The initial conditions of (3) is

x(0) = x 0 ∈ R N and x& (0) = x& 0 ∈ R N (4)

2.2 Review of State-Space Formalism

Obviously, Eq. (3) describes a non-viscous damped dynamic system. The state-space
method has been used extensively in the modal analysis of the non-proportional vis-
cous damped. We expect to extend the state-space approach to this non-viscous
damped system and simplify the expressions of damping and equation, in order to
find a step-by-step integral algorithm to analyse the system dynamic responses. Re-
cently Wagner and Adhikari have proposed a state-space method for this damped
system [16]. Here we briefly review their main results.

Case A: All Ck Matrices are of Full Rank

We introduce the internal variables

y k ( t ) = ∫ µ k e − µk (t −τ ) x& (τ )dτ ∈ R N ,
t
∀k = 1,K , n (5)
0

Differentiating (5) one obtains the evolution equation

y& k ( t ) + µ k y k ( t ) = µk x& ( t ) (6)

Using additional state-variables v ( t ) = x& ( t ) , Eq. (3) can be represented in the first-
order form as

Bz& ( t ) = Az ( t ) + r ( t ) (7)
224 H. Shen and Z. Duan

where
⎡ ∑ n Ck M −C1 µ1 L −Cn µn ⎤
⎢ k =1 ⎥
⎢ M O O O O ⎥
B = ⎢ −C µ O C1 µ12 O O ⎥∈R
⎥ m× m
(8)
⎢ 1 1
⎢ M O O O O ⎥
⎢ ⎥
⎣⎢ −Cn µ n O O O Cn µn2 ⎦⎥

⎡ −K O O O O ⎤
⎢O M O O O ⎥
⎢ ⎥
A = ⎢ O O −C1 µ1 O O ⎥ ∈ R m× m (9)
⎢ ⎥
⎢O O O O O ⎥
⎢⎣ O O O O −Cn µ n ⎥⎦

⎧f ( t ) ⎫ ⎧ x (t ) ⎫
⎪ ⎪ ⎪ ⎪
⎪⎪ 0 ⎪⎪ ⎪ v (t ) ⎪
⎪ ⎪
r (t ) = ⎨ 0 ⎬ ∈ Rm , z ( t ) = ⎨ y1 ( t ) ⎬ ∈ R m (10)
⎪ M ⎪ ⎪ M ⎪
⎪ ⎪ ⎪ ⎪
⎪⎩ 0 ⎪⎭ ⎩⎪ y n ( t ) ⎪⎭
where A and B are the system matrices in the extended state-space. They are sym-
metric matrices, and the order of the system is m = 2 N + nN ; z ( t ) is the extended
state-vector, and r ( t ) is the force vector in the extended state-space, O is a n-order
null matrix, rank ( Ck ) = N , ∀k = 1,K , n .
It can be seen that, when µ → ∞ , Eq. (5) change into y ( t ) = x& ( t ) , Eq. (3) will
change into a equation of viscous damped system.

Case B: Ck Matrices are of Rank Deficient

When Ck matrices are rank deficient, we should transform Ck into lower order full
rank matrices; thus, lots of memory storage and time of calculation will be saved.
Let us introduce a matrix R k ∈ R N × rk

R Tk Ck R k = d k ∈ R rk × rk (11)

where d k is a diagonal matrix consisting of only the nonzero eigenvalues of Ck , thus,

the columns of matrix R k are the eigenvectors of Ck . Define a set of variables of
reduced dimension y% k ( t ) ∈ R N ×rk using R k as
 A Time-History Analysis Algorithm of a Non-viscously Damped System 225

y k ( t ) = R k y% k ( t ) (12)

Eq. (3) can be represented in a first-order form as

% &% ( t ) = Az
Bz % % ( t ) + r% ( t ) (13)

where
⎡ ∑ n Ck M −C1R1 µ1 L −C n R n µ n ⎤
⎢ k =1
⎥
⎢ M O N ,N O N , r1 L O N , rn ⎥
⎢ ⎥
B% = ⎢ −R1T C1 µ1 O TN , r1 R1T C1 R1 µ12 L O r1 , rn ⎥∈R
m% × m%
(14)
⎢ ⎥
⎢ M M M O M ⎥
⎢ − R nT Cn µn O TN , rn O Tr1 , rn L R n Cn R n µn ⎥⎦
T 2
⎣

⎡ −K ON ,N O N , r1 L O N , rn ⎤
⎢ ⎥
⎢O N , N M O N , r1 L O N , rn ⎥
% = ⎢ OT
A OTN , r1 −R1T C1R1 µ1 L O r1 , rn ⎥ ∈ R m×m
(15)
⎢ N , r1 ⎥
⎢ M M M O M ⎥
⎢ T ⎥
⎢⎣O N , rn O TN , rn O Tr1 , rn L − R Tn Cn R n µ n ⎥⎦

⎧f ( t ) ⎫ ⎧ x (t ) ⎫
⎪ ⎪ ⎪ ⎪
⎪ 0N ⎪ ⎪ v (t ) ⎪
⎪ ⎪ ⎪ ⎪
r% ( t ) = ⎨ 0r1 ⎬ ∈ R m% , z% ( t ) = ⎨ y% 1 ( t ) ⎬ ∈ R m% (16)
⎪ M ⎪ ⎪ M ⎪
⎪ ⎪ ⎪ ⎪
⎪⎩ 0rn ⎪⎭ ⎪⎩ y% n ( t ) ⎪⎭

n
It can be seen that the order of the system is m% = 2 N + ∑ rk , and each symbolic sig-
k =1

nificance is same as case A. when All Ck Matrices are of Full Rank, each R k matrix
can be chosen as the identity matrix and (13) is same to (7), so we can consider full
rank as a special case.

2.3 Introductions of Analysis Method

After obtaining the system matrices of state-space, the eigenvalue Λ and eigenvec-
tor Ψ of system matrices A and B (or A % and B% ) are solved through the solution
⎧Ψ ⎫
method of two order eigenvalue problem and satisfied equation ( AΛ + B ) ⎨ ⎬=0
⎩ ΨΛ ⎭
where Λ is a diagonal matrix; then the common mode superposition method is ap-
plied to solve the system dynamic time-history analysis considering the initial condi-
tion. The time-history analysis result calculated by this method under the ideal state of
226 H. Shen and Z. Duan

complete model is the theoretical accurate solution, but the application of the method
is limited by inefficiency calculation and demanding computing time.
Adhikari presented the direct time-domain integration method in order to solve the
shortcoming of wasting large computing time of the mode superposition method [18].
Through direct integration of (7) or (13) at every time step, the time-domain recursion
expression of the state vector z ( t ) or z% ( t ) is obtained and the explicit recursive
equations of the displacement and velocity are derived. The method can significantly
increased the efficiency of the system time history analysis due to avoiding the large
scale calculation of entire system matrices and state vector, but the direct integration
approximate treatment of state equation induces decreasing calculation accuracy
which will be discussed in the following numerical example.
Considering computational efficiency and calculation accuracy, the precise time-
integration method presented by Mr. Zhong [10] is applied to treat and solve the system
state space equation. The basic idea of the precise time-integration method is to change
the conventional two order system equation into one order expression, and the general
solution of the system equation is subsequently transformed into the exponential matrix
and then the 2N algorithm is adopted to calculate the exponential matrix to complete the
system time history analysis by applying the additional theorem of exponential function.
The errors of the precise time-integration method are only originated from the exponen-
tial matrix expansion besides the arithmetic errors of the matrix multiplication. The
more accuracy analysis indicates that the exponential matrix expansion is very small for
a damped vibration system [11], so the numerical results of the precise time-integration
method can meet the exact solution; moreover, the main body of this method is the add
operation applied for solving the exponential matrix and the method can embody the
technical characteristic of the computes and increase the computational efficiency of the
algorithm. However, there is an inverse matrix in the precise time-integration equation
for the nonhomogeneous dynamic equation of the structure under external force
[13][14] and there are problems such as loss calculation accuracy, stabilization and non-
existence of inverse matrix in the process of matrix inversion. The problems always
exist in the multi-DOF system and are particularly serious for matrices with high dimen-
sions. In this paper, a modified precise time-integration method without the inversion of
matrix is proposed combining with the basic principles of the precise time-integration
and Gauss-Legendre integration.

3 Gauss Precise Integration

Considering the general case when the Ck matrices are of rank deficient, rearrange
the state-space Eq. (13) as
& = HV + F
V (17)

where V = z% (τ ) , H = B% -1 A
% , F = B% -1r% (τ ) . In order to describe the algorithm, the
exponential matrix is defined as

e Hτ ≈ I + Hτ + ( Hτ ) 2 + ( Hτ ) 6 + ... + ( Hτ ) n ! + ...
2 3 n
(18)
 A Time-History Analysis Algorithm of a Non-viscously Damped System 227

Multiplying both sides of (17) by e −Hτ and integrating it from t to tk after organiza-
tion yields

e − Hτ ( V
& - HV ) = e − Hτ F ， ∫ d (e t

tk
− Hτ
V ) = ∫ e − Hτ F (τ ) dτ
t

tk
(19)

then

V (t ) = e Vk + e Ht ∫ e − Hτ F (τ ) dτ
H ( t − tk ) t

tk
(20)

where V and Vk represent the solution vector of t and tk , respectively.

Introducing Gauss-Legendre quadrature
b−a m a+b b−a
f ( x )dx = ∑ tj ) + E ( f )
b
∫a 2 j =0
Lj f (
2
+
2
(21)

where t j is the coordinate of integral point, L j is a coefficient of Gauss-Legendre

quadrature, E ( f ) is the Residual error.
Let τ = t − tk . Substituting (21) into (20) yields
τ
τ (1− t j ) ⎛ τ ⎞
F ⎜ tk + (1 + t j ) ⎟ + O (τ 2 n + 2 )
m
V ( t ) = e Hτ Vk + ∑L e
H
2
j
2 j =0 ⎝ 2 ⎠
(22)
τ ⎛ τ ⎞
= Tτ Vk + ∑ L j Tj F ⎜ tk + (1 + t j ) ⎟ + O (τ 2 n + 2 )
m

2 j =0 ⎝ 2 ⎠
τ
H (1− t j )
where Tτ = e Hτ , Tj = e 2
. Eq. (22) is the step-by-step integral recursion expres-
sion of solution vector Vt from solution vector Vk of tk . It suggests that, there is not
any inverse matrix in the expression. Under the acting of discrete load (such as earth-
quake wave), the cubic spline interpolation is used to calculate the F ( t j ) of the
Gauss integral point.
Now, let’s discuss the precise computation of exponential matrix Tτ , the additional
theorem of the exponential function is given by the identity

Tτ = e Hτ = ( e Hτ )
M M
(23)

＝ ，
where M = 2 N , generally N 20 Because τ should be a small time interval,
∆t = τ M is an extremely small time interval. The exponential matrix Tτ departs
from the unit matrix I into a very small extent. Hence it should be distinguished as

e H∆t ≈ I + H∆t + ( H∆t ) 2 = I + Ta

2
(24)

obviously, for computing the matrix Tτ , Eq. (24) should be factored as

228 H. Shen and Z. Duan

Tτ = [ I + Ta ] = [ I + Ta ]
2 N −1
× [ I + Ta ]
2N 2 N −1
(25)

Such factorization should be iterated N times, and the high-precision value of Tτ will
be obtained, so will the Tj . By substituting them into (22), we can realize an efficient
high-precision time-history analysis of this non-viscously damping system.

4 Numerical Example
Here a two-dimension and three-story frame is used. It has 12 DOF, each node has a
rotational DOF and each story has a lateral DOF. This example only is applied to test
the performance of the proposed method, and the damping model will not be dis-
cussed in detail. Therefore, the non-viscously damping model of system can be built
by one exponential function, and the motion equation of system is described in (3)
(the value of n takes 1) and mechanical model is shown in Fig 1. Otherwise, to test the
calculated accuracy and the method of dealing with the discrete load of the proposed
method, it is necessary to consider the displacement of the second floor under the
acting of sinusoidal wave sin(π t ) and El-Centro wave respectively.
For the numerical calculation we have assume
E = 3 × 107 N / m 2 , mcolumn = 300kg , mbeam = 100kg , mst or y = 2000kg ,
Acolumn = 0.6 × 0.6m 2 , Abeam = 0.2 × 0.4m 2 µ1 = 1 γ 1Tmin , γ 1 = 0.05 , Tmin = 2π ωmax ,
ξ1ω2 − ξ 2ω1 ξ ω −ξ ω
C1 = α M + β K , α = 2ω1ω2 , β = 2 2 22 12 1 , ξ1 = 0.02 , ξ 2 = 0.05 .
ω2 − ω1
2 2
ω2 − ω1

20
MSM
GPIM
15

10
U2(mm)

-5

-10

-15
0 5 10 15 20 25 30
Time(s)

Fig. 1. Twelve-DOF steel frame with non- Fig. 2. Horizontal displacement of U 2 under
viscous damping El-Centro wave

Firstly, three types of methods are used to calculate the displacement of structure
under the acting of sinusoidal wave sin(π t ) , and the time steps are taken 0.02s and
0.5s, respectively. The number of Gauss integral points is taken 2, that is, m=2. Here
the results obtained by the modal superposition method (MSM) [16] are deemed as
 A Time-History Analysis Algorithm of a Non-viscously Damped System 229

the exact solution. If the step takes small value, the results obtained by all the three
methods are similar, as shown in Fig. 3a. If the big step is taken, the results obtained
by the direct time-domain integration method (DTIM) and precise time-integration
method (PIM) generate big errors, while the proposed method (GPIM) can still obtain
the exact results, as shown in Fig. 3b. It is obvious that the order of accuracy of GPIM
depends on the number of Gauss integral point and the size of step τ , it is selectable.
However, no matter whether the function fitting or the numerical calculation,
bothDTIM and PIM methods may generate big integral error when these two methods
are applied to analyse the structural response under the acting of discrete load, since it
is inevitable to calculate the integration and difference of loading function in these
two methods. The proposed method can overcome the disadvantages of the above two
methods because there are no integral and differential items in recursion expression
and the cubic spline interpolation is used to calculate the Gauss integral point of load.
The displacement of the second floor under the acting of EI-Centro wave is calculated
by the proposed method, and the results are shown in Fig. 2.

0.4 0.4
MSM DTIM PIM GPIM MSM DTIM PIM GPIM
0.3 0.3

0.2 0.2

0.1 0.1
U2(mm)
U2(mm)

0 0

-0.1 -0.1

-0.2 -0.2

-0.3 -0.3

-0.4 -0.4
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time(s) Time(s)

(a) h=0.02s (b) h=0.5s

Fig. 3. Horizontal displacement of U 2 under Sine wave

It is obvious that the calculation accuracy and amount of the proposed method de-
pend on the number of Gauss integral point and the size of step τ . Comparing with
conventional precise time-integration methods, the proposed method aggravates the

＝
calculating burden of exponential matrix, while the calculating accuracy is improved
obviously when only m 2 , so the calculation amount is not significantly increased.
On the other hand, the relative big integral step can be implemented, which can
lighten the calculating burden evidently. As a result, the proposed method could im-
prove both the calculating accuracy and the efficiency.

5 Conclusions
The time-history analysis method of a convolution integral non-viscously damped system
is studied in this paper, based on the state-space form of this non-viscous damped system
given by Adhikari and Wagner. By combining basic principles of the precise integration
230 H. Shen and Z. Duan

and the Gauss-Legendre quadrature, a new algorithm to analyze the dynamic response of
this damping system in time domain is proposed. Because there are no inverse matrix
items in the proposed method, the advantage is more significant when the method is
applied to large structures with multi-DOFs in which nonsingular phenomenon of system
matrix occurs easily. A numerical example shows that the proposed method can both
ensure the high calculating accuracy and lighten the calculating burden by increasing the
integral step. For the discrete load, it is more effective and reliable to operate.

References
1. Rayleigh, L.: Theory of Sound, 2nd edn., vol. 2. Dover Publications, New York (1877);
(1945 re-issue)
2. Dong, J., Deng, H., Wang, Z.: Studies on the damping models for structural dynamic time
history analysis. World Information on Earthquake Engineering 16(4), 63–69 (2000)
(in Chinese)
3. Liang, C., Ou, J.: Relationship between Structural Damping and Material Damping. Earth-
quake Engineering and Engineering Vibration 26(1), 49–55 (2006) (in Chinese)
4. Woodhouse, J.: Linear damping models for structural vibration. Journal of Sound and
Vibration 215(3), 547–569 (1998)
5. Maia, N.M.M., Silva, J.M.M., Ribeiro, A.M.R.: On a general model for damping. Journal
of Sound and Vibration 218(5), 749–767 (1998)
6. Li, Q.S., Liu, D.K., Fang, J.Q., Jeary, A.P., Wong, C.K.: Damping in buildings: its neural
network model and AR model. Engineering Structures 22, 1216–1223 (2000)
7. Adhikari, S.: Damping modelling using generalized proportional damping. Journal of
Sound and Vibration 293, 156–170 (2006)
8. Wilson, E.L.: Nonlinear dynamic analysis of complex structures. Earthquake Engineering
and Structural Dynamics 1(3), 241–252 (1973)
9. Newmark, N.M.: A method of computation for structural dynamics. Journal of Engineer-
ing Mechanics 85(3), 249–260 (1959)
10. Zhong, W.: On precise time-integration method for structural dynamics. Journal of Dalian
University of Technology 34(2), 131–136 (1994) (in Chinese)
11. Zhong, W.: On precise integration method. Journal of Computational and Applied Mathe-
matics 163, 59–78 (2004)
12. Zhong, W., Zhu, J., Zhong, X.: A precise time integration algorithm for nonlinear systems.
In: Proc. of WCCM-3, vol. 1, pp. 12–17 (1994)
13. Qiu, C., Lu, H., Cai, Z.: Solving the problems of nonlinear dynamics based on Hamilto-
nian system. Chinese Journal of Computational Mechanics 17(2), 127–132 (2000)
(in Chinese)
14. Lu, H., Yu, H., Qiu, C.: An integral equation of non-linear dynamics and its solution
method. Acta Mechanica Solida Sinica 22(3), 303–308 (2001) (in Chinese)
15. Biot, M.A.: Linear thermodynamics and the mechanics of solids. In: Proc. of the third US
national congress on applied mechanics, pp. 1–18. ASME Press, New York (1958)
16. Wagner, N., Adhikari, S.: Symmetric state-space formulation for a class of non-viscously
damped systems. AIAA J. 41(5), 951–956 (2003)
17. Adhikari, S., Woodhouse, J.: Identification of damping: Part 1, viscous damping. Journal
of Sound and Vibration 243(1), 43–61 (2001)
18. Adhikari, S., Wagner, N.: Direct time-domain integration method for exponentially
damped linear systems. Computers and Structures 82, 2453–2461 (2004)
 Application of LSSVM-PSO to Load Identification in
Frequency Domain

Dike Hu, Wentao Mao, Jinwei Zhao, and Guirong Yan

MOE key lab for strength and vibration, Xi’an Jiaotong University,
Xi’an, China
{hdkjll,maowt.mail,zjw76888,yanguir}@gmail.com

Abstract. It is important to identify loads on aircraft to facilitate the designing

of aircraft and ground environmental experiments. A new approach of load
identification in frequency domain based on least squares support vector ma-
chine (LSSVM) and particle swarm optimization (PSO) was proposed. The
principle of load identification using LSSVM was derived and the correspond-
ing model of relationship between loads and responses was constructed using
LSSVM. To get better performance of identification, PSO was adopted to find
optimal hyper-parameters of LSSVM with the best generalization ability of the
identification model. The results of numerical simulation and random vibration
experiments of simplified aircraft show that the proposed approach can identify
not only singlesource load, but also multisource loads effectively with high pre-
cision, which demonstrate the proposed approach is greatly applicable to engi-
neering project.

Keywords: load identification, LSSVM, PSO, inverse problem.

1 Introduction
There are some loads such as vibration, noise etc. acting on aircraft during the trans-
portation, launch, and flight phase. It is commonly recognized that the acquisition of
these loads facilitate not only the designing of strength and rigidity of aircraft and
reliability of internal instruments, but also ground environmental experiments. In
practice, it is difficult to obtain accurate loads directly by experiments on the ground
or theoretical analysis [1] because of the complexity of aircraft structure and bound-
ary conditions of the structure and the corresponding loading conditions. However,
the responses of aircraft structure such as acceleration, displacement, strain, etc. can
be easily measured by the sensors in applications. Usually, it is using the structural
responses that loads are identified. This process is frequently referred to as “load
identification”.
Because many problems are treated in frequency domain in aerospace industry, the
load identification problem in frequency domain is more concerned. The basic idea of
the traditional frequency approach of load identification is that [2] the transfer func-
tion model of system is established and the loads are identified through the inverse of
the transfer function and the responses. However the transfer function matrix is often
ill-posed in resonance region, as a result, the precision and stability of load identifica-
tion are not as good as expected.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 231–240, 2009.
© Springer-Verlag Berlin Heidelberg 2009
232 D. Hu et al.

It is well considered that the relationship between loads and responses only de-
pends on the structure itself and boundary conditions [1], and is inherent once the
structure and the boundary conditions are determined. Thus the load identification
process can be understood as seeking the relationship between loads and responses
when the structure and the boundary conditions are determined.
The support vector machine (SVM) is a new tool to find the complex relationship
between variables in recent years. It can get the best compromise between the com-
plexities of the model and the learning ability according to the limited samples. The
inner product operation in high-dimensional space is transformed into the kernel func-
tion operation in low-dimensional space by the kernel function method, which effec-
tively solves the "dimension disaster" problem. The Least Squares Support Vector
Machine (LSSVM) is a new SVM model proposed by Suykens[3]. The LSSVM uses
least squares linear constraints as a loss function, which simplifies the convex optimi-
zation problem in the traditional SVM to a set of linear equations in LSSVM. It has
many advantages [3]: operation simple, fast convergence, high precision, etc. and is
widely used in estimating nonlinear function, etc.
Taking advantages of the SVM and LSSVM described above, this paper firstly use
LSSVM to search for the relationship between loads and responses of aircraft in fre-
quency domain, and establish a model of load identification. Secondly, because the
performance of LSSVM is determined by the hyper-parameters (the regularization
parameter and kernel parameters are called hyper-parameters) of LSSVM, the particle
swarm optimization (PSO) [5] is used to find optimal parameters based on the best
generalization ability of the load identification model. Finally, the identification of
load of the simulation model and random vibration experiments are conducted with
LSSVM-PSO.

2 Principle of Load Identification Based on LSSVM-PSO in

Frequency Domain

2.1 Identification Model of SVM in Frequency Domain

For linear deterministic (or stochastic) system, the relationship between inputs
F (ω ) (or S FF (ω ) ) and outputs x(ω ) (or S xx (ω ) ) can be described as the following
function:
x(ω ) = H (ω ) F (ω )
(1)
S xx (ω ) = H (ω ) S FF (ω ) H H (ω )

where H (ω ) is the transfer function, described as the following function:

H (ω ) = ( K − ω 2 M + iω C ) −1 (2)

where M , K and C are mass matrix, stiffness matrix and damping matrix.
When the responses x(ω ) (or S xx (ω ) ) are known, according to Eq.(1), the loads can
be obtained as the following form:
 Application of LSSVM-PSO to Load Identification in Frequency Domain 233

F (ω ) = H (ω )−1 x(ω )
(3)
S FF (ω ) = [ H H (ω ) H (ω )]−1 H H (ω ) S xx (ω ) H (ω )[ H H (ω ) H (ω )]−1

According to Eq.(3), the identification of loads need the inverse of transfer function,
but if the condition number of H (ω ) is large, it would lead to numerical instability of
load identification. Because there are noise in the measurement and some nonlinear
factors in the structure, the relationship between loads and responses is often nonlin-
ear, written as the following function:
F (ω ) = f (ω , x(ω ))
(4)
S FF (ω ) = f (ω , S xx (ω ))
The SVM can transform the nonlinear problem in low-dimensional space into a linear
problem in high-dimensional space by a nonlinear mapping function. According to
the theory of SVM [6], if the x(ω ) (or S xx (ω ) ) are the inputs of SVM, F (ω ) (or S FF (ω ) )
are the outputs of SVM, and then the nonlinear function f (⋅) can be viewed as a
regression model of SVM. Application SVM to load identification is seeking the
relationship between loads x(ω ) (or S xx (ω ) ) and responses F (ω ) (or S FF (ω ) ), and estab-
lishing the regression model of f (⋅) .

2.2 Model Selection of LSSVM Based on PSO

LSSVM is a new SVM model, whose advantage has been described above. The re-
gression of LSSVM can be formulated as following: Given a training set { xi , yi }li=1 ,
where xk ∈ R n is the kth input pattern and yk ∈ R is the kth output pattern, the follow-
ing optimization problem can be written as follows:
1 C l
min J ( w , ξ ) = || w 2 || + ∑ (ξi ) 2
2 2 i =1 (5)
s.t : yi = w • φ ( xi ) + b + ξi , i = 1, 2,..., l
T

where φ (•) is a nonlinear function which maps the input space into a higher dimen-
sional space. C is the regularization parameter and ξi is the error function.
In order to solve the optimization problem above, constructing Lagrange can be
written as follows:
l
L( w , b, ξ , α ) = J ( w , ξ ) − ∑ α i {w T φ ( x i ) + b + ξi − yi } (6)
i =1

where α i (α i ∈ R) is the Lagrange multiplier.

Derivation of L( w , b, ξ ,α ) and elimination of w, ξi , the following set of linear equa-
tions can be obtained:
⎡ K + C −1I I ⎤ ⎡α ⎤ ⎡ y ⎤
⎢ T ⎥⎢ ⎥ = ⎢ ⎥ (7)
⎣ I o ⎦ ⎣b ⎦ ⎣ 0 ⎦
234 D. Hu et al.

where I = (1,1,...,1)T , y = ( y1 , y2 ,..., yl )T , α = (α1 , α 2 ,..α l )T K = {Kij = K ( xi , x j )}li , j =1 is the kernel ma-
trix. where K ij ( x, y ) is kernel function, described as the following function: .

K ij ( x, y ) = φ ( xi )T • φ ( yi ) (8)

α and b can be obtained by solving the set of linear Eq.(7). Then the decision func-
tion of LSSVM can be obtained as following:
l
y ( x) = f (⋅) = ∑ α i K ( x, xi ) + b (9)
i =1

According to Eq.(8) and Eq.(9), the final result is not related to the specific mapping
function, but related to the inner product of kernel function. Generally, the kernel
function is chosen firstly, then α and b can be calculated, and the regression predic-
tion can be achieved according to Eq.(9).
Commonly practicable kernel functions including:
2
Gauss kernel: K ( x , x' ) = exp(- x - x' /σ)

Polynomial kernel: K ( x , x' ) = (( x ⋅ x' )+ c)d

Sigmoid kernel: K ( x , x' ) = tanh(κ ( x ⋅ x' ) + ν )

As analyzed above, the hyper-parameters are the key of the achievement of regression
prediction. To evaluate the performance of the hyper-parameters, leave-one-out and
cross validation are commonly used. However both of these methods require high
computational cost when the training set is large. The leave-one-out method has been
improved in literature [7], and a new simple generalization bound is obtained, which
is the unbiased estimate of the true leave-one-out generalization ability. This bound
can be described as following:

{ )}
l
PRESS (θ ) = ∑ ri(
2
−i
(10)
i =1

^ (− i ) αi ⎡ K + C −1I 1⎤
where ri( −i ) = yi − y i = ,Q=⎢ ⎥.
Qii−1 ⎣⎢1
T
0 ⎦⎥
The aim of model selection is to choose the best hyper-parameters by optimizing
the generalization ability of LSSVM. In practice, users usually adopt grid search
method and gradient-based methods. The former requires re-training the model many
times, and the latter commonly falls into the local minimum. As a kind of popular
method, swarm intelligence algorithm has been proved to be an efficient and robust
method for many non-linear, non-differentiable problems. In this paper, particle
swarm optimization (PSO) [5] is applied to select the optimal hyper-parameters based
on the best generalization ability of the load identification model. The procedure of
model selection using PSO is presented as follows:
 Application of LSSVM-PSO to Load Identification in Frequency Domain 235

Algorithm 1. Model Selection using PSO and Generalization Criteria of LSSVM

Initialize: I : the number of iterations, m : swarm size, swarm: S = {x1 , x 2 ,L , x m } ,
present location and fitness value of each particle. The equation (10) is taken as
the fitness function. i = 0
while i < I
for all x1,L,m ∈ S
Calculate velocity and Update its position according to the following equation:
v j ( i ) = w * v j ( i − 1) + c1 * r1 * ( p jbest − x j (i − 1) ) + c2 * r2 * ( gbest − x j (i − 1) )
x j (i) = x j (i − 1) + v j (i )
where w ∈ [ 0,1]
is the adaptive inertia weight;
c1 , c2 ∈ R are the individual/social behavior weights; r1 , r2 ∈ [ 0,1] are uniformly dis-
tributed random numbers; p jbest is the best previous position of the jth particle
and gbest is the best particle in the swarm.
Evaluate the goodness of x j
Update p jbest (if necessary)
end for
Update gbest (if necessary)
i ++
end while

In algorithm 1, xj contains hyper-parameters, and gbest is the smallest PRESS (θ ) . The

optimal hyper-parameters can be obtained through algorithm 1.

2.3 Load Identification Process with LSSVM-PSO

The load identification approach with LSSVM-PSO is shown in Figure 1.

x(ω ) System F (ω )
under test

data LS-SVM
preprocessing model F% (ω )
selection
based on
data PSO
preprocessing

Fig. 1. An approach for load identification using LSSVM-PSO

An approach for load identification using LSSVM-PSO shown in Figure 1 contains

the following four sub-processes. Firstly, preparing for loads and responses data
required for training, i.e. the structure is added with known loads, and the structural
responses and loads are obtained. This step corresponds to the dynamic calibration proc-
ess of the actual system. White noise is often chosen as loads in the actual dynamic
236 D. Hu et al.

calibration process, because all the characteristics of the system can be excited by white
noise. Secondly, training the LSSVM based on PSO. With the responses as inputs of
LSSVM, and loads as outputs of LSSVM, the PSO is used to find optimal parameters of
LSSVM based on the best generalization ability of the relationship between loads and
responses. Then the model of load identification is established. Thirdly, measuring the
responses in the operation process of aircraft. Finally, calculating the loads by putting
the responses obtained in the third phase into the model established in the second phase.

3 Application of LSSVM-PSO to Load Identification of Flight

Simulation Model
Because the identification of flight loads with LSSVM and PSO is a new study, there
are many factors that need to be studied before conducting flight load identification.
In order to verify the feasibility of the proposed method, a simulation model is cho-
sen, which is simplified as following: without considering the internal structure of
aircraft, the aircraft is reduced to a cylindrical structure including a cylindrical shell
and a fixture. The cylindrical structure is activated with only one concentrated force.
The finite element model is shown in Figure 2. The concentrated force is added near
the centre of the fixture. The parameters of the finite element model are described
as following: the external diameter, inner diameter, and height of the cylindrical shell
is 370mm,360mm and 370mm respectively; the diameter and height of the fixture is
380mm and 17.5mm respectively; the density, modulus of elasticity and Poisson ratio
is 7850kg / m3 , 2.1× 1011 Pa and 0.3 respectively for both of them.
An axial white noise force is added in the concentrated force point, which can ex-
cite all the characteristics of the system, meanwhile, the acceleration response of one
of the cylindrical shell structure’s nodes is obtained. The same process repeated 29
times with the same force point and the same response point but the different ampli-
tude spectrum of white noise force. 29 groups of forces and responses are obtained. It
is a white noise force, so the amplitude spectrum in frequency domain is a flat spec-
trum, but the phase is random. Only the size of amplitude spectrum is changing and
the shape of amplitude spectrum is always flat among the 29 groups of white noise
force.
The amplitude spectrums of load and acceleration response in a group are con-
structed as a training sample, where the acceleration response is the input of LSSVM-
PSO, and the load is the output of LSSVM-PSO. The gauss kernel is chosen to be the
kernel function. According to Eq.(4), the relationships between loads and responses
are one to one with frequencies, i.e. the relationships are different in different fre-
quencies. Thus the influence of dates in different frequencies does not need to be
considered, and the learning process can be implemented at the corresponding fre-
quency. There are 2000 frequency points in the simulation data, so the total times of
learning process is 2000. The relationships are established through the learning of 29
training samples with LSSVM-PSO, and the model of load identification is obtained.
Another force is added in the same force point, whose shape is trapezoidal and pa-
rameters are different from the 29 white noise force described above. The amplitude
spectrums of trapezoidal spectrum force and acceleration response are obtained. The
acceleration response is inputted into the model of load identification established
 Application of LSSVM-PSO to Load Identification in Frequency Domain 237

above, and the force can be identified, i.e. the amplitude spectrum of load is calcu-
lated according to Eq.(10) with the parameter obtained from the training process. The
amplitude spectrum of trapezoidal spectrum force is used to verify the identification
results. The load identification results are shown in Figure 4(The blue solid line is
actual measured values and the red dot Line is identified values).

Fig. 2. The finite element model of cylindrical Fig. 3. The identification results
structure

1 n
∑ ⎡ Preal (ω j ) − Ppred (ω j )⎤⎦
2
The mean square error RMSE = and average relative error
n j =1 ⎣
1 n
Er = ∑ Pi , real − Pi , pred / Pi ,real
n i =1
are chosen to judge the identification error. The identifica-
tion error of trapezoidal spectrum force are: RMSE = 0.0027 , Er = 0.0082% .
As shown in Figure 3, the blue solid line almost coincides with the red dot Line,
i.e. the amplitude spectrum of identified load corresponds with the measured load
spectrum. Because there is no noise in simulation data and the only nonlinear factor of
numerical error is very poor, the mean square error and average relative error are both
very small in the identification error. The simulation tests demonstrate that the pro-
posed approach based on LSSVM-PSO is an effective way to identify load in fre-
quency domain, and the identification accuracy is high.

4 Load Identification of Random Vibration Experiments of

Cylindrical Structure
This section uses the proposed approach to identify multisource loads of random
vibration experiments of the cylindrical structure. According to Eq.(4), each load is
only related to responses and is nothing to do with other loads. In this way the multi-
source loads identification problem can be transformed into a number of singlesource
load identification problems. Each load identification model can be established by
means of that all the responses are the inputs of LSSVM-PSO, and one of loads is the
output of LSSVM-PSO. Another load identification model can also be established by
the same inputs and the different output of another load. The process repeats times as
many as the number of loads. In this way the multisource loads identification problem
is solved.
238 D. Hu et al.

The cylindrical structure system of random vibration experiments is composed of a

cylindrical shell, a fixture and a vibration shaker. The fixture is connected with the
cylindrical shell by 18 bolts, and the whole cylindrical structure is fixed on the vibra-
tion shaker through four force sensors. The forces measured by the four force sensors
are the actual loads of the cylindrical structure. The system of random vibration ex-
periments is shown in Figure 5.

(a) (b)
Fig. 5. The system of random vibration experiments: (a) The four force sensors; (b) Scheme of
the whole system

The driver spectrum shown in Figure 6 is added to drive the vibration shaker. 29
groups of data are obtained by changing the driver spectrum parameter values of
a,b,c,d,e,f,g, where a is fixed on 5Hz and g is fixed on 2000Hz. Each group’s data
contains the power spectral densities (PSD) of four acceleration responses whose
positions are located on the cylindrical structure, four forces and drive current.
It brings nonlinear effects during the vibration process because of the connection of
bolts. Moreover, the signals are influenced by the noise of the blower of the vibration
shaker, while these also bring nonlinear effects. The relationship of PSD between
drive current and one of the four acceleration responses in a specific frequency
(100Hz) among 29 groups’ data is shown in Figure 7. As shown in Figure 7, the rela-
tionship between response and drive current is nonlinear, which denotes the relation-
ship between responses and forces is nonlinear. The methods based on linear system
used to identify loads will lead to large identification error in this case.

Fig. 6. The driver spectrum Fig. 7. The relationship between current and
response
 Application of LSSVM-PSO to Load Identification in Frequency Domain 239

The gauss kernel is also chosen to be the kernel function. The four PSDs of four
acceleration responses are used as inputs of LSSVM-PSO and the PSD of each force
is used as output of LSSVM-PSO. Four identification models of load identification
are established, where each model is used to identify one force.
Another drive spectrum whose parameters are different from the 29 drive spectrum
described above is used to drive the vibration shaker. The PSDs of four acceleration
responses at the same points and four forces are obtained. The acceleration responses
are inputted into each model of load identification established above to identify loads
and the forces are the actual loads of cylindrical structure to compare with the identi-
fied forces to verify the validity of the proposed approach and identification accuracy.
The identification process is the same as the process described at section 3. The iden-
tification results are shown in Figure 8. The mean square error and average relative error
as described above and energy relative error Egrms = grmsreal − grms pred / grmsreal are chosen
to judge identification error. The identification errors are shown in Table 1.

(a) (b) (c) (d)

Fig. 8. The identification results of four forces: (a):Force1; (b):Force2; (c):Force3; (d):Force4

Table 1. The identification error of four forces

Force 1 Force 2 Force 3 Force 4

RMSE 0.1436 0.1830 0.2595 0.1515
Er 21.35% 22.47% 18.50% 21.75%
Erms 0.11% 0.11% 0.077% 0.086%

As shown in Figure 8, the blue solid line almost coincides with the red dot line, the
location and size of peak is also consistent and there are no false peaks in the identifi-
cation curves, i.e. the PSD of identified load corresponds with the measured PSD of
load. These demonstrate that the proposed approach based on LSSVM-PSO can iden-
tify the four forces with effective results. The four forces are not exactly same that the
individual peaks don’t appear in some forces, because the cylindrical structure is
deformed during the vibration process. As can be seen from the identification error in
Table 1, because the system is nonlinear seriously and the signals are influenced by
the noise greatly, the mean square error and average relative error are both larger than
the results of the simulation data, but the experimental identification errors are still
small in the bound of engineering permissible error .The identified results are almost
according to the measured values in terms of energy, which are import to experiment.
The experiments demonstrate that the proposed approach based on LSSVM-PSO is an
effective way to identify multisource loads in frequency domain, and the identifica-
tion accuracy is high.
240 D. Hu et al.

5 Conclusions and Final Remarks

The identification of loads on aircraft is a complex problem. In this paper, a new ap-
proach using LSSVM is proposed to seeking the relationship between loads and re-
sponses and establishing the load identification model. As the performance of
LSSVM is determined by the proper hyper-parameters, how to obtain the proper hy-
per-parameters is treated as an optimization problem and the PSO is used to solve this
problem based on the best generalization ability. To demonstrate the effectiveness of
the proposed method, we firstly simplify an aircraft to a cylindrical structure. The
results of numerical simulation show that the proposed approach can identify single-
source load with high precision. Moreover, the multisource loads identification prob-
lem is also transformed into a number of singlesource load identification problems in
frequency domain. The results of random vibration experiments of cylindrical struc-
ture show that the proposed approach can identify multisource loads and the experi-
mental identification errors are satisfying the demands of engineering permissible
error. The approach based on LSSVM-PSO proposed in this paper can identify not
only singlesource load, but also multisource loads in frequency domain, which supply
a new idea for load identification. The approach was discussed in frequency domain,
and can also be used for load identification in time domain. However, the position of
loads is determined previously in the proposed approach, the future study will be
developed to make this approach more effective to identify the loads of unknown
positions.

References
1. Cao, X., Sugiyama, Y., Mitsui, Y.: Application of artificial neural networks to load identifi-
cation. Computers and Structures 69, 63–78 (1998)
2. Xu, Z.-y., Liao, X.-h.: Dynamic State Loading Identification and Its Development. Journal
of Changzhou Institute of Technology 19, 13–18 (2006)
3. Suykens, J.A.K., Vandewalle, J., Moor, D.B.: Optimal Control by Least Square Support
Vector Machines. Neural Networks 14, 23–25 (2001)
4. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of the IEEE Inter-
national Conference on Neural Networks, Perth, Australia, pp. 1942–1984 (1995)
5. Vapnik, V.: The Nature of Statiscal Learning Theory. Springer, New York (1999)
6. Cawley, G.C.: Leave-one-out cross-validation based model selection criteria for weighted
LS-SVMs. In: Proceedings of the International Joint Conference on Neural Networks,
Vancouver, BC, Canada, pp. 2970–2977 (2006)
 Local Model Networks for the
Optimization of a Tablet Production Process

Benjamin Hartmann1 , Oliver Nelles1 , Aleš Belič2 ,

and Damjana Zupančič–Božič3
1
University Siegen, Department of Mechanical Engineering,
D–57068 Siegen, Germany
{benjamin.hartmann,oliver.nelles}@uni-siegen.de
2
University Ljubljana, Department of Electrical Engineering,
SLO–1000 Ljubljana, Slovenia
ales.belic@fe.uni-lj.si
3
KRKA d.d., SLO–8501 Novo Mesto, Slovenia
damjana.zupancic@krka.biz

Abstract. The calibration of a tablet press machine requires compre-

hensive experiments and is therefore expensive and time-consuming. In
order to optimize the process parameters of a tablet press machine on the
basis of measured data this paper presents a new approach that works
with the application of local model networks. Goal of the model-based
optimization was the improvement of the quality of produced tablets,
i.e. the reduction of capping occurence and the variation of the tablet
mass as well as the variation of the crushing strength. Modeling and op-
timization of the tablet process parameters show that it is possible to
find process settings for the tabletting of non-preprocessed powder such
that a sufficient quality of the tablets can be achieved.

1 Introduction

The pharmeceutical industry is increasingly aware of the advantages of imple-

menting a quality-by-design (QbD) principle, including process analytical tech-
nology (PAT), in drug development and manufacturing [1], [2], [3]. Although
the implementation of QbD into the product development and manufacturing
inevitably requires large resources, both human and financial, large-scale pro-
duction can be established in a more cost-effective manner and with improved
efficiency and product quality.
Tablets are the most common pharmaceutical dosage form prepared by com-
pression of a dry mixture of powders consisting of active ingredient and ex-
cipients into solid compacts. The process of tabletting consists of three stages:
a) The powder mixture is filled into the die; b) compaction, where the pow-
der is compressed inside a die by two punches, resulting in plastic and elastic


This work was supported by the German Research Foundation (DFG), grant NE
656/3–1.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 241–250, 2009.


c Springer-Verlag Berlin Heidelberg 2009
242 B. Hartmann et al.

deformation and/or particles fragmentation, and c) ejection, where the tablet

is ejected from the die and elastic recovery of the tablet may occur. An inten-
sive elastic recovery can lead to separating the upper part of a tablet from the
tablet body (capping). The mechanical behavior of powders during tabletting
and the quality of tablets depend on the powder characteristics (formulation)
and the tabletting parameters on the tablet press machine [4], [5], [6], [7]. In
order to assure a high and repeatable quality of products the processes and for-
mulations should be optimized. This is particularly important on high capacity
rotary tablet presses which can produce a few hundred thousand tablets per
hour and are very sensitive to product and process variables. A special challenge
is the large scale production of products in form of tablets where the active drug
loading is high (> 50% of the total mass of a tablet). Active drugs as organic
molecules very often have inappropriate physical properties which prevent using
direct compression of powder mixtures, which is the most economical produc-
tion method (shortest process time, minimum cleaning and validation need, low
energy consumption, environmentally friendly). The problem of capping is often
observed in production of tablets with high drug loading. This problem can be
solved either by formulation or process optimization. To find the optimal set
of the parameters (main compression force, precompression force, compression
speed ...) on the tablet press, it is possible to make a few tablets for each set
of parameters, and then the operator can decide which settings should work
best, and then starts the production. Fine machine setup optimization must
be performed, due to some batch to batch variations of raw materials or some
other variables. During the setup procedure, many of the tablets do not pass
the quality control and must hence be discarded, which can be very expensive,
regarding the price of the raw material and the number of trials, needed to find
optimal settings. To optimize the quality of tablets and to reduce the number of
faulty tablets, modeling and simulation procedures can be used, where a model
of quality parameters with respect to process parameters must be developed.
The raw materials used in tests should cover all the expected characteristics of
raw materials that can be expected to appear in production. With the model,
optimal settings of the machine can be found with respect to raw material char-
acteristics, without further loss of raw material. The aim of the present study
was the optimization of the tabletting process in order to diminish capping oc-
currence and the variation of tablet mass and crushing strength. Optimization
was performed for the product where 70% of the tablet weight represent active
ingredients, on the high capacity rotary-tablet press, in the standard production
environment. Local model networks were used to model the relation between
quality and process parameters.

2 Production of Model Tablets

Due to the high drug content and chosen manufacturing procedure (direct com-
pression) changes in raw material characteristics can result in variable crushing
strength of tablets, large tablet mass variation and most problematic, intense
 Local Model Networks for the Optimization of a Tablet Production Process 243

capping problems. If problems in direct tabletting are too severe, dry granula-
tion can be used as an alternative to improve the particle compression character-
istics. Dry granulation means aggregation of smaller particles using force. The
compacts are milled and sieved afterwards. Next step is tabletting. Three types
of mixtures for tabletting were prepared:
A mixture for direct tabletting (Type: “Direkt” - 1 sample),
B granulate prepared by slugging (precompression on a rotary-tablet press),
using different parameters of tabletting speed and compression force (Type:
“Briket” - 4 samples),
C granulate prepared by precompressing on a roller compactor using different
parameters of compacting speed and force (Type: “Kompakt” - 4 samples).
The composition of all powder mixtures was the same, milling and sieving of
compacts was performed on the same equipment. After dry granulation, the par-
ticle size distribution is normally shifted toward bigger particles, also capping
occurence is often decreased due to a change in particle characteristics (flowa-
bilty and compression characteristics). Nine samples were tableted on a rotary-
tablet press using different combinations of the following parameters: compres-
son speed v, precompression force pp, main compression force P. Each of these
parameters were set at three levels. Tablets were evaluated according to capping
occurence, crushing strength and tablet mass variability [8].

2.1 Capping Coefficient

The capping occurency (capping coefficient, CC) was calculated as the ratio
between the number of tablets that have a tendency to cap and the number of
tablets that have been observed. It was observed visually on each tablet after a
classical tablet hardness test (Erweka). A Tablet was considered to have capping
tendency if the upper part of the tablet falled off from the tablet body or if a
significant shape (a step-like form) was formed on the breaked surface of the
tablet.

2.2 Experiments on the Rotary-Tablet Press

For experiments on the rotary-tablet press three groups of raw materials were
used (“Direkt”, “Briket”, “Kompakt”). The types Kompakt and Briket were
further divided into four sub-groups depending on the combination of compres-
sion force and speed during slugging or roller compaction. For each of these
raw materials several combinations of settings on the rotary-tablet press were
set (tabletting speed, main compression force, precompression force). Produced
tablets were controlled afterwards: CC, average crushing strength, standard devi-
ation of crushing strength, average tablet mass and standard deviation of tablet
mass. For each raw material and parameter setting, 10 tablets were made and
analyzed. All together 76 data points were measured.
244 B. Hartmann et al.

3 Modeling

To model the relation between process parameters, raw material characteristics,

and quality of the tabletting process (CC, standard deviation of mass, standard
deviation of crushing strength) it was not reasonable to use physical laws that
describe the operation of the tabletting machine, considering the aim. The sys-
tem characteristics depend on many stochastic processes (particle rearrangement
and deformation type, friction, bondforming between particles, ...), therefore, the
model would be too complex for process parameter optimization. On the other
hand, there is a relatively large database available, hence, it is more appropriate
to use data-based computational intelligence methods instead.

3.1 Local Model Networks

To model the system characteristics a local model network approach [9], [10],
[11] was used with two to four inputs, depending on the output. For each output
(CC, standard deviation of crushing strength, standard deviation of mass) a
separate neuro-fuzzy model was constructed.
The output ŷ of a local model network with p inputs u = [u1 u2 · · · up ]T can
be calculated as the interpolation of M local model outputs ŷi , i = 1, . . . , M [12],

M
ŷ = ŷi (u)Φi (u) (1)
i=1

where the Φi (·) are called interpolation or validity or weighting functions. These
validity functions describe the regions where the local models are valid; they
describe the contribution of each local model to the output. From the fuzzy logic
point of view (1) realizes a set of M fuzzy rules where the Φi (·) represent the
rule premises and the ŷi are the associated rule consequents. Because a smooth
transition (no switching) between the local models is desired here, the validity
functions are smooth functions between 0 and 1. For a reasonable interpretation
of local model networks it is furthermore necessary that the validity functions
form a partition of unity:
M
Φi (u) = 1 . (2)
i=1

Thus, everywhere in the input space the contributions of all local models sum up
to 100%. In principle, the local models can be chosen of arbitrary type. If their pa-
rameters shall be estimated from data, however, it is extremely beneficial to choose
a linearly parameterized model class. The most common choice are polynomials.
Polynomials of degree 0 (constants) yield a neuro-fuzzy system with singletons
or a normalized radial basis function network. Polynomials of degree 1 (linear)
yield local linear model structures, which is by far the most popular choice. As
the degree of the polynomials increases, the number of local models required for
a certain accuracy decreases. Thus, by increasing the local models’ complexity,
 Local Model Networks for the Optimization of a Tablet Production Process 245

at some point a polynomial of high degree with just one local model (M = 1) is
obtained, which is in fact equivalent with a global polynomial model (Φ1 (·) = 1).
Besides the possibilities of transferring parts of mature linear theory to the
nonlinear world, local linear models seem to represent a good trade-off between
the required number of local models and the complexity of the local models
themselves. This paper deals only with local models of linear type:
ŷi (u) = wi,0 + wi,1 u1 + wi,2 u2 + . . . + wi,p up . (3)
However, an extension to higher degree polynomials or other linearly parame-
terized model classes is straightforward.

3.2 Data Pre-processing

Each raw material batch has a unique composition of particles, since its pro-
duction cannot be exactly repeated. The raw material characteristics can be
charaterized by the particle size distribution. The model outputs are a function
of the tablet composition, the main compression pressure P, the pre-compression
pressure pp, and the tabletting speed v. The tablet composition and the raw ma-
terial characteristics, respictively, are given by the particle size distribution. The
particle sizes were separated in the following 8 subgroups: 0-0.045, 0.045-0.071,
0.071-0.125, 0.125-0.25, 0.25-0.5, 0.5-0.71, 0.71-1.0, and 1.0-1.25mm. As training
data we had 9 different tablet compositions available. Figure 1 shows the cor-
relation between the standard deviation σ of the particle size dirstribution and
the mean values µ. Due to the sparse training data we assumed that the raw
material characteristic is uniquely desribed with the mean value µ of the particle
size distribution. For simplicity the inputs and outputs had been normalized to
the interval [0, 1].

3.3 Model Structure

Due to available prior knowledge it can be assumed that the three different
model outputs which represent the tablet quality depend on different sets of

Fig. 1. Correlation between standard deviation σ of the particle size dirstribution and
the mean values µ
246 B. Hartmann et al.

Fig. 2. Model structure. One model for each output dimension.

input variables. Therefore it is advantageous to build up three individual MISO

models each with the relevant inputs. Using a single MIMO model would discard
this knowledge and thus would yield suboptimal results, compare Fig. 2. The
task is to model the following three output dimensions: The capping coefficient
CC, the standard deviation of the tablet mass σm and the standard deviation
of the crushing strength σcs , respectively.

4 Results
First, the local model networks were designed and validated, next, the models
were used to find the optimal process parameters. Furthermore we investigated
the robustness of the optimization in terms of varying training data.

4.1 Modeling
The training of the local model networks was performed with the LOLIMOT al-
gorithm proposed in [9]. One advantage of this training algorithm among others
is that the training has only to be carried out once to get reliable and repro-
ducible results. For final training the whole data set was used, since, considering
the rather small dimensionality of the problem, the nonlinear input-output rela-
tionship could be visually monitored in order to check for potential overfitting.
The identified relations can be seen in Figs. 4 to 7. For modeling the standard
deviation of the mass σm 3 local models (LMs) were used. The standard devi-
ation of the crushing strength σcs was modeled with 4 LMs and the capping
coefficient CC with 3 LMs.

4.2 Optimization
Once the models for the tabletting process parameters have been built, the ques-
tion arises: How they should be utilized for optimization? Figure 3 illustrates the
used optimization scheme. Basically, the manipulated variables that are model
inputs are varied by an optimization technique until their optimal values are
 Local Model Networks for the Optimization of a Tablet Production Process 247

Fig. 3. Optimization of the tablet press parameters

found. Optimality is measured in terms of a loss function that depends on the

tabletting process parameters, i.e. the model outputs. When the operating point
dependent optimal input values are found, they can be stored in a static map
that represents a feedforward controller, which can be realized either as a look-up
table (since it has only two inputs) or as a neural network. To convert the multi-
objective optimization problem into a single-objective problem the loss function
was formulated as follows:

J (σm , σcs , CC) = w1 σm + w2 σcs + w3 CC

−→ min with 0 ≤ P ≤ 1, 0 ≤ pp ≤ 1 . (4)
P,pp

This method is called the weighted-sum method, where the coefficients wi of the
linear combination are called weights [13], [14]. All relevant model outputs are
directly incorporated and the loss function is chosen as a weighted sum of the
tablet process parameters. Due to expert knowledge the parameters were set to
w1 = 0.25, w2 = 0.25 and w3 = 0.5. For optimization the BFGS quasi-Newton
algorithm [15] was used. The results are summarized in Table 1. Figure 6 shows
the shapes of the loss functions for the three tablet types “Direkt”, “Kompakt”
and “Briket”. Figure 8 illustrates the variation of the optimal process param-
eters, if single data points are omitted from the training dataset. Because of
the sparse dataset, omitting critical data points can lead to different shapes of
the loss function. Therefore outliers can be observed in the boxplots. However,
the comparison of the median values with the results generated with the whole
dataset (Table 1) is satisfactory.
248 B. Hartmann et al.

Fig. 4. Model plot that shows the relations between the inputs mean value of the
tablet particle sizes µ, main compression force P, pre-compression force pp, the tablet-
ting speed v and the output standard deviation of tablet mass σm . The red dots are
measured data points.

Fig. 5. Model plot that shows the relations between the inputs mean value of the tablet
particle sizes µ, main compression force P , pre-compression force pp and the output
standard deviation of the crushing strength σcs

Fig. 6. Loss functions with respect to the raw material parameter µ = 0, 0.86, 1, main
compression force P , pre-compression force pp and tabletting speed v = 1
 Local Model Networks for the Optimization of a Tablet Production Process 249

Fig. 7. Model plot that shows the relations between the inputs mean value of the tablet
particle sizes µ, main compression force P and the output capping coefficient CC

Fig. 8. Boxplots that illustrate the variation of the optimal process parameter values
in case of omitting single data points

Table 1. Results of the optimization

µ P pp
0 0.21 0.69
0.5 0.04 0.73
0.86 0.01 0.70
1 0.03 0.74

5 Conclusion

This paper presents a local model network approach for the model-based opti-
mization of the parameters of a tablet press machine. Several experiments on
the tablet press machine with different tablet and process parameters made it
possible to create a well generalizing model. Due to the high flexibility and ro-
bustness with respect to the sparse and noisy data, local model networks are well
suited for this kind of modeling problem. Model plots show the applicability of
250 B. Hartmann et al.

the selected approach. Modeling and simulation indicate that it is possible to

find process settings for the tabletting of non-preprocessed powder such that a
sufficient quality of tablets can be achieved.

References
1. Frake, P., Greenhalgh, D., Gierson, S., Hempenstall, J., Rudd, D.: Process control
and end-point determination of a fluid bed granulation by application of near infra-
red spectroscopy. International Journal of Pharmaceutics 1(151), 75–80 (1997)
2. Informa: PAT - Quality by design and process improvement, Amsterdam (2007)
3. Juran, J.: Juran on Quality by Design. The Free Press, New York (1992)
4. Sebhatu, T., Ahlneck, C., Alderborn, G.: The effect of moisture content on the
compression and bond-formation properties of amorphous lactose particles. Inter-
national Journal of Pharmaceutics 1(146), 101–114 (1997)
5. Rios, M.: Developments in powder flow testing. Pharmaceutical Technology 2(30),
38–49 (2006)
6. Sorensen, A., Sonnergaard, J., Hocgaard, L.: Bulk characterization of pharmaceu-
tical powders by low-pressure compression ii: Effect of method settings and particle
size. Pharmaceutical Development and Technology 2(11), 235–241 (2006)
7. Zhang, Y., Law, Y., Chakrabarti, S.: Physical properties and compact analysis of
commonly used direct compression binders. AAPS Pharm. Sci. Tech. 4(4) (2003)
8. Belič, A., Zupančič, D., Škrjanc, I., Vrečer, F., Karba, R.: Artificial neural networks
for optimisation of tablet production
9. Nelles, O.: Axes-oblique partitioning strategies for local model networks. In: In-
ternational Symposium on Intelligent Control (ISIC), Munich, Germany (October
2006)
10. Hartmann, B., Nelles, O.: On the smoothness in local model networks. In: American
Control Conference (ACC), St. Louis, USA (June 2009)
11. Hartmann, B., Nelles, O.: Automatic adjustment of the transition between local
models in a hierarchical structure identification algorithm. In: IFAC Symposium
on System Identification (SYSID), Budapest, Hungary (August 2009)
12. Nelles, O.: Nonlinear System Identification. Springer, Berlin (2001)
13. Chong, E., Zak, S.: An Introduction to Optimization. John Wiley & Sons, New
Jersey (2008)
14. Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press,
Cambridge (2004)
15. Scales, L.: Introduction to Non-Linear Optimization. Computer and Science Series.
Macmillan, London (1985)
 Retracted: Implementation of On/Off Controller for
Automation of Greenhouse Using LabVIEW

R. Alimardani1, P. Javadikia1, A. Tabatabaeefar1, M. Omid1, and M. Fathi2

1
Department of Agricultural Machinery Engineering, Faculty of Biosystems Engineering,
University of Tehran, 31587-77871 Karaj, Iran
2
Department of Electronic and Informatics, University of Siegen, 57074 Siegen, Germany
{Pjavadikia,atabfar,omid,rmardani}@ut.ac.ir
fathi@informatik.uni-siegen.de

Abstract. The present study is concerned with the control and monitoring of
greenhouse air temperature, humidity, Light intensity, CO2 concentration and

ed
irrigation. A computer-based control and monitoring system was designed and
tested and to get this target is used Supervisory Control & Data Acquisition
(SCADA) system. The end product is expected to give the farmer or end user a
kiosk type approach. Entire greenhouse operation is governed and monitored
through this kiosk. This approach is fairly novel considering the unified system
ct
design and the SCADA platform, NI LabView 7.1.

Keywords: Control, Lab View, Automation, on/off, Greenhouse.

ra

1 Introduction
et

Environmental conditions have a significant effect on plant growth. All plants require
certain conditions for their proper growth. Therefore, it is necessary to bring the envi-
ronmental conditions under control in order to have those conditions as close to the
idea as possible. To create an optimal environment, the main climatic and environ-
R

mental parameters need to be controlled. These parameters are nonlinear and extremely
interrelated, rendering the problem of management of a greenhouse rather intractable
to analyze end control through the classical control methods. An automated manage-
ment of a greenhouse brings about the precise control needed to provide the most
proper conditions of plant growth. Greenhouse Automation has considerably evolved
in the past decade with the onus shifting to factors like flexibility, time-to-market and
upgradeability. SCADA systems, hence, form an interesting platform for developing
systems for greenhouse automation. It allows a ‘supervisory level’ of control over the
greenhouse, facilitates data acquisition and creates a developer-friendly environment
for implementing algorithms.
The five most important parameters to consider when creating an ideal greenhouse
are temperature, relative humidity, soil moisture, light intensity and Co2 concentration.
In order to design a successful control system, it is very important to realize that the five
parameters mentioned above are nonlinear and extremely interdependent [2-1]. The
computer control system for the greenhouse includes the following components [5]:

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 251–259, 2009.
© Springer-Verlag Berlin Heidelberg 2009
252 R. Alimardani et al.

1. Acquisition of data through sensors

2. Processing of data, comparing it with desired states and finally deciding what
must be done to change the state of the system
3. Actuation component carrying the necessary actions.

This paper describes a solution to the second part of the system. The information is
obtained from multi-sensor stations and is transmitted through serial port to the
computer. It will then be processed and the orders will be sent to the actuation
network.

2 Material and Method

The original system includes four main subsystems, namely, the Temperature and
Humidity, Irrigation, light intensity and co2 concentration. The system program con-
tains the running algorithm, which governs the whole subsystems. The real time Data

ed
transmit from microcontroller to Lab View through serial port of the computer and the
system to the microcontroller. The designed system has follows.
Real time soil moisture
Real time Co2
ct
Real time lighting
Real time temperature
Real time humidity
ra
The seven actuators are as follows.
Water valve
Ventilation Fans
et

Cooler
Circulate Fans
Sprayer
Heater
R

Lamps and Shade

The system can be set up for any type of plant and vegetation by simply setting cer-
tain parameters on the front panel of Lab View (fig. 1.). The entire system is simu-
lated in Lab View by National Instruments. The main block-diagram of an automated
greenhouse is shown in fig. 2.

2.1 Temperature and Humidity System

The temperature and humidity system consists of different subsystems as follows.

Special attention must be given to these subsystems for the proper functioning of the
overall system.
Implementation of On/Off Controller for Automation of Greenhouse Using LabVIEW 253

ed
Fig. 1. Front panel of the main system
ct
ra
et
R

Fig. 2. Block diagram of the main system

2.1.1 Heating Subsystem

Temperature control is necessary for attaining high crop growth, yield and quality.
Extreme temperatures may induce stress and associated damage to the plasmatic
structures or the photosynthetic apparatus of the plant. Less extreme suboptimal tem-
peratures may delay plant development and affect other plant characteristics such as
254 R. Alimardani et al.

dry matter distribution [6]. Generally, the protection given by the greenhouse is suffi-
cient to allow the development of crops during winter without the use of heating sys-
tems. However, a greenhouse with automated heating facilities presents advantages
like increased production speed, possibility of producing products out of season and
better control of producing products out of season and better control of diseases[5].
Uniform crop growth is very important for most production systems and the heating
and ventilation systems have a major impact on producing uniform crops [7].

2.1.2 Ventilation and Cooling Subsystem

Ventilation systems can be either mechanical or natural (i.e., without the use of fans
to move air through the greenhouse) [7]. whenever the natural ventilation is insuffi-
cient, it is necessary to force it, which is done using ventilators. The cooling effi-
ciency can be increased combining the natural and mechanical ventilation systems
with air humidifiers. A general type of cooling system uses a porous pad installed in
one top side of the greenhouse, which maintained wet. On the opposite side, an ex-
haust fan is installed. The air admitted through the pad becomes cooler by evaporation

ed
effect [5].

2.1.3 Humidity Subsystem

The fog system is formed by suspended pipes on the greenhouse structure, spraying
tiny drops of water into the greenhouse, contributing to increase air humidity and
ct
decrease the air temperature. Dehumidifier system is used for decreasing the air
humidity [5].
ra
et
R

Fig. 3. Front panel and block diagram of the Temperature system

Implementation of On/Off Controller for Automation of Greenhouse Using LabVIEW 255

2.1.4 System Design and Simulation

The simulated system in LAbVIEW is described in Fig. 3 and 4 for temperature and
humidity systems. The TEMPERATURE PALLETE shows the current temperature
and the idea ranges of temperature for desired crop, while the HUMIDITY PALLETE
shows the current humidity and the idea ranges of humidity. If the temperature falls
below a set value, then heating system is turned on. Similarly, if the humidity rises
above a set value, then the dehumidifier is turned on. The method is an on-off control
with a certain dead band which can be arbitrarily defined by the user.

2.2 Lighting System

Using the supplemental lighting system is a common way for greenhouse lighting.
However it can be done either with photoperiod lighting system or through walkway
and security lighting. Greenhouse lighting systems allow us to extend the growing
season by providing plants with an indoor equivalent to sunlight [3]. Front panel and
block diagram of the light system is shown in figure 5.

ed
ct
ra
et
R

Fig. 4. Front panel and block diagram of the Humidity system

2.3 Co2 System

The fifth major factor, namely the Co2 concentration, plays a very important role in the
photosynthesis process. The average Co2 concentration in the atmosphere is approxi-
mately 313 ppm, which is enough for effective photosynthesis. A problem arises when a
greenhouse is kept closed in autumn or/and winter in order to retain the heat, when not
enough air is circulated to have the appropriate Co2 concentration [1, 4]. In order to
improve the growing of herbs inside the greenhouse, it is necessary to increase Co2
concentration in company with favorable conditions of temperature and light.
256 R. Alimardani et al.

ed
ct
Fig. 5. Front panel and block diagram of the Light system
ra

Front panel and block diagram of the Co2 system is shown in figure 6.
et
R

Fig. 6. Front panel and block diagram of the Co2 system

Implementation of On/Off Controller for Automation of Greenhouse Using LabVIEW 257

2.4 Irrigation System

The irrigation system consists of a soil moisture control module. For the desired plant,
the optimum value of soil moisture is set through the front panel. The values of soil
moisture are measured through sensors in real time which is then displayed in the
front panel. Depending on this condition, the flow control valve is handled. The RT
values show the Real-Time values as obtained from the sensors.

3 Results and Discussion

Performance of the system was evaluated by installing it in the greenhouse. A number
of experiments were carried out during the 2008 autumn season. In the next section,
one of the experiments which were conducted from 10 to 12 a.m. on December 7,
2008, in the city of Karaj, is presented and discussed. The system performance was
tested for two modes of operations: (a) uncontrolled mode of operation, and (b) con-
trolled mode. During the first day the system was tested as a data-acquisition system,

ed
i.e. only data from the sensors were recorded and monitored on the screen with no
operation of fans, sprayer or other installed environmental systems in the greenhouse.
For the second day of the experiment, inside air temperature was controlled by the
system.
ct
Fig. 7 shows the results for the uncontrolled experiment. The data were recorded
with no operation of fans and sprayer. These show comparison between the inside and
the outside condition. The results indicate that the outside temperature is always,
about 15oC, less than the temperature inside the greenhouse, because the solar radia-
ra

tion entered the greenhouse through transparent poly carbonate. This is a further
confirmation of greenhouse effect. Also, it was found that the rate of change of tem-
perature in the upper part, i.e. near the cover is higher than that of the height of plants.
This rise of vertical temperature is due to receiving solar incident radiation. During a
et
R

Fig. 7. Measured temperatures and relative humidity in no-control operation

258 R. Alimardani et al.

temperature decrease, it was found that the heat exchange near the cover is occurring
more rapidly. The fluctuations of temperature and to a lesser extend of RH in the
model greenhouse during observation time were affected by natural conditions such
as surface evaporation within the greenhouse, solar radiation and ambient tempera-
ture. For the controlled mode of operation, two separate experiments are carried out:
(i) temperature control only (Fig. 8a), and (ii) both temperature and humidity control
(Fig. 8b). When the controller was put into operation, it was found that
Tout<Tmid<Tup.
This could be partly due to the use of re-circulating fans.

ed
ct
ra
et
R

Fig. 8. Measured temperatures and relative humidity under the action of controller: (a) Tem-
perature control (up), and (b) Temperature and RH control (down)

4 Conclusions
This paper described the design and simulation and Implementation of a fully automated
greenhouse using LabVIEW. For better simulation results, the entire five major interre-
lated environment variables in a greenhouse Temperature-Humidity, soil moisture, light
Implementation of On/Off Controller for Automation of Greenhouse Using LabVIEW 259

intensity and Co2 concentration were considered altogether. The resulting system was
designed in such a way that the system could be easily controlled in such a way that the
system could be easily controlled and monitored by an amateur user who might have no
or little technical background on the subject, the main advantages of this simulation are
more facilities in the entire system. By means of this simulation, the optimal level of
environment and growth factors inside the greenhouse can be achieved. Once the pro-
posed system is designed, standardized, and implemented, it provides and convenient
control over the greenhouse management in order to increase efficiency.

References
1. Putter, E., Gouws, J.: An Automatic Controller for a Greenhouse Using a Supervisory Ex-
pert System. In: Electro technical Conference, Melecon 1996, 8th Mediterranean, pp. 1160–
1163 (1996)
2. Fourati, F., Chtourou, M.: A Grennhousr Control with Feed-Forward and Recurrent Neural

ed
Networks. Simulation Modeling Practice and Theory 15, 1016–1028 (2007)
3. Bowman, G.E., Weaving, G.S.: A Light Modulated Greenhouse Control System. Journal of
Agriculture Engineering Research 15(3), 255–258 (1970)
4. Klaring, H.P., Hauschild, C., Heibner, A., Bar-Yosef, B.: Model-Based Control of Co2
Concentration in Greenhouse at Ambient Levels Increases Cucumber Yield. Agriculture for
Meteorol. 143, 208–216 (2007)
ct
5. Metrolh, J.C., Serodio, C.M., Couto, C.A.: CAN Based Actuation System for Greenhouse
Control. Indusrial Elect., 945–950 (1999)
6. Korner, O., Challa, H.: Design for an Improved Temperature Integration Concept in Green-
ra
house Cultivation. Computer and Electronic in Agriculture 39, 39–59 (2003)
7. Roberts, W.J.: Creating a Master Plan for Greenhouse Operation, Rutgers, The State Uni-
versity of New Jersey (2005)
et
R
Structure Design of the 3-D Braided Composite Based on
a Hybrid Optimization Algorithm

Ke Zhang

School of Mechanical and Automation Engineering,

Shanghai Institute of Technology
120 Caobao Road, Shanghai 200235, China
zkwy2004@126.com

Abstract. Three-dimensional braided composite has the better designable

characteristic. Whereas wide application of hollow-rectangular-section three-
dimensional braided composite in engineering, optimization design of the
three-dimensional braided composite made by 4-step method were intro-
duced. Firstly, the stiffness and damping characteristic analysis of the com-
posite is presented. Then, the mathematical models for structure design of the
three-dimensional braided composite were established. The objective func-
tions are based on the specific damping capacity and stiffness of the compos-
ite. The design variables are the braiding parameters of the composites and
sectional geometrical size of the composite. The optimization problem is
solved by using ant colony optimization (ACO), contenting the determinate
restriction. The results of numeral examples show that the better damping and
stiffness characteristic could be obtained. The method proposed here is useful
for the structure design of the kind of member and its engineering application.

Keyword: Composite, Structure Design, Ant Colony Algorithm.

1 Introduction
At present, textile composites are being widely used in advanced structures in avia-
tion, aerospace, automobile and marine industries [1]. Textile composite technology
by preforming is an application of textile processes to produce structured fabrics,
known as performs. The preform is then impregnated with a selected matrix material
and consolidated into the permanent shape. Three-dimensional (3-D) braiding method
which was invented in 1980s offers a new opportunity in the development of ad-
vanced composite technology. The integrated fibre network provides stiffness and
strength in the thickness direction, thus reducing the potential of interlaminated fail-
ure, which often occurs in conventional laminated composites. Other distinct benefits
of 3-D textile composites include the potential of automated processing from preform
fabrication to matrix infiltration and their near-net-shape forming capability, resulting
in reduced machining, fastening, and scrap rate [2]. The direct formation of the struc-
tural shapes eliminates the need for cutting fibres to form joints, splices, or overlaps
with the associated local strength loss, and simplifies the laborious hand lay-up com-
posite manufacturing process.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 260–269, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Structure Design of the 3-D Braided Composite 261

In comparing with two dimensional composite laminates, a key property of the

new innovative 3-D braided composite is its ability to reinforced composites in the
thickness direction. Braiding with continuous fibres or yarns can place 3-D rein-
forcements in monocoque structural composites, it includes multi-directional fibre
bundle which interconnect layers. Since the braiding procedure dictates the yarn
structure in the preform and the yarn structure dictates the properties of the compos-
ite, designing the braiding procedure to yield the desired structural shape that is
endowed with the desired properties is an important element in textile composite
technology [3]. Thus, it is feasible to design the textile structural composites with
considerable flexibility in performance based upon a wide variety of preform geome-
tries and structure parameter.
The ant colony optimization (ACO) algorithm is a novel simulated ecosystem evo-
lutionary algorithm. It takes inspiration from the observations of ant colonies foraging
behavior with which ants can find the shortest paths from food sources to their nest
[4]. Preliminary study has shown that the ant colony algorithm is very robust and has
great abilities in searching better solutions. ACO algorithm has been successfully
used to solve many problems of practical significance including the Quadratic As-
signment Problem [5], Traveling Salesman Problem [6], Single Machine Total
Weighted Tardiness Problem [7].
Whereas wide application of hollow-rectangular-section 3-D braided composite in
engineering, the 3-D braided composite made by 4-step method is taken as the studied
object in this paper. The 3-D braided composite description is given in section 2. The
stiffness and damping analysis of the composite were introduced in Section 3. And
model of structure optimization designs for the composite were proposed. A hybrid
algorithm based on ACO algorithm for solving the optimization problem is presented
in section 4. Simulation results and the discussion of the results are presented in sec-
tion 5. Finally, conclusions are given in Section 6.

2 3-D Braided Composite Description

The 3-D braided fibre construction is produced by a braiding technique which inter-
laces and orients the yarns by an orthogonal shedding motion, followed by a compact-
ing motion in the braided direction. The basic fibre structure of 4-step 3-D braided
composite is four-direction texture. An idealized unit cell structure is constructed
based upon the fibre bundles oriented in four body diagonal directions in a rectangular
parallelepiped which is shown schematically in Fig.1 [8][9]. The yarn orientation
angles α and β are so-called braided angle. From the geometry of a unit cell associ-
ated with particular fibre architecture, different systems of yarn can identified whose
fibre orientations are defined by their respective interior angle α and cross sectional
angle β , as previously show in Fig.1. According to requirement of braid technology,
the range of α and β are commonly between 20°and 60°. In Fig.1, geometric parame-
ters of cross section of composite are b1, b2, h1, and h2 respectively.
262 K. Zhang

Fig.1. The unit cell structure of a 3-D braided composite

3 Optimization Design Model of 3-D Braided Composite

3.1 The Stiffness Properties of 3-D Braided Composite

A model for grain system interruption has been chosen wherein the stiffness for each
system of yarns are superimposed proportionately according to contributing volume
of four yarn system to determine the stiffness of the composite. Let volume percent of
fibre of the composite be V f , volume fraction of each system of braiding yarn is equal
to V f / 4 [10]. Assuming each system of yarn can be represented by a comparable
unidirectional lamina with an elastic matrix defined as C and a coordinate transforma-
tion matrix as T, the elastic stiffness matrix Q of this yarn system in the longitudinal
direction of the composite can be expressed as Q = TCT T .
Here, we ignore approximately the effect of stress components σ y , σ z and τ yz , and
only consider the effect of σ x , τ zx and τ xy .
The constitutive relations of braiding yarn for 3-D braided composite are gotten as
follows [11]

[ N ]i = [ K ]i [ε ]i (i = 1,2,3,4) . (1)

By stiffness matrix of each system, according to composition principle of composite

mechanics, we can obtain the general stiffness matrix of 3-D composite ：
1 4
[K ] = ∑ [ K ]i , (2)
4 i =1
 Structure Design of the 3-D Braided Composite 263

⎡ K11 0 K14 ⎤
0
⎢ ⎥
0 K 22 K 23
0 ⎥
where λ i = V f i / V f , [K ]i =⎢ , K11 , K 22 , K 33 and K 44 are
⎢ 0 K 32 0 ⎥
K 33
⎢ ⎥
⎣ K 41 0 K 44 ⎦
0
tensile, flexural, torsional and shear stiffness coefficient respectively, the others are
the coupling stiffness coefficient..

3.2 The Damping Properties of 3-D Braided Composite

By definition, specific damping capacity can be expressed as

Ψ = ∆U/U m , (3)

where ∆U is dissipation energy in anyone stress cycle, U m is maximum strain energy

in the cycle.
Adams and Bacon had presented cell damping model of laminated composite in
reference [12], they think that per unit area dissipation energy of each layer can be
broken down into three component relate to direct strain, thus
∆U = ∆U x + ∆U y + ∆U xy . (4)

Considering the influence of transverse shear deformation, reference [13] has modi-
fied equation (6). In equation (7), they added to two terms dissipation energy resulted
by transverse shearing strain, namely
∆U = ∆U x + ∆U y + ∆U xy + ∆U yz + ∆U xz . (5)

As before, 4-step 3-D braided composite can be regard as superimpose of four unidi-
rectional fibre system, and each yarn system is not in same plane. Thus, unit energy
dissipation of each system should break down into six components,
∆U = ∆U x ′ + ∆U y ′ + ∆U z ′ + ∆U x ′y ′ + ∆U y ′z ′ + ∆U x ′z ′ , (6)

where x ′y ′z ′ is fibre coordinate system of each system.

Hereinafter, suppose composite only have the action of bending moment M x . Gen-
eral dissipation energy of composite is the superimpose by dissipation energy of each
system according to volume fraction [11] ：
1 4 1 4 l'
∆U = ∑ ∆U k = ∑ ∆Wk ∫0
/2
M x2 dx . (7)
4 k =1 4 k =1
where l’ is length of the composite. And maximum strain energy of composite is as
' 1 l' / 2 2
U m = ∫0l / 2 M x k x dx = ∫ M x dx . (8)
k 22 0
264 K. Zhang

Therefore, specific damping capacity of the composite can be obtained as follows

k 22 4
Ψ= ∑ ∆Wk . (9)
4 k =1

3.3 Optimization Design Model of 3-D Braided Composite

We choose specific damping capacity and some stiffness coefficient simultaneously

as maximum optimization objective function, structure parameter of materials (α, β)
and geometric parameter of section (b1, b2, h1, h2) as design variables, constraint
condition are numeric area of structure parameters of materials on technological re-
quirements. Here, we design optimization mathematical models of the 3-D braided
composite as follows,
f=max ( η1Ψ + η 2 K11 + η 3 K 22 + η 4 K 33 + η5 K 44 )

s.t. α min ≤ α ≤ α max , β min ≤ β ≤ β max

(h2 / b2 ) min ≤ (h2 / b2 ) ≤ (h2 / b2 ) max (10)

(b1 / b2 ) min ≤ (b1 / b2 ) ≤ (b1 / b2 ) max

(h1 / h2 ) min ≤ (h1 / h2 ) ≤ (h1 / h2 ) max

Where η1 + η 2 + η 3 + η 4 + η5 = 1 , and η1 to η 5 are respectively weight coefficients for

specific damping capacity and flexural stiffness coefficient in the unitive objective
function. Of course, we also obtain other optimization model according to practical
requirement. For the non-linear optimization design problem, optimization solution
can be obtained expediently by using the facilities of ant colony optimization.

4 Hybrid Optimization Algorithm

4.1 Optimization Model

Ant colony optimization is a kind of discrete algorithm. There for the optimization
problem must be converted in to a discrete one before can be solved by an ant colony
algorithm.
First discretize each variable in its feasible region, i. e., divide the variables to be
optimized in its value range. The number of the division relates with the range of the
variable and the complexity of the problem, etc. Suppose the variable to be optimized
is divided into q i nodes, denote the node value as x ij where i is variable number
( i = 1,2, K n ) and j is node number of x i ( j = 1,2, K q ij ).
After the division, a network which composed of variables and their division is
constructed as depicted in Fig. 2, where “●”is the node of each variable division. The
number of the division of each variable can be unequal and the division can be linear
or nonlinear.
 Structure Design of the 3-D Braided Composite 265

Suppose using m ants to optimize n variables, for ant k at time t = 0 , set the ant at
origin O. Then the ant will start searching according to the maximum transition prob-
ability. The searching starts from x 1 and goes on in sequence. Each ant can only se-
lect one path in its division to do transition. The ant k transit from xi to xi+1are called a
step and is denoted as t . The ant k starts searching from the first variable x 1 to the last
variable xn is called an iteration and is denoted as NC. An iteration includes n steps.

x
Fig. 2. Variable division and ant search path

At time t, the transition probability of ant k transit from node h of x i to node l of

xi+1 is defined as equation (11)

phlk (t ) =
[τ hl (t )]α [η hl ]β ,
q i +1

∑ [τ ] [η hj ]
α β (11)
hj (t )
j =1

where τ hl (t ) is the pheromone level of node l of x i + 1 at time t. η hl ( NC ) is the

visibility of node l of xi+1 in NC iteration. The denominator represents the sum of
pheromone level multiple visibility of all node of xi+1 at time t. Given α , β > 0 are
heuristic factors, they will control the relative importance of pheromone level τ hl (t )
and visibility η hl ( NC ) respectively. While searching, the ant will select the node
which has the maximum transition probability node to transit.
The visibility of node j of variable x i in iteration NC is defined as equation (12)

( ri upper − ri lower ) − xij − x*ij ( NC − 1)

η ij ( NC ) = , (12)
ri upper − ri lower

where i is variable number ( i = 1,2, K n ) and j is node number of x i ( j = 1,2, K q ij ).

riupper , rilower are the upper and lower limit value of range of variable x i respectively.
266 K. Zhang

* ( NC − 1 ) is the optimal value of the variables (parameter values of datum axes)

x ij

in iteration ( NC − 1) . In the first iteration, x * ( NC − 1 ) can be obtained by using the

ij
least-square method. After that ( NC > 1 ), x * ( NC − 1 ) will be the node value of
ij
the variable relevant the optimal path produced in last iteration.
At the initial time t = 0 , the pheromone levels of each node are equal, i. e.,
τ ij (0) = c and all of the ants are at origin O. After n time units, all of the ants will
crawl from start node to terminal node and the pheromone levels of each node can be
adjusted according following formation.

τ ij (t + n) = ρτij (t) + ∆τ ij , (13)

where i is variable number ( i = 1,2, K n ) and j is node number of vari-

able x i ( j = 1,2, K q ij ). (1 − ρ ) represents the decay of pheromone level of node xij
from t to time t + n . The evaporation rate ρ (1 < ρ < 1) in equation (13) is an impor-
tant parameter which is used to ensure that pheromones on unvisited paths evaporate
over time and that those paths are affixed low visiting probabilities in the future. A
judicious compromise between pheromone evaporation and deposition is necessary.

m
∆τ ij = ∑ ∆τ
k =1
k
ij , (14)

where ∆τ ijk (t ) is the pheromone of node xij left by ant k in this iteration. It can be
calculated according to equation (15)

，
⎧Q f k if kth ant pass through this node in this iteration
∆τ ijk = ⎨
⎩ ，
0 else
, (15)

where f k is the value of objective function of the ant k in this iteration. It is calcu-
lated with equation (10). Q is a positive constant.

4.2 Search Algorithm Hybridization

Being a global search method, the ACO algorithm is expected to give its best results
if it is augmented with a local search method that is responsible for fine search. The
ACO algorithm is hybridized with a gradient based search. The ‘‘candidate’’ solution
found by the ACO algorithm at each iteration step is used as an initial solution to
commence a gradient-based search using ‘‘fmincon’’ function in MATLAB [14],
which based on the interior-reflective Newton method. The ‘‘fmincon’’ function can
solve the constraint problem here.
 Structure Design of the 3-D Braided Composite 267

4.3 Hybrid Optimization Algorithm

The procedure of hybrid optimization algorithm is shown as follows.

Step1: Decide the number of variable and discrete the variables by dividing them.
Each variable is divided into q i nodes which constitute the number vector q . Node
value is x ij .
Step2: Given the ant number m and define a one dimensional vector path k for
ant k which has n elements. The one dimensional vector path k is called path table
and used to save the path that the ant k passed through.
Step3: Initialization. Let time counter t = 0 and the maximum iteration number
is NC max . Let the pheromone level of each node τ ij (0) = c , ∆ τ ij = 0 . Set the m
ants at origin O. According equation (12), calculate the visibility η ij (1) used in the
first iteration.
，
Step4: Let variable counter i = 1 variable division number counter q = 1 .
Step5: Use equation (11) to calculate the transition probability of ant k transit from
node h of variable x i to node l of variable x i + 1 and select the maximum probability
to do transition. Then save node l of x i + 1 to the ith element of path k . Transit all
ants from node of variable x i to node of variable x i + 1 .
Step6: Let i = i + 1 , if i ≤ n , then go to Step 5, else, go to Step 7.
Step7: Based on the path that ant k walked, i. e., path k , calculate the objective
function f k of ant k according to objective functions. Save the node values relevant
optimal path in this iteration.
Step8: Let t ← t + 1 , NC ← NC + 1 .Update the pheromone level of each node ac-
cording to equations (13), (14), (15) and clear all elements in path k ( k = 1 ~ m ).
Calculate visibility of each node according to equation (12).
Step9: Let the above solution as an initial solution, and carry out a gradient-based
search using ‘‘fmincon’’ function in MATLAB Optimization Toolbox.
Step10: If NC < NC max and the whole ant colony have not converged to the same
path yet, then set all of the ants at the origin O again and go to Step 4; If
NC < NC max and the whole ant colony have converged to the same path then the
iteration complete and the optimal path is the optimal solution.

5 Numerical Examples
In order to test the validity of the proposed optimization design procedure, numerical
examples have been preformed. Suppose braiding yarn system (four-direction) of 3-D
braided composite be homogeneous carbon fiber, where performance of component
：
material are Ef1=258.2GPa, Ef2=18.2 GPa, Gf21=36.7 GPa, µf21=0.25, µm=0.35,
ΨL=0.0045, ΨT=0.0422, ΨLT=0.0705, ΨTT =0.0421, Em=3.4GPa, Gm=1.26GPa [11].
The performance of unidirectional composite for each system can be calculated by
268 K. Zhang

performance of component material according to mesomechanics. The perimeter of

rectangular cross section is equal to 0.08m, and length of the composite is 0.5m.
Here, we choose common constraint condition: αmin=20°, αmax=60°, βmin=20°,
βmax=60°, (h2 / b2 ) min =0.4, (h2 / b2 ) max =2.5, (b1 / b2 ) min =0.2, (b1 / b2 ) max =0.8,
(h1 / h2 ) min =0.2, (h1 / h2 ) max =0.8. Two examples are performed as follows.

(1) Example 1: Let η1=1, η2=η3=η4=η5=0, maximum of damping Ψ can be calcu-

lated by using hybrid optimization algorithm presented, and the optimization results
according to optimization model are shown in Table 1.
(2) Example 2: Let η3=1, η1=η2=η4=η5=0, maximum of stiffness K 22 can be cal-
culated and the optimization results are shown in Table 1.
(3) Example 3: Let η1=η3=0.5, η2=η4=η5=0. Having eliminated difference of or-
der of magnitude for Ψ and K22, the optimization results are shown in Table 1.
As shown in Table 1, the optimal objectives of hollow-rectangular-section 3-D
braided composite can be improved observably by optimization design for structure
parameter of materials and geometric parameter of section.
From the optimization results of Example 1 and 2, if only make a search for
maximizing specific damping capacity or stiffness coefficients of the composite,
damping or stiffness characteristic of the hollow-rectangular-section 3-D braided
composite can be improved obviously . But it may influence on flexural stiffness or
，
damping of composite respectively (may result in a few reducing) thus Example 1
and 2 exist any limitation and shortage. The optimization results of Example 3 show
that both of damping and stiffness may attain better maximum at the same time. The
optimization effects denote that the optimization models are propitious to engineering
application.

Table 1. The results of preliminary design and optimization design

β K 22
α h2 / b2 b1 / b 2 h1 / h2 Ψ
(GPa·m2)
30° 30° 1.0 0.5 0.5 0.0027 2.51×10-5
Preliminary
30° 30° 1.0 0.5 0.5 0.0027 2.51×10-5
design
30° 30° 1.0 0.5 0.5 0.0027 2.51×10-5
36° 52° 1.04 0.35 0.34 0.0106 1.59×10-5
Optimal
23° 28° 1.12 0.29 0.28 0.0022 1.87×10-4
design
27° 45° 1.29 0.22 0.22 0.0079 3.92×10-5

6 Conclusions
According to the stiffness and damping properties of hollow-rectangular-section 3-D
braided composite, the mathematical model of structure design of the 3-D braided
composite was proposed. The results of numeral examples show that the better damp-
ing and stiffness characteristic could be obtained by using optimal design, contenting
 Structure Design of the 3-D Braided Composite 269

the determinate restriction. The method proposed here is useful for the design and
engineering application of the kind of member.

References
1. Huang, G.: Modern Textile Composites. Chinese Textile Press, Beijing (2000)
2. Zheng, X.T., Ye, T.Q.: Microstructure Analysis of 4-Step Three-Dimensional Braided
Composite. Chinese Journal of Aeronautics 16(3), 142–149 (2003)
3. Wang, Y.Q., Wang, A.S.D.: On the Topological Yarn Structure of 3-D Rectangular and
Tubular Braided Preforma. Composites Science and Technology 51, 575–583 (1994)
4. Colorni, A., Dorigo, M., Maniezzo, V.: Distributed Optimization by Ant Colonies. In: The
First European Conference on Artificial Life, pp. 134–142 (1991)
5. Gambardella, L.M., Taillard, E.D., Dorigo, M.: Ant Colonies for the Quadratic Assign-
ment Problem. J. Oper. Res. Soc. 50, 167–176 (1999)
6. Gamardella, L.M., Bianchi, L., Dorigo, M.: An Ant Colony Optimization Approach to the
Probabilistic Traveling Salesman Problem. In: Guervós, J.J.M., Adamidis, P.A., Beyer,
H.-G., Fernández-Villacañas, J.-L., Schwefel, H.-P. (eds.) PPSN 2002. LNCS, vol. 2439,
p. 883. Springer, Heidelberg (2002)
7. Dorigo, M., Stuzle, T.: Ant Colony Optimization. MIT Press, Cambridge (2004)
8. Yang, J.M.: Fiber Inclination Model of Three-dimensional Textile Structural Composites.
Journal of Composite Materials 20, 472–476 (1986)
9. Li, W., Hammad, M., EI-Shiekh, A.: Structural Analysis of 3-D Braided Preforms for
Composites, Part 1:Two-Step Preforms. J. Text. Inst. 81(4), 515–537 (1990)
10. Cai, G.W., Liao, D.X.: The Stiffness Coefficients of Rectangular Cross Section Beams
with 2-step Three-dimensional Braided Composite Beam. Journal of Huazhong University
of Science and Technology 24(12), 26–28 (1996)
11. Cai, G.W., Zhou, X.H., Liao, D.X.: Analysis of the Stiffness and Damping of the 4-step
Three-dimensional Braided Composite Links. Journal of Mechanical Strength 21(1),
18–23 (1999)
12. Adams, R.D., Bacon, D.G.C.: Effect of Orientation and Laminated Geometry on the
Dynamic Properties of CFRP. J. Composite Materials 7(10), 402–406 (1973)
13. Lin, D.X., Ni, R.G., Adams, R.D.: Prediction and Measurement of the Vibrational Damp-
ing Parameters of Carbon and Glass Fibre-Reinforced Plastics Plates. Journal of Compos-
ite Material 18(3), 132–151 (1984)
14. Grace, A.: Optimization Toolbox for Use with MATLAB, User’s Guide. Math Works Inc.
(1994)
 Robot Virtual Assembly Based on Collision Detection in
Java3D

Peihua Chen1,∗, Qixin Cao2, Charles Lo1, Zhen Zhang1, and Yang Yang1
1
Research Institute of Robotics, Shanghai Jiao Tong University
cph@sjtu.edu.cn, charleslo77@gmail.com, zzh2000@sjtu.edu.cn,
iyangyang186@yahoo.com.cn
2
The State key Laboratory of Mechanical System and Vibration,
Shanghai Jiao Tong University
qxcao@sjtu.edu.cn

Abstract. This paper discusses a virtual assembly system of robots based on

collision detection in Java 3D. The development of this system is focused on
three major components: Model Transformation, Visual Virtual Assembly, and
an XML output format. Every component is presented, as well as a novel and
effective algorithm for collision detection is proposed.

Keywords: Virtual Assembly, Collision detection, Service robot, Java 3D,

VRML.

1 Introduction
Virtual Assembly (VA) is a key component of virtual manufacturing [1]. Presently,
virtual assembly serves mainly as a visualization tool to examine the geometrical
representation of the assembly design and provide a 3D view of the assembly process
in the field of virtual manufacturing [2-5]. However, in service robots’ simulation, we
also need to assemble many types of robots in the virtual environment (VE). There are
mainly two methods of virtual assembly in these systems, e.g., assembling every
single part together via hard-coding using link parameters between each part, and
another method is assembling a virtual robot in an XML(eXtensible Modeling Lan-
guage), both of which are not visual and complicated.
This paper focuses on the visual implementation of making virtual assembly of service
robots in Java3D. We choose Java JDK 1.6.0 and Java 3D 1.5.2 as the platform. First we
discuss the system architecture and methodology. This is followed by its implementation.
After that, an application of this system will be presented and conclusions obtained.

2 System Architecture
2.1 System Architecture
The architecture of this system is presented in Fig.1. It includes three parts: Model
Transformation, Visual Virtual Assembly (VVA), and XML Output. Firstly, after we
∗
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 270–277, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Robot Virtual Assembly Based on Collision Detection in Java3D 271

create the CAD models of the robot, we export these models as IGES [7] format to be
imported into 3DS Max. In 3DS Max, we transform the Global Coordinate System
(GCS) of the IGES model into the coordinate system according to the right-hand rule,
and then export the models in VRML format. Then, we load the VRML files into Java
3D. After which, we set the scale factor of the model to 1:1. Then, we set the target
model as static, and move the object model towards the target model until the colli-
sion detection between them is activated. After which, we attach the object model to
the target model, which means appending the object node to the target node. If the
virtual assembly is not completed, then import another VRML file into the VE for
assembling; if it is finished, then output the assembly result into an XML file.

Fig. 1. System Architecture of VA for robots

2.2 Scene Graph Architecture

The use of scene graph architecture in Java 3D can help us organize the graphical ob-
jects easily in the virtual 3D world. Scene graphs offer better options for flexible and
efficient rendering [8]. Fig.2 illustrates a scene graph architecture of Java 3D for loading
the VRML models into the VE for scaling, translation and rotation transformations. In
Fig.2, BG stands for BranchGroup Node, and TG, Transform -Group Node.
As shown in Fig.2, the Simple Universe object provides a foundation for the entire
scene graph. A BranchGroup that is contained within another sub-graph may be re-
parented or detached during run-time if the appropriate capabilities are set. The Trans-
formGroup node specifies a single spatial transformation, via a Transform3D object,
that can position, orient, and scale all of its children [9], e.g., moveTG, rotTG,
scaleTG, etc., as shown in Fig.2. The moveTG handles the translations of the VRML
model, rotTG is for rotations, and scaleTG is for scaling.
272 P. Chen et al.

Fig. 2. Scene graph architecture in Java 3D

3 Implementation

3.1 Loading VRML Models

Java 3D provide support for runtime loaders. In this paper, we use the VrmlLoader
class in package j3d-vrml97 to load the VRML files. In our project, we rewrite a
vrmlload class with the construction method of vrmlload (String filename). After
selecting a VRML file, the model will be seen in the 3D Viewer of our project, as
shown in the left picture of Fig.3. The coordinate system in 3D Viewer is right-hand,
with the orientation semantics such that +Y is horizontal to the right, +X is directly
towards the user, and +Z is the local gravitational direction upwards.

Fig. 3. A repositioned VRML model and its control panel

 Robot Virtual Assembly Based on Collision Detection in Java3D 273

Fig.3 also shows how a model’s position, orientation, and size can be adjusted.
Here it shows a robot body that has been moved and rotated. In the control panel of
the model, there are several operations. They are: translations of x, y, and z axes;
rotations of the model around x, y, and z axes; scaling the model; displaying “Pos(x,
y, z)” and “Rot(x, y, z)” at runtime, as shown in Fig.3. In addition, when another
model is loaded, a corresponding control panel will be created at the same time.

3.2 Algorithm for Collision Detection

In the VVA system, collision detection plays an important role. Every node in the
scene graph contains a Bounding field that stores the geometric extent of the node
[10]. Java3D offers two classes "WakeupOn -CollisionEntry" and “WakeupOnColli-
sion -Exit" to carry out collision detection for 3D models. When any objects in Java
3D scene collide with other bounds, "Wakeup -OnCollisionEntry" is triggered and
when collision is released, "WakeupOnCollision -Exit" is triggered. So we can use
them as triggering conditions and we are able to obtain the control behavior after
collision occurs.
This paper proposes a novel and effective algorithm to detect collisions. VRML
file format uses IndexedFaceSet node to describe shapes of faces and triangle patches
to reveal all types of shapes in 3D models. IndexedFaceSet includes a coord field that
contains a series of spatial points, which can be used in the coordIndex field to build
the surfaces of 3D models. We conserve the points' position, and establish space for
collision detection using the following method.
BoundingPolytope is chosen to use more than three half-space [11] to define a
convex and closed polygonal space of a VRML model. We make use it to build the
collision detection space. The function that defines every half-space α is: Ax + By +
Cz + D ≤ 0, where A, B, C, and D are the parameters that specify the plane. The pa-
rameters are passed into the x, y, z, and w fields, respectively, of a Vector4d object in
the constructor of BoundingPolytope (Vector4d[] planes). Thus we must make out the
parameters A, B, C, and D of every triangle plane. The following method is to com-
pute the value of A, B, C, and D. We suppose that one triangle is made up of three
uuuuv uuuuv
points: L, M, and N, shown as in Fig.4. And we obtain vector LM and MN based on
a vertex sequence.

Fig. 4. Triangle plane in the WCS

274 P. Chen et al.

uuuuv
LM = ( X M − X L )i + (YM − YL ) j + ( Z M − Z L ) k . (1)
uuuuv
MN = ( X N − X M )i + (YN − YM ) j + ( Z N − Z M )k
. (2)
uuuuv uuuuv
The normal vector of plane α is the vector product of LM and MN . That is
⎡ i j k ⎤
uuv uuuuv uuuuv .
n f = LM × MN = ⎢⎢ ( X M − X L ) (YM − YL ) ( Z M − Z L ) ⎥⎥ (3)
⎣⎢( X N − X M ) (YN − YM ) ( Z N − Z M ) ⎦⎥
uuv
Once we obtain normal vector n f , we get values of A, B, C. Then we can obtain the
value of constant D if we substitute any point’s coordinate into the equation. After
transmitting all plane parameters to one group of Vector4d objects, we can use
BoundingPolytope to build a geometrical detection space. Then, we invoke setColli-
sionBounds(Bounds bounds) to set the collision bounds. After setting a collision
bound, Java3D can carry out collision detection in response to the setting bounds.

3.3 Visual Virtual Assembly

While assembling the object model to the target model, we have to rotate the object
model around the z-axis, y-axis, and x-axis in this order to correspond with the target
Local Coordinate System (LCS). Then we need to move along the x, y, and z axes in
both the positive and negative directions. In Fig.5, we present how we assemble the
model J4 to model J3. Fig.5 (a) shows the initial positions of the two VRML models.
Fig.5 (b) presents the state after the second model (J4) is rotated about the x-axis by -
90°. We can also see from this picture that we don’t need to make any translations

J3

J4

(a) (b) (c)

(d) (e) (f)

Fig. 5. Example of assembling one VRML model to another
 Robot Virtual Assembly Based on Collision Detection in Java3D 275

along y-axis anymore, because the second model’s y-axis is already corresponding to
the first model. From Fig.5(c) and Fig.5 (d), we can obtain the offset values along +x
direction and –x direction through the collision detection between the two models. As
we have set the threshold value of collision detection at 0.5mm, thus when the dis-
tance between the two models is smaller than 0.5, the collision detection will be acti-
vated and the actual offset value will be obtained. Here, the offset value in the +x
direction is 140.5-0.5=140(mm), and the offset value in the –x direction is -
80.5+0.5=-80(mm), as shown in Fig.5 (d). So, in order to make the x-axis of the sec-
ond model correspond to the first one, the total offset value along the x-axis is: (140 +
(- 80)) / 2 = 30(mm), and the result can be seen in Fig.5 (e). Then move along the z-
axis, and the operations are similar with the previous method. The final result of this
virtual assembly is presented in Fig.5 (f).
After we have moved the second model to the correct position, we attach the sec-
ond model’s node to the first model’s node.

3.4 Save the Virtual Assembly Result in XML Files

After we have realized the virtual assembly of the robot, we save the virtual assembly
result as XML files, as shown in Fig.1. In this paper, we use JAXB (Java Architecture
for XML Binding) [12] to access and process XML data. JAXB makes XML easy to
use by compiling an XML schema into one or more Java technology classes. The
structure of the XML schema of our project is presented in Fig.6, and the output data
could be referred to, in this structure.

Fig. 6. Structure of the XML schema

4 Application
In this paper, we create the CAD models of the real robot as shown in Fig.7, and then
translate them to VRML files according to the right-hand rule.
276 P. Chen et al.

Fig. 7. A real robot with left arm

Then, we start the visual virtual assembly in Java 3D, as shown in Fig.8. After the
virtual assembly is completed, we save the virtual assembly result to an XML file.
Then we run the program again, and open the XML file just created, as shown in
Fig.9 (a). At the same time, we can control the motion of every joint through the cor-
responding control panel, as shown in Fig.9 (b).

Fig. 8. Virtual assembly of the robot

(a) Open the XML (b) Control the mo-

file in Jave 3D tions of the robot

Fig. 9. Open the XML file stored the VA data and Control the motions of the robot
 Robot Virtual Assembly Based on Collision Detection in Java3D 277

5 Conclusions
From the application in section 5, Virtual Assembly of robots is carried out based on
Collision Detection, and it proves to work well. Virtual Assembly of robots based on
Collision Detection in Java3D allows for visual feedback, and makes the process
much easier for the end-users. Through the virtual assembly, users can obtain the
assembly result file in XML format that complies with a special XML schema. Then
the users can easily load the new assembled robot via the XML file, and operate every
joint of the robot, e.g., developing a program for workspace and kinematic analysis.

Acknowledgments. This work was supported in part by the National High Technol-
ogy Research and Development Program of China under grant 2007AA041703-1,
2006AA040203, 2008AA042602 and 2007AA041602; the authors gratefully ac-
knowledge the support from YASKAWA Electric Cooperation for supporting the
collaborative research funds and address our thanks to Mr. Ikuo agamatsu and Mr.
Masaru Adachi at YASKAWA for their cooperation.

References
1. Jayaram, S., Connacher, H., Lyons, K.: Virtual Assembly Using Virtual Reality Tech-
niques. CAD 29(8), 575–584 (1997)
2. Choi, A.C.K., Chan, D.S.K., Yuen, A.M.F.: Application of Virtual Assembly Tools for
Improving Product Design. Int. J. Adv. Manuf. Technol. 19, 377–383 (2002)
3. Jayaram, S., et al.: VADE: a Virtual Assembly Design Environment. IEEE Computer
Graphics and Applications (1999)
4. Jiang-sheng, L., Ying-xue, Y., Pahlovy, S.A., Jian-guang, L.: A novel data decomposition
and information translation method from CAD system to virtual assembly application. Int.
J. Adv. Manuf. Technol. 28, 395–402 (2006)
5. Choi, A.C.K., Chan, D.S.K., Yuen, A.M.F.: Application of Virtual Assembly Tools for
Improving Product Design. Int. J. Adv. Manuf. Technol. 19, 377–383 (2002)
6. Ko, C.C., Cheng, C.D.: Interactive Web-Based Virtual Reality with Java 3D, ch. 1. IGI
Global (July 9, 2008)
7. http://ts.nist.gov/standards/iges//
8. Sowizral, H., Rushforth, K., Deering, M.: The Java 3D API specification, 2nd edn. Addi-
son-Wesley, Reading (2001)
9. Sun Microsystems. The Java 3DTM API Specification, Version 1.2 (March 2000)
10. Selman, D.: Java 3D Programming. Manning Publications (2002)
11. http://en.wikipedia.org/wiki/Half-space
12. Ort, E., Mehta, B.: Java Architecture for XML Binding (JAXB). Sun Microsystems (2003)
 Robust Mobile Robot Localization by Tracking Natural
Landmarks

Xiaowei Feng, Shuai Guo, Xianhua Li, and Yongyi He

College of Mechatronics Engineering and Automation, Shanghai University,

Shanghai 200444, China
xwfeng1982@163.com

Abstract. This article presents a feature-based localization framework to use

with conventional 2D laser rangefinder. The system is based on the Unscented
Kalman Filter (UKF) approach, which can reduce the errors in the calculation
of the robot’s position and orientation. The framework consists of two main
parts: feature extraction and multi-sensor fusing localization. The novelty of
this system is that a new segmentation algorithm based-on the micro-tangent
line (MTL) is introduced. Features, such as lines, corners and curves, can be
characterized from the segments. For each landmark, the geometrical parame-
ters are provided with statistical information, which are used in the subsequent
matching phase, together with a priori map, so as to get an optimal estimate of
the robot pose. Experimental results show that the proposed localization method
is efficient in office-like environment.

Keywords: Localization, feature extraction, Unscented Kalman Filter, mobile

robot.

1 Introduction
Localization is the key for autonomous mobile robot navigation systems. It is well
known that using solely the data from odometry is not sufficient since the odometry
provides unbounded position error [1]. The problem gives rise to the solution that
robot carries exteroceptive sensors (sonar, infrared, laser, vision, etc.) to get land-
marks and uses this operation to update its pose estimation.
Instead of working directly with raw scan points, feature based localization first
transforms the raw scans into geometric features. Feature based method increase the
efficiency and robustness of the navigation and makes the data more compact. Hence,
it is necessary to automatically, and robustly, detect landmarks for data association,
and ultimately localization and mapping purposes for robots. This approach has been
studied and employed intensively in recent research on Robotics [2]. One of the key
problems of feature-based navigation is the reliable acquisition of information from
the environment. Due to many advantages of laser rangefinder, such as good reliabil-
ity, high sampling rate and precision, it is wildly used in mobile robot navigation.
UKF[3] is a nonlinear optimal estimator based on given sampling strategy without
the linearization steps required by the Extended Kalman Filter(EKF). And it is not

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 278–287, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Robust Mobile Robot Localization by Tracking Natural Landmarks 279

restricted to Gaussian distributions. It has been improved that with appropriate sigma
point selection schemes, the UKF performs better estimation by exploiting more in-
formation such as the first three moments of an arbitrary distribution or the first four
non-zero moments of a Gaussian distribution. In proposed system, UKF fuses the
information from odometry with the information from range images to get robot’s
position estimation.
In this paper, we describe a feature based localization system, which is composed
of two main parts: feature extraction and probabilistic localization. For feature extrac-
tion, we describe a geometrical feature detection method for use with conventional 2D
laser rangefinder. The contribution of this paper is that our method employs the vari-
ance function of local tangent associated to laser range scan readings to track multiple
types of natural landmarks, such as line, curve and corner, instead of calculating cur-
vature function. The extracted features and their error models are used, together with
a priori map, by an UKF so as to get an optimal estimate of robot’s current pose vec-
tor and the associated uncertainty.

2 Scan Data Clustering and Segmentation

2.1 Adaptive Clustering

Clustering is a key procedure for landmark extraction process. It can roughly segment
the range image to some consecutive point clusters, which will also enhance the effi-
ciency of the extraction procedure. Range readings belong to the same segment while
the distance between them is less than a given threshold [4].

Dmax = rl −1 (sin ∆φ sin(λ − ∆φ )) + 3σ r (1)

Where ∆φ is the laser angular resolution, λ is an auxiliary constant parameter and

σ r the range variance. In our experiments, the parameter values are σ r =0.01 m and
λ=10°. Clusters having too few numbers of points, such as isolated range readings, are
removed as big noise.

2.2 Segmentation Based on MTL

For range data acquired from clustering process, the relationship of adjacent points
is, however, totally dependent on the shape of the scanned environment. The differ-
ence between neighboring points can be depicted by a function of geometry pa-
rameters, such as the direction of the local tangent or curvature of objects in the
environment, etc.
The tangent of the points belong to the same features is supposed to be same or
change regularly, see Fig.1(a). In this work, we employ the variance function of local
tangent associated to laser range scan images to break the segments into several
unique features. The tangent of each scan point can be calculated by line-regression of
k-neighboring points. In our experiment, k is 7.
280 X. Feng et al.

(a) (b)
Fig. 1. a) MTL of range points; b) Features extracted associated to (a): -line, - corner and
○ -curve

The variance between two consecutive tangents which belong to point p (i ) and
p (i + 1) is

var = arctan( (k (i) − k (i +1)) (1+ k (i) ⋅ k (i +1)) ) (2)

Where k (i ) and k (i + 1) are the slop of two tangents for p (i ) and p (i + 1) . Fig.2 is
the tangent variance function with respect to Fig.1(a).

Fig. 2. Tangent variance function with respect to Fig.1(a)

The tangent variance function is searched for geometric features:

1) Line segments. Therefore, they are those sets of consecutive range readings with
variance values close to zero. And they have a certain minimum width ( lmin =5
range points in our experiment).
2) Curve segments. The variance function associated to a curve segment is detected
wherever a region of the function has a constant value larger than a preset value.
This value can be found over a certain minimum width and the maximum in the
middle of the region differs only by a certain amount from the values at the end.
3) Corners. The corner is always defined by a value associated to a local peak of the
variance function. It is a region between two segments which have been marked
as line or curve and also has a certain minimum width, which can filter out spuri-
ous peak value due to random noise which has a relative low value.
 Robust Mobile Robot Localization by Tracking Natural Landmarks 281

Fig.1(b) shows the detected line segments, corners and curve segments associated to
Fig.1(a).

3 Landmark Extraction

3.1 Line Segments

The line parameters are (θ , d ) from the line equation x cosθ1 + y sinθ1 = d1 . Given n
points p(ri , φi ) , the regression parameters are given by following expressions[5]:

1 −2∑ i ( y − yi )( x − xi ) 1 N ,
θ = arctan( ) = arctan
2 ∑ i ⎡⎣( y − yi )2 − ( x − xi )2 ⎤⎦ 2 D (3)
d = x cos θ + y sin θ .

Where x = ∑ r cos φ
i i n and y = ∑ ri sin φi n .
The parameters θ and d are a function of all the measured points pi , i = 1 n
belonging to equation (3). Assuming that the individual measurements are independ-
ent, the covariance matrix of parameters (θ , d ) can be calculated as

Cθ ,d = ∑ i J i Cxyi J iT .
n
(4)

Where uncertainty of a measured point pi is represented by a covariance matrix C xyi ,

and the terms of Jacobian matrix are obtained as follows
∂θ ∂xi = (( y − yi ) D + ( x − xi ) N ) ( N 2 + D 2 ) ,
∂θ ∂yi = (( x − xi ) D + ( y − yi ) N ) ( N 2 + D2 ) ,
(5)
∂d ∂xi = cos θ n + ( y cos θ − x sin θ )(∂θ ∂xi ) ,
∂d ∂yi = sin θ n + ( y cos θ − x sin θ )(∂θ ∂yi ) .
Where N and D are the numerator and denominator of the expression (3).

3.2 Curve Segments

A circle can be defined by the equation ( x − x0 ) + ( y − y0 ) 2 = r 2 , where ( x0 , y0 )

2

and r are the center and the radius of the circle, respectively. For a circle fitting prob-
lem, the data set ( x, y ) is known and the circle parameters need to be estimated.
Assuming that we have obtained M measurements p( xi , yi ) , our objective is to
find p( x0 , y0 , r ) subjecting to fi ( x0 , y0 , r) = ( xi − x0 )2 + ( yi − y0 )2 − r 2 = 0
where i = 1 M .
282 X. Feng et al.

We use the Levenberg-Marquardt[6] method to solve this nonlinear least-squares

problem.

( AkT Ak + λk I )∆pk = − AkT f k . (6)

A = [∇f1 , ∇f 2 , , ∇f M ] is the Jacobian matrix, and ∆pk = pk +1 − pk and

T
Where
pk is the estimate of p( x0 , y0 , r ) at the k − th iteration. The initial value of
( x0 , y0 , r ) is got by using the first, last and middle points. According to[7], an esti-
mation of the curve segment uncertainty can be approximated by three points of the
curve segment.

3.3 Corners

Then, real and virtual corners can be obtained from the intersection of the previously
detected line segments. Once a corner is detected, its position ( xc , yc ) is estimated
as the intersection of the two lines which generate it. Given two lines
x cos θ1 + y sin θ1 = d1 and x cos θ 2 + y sin θ 2 = d 2 , the equations to calculate
corner ( xc , yc ) is:

⎧ xc = (d1 sin θ 2 − d 2 sin θ1 ) (sin(θ 2 − θ1 ))

⎨ . (7)
⎩ yc = ( d 2 cos θ1 − d1 cos θ 2 ) (sin(θ 2 − θ1 ))
Where (d1 , θ1 ) and (d 2 ,θ 2 ) can be get from previous line extraction process. And
the variance of ( xc , yc ) is

C xc yc = JCd1θ1d2θ2 J T . (8)

Where J is the Jacobian matrix of equation (7) to (d1 ,θ1 , d 2 , θ 2 ) . And Cd1θ1d 2θ2 is
the covariance matrix of line parameters ( d1 , θ1 , d 2 , θ 2 ) . For convenience, the line
parameters ( d1 , θ1 ) and ( d 2 , θ 2 ) for two different tangent lines are supposed to be
independent. Then Cd1θ1d 2θ 2 is simplified to be diag (Cd1θ1 , Cd2θ2 ) .

4 Natural Landmark-Based Localization

Wheel slippage causes a drift in the mobile robot, which can be quite disastrous in
mobile robot localization and navigation. To provide noise rejection and develop a
model dependant estimation of position and orientation, an Unscented Kalman Filter
(UKF) is applied in our research.
 Robust Mobile Robot Localization by Tracking Natural Landmarks 283

In this application the initial location of the robot is known, and the robot has a
priori map of the locations of geometric landmark pi = [ p x , p y ]i . Each landmark is
assumed to be known. At each time step, observations z j (k ) of these Landmarks are
taken. UKF combines observations with a model of the system dynamics which are
the kinematics relationships that express the change in position of the robot as a func-
tion of the displacement of two wheels to produce an optimal estimate of the robot's
position and orientation.

4.1 Kinematic Model

The system model describes how the vehicle’s position x(k ) changes with time in
response to a control input u ( k ) = [ sl , sr ]k and a noise disturbance v( k ) . It has
the form:
x(k + 1) = f ( x(k ), u (k )) + v(k ) , v(k ) ∼ N (0, Q(k )) . (9)

Where f ( x(k ), u (k )) is nonlinear state transition function. And noise v(k ) is as-
sumed to be zero mean Gaussian with variance Q ( k ) . According to the kinematics of
a two wheel differential drive mobile robot using in our experiment, the translation
function then has the form:

⎡ x(k ) ⎤ ⎡ ∆s ⋅ cos(θ + ∆θ 2) ⎤
f (⋅, ⋅) = ⎢⎢ y (k ) ⎥⎥ + ⎢⎢ ∆s ⋅ sin(θ + ∆θ 2) ⎥⎥ .
(10)
⎢⎣θ (k ) ⎥⎦ ⎢⎣ ∆θ ⎥⎦

Where ∆s = ( sr + sl ) 2 and ∆θ = ( sr − sl ) b and b is the width between

two wheels. The control input u ( k ) = [ sl , sr ]k are the displacement of two
wheels, which can be obtained from odometry.

4.2 Observation Model

The observation model describes how the measurements z j ( k ) are related to the ve-
hicle’s position and has the form:

z j (k ) = h( x(k ), pi ) + w j (k ) , w j (k ) ∼ N (0, R j (k )) . (11)

The measurement function h( x(k ), pi ) expresses an observation z j (k ) from the

sensor to the target as a function of the robot’s location x (k ) and the landmark coor-
dinates pi . w j ( k ) is a zero-mean, Gaussian measurement noise with covari-
ance R j ( k ) . Both the distance and orientation information measured by the sensors
284 X. Feng et al.

are relative to the robot itself, therefore the observation equation is the function of the
robot state and the landmark:

⎡ ( p − x (k )) 2 + ( p − y (k )) 2 ⎤
⎢ x y
⎥
h(⋅, ⋅) = ⎢ p y − y (k ) ⎥ . (12)
⎢ arctan( ) − θ (k ) ⎥
⎣ p x − x(k ) ⎦

4.3 Estimation Cycle

First, using the system model and control input u (k ) , we predict the robot’s new
location at time step k+1:

x (k + 1) = ∑ i = 0 Wi ⋅ χ ik +1 .
2n
(13)

Where χ ik +1 = f ( χˆ ik , u (k )) in which Wi is the mean weight with respect to the

i − th sigma point χ i
k +1
[3], and x (k + 1) is the expectation of robot predicted pose.
And the variance associated with this prediction is

PX k +1 = ∑ i = 0 Wi ( χik +1 − x (k + 1))( χik +1 − x (k + 1))T .

2n
(14)

Next, we use this predicted robot location to generate predicted observations of

landmark p j , j = 1, , M .

z j (k + 1) = ∑ i =0 Wiζ i k, j+1 .
2n
(15)

Where ζ i , j = h( χik +1 , p j ) .
k +1

The next step is to actually take a number of observations z j (k + 1) of different

landmarks and compare these with our predicted observations. The difference be-
tween a prediction z j ( k + 1) and an observation z j ( k + 1) is termed the innovation,
and it is written as ν j ( k + 1) = z j (k + 1) − z j (k + 1) .
The variances of these innovations S j ( k + 1) , can be estimated from

2n
S j (k + 1) = ∑ Wi (ζ i k, j+1 − z ( k + 1))(ζ i k, j+1 − z ( k + 1))T + R j (k + 1) . (16)
i =0

Where R j (k + 1) is covariance of additive measurement noise for observations

z j (k + 1) .
 Robust Mobile Robot Localization by Tracking Natural Landmarks 285

The innovations are obtained by matching observations to predicted targets. The

matching criterion is given as validation regions in

v j (k + 1) S −j 1 (k + 1)vTj (k + 1) ≤ χ 2 . (17)

This equation is used to test each sensor observation z j (k + 1) for membership in

the validation gate for each predicted measurement. When a single observation falls in
a validation gate, we get a successful match.
Finally, we use innovation ν j ( k + 1) to update the vehicle location vector and as-
sociated variance. We then utilize the standard result that the Kalman gain can be
written as
K kj +1 = PX k +1 PZ−k1+1 . (18)
k +1Z j j

Where
2n .
PZ k +1 = ∑Wi (ζ i k, j+1 − z j ( k + 1))(ζ i k, j+1 − z j (k + 1))T (19)
j
i =0

2n . (20)
PX k +1 = ∑Wi ( χik +1 − x (k + 1))(ζ ik, j+1 − z j (k + 1))T
k +1Z j
i =0

Then the updated vehicle position estimate is

.
xˆ(k + 1) = x (k + 1) + ∑ i =0 K k +1 ⋅ν j ( k + 1)
M
(21)

With associated variance

.
= PX k +1 − ∑ j =0 K kj +1 ⋅ PZ−k1+1 ⋅ ( K kj +1 )T
M
PXˆ (22)
k +1 j

5 Experimental Results
In this section, we use the mobile robot RoboSHU which is equipped with a laser
rangefinder LMS200 as exteroceptive sensor to test the localization capability of our
method. The rangefinder is configured to have a view of 180 degrees with a resolution
of nearly 0.5°. The robot is running under a RTOS in a 1.6 GHz PC and can be con-
trolled by a remote control module via wireless network. All Algorithms are pro-
grammed in C++. The average processing times per scan is 0.3s, being suitable for
real time applications. The localization experiment is carried out in a laboratory envi-
ronment (see Fig.4).
For simplicity, we assumed time-invariant, empirically estimated covariance matri-
ces Q(k ) = Q and R j ( k ) = R , j = 1, , M , given by R = diag (50,50, 0.01) ,
Q(1,1) = Q(2, 2) = 10 , and Q(1, 2) = Q(2,1) = 0 . The initial pose ( x0 , y0 , θ 0 ) is
assumed Gaussian with initial covariance matrices P0 = diag (100,100, 0.025) .
286 X. Feng et al.

Fig. 3. Pose errors computed with the ground truth and the 1σ confidence bounds computed
using the estimate covariance of system in x, y and θ are shown

Fig.3 shows the pose estimate errors of our UKF localization method, respectively
in coordinates ( x, y ) and orientation θ . The 1σ confidence bounds for the estimates
are also superimposed. It can be seen that the errors are not divergent.

Fig. 4. Localization experiment: odometry(blue line), estimated trajectory(red line), real

data(black line)

Fig.4 represents the trajectory estimation obtained using our localization method.
The trajectory predicted by the odometry deviates from the true robot’s path as time
increases. The UKF is capable of tracking the robot’s position fairly accurate, even
though the trajectory includes abrupt turns. We can see that the robot can localize
itself by fusing the sensor information so as to navigate successfully. This experiment
proves the validity of the localization algorithm based on UKF.

6 Conclusions
This paper describes the implementation and obtained results of an UKF based Local-
ization framework using a 2D laser rangefinder. A segmentation based on the
variance of the local tangent value is introduced. This algorithm can provide line
segments, corners and curve segments for mobile robot localization. Landmarks ex-
tracted from these segments are not only characterized by their parameters but also
uncertainties. The system is based on an UKF that utilizes matches between observed
geometric Landmarks and a priori map of Landmarks locations to provide optimal
pose estimation for mobile robot. The accuracy and robustness of the proposed
method was demonstrated in an indoor environment experiment.
 Robust Mobile Robot Localization by Tracking Natural Landmarks 287

Acknowledgement
We are grateful for the financial support from the State High-Tech Development Plan
(863 program) of China under contract No. 2007AA041604.

References
1. Iyengar, S., Elfes, A.: Autonomous Mobile Robots, vol. 1, 2. IEEE Computer Society Press,
Los Alamitos (1991)
2. Jensfelt, P., Christensen, H.: Laser Based Position Acquisition and Tracking in an Indoor
Environment. In: IEEE Int. Proc. on Robotics and Automation, vol. 1 (1998)
3. Julier, S.J., Uhlmann, J.K., Durrant-Whyte, H.F.: A new approach for filtering nonlinear
systems. In: Proc. Am. Contr. Conf., Seattle, WA, pp. 1628–1632 (1995)
4. Arras, K.O., Tomatis, N., Jensen, B.T., Siegwart, R.: Multisensor on-the-fly localization:
Precisionand reliability for applications. Robotics and Autonomous Systems 34, 131–143
(2001)
5. Arras, K.O., Siegwart, R.: Feature extraction and scene interpretation for map based naviga-
tion and map building. In: Proc. of SPIE, Mobile Robotics XII, vol. 3210 (1997)
6. Nash, J.C.: Compact numerical methods for computers: linear algebra and function minimi-
zation. Adam Hilger Ltd. (1979)
7. Nunez, P., Vazquez-Martin, R., del Toro, J.C., Bandera, A., Sandoval, F.: A Curvature
based Method to Extract Natural Landmarks for Mobile Robot Navigation. In: IEEE Int.
Symposium on Intelligent Signal Processing 2007, October 2007, pp. 1–6 (2007)
 Multi-Robot Dynamic Task Allocation Using Modified
Ant Colony System

Zhenzhen Xu, Feng Xia, and Xianchao Zhang

School of Software, Dalian University of Technology, 116620 Dalian, China

njxzz@hotmail.com, dr.fxia@hotmail.com, xczhang@dlut.edu.cn

Abstract. This paper presents a dynamic task allocation algorithm for multiple
robots to visit multiple targets. This algorithm is specifically designed for the
environment where robots have dissimilar starting and ending locations. And the
constraint of balancing the number of targets visited by each robot is considered.
More importantly, this paper takes into account the dynamicity of multi-robot
system and the obstacles in the environment. This problem is modeled as a con-
strained MTSP which can not be transformed to TSP and can not be solved by
classical Ant Colony System (ACS). The Modified Ant Colony System (MACS)
is presented to solve this problem and the unvisited targets are allocated to ap-
propriate robots dynamically. The simulation results show that the output of the
proposed algorithm can satisfy the constraints and dynamicity for the problem of
multi-robot task allocation.

Keywords: multi-robot, task allocation, ant colony, MTSP.

1 Introduction
The multiple robots system has been widely used in the planetary exploration, seafloor
survey, mine countermeasures, mapping, rescue and so on. A typical mission is visiting
the targets locating on different positions in a certain area. Multi-robot task allocation
(MRTA) is a critical problem in multi-robot system. This research introduces a dy-
namic task allocation algorithm for multiple robots to visit multiple targets.
The existing research on multi-robot task allocation for targets visiting is limited and
immature. The waypoint reacquisition algorithm provides a computationally efficient
task allocation method [1]. This approach is accomplished by employing a cluster-first,
route-second heuristic technique with no feedback or iterations between the clustering
and route building steps. However, it will generally not find the high-quality solutions.
It also may leave some waypoints (i.e. targets) unvisited. The planning of underwater
robot group is modeled as a MTSP and implemented by genetic algorithm [2]. The
author of this paper has studied the multi-objective task allocation problem for
multi-robot system in the literature [3]. However, all the proposed algorithms do not
take into account the dynamicity of multi-robot system and the obstacles in the envi-
ronment. Due to the dynamicity of multi-robot system, when some robot is damaged by
a destructive target, the system must reallocate the targets unvisited to the remaining
robots on-line. Meanwhile, each robot will change its path during the procedure of
avoiding obstacles, thus the initial schedule of targets visiting may not be optimal and
rescheduling should be considered.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 288–297, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Multi-Robot Dynamic Task Allocation Using Modified Ant Colony System 289

In this paper, the task allocation problem for multiple robots is modeled as a con-
strained MTSP. The MTSP is an NP-hard problem in combinatorial optimization. It is a
generalization of the well-known TSP where more than one salesman is allowed to be
used in the solution. Although there is a wide body of the literature for the TSP, the
MTSP has not received the same amount of attention. To find a guaranteed optimal
solution to the MTSP using exhaustive searches is only feasible for very small number
of cities. Many of the proficient MTSP solution techniques are heuristic such as evo-
lutionary algorithm [4], simulated annealing [5], tabu search [6], genetic algorithms [7],
neural networks [8] and ant systems [9].
Due to the requirement of targets visiting for multiple robots, robots typically have
dissimilar starting and ending locations and the workload of each robot should be
balanced. This problem belongs to the fixed destination multi-depot MTSP [10] and the
constraints of balancing the targets number visited by each vehicle should also be taken
into account. However, the heuristic methods mentioned above can not deal with this
constrained MTSP directly. The classical Ant Colony System (ACS) algorithm is used
to solve MTSP which can be transformed to TSP widely. In ACS, each ant constructs a
solution from a random starting point and visits all targets one by one. However, it can
not solve the fixed destination multi-depot MTSP with the constraints mentioned
above. In this paper, a new Modified Ant Colony System (MACS) algorithm is pre-
sented to solve this constrained MTSP and applied to the multi-robot dynamic task
allocation problem.
The reminder of this paper is organized as follows. First, the multi-robot task allo-
cation problem is formulated by integer linear programming formulation. In Section 3,
we give a detail description of the Modified Ant Colony System (MACS). Multi-robot
dynamic task allocation using MACS is introduced in section 4. Finally, simulation
results show that the proposed algorithm can find optimized solutions which can satisfy
the constraints and dynamicity for the problem of multi-robot dynamic task allocation.

2 Multi-Robot Task Allocation Problem Statement

2.1 Integer Linear Programming Formulation

The multi-robot task allocation problem can be stated as follows: Given n targets with
random locations in a certain area, let there be m robots which typically have dis-
similar starting and ending locations outside the area. The m robots must visit the n
targets, and each target must be visited exactly once by only one robot. In order to
saving energy and time, the total distance of visiting all targets must be minimized. In
addition, the targets number visited by each robot should be average due to the re-
quirement of workload balancing.
We propose the following integer linear programming formulation for the con-
strained MTSP defined above. The distance objective function f (x ) can be described
as follows:
m ni −1
f ( x) = ∑ ( ∑ d (Ti k , Ti k +1 ) + d ( S i , Ti1 ) + d (Ti ni , Ei )) (1)
i =1 k =1
290 Z. Xu, F. Xia, and X. Zhang

where Ti k is the k th target visited by robot i , d (Ti k , Ti k +1 ) is the distance between

Ti k and Ti k +1 , Si is the starting depot of robot i , Ei is the ending depot of robot i ,
ni is the target number assigned to robot i .
The task number constraint g (x) is defined as follows:

g ( x) = max(ni ) − min( ni ) ∈ {0,1} (2)

The constrained task allocation problem for multiple robots is formulated as follows:
min f ( x) (3)

subject to g ( x ) (4)

2.2 Colony Architecture of Multi-Robot System

The multi-robot system discussed in this paper is constructed by the hierarchical colony
architecture [11]. Fig. 1 shows the logical hierarchical model of multi-robot system.

Fig. 1. The hierarchical colony architecture of multi-robot system

In the hierarchical colony architecture, a leader-follower organization structure is

adopted. The leader robot supervises the whole system to enhance global coordination
and supports the capability of self-organization in case of partial or total failure of one
or several robots. Meanwhile, each robot is an autonomous agent and distributed in
physics. It can autonomously execute individual tasks and communicate with other
robots when necessary. Thus, the hierarchical colony architecture provides multi-robot
system with a framework for interaction and coordination among member robots.
The leader robot can be dynamic chosen with any effective strategy. An index-based
method is proposed here according to the predefined unique index of each robot. The
robot with the index i ˄ 1 i  {1ˈ2ˈ, m  1}˅is selected to be the leader when the
former leader robot with the index i˄i  {1ˈ 2ˈ, m  1}˅is damaged and out of work.
 Multi-Robot Dynamic Task Allocation Using Modified Ant Colony System 291

3 Modified Ant Colony System

In this section, a Modified Ant Colony System (MACS) algorithm which is extended
from the classical Ant Colony System (ACS) [12] for solving multi-robot task alloca-
tion is proposed. Some modifications are adopted at the steps of initialization and
solution construction respectively.

3.1 Initialization

Due to the characteristic of fixed destination multi-depot MTSP, the MACS algorithm
is different from the classical ACS in initialization. The dissimilar starting depots and
ending depots of robots should be considered. Instead of putting all ants on the targets
randomly in ACS, the MACS put all ants on the starting depots or ending depots of
robots randomly. That is, we put all ants on the S i (i = 1,2, L , m) or Ei (i = 1,2,L, m)
randomly. All ants will start from one depot and search at the same space.
Similarly, initialization of the pheromone matrixes and cost matrixes of ant colony
are modified. The pheromone and the cost from one depot to all targets should be
calculated and stored.

3.2 Solution Construction

As far as the multiple depots and workload balancing are concerned, the method of
solution construction needs to be modified. According to the theory of classical ACS,
each ant runs the steps of solution construction and pheromone updating until all targets
have been visited. In MACS, there are three modifications during the procedure of
solution construction.
(1) The task number allocation phase is employed to realize the task number con-
straint. Here, the task number is equal to the number of targets assigned to each robot to
visit. This phase focuses on assigning the number of task while the task allocation and
optimization will be performed at the following phase. If the number of targets n can
be divided exactly by the number of robots m , the task number ni for robot
i (i = 1,2,L m) should be n / m . Otherwise, let s be the quotient and r be the re-
mainder, then ni can be defined mathematically as:

⎧ s + 1, if (u > u0 and v < r ) or (u ≤ u0 and w ≥ m − r )

ni = ⎨ (5)
⎩ s, if ( u ≤ u0 and w < m − r ) or (u > u0 and v ≥ r )

where u is a random uniform variable [0,1] and the value u0 is a parameter to influ-
ence the target number of different robots. v denotes the total number of robots whose
target number is s + 1 and w denotes the total number of robots whose target number
is s . After this phase, the task number for each robot to visit is clear and we can
m
see ∑n
i =1
i = n.
292 Z. Xu, F. Xia, and X. Zhang

(2) Each ant starts from the initial position, and then selects unvisited targets to
construct a route. In MACS, the procedure of solution construction is a little complex
compared with the ACS. Suppose that an ant starts from S p , p ∈ {1,2,L, m} , if the
number of targets which the ant has visited is equal to n p , then the ant returns to E p .
After that, the ant chooses an unvisited S i (i = 1,2, L p − 1, p + 1,L, m) , departs from
that starting depot and selected unvisited targets iteratively. The targets which have
been visited by an ant are recorded in the ant’s tabu table. A valid solute is constructed
by an ant until all the targets, starting depots and ending depots are visited. We define
the ant number is a , thus there are a solutions constructed by a ants.
Every ant selects the next city independently. The rules of moving from target i to
target j for ant k can be formulated as follows:

⎪
j = ⎨ l∈Nik
{
⎧ arg max τ il [η il ]β , } if q ≤ q0
(6)
⎪⎩ J , otherwise

If q is larger than q0 , the probability of moving from target i to target j for ant k
can be formulated as follows:

pijk =
[τ ij ]α [ηij ]β , if j ∈ N ik
∑ [τ il ]α [ηil ]β (7)
l∈N ik

where τ il is equal to the amount of pheromone on the path between the current target i
and possible target l . η il is the heuristic information which is defined as the inverse of
the cost (e.g. distance) between the two targets. q is a random uniform variable [0,1]
and the value q0 is a parameter to determine the relative influence between exploita-
tion and exploration. α and β are the parameters whose value determine the relation
between pheromone and heuristic information. N ik denotes the targets which ant k has
not visited.
(3) The objective function values of the routes constructed by all ants is computed
according to Eq.1 and sorted in increasing order. In order to improve future solutions,
the pheromone on the best solution found so far must be updated after all ants have
constructed their valid routes. The global pheromone updating is done using the fol-
lowing equation:
τ ij = (1 − ρ )τ ij + ρ∆τ ijbest , ∀(i, j ) ∈ T best (8)

where ρ (0 < ρ ≤ 1) is a parameter that controls the speed of global pheromone

evaporation. T best denotes the best solution found so far, and ∆τ ijbest is defined as the
inverse of the objective value of T best . In addition, each ant should execute local
 Multi-Robot Dynamic Task Allocation Using Modified Ant Colony System 293

pheromone updating as soon as it selects a new target j according to the following

equation:
τ ij = (1 − ξ )τ ij + ξτ 0 (9)

τ 0 = 1 / nC nn (10)

where ξ (0 < ξ < 1) represents the speed of local pheromone evaporation. τ 0 is the
initial value of pheromone. C nn is the objective value computed with the Nearest
Neighbor Algorithm. The global updating belongs to a positive feedback mechanism,
while the local updating can increase the opportunity to choose the targets which have
not been explored and avoid falling into the state of stagnation.
After the MACS have met the end condition (e.g. the pre-specified number of it-
erations, denoted as N c ), the optimal solution is obtained. The time complexity of
MACS is O ( N c ⋅ a ⋅ n 2 ) which is equal to the time complexity of the classical ACS.

4 Multi-Robot Dynamic Task Allocation Using MACS

Multi-robot dynamic task allocation consists of two phases: initial task allocation and
online task reallocation. The MACS is a general method and can be applied in both
phases. The corresponding optimal results will be obtained with appropriate input data
at different conditions.

4.1 Initial Task Allocation

In the first phase, the MACS algorithm is used to solve constrained MTSP and all
targets are allocated to appropriate robots initially. The MACS is carried out with the
program language of c and embedded in the control system software of each robot.
Initial task allocation should be run by a leader robot with centralized mode. The result
of allocation will be sent to each robot in the multi-robot system by communication.

4.2 Online Task Reallocation

In the second phase, the robots set off to visit individual targets and run local path
planning algorithm to avoid obstacles.
Task reallocation with centralized mode. The leader robot is responsible for colony
supervision online to conduct failures. All the follower robots will send state messages
to leader robot whenever a timer signal arrived during the procedure of executing tasks.
If one robot has been found losing touch with the leader robot. Then it is considered
encountering damage and lost. In order to avoid leaving some targets unvisited, the
leader robot will reallocate the tasks unvisited to other remaining robots with central-
ized mode. The MACS algorithm only runs in the control system of the leader robot at
this situation. The inputs of algorithm are the current positions of remaining robots and
the positions of targets unvisited. The output of task reallocation will be sent to the
294 Z. Xu, F. Xia, and X. Zhang

remaining robots. Each robot which receives this message will update its task list and
visit the new target.
Task rescheduling with distributed mode. In order to adapt the unpredictable envi-
ronment, each robot will reschedule its targets list during the procedure of avoiding
obstacles with distributed mode. The MACS algorithm runs in the control system of
each robot at this situation. The inputs of algorithm are the current position of the robot
avoiding obstacles and the positions of targets unvisited by this robot. The result of task
rescheduling is the targets list with new order which is probably different with the
former order of visiting.

5 Simulation Experiments
In this section, we discuss the parameter settings for the proposed task allocation al-
gorithm and present simulation results. According to the MACS algorithm, the pa-
rameters are set as follows: a = n , α = 1 , β = 2 , q0 = 0.9 , u0 = 0.5 , ρ = 0.5 ,
ξ = 0.1 , N c = 200 . The algorithm was tested in the scenario which contains 20 targets
( n = 20 ) and 3 robots ( m = 3 ). The targets positions and the obstacles are generated
randomly. Assumed that the respective starting and ending depots coordinates of the
three robots are S1 = ( 20,0) , S 2 = (50,0) , S 3 = (80,0) , E1 = (40,100) , E2 = (50,100) ,
E3 = (60,100) . Two typical experiments were carried out to verify the validity and
rationality of the proposed task allocation algorithm.
Fig. 2 shows the results of experiment A which focuses on verifying task reallocation
with centralized mode. The targets layout and the result of initial task allocation are
shown in Fig. 2 a). As we have seen, each robot obtains a valid route, and the task
number constraint is satisfied. The leader robot is R1 and it supervises the whole
multi-robot system.
In Fig. 2 b), the robot R2 is damaged and out of work, then the leader robot R1 detects
this failure. The MACS is applied to task reallocation by robot R1 with the input of
coordinates of targets unvisited and current positions of remaining robot (R1 and R3).
The result of online task reallocation is illustrated in Fig. 2 c). The unvisited targets are
allocated to R1 and R3 autonomously.
The coordinates of targets in this experiment are shown in Table 1. However, the
process of targets visiting with multi-robot system is unpredictable and the layout of
targets is randomly. There are many other simulation experiments with various layouts
of targets, and the similar valid solutions are obtained. This experiment results show
that the leader robot can perform online task reallocation with centralized mode using
MACS algorithm.
Experiment B focuses on the procedure of task rescheduling with distributed mode.
Assumed that there are two obstacles with the figure of square in the area, the layouts of
targets and obstacles can be seen in fig. 3 a). The two obstacles have the size of
15m × 10m and 20m × 5m respectively. The coordinates of targets in experiment B are
shown in Table 1. According to the initial task allocation, the paths of robot R2 and R3
are across the obstacles.
 Multi-Robot Dynamic Task Allocation Using Modified Ant Colony System 295

100 100

90 90

80 80

70 70

60 60
R3
Y(m)

Y(m)
50 50 R2
R1
40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
X(m) X(m)

a) b)

100

90

80

70

60
task reallocation
Y(m)

50

40

30

20

10

0
0 10 20 30 40 50 60 70 80 90 100
X(m)

c)

Fig. 2. Results in experiment A

R2 or R3 will run local path planning and avoid the obstacles as soon as it detects the
obstacle during the procedure of target visiting. Meanwhile, each robot independently
runs MACS online to optimize the schedule of targets visiting with the input of its
current position and the coordinates of targets unvisited in its task list. The results of
task rescheduling of R2 and R3 are shown in Fig. 3 b). The travel distance of robot
according to the new visiting order of targets is reduced obviously.

Table 1. Coordinates of targets

Experiment Coordinates
{52,82}, {18,40}, {47,90}, {37,32}, {37,86}, {79,69}, {57,37}, {69,69},
A {58,10}, {82,70}, {13,50}, {56,43}, {96,41}, {87,86}, {97,52}, {94,89},
{71,27}, {29,56}, {92,14}, {33,71}
{28,22}, {59,74}, {46,75}, {47,70}, {77,55}, {42,44}, {67,38}, {66,69},
B {35,60}, {21,60}, {20,51}, {48,46}, {60,66}, {15,19}, {37,26}, {63,63},
{22,30}, {50,69}, {17,43}, {61,88}
296 Z. Xu, F. Xia, and X. Zhang

The task rescheduling method allows the bulk of the computation to be distributed
among the robots on their respectively targets list and further reduces the computational
burden.
The computing time of MACS algorithm under the condition mentioned above is
less than 1s and can satisfy the requirement of real-time for robot control system. The
simulation results show that the output of the proposed algorithm can solve the con-
straints and dynamicity for the problem of multi-robot dynamic task allocation.

100 100

90 90

80 80

70 70

60 60
task
Y(m)

Y(m)
50 50 rescheduling

40 40

30 30

20 20

10 10

0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
X(m) X(m)

a) b)

Fig. 3. Results in experiment B

6 Conclusion

In this paper, we considered the situation which is more appropriate in the real
multi-robot task allocation problem for task visiting. Besides the objective of mini-
mizing the total distance of multiple robots, we also took into account the constraints of
multiple depots and workload balancing.
The task allocation problem for multiple robots was modeled as a constrained
MTSP. This is also a fixed destination multi-depot MTSP. As many proficient heuristic
methods to solve MTSP can not deal with such problem directly, this research proposed
Modified Ant Colony System (MACS) method to solve this constrained MTSP.
The proposed MACS algorithm was used to solve the dynamic task allocation for
multiple robots. Combined with the hierarchical colony architecture and the dynamicity
of multi-robot system, the MACS algorithm was not only applied to reallocate targets
unvisited with centralized mode when some robot encountered damage but also carried
out to reschedule individual targets list with distributed mode when some robot avoided
obstacles. This research made the first attempt to solve the dynamic task allocation
problem.
The simulation results show that the output of the proposed MACS algorithm can
not only solve the constraints of multi-robot task allocation problem, but also satisfy the
requirement of dynamicity for the problem of online task reallocation.
 Multi-Robot Dynamic Task Allocation Using Modified Ant Colony System 297

References
1. Stack, J.R., Smith, C.M., Hyland, J.C.: Efficient reacquisition path planning for multiple
autonomous underwater robots. In: Ocean 2004 - MTS/IEEE Techno-Ocean 2004 Bridges
across the Oceans, pp. 1564–1569 (2004)
2. Zhong, Y., Gu, G.C., Zhang, R.B.: New way of path planning for underwater robot group.
Journal of Harbin Engineering University 24(2), 166–169 (2003)
3. Xu, Z.Z., Li, Y.P., Feng, X.S.: Constrained Multi-objective Task Assignment for UUVs
using Multiple Ant Colonies System. In: The 2008 ISECS International Colloquium on
Computing, Communication, Control, and Management, pp. 462–466 (2008)
4. Fogel, D.B.: A parallel processing approach to a multiple traveling salesman problem using
evolutionary programming. In: Proceedings of the fourth annual symposium on parallel
processing, pp. 318–326 (1990)
5. Song, C., Lee, K., Lee, W.D.: Extended simulated annealing for augmented TSP and
multi-salesmen TSP. In: Proceedings of the international joint conference on neural net-
works, vol. 3, pp. 2340–2343 (2003)
6. Ryan, J.L., Bailey, T.G., Moore, J.T., Carlton, W.B.: Reactive Tabu search in unmanned
aerial reconnaissance simulations. In: Proceedings of the 1998 winter simulation confer-
ence, vol. 1, pp. 873–879 (1998)
7. Tang, L., Liu, J., Rong, A., Yang, Z.: A multiple traveling salesman problem model for hot
rolling scheduling in Shangai Baoshan Iron & Steel Complex. European Journal of Opera-
tional Research 124, 267–282 (2000)
8. Modares, A., Somhom, S., Enkawa, T.: A self-organizing neural network approach for
multiple traveling salesman and robot routing problems. International Transactions in
Operational Research 6, 591–606 (1999)
9. Pan, J.J., Wang, D.W.: An ant colony optimization algorithm for multiple traveling sales-
man problem. In: Proceedings of the first international conference on innovative computing,
information and control (2006)
10. Kara, I., Bektas, T.: Integer linear programming formulations of multiple salesman prob-
lems and its variations. European Journal of Operational Research, 1449–1458 (2006)
11. Xu, Z.Z., Li, Y.P., Feng, X.S.: A Hierarchical control system for heterogeneous multiple
uuv cooperation task. Robot 30(2), 155–159 (2008)
12. Dorigo, M., Gambardella, L.M.: Ant colony system: A cooperative learning approach to the
traveling salesman problem. IEEE Transactions on Evolutionary Computation 1(1), 53–66
(1997)
 A Novel Character Recognition Algorithm Based
on Hidden Markov Models

Yu Wang, Xueye Wei, Lei Han, and Xiaojin Wu

School of Electronics and Information Engineering, Beijing JiaoTong University,

Beijing 100044, China
bjtuwangyu@126.com

Abstract. In this paper, a novel character recognition algorithm based on the

Hidden Markov Models is proposed. Several typical character features are ex-
tracted from every character being recognized. A novel 1D multiple Hidden
Markov models is constructed based on the features to recognize characters. A
large number of vehicle license plate characters are used to test the perform-
ance of the algorithm. Experimental results prove that the recognition rate of
this algorithm is high aiming at different kinds of character.

Keywords: Character recognition, Hidden Markov Models.

1 Introduction
The optical character recognition (OCR) has become an important research orienta-
tion in modern computer application domain. The character has many applications,
such as the vehicle license plate character recognition, information retrieval and un-
derstanding for video images, the character recognition for bills and documents and
etc. In the past years, many researches have been devoted to character recognition.
Independent component analysis character recognition in [1] recognizes characters
using a target function to reconstruct the character being recognized and analyzing the
errors; character recognition based on neural network in [2,3] recognize characters
with strong anti interference performance and brief program; character recognition
based on nonlinear active shape models in [4] has good performance when the charac-
ter is nonrigid and with normal variation rules. Actually, there are lots of different
uncertain information in character recognition such as the variable character shapes,
the complex character pattern, so the algorithms in [1-4] have respective limitations in
their applications.
It is a good method to adopt the statistical model building in character recognition
because of the uncertain information. The Hidden Markov Models (HMMs) has supe-
riority performance to process the serialization dynamic non-stationary signal. It has
been widely used in speech recognition [5]. The HMMs has great potential in the
character recognition because it has better flexibility in variable pattern processing.
One-dimension (1D) HMMs is used to recognize character in [6]. A two-dimension
(2D) HMMs is used in [7]. Because a 2DHMMs has a heavier computation than the
1DHMMs, a pseudo 2DHMMs with less computation is applied to recognize charac-
ters in [8].

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 298–305, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 A Novel Character Recognition Algorithm Based on Hidden Markov Models 299

The algorithm proposed in this paper is based on the Freeman chain code. It ex-
tracts multiple character features from each character. A 1D-multiple-HMMs is built
according these features to recognize characters. A large number of vehicle license
plate characters are used to test the performance of the algorithm. The experimental
results indicate this algorithm has a good recognition performance aiming at different
kinds of character.

2 Hidden Markov Models

A hidden Markov model (HMM) is a statistical model in which the system being
modeled is assumed to be a Markov process with unobserved state. The Hidden
Markov Model is a finite set of states, each of which is associated with a (generally
multidimensional) probability distribution. Transitions among the states are governed
by a set of probabilities called transition probabilities. In a particular state an outcome
or observation can be generated, according to the associated probability distribution.
It is only the outcome, not the state visible to an external observer and therefore states
are ``hidden'' to the outside; hence the name Hidden Markov Model.
An HMM can be defined as follows:
λ = (π , A, B ) (1)

where π is the initial state distribution, A is the set of state transition probabilities
and B is the probability distribution in each of the states.
Character recognition involves two basic problems of interest in HMM: the learn-
ing problem (construct models) and the evaluation problem (recognize character).
The learning problem is how to adjust the HMM parameters, so that the given set
of observations (called the training set) is represented by the model in the best way for
the intended application. Thus it would be clear that the ``quantity'' we wish to opti-
mize during the learning process can be different from application to application. In
other words there may be several optimization criteria for learning, out of which a
suitable one is selected depending on the application. This problem can be solved by
Baum-Welch algorithm in [5].
The evaluation problem is how to recognize characters by HMM. Give the HMM
model λ and the observation sequence X = ( X 1 , X 2 ,..., X T ) to calculate the probabil-
ity of P = ( X | λ ) . This problem can be solved by forward-backward algorithm in [5].

3 Model Building

This proposed algorithm is based on the Freeman chain code. Several different fea-
tures of every character are extracted. Slope analysis method is used to extract chain
code of each feature than has been projected. These extracted chain code are used as
observation sequence to establish model. Freeman 8-direction chain code is shown in
figure 1.
300 Y. Wang et al.

45$

Fig. 1. Freeman 8-direction chain code

3.1 Model Initialization

The number of observation symbols is M = 8 because the values of Freeman 8-

direction chain code are (1, 2,3, 4,5, 6, 7,8) which are used as observation sequence.
Considering many redundancy situations that exist in some characters, a typical leap
type left-right 8-states Markov chain model is selected as the basic structure of every
character. This Markov chain is shown is figure 2.

Fig. 2. Left-right 8-states Markov chain model

Each state in the model has a self-circulation process to absorb redundancy strokes.
Each state can transfer to the first and the second state behind it to describe the want-
ing strokes of characters. The initial state distribution π and the set of state transition
probabilities A can be defined by this model.
We suppose the probability of different symbols in observation sequence is the
same, so observation probability in each of the states can be initialized as 1/ M .

3.2 Character Features Extraction

3.2.1 Horizontal and Vertical Projection Chain Code

Some characters are simple, others are complex. Characters’ vertical direction is
complex which has lots of horizontal strokes. Characters’ horizontal direction is com-
plex which has lots vertical strokes. The sole horizontal projection and the sole verti-
cal projection can not well reflect all features of character individually. Horizontal
and vertical projection should be extract separately to express features of characters.
The projection modes are shown in figure 3.
After the extraction of horizontal and vertical projection image, Freeman chain
codes of projection image contour can be extracted utilizing the slope method [9]
 A Novel Character Recognition Algorithm Based on Hidden Markov Models 301

(a) (b)

Fig. 3. Horizontal and Vertical Projection mode

(a) (b)

Fig. 4. Character A and its horizontal projection contour

from the left of the contour. These chain codes can be used qua the horizontal and
vertical features observation sequence. A typical 16 × 16 character ``A`` and its hori-
zontal projection contour image are shown in figure 4. Its chain code is (122…63). In
the similar way, vertical projection contour chain codes can be got.

3.2.2 Target Pixels’ Centroid Distance Projection Chain Code

Many characters have translations, rotation and scale variation in different extent
affine variations, so how to extract appropriate features for computer recognition is
hard. In this paper, a novel character feature extraction method, the binary image
character centroid distance projection, is proposed. The contour image of character
centroid distance projection can be used to extract chain codes to express the observa-
tion sequence of character features. This method can express character features al-
though the character is in the affine variations.
Suppose that the origin of image f ( x, y ) is in the centroid position (a,b) and the
image f ( x, y ) is transformed from image f ( x ' , y ' ) in the transformation
，
x = x ' − a y = y ' − b . The coordinate of centroid (a,b) can be denoted as:

∫∫ x f ( x , y )dx dy
' ' ' ' '

a= (2)
∫∫ f ( x , y )dx dy
' ' ' '

∫∫ y f ( x , y )dx dy
' ' ' ' '

b= (3)
∫∫ f ( x , y )dx dy
' ' ' '

An image can be got from the image f ( x, y ) after the affine variations
302 Y. Wang et al.

⎡x ⎤ ⎡ x⎤ ⎡d ⎤
'

(4)
⎢ y ⎥ = A ⎢⎣ y ⎥⎦ + ⎢⎣ d ⎥⎦
a 1

⎣ ⎦
'

b 2

The image centroid position is not change can be proved.

PROOF: expression (5) can be got from the expression (4)

⎡x −d ⎤ ⎡x⎤
'

⎢y ⎥ = A ⎢⎣ y ⎥⎦ (5)
a 1

⎣ −d ⎦
'

b 2

Translations have no influence to the centroid position, so the translations’ influence

can be neglected. The centroid of image after affine variations can be expressed as:

∫∫ xf ( x , y )dxdy
' '

=
' a b
a (6)
∫∫ f ( x , y )dxdy
' '
a b

∫∫ yf ( x , y )dxdy
' '

b
'
=
a b
(7)
∫∫ f ( x , y )dxdy
' '
a b

⎡x ⎤ ⎡ x −d ⎤
'' '

⎢y ⎥ = ⎢y ⎥ , so
a a 1

Let
⎣ ⎦ ⎣ −d ⎦
'' '

b b 2

⎡x⎤ = ⎡ x ⎤ ⎡k ⎤ ⎡x ⎤
'' ''

−1
k
⎢ y ⎥ = ⎢⎣ k
11 12
(8)
⎥⎦ ⎢⎣ y ⎥⎦
a a

⎢⎣ y ⎥⎦ A
⎣ ⎦
''
k
''

b 21 22 b

Hence

∫∫ xf ( x , y )dxdy
'' ''

a =
' a b

∫∫ f ( x , y )dxdy
'' ''
a b

∫∫ (k x + k y ) f ( x , y )dx dy
'' '' '' '' '' ''

=
11 a 12 b a b a b

∫∫ f ( x , y )dx dy
'' '' '' ''
a b a b
(9)
∫∫ x f ( x , y )dx dy
'' '' '' '' ''

=k
a a b a b

∫∫ f ( x , y )dx dy
11 '' '' '' ''
a b a b

∫∫ y f ( x , y )dx dy
'' '' '' '' ''
b a b a b
+k
∫∫ f ( x , y )dx dy
12 '' '' '' ''
a b a b

=0
 A Novel Character Recognition Algorithm Based on Hidden Markov Models 303

2 2

5 2 2

(a) (b)

Fig. 5. 5 × 5 window and targets’ centroid distance

In a similar way b = 0 .
'

From the expressions above, we can get that the centroid position of image after af-
fine variations is not changed. So the character target pixels’ centroid distance ex-
presses the features of character. Figure 5 shows the window 5 × 5 with centroid as
the origin and the target pixels’ centroid distance image.

3.3 Model Building

An observation sequence set for every character X = { X 1 , X 2 , X 3 } and sub-models for

different character features λ1 , λ2 and λ3 can be got from these character features
Freeman chain codes. The system model λ = {λ1 , λ2 , λ3 } can be got. Each system
λ = {λ1 , λ2 , λ3 } corresponding to the character is saved in database for recognition.

4 Character Recognition

The character optimal model λ = {λ1 , λ2 , λ3 } can recognize characters after parameters
estimation. Three observation sequences can be got for each character being recognized.
In character recognition, forward-backward algorithm is used to calculate the probabil-
ity of each sub-model according to the corresponding observation sequence. Three val-
ues of probability P1 , P2 and P3 where P1 = P ( X 1 | λ1 ) , P2 = P ( X 2 | λ2 ) , P1 = P ( X 1 | λ1 ) . So
the probability P of each character being recognized for each character model in data-
base using the weighted method.
P = P ( X | λ ) = w1 P1 + w2 P2 + w3 P3 (10)

where w1 = w2 = w3 = 1/ 3 . The character model in database with the biggest prob-

ability P is the character recognized.

5 Experiment Results Analysis

In order to testify the character recognition performance of the proposed algorithm, a
lot of vehicle license plate characters are used as the experimental images. All the
304 Y. Wang et al.

Fig. 6. Experimental plates and the characters recognition results

experimental character images have been preprocessed: noise removal, image binari-
zation and character normalization. Some vehicle license plate used for experiment
and the characters recognition result are shown in figure 6. Results of recognition for
different kinds of characters are shown in table 1.

Tabel 1. Results of recognition for different kinds of characters

Recognition
Characters Recognized
Rate
Chinese
197 180 91.3%
character
Alphabet 366 335 91.5%
Numeral 816 770 94.3%
Total 1379 1285 93.1%

The results in table 1 show that the proposed algorithm has a high recognition rate
although there are three different kinds of characters and the results in figure 6 show
that characters can be well recognized no matter the characters are inclined or fouled.

6 Conclusion
The character recognition algorithm based on the HMM is a worthy studying subject.
In this paper, a recognition algorithm based on the Freeman chain code and HMM is
proposed. A 1D-multiple-HMMs is built according multiple features which are ex-
tracted from the character to recognize characters. A large number of vehicle license
plate characters are used to test the performance of the algorithm. The experimental
results show that this proposed algorithm has a well performance in recognition for
different kinds of characters.
 A Novel Character Recognition Algorithm Based on Hidden Markov Models 305

References
[1] Min, L., Wei, W., Xiaomin, Y., et al.: Independent component analysis based on licence
plate recognition. Journal of Sichuan University 43(6), 1259–1264 (2006)
[2] Qingxiong, Y.: Realization of character recognition based on neural network. Information
Technology (4), 92–95 (2005)
[3] Tindall, T.W.: Application of neural network technique to automatic license plate recog-
nition. In: Proceedings of European Covention on Security and Detection, Brighton,
Enaland, pp. 81–85 (1995)
[4] Shi Da-ming Gunn, S.R., Damper, R.I.: Handwritten Chinese radical recognition using
nonlinear active shape models. IEEE Trans. on Pattern Analysis and Machine
Intelligence 25(2), 277–280 (2003)
[5] Rabiner, L.R.: A tutorial on hidden Markov models and select applications in speech rec-
ognition. Proc. of IEEE 77(2), 257–286 (1989)
[6] Awaidah, S.M., Mahmoud, S.A.: A multiple feature/resolution scheme to Arabic (Indian)
numerals recognition using hidden Markov models. Signal Processing 89(6), 1176–1184
(2009)
[7] Yin, Z., Yunhe, P.: 2D-HMM based character recognition of the engineering drawing.
Journal of Computer Aided Design and Computer Graphics 11(5), 403–406 (1999)
[8] Li, J., Wang, J., Zhao, Y., Yang, Z.: A new approach for off-line handwritten Chinese
character recognition using self-adaptive HMM. In: The Fifth World Congress on Intelli-
gent Control and Automation, 2004 (WCICA 2004), vol. 5, pp. 4165–4168 (2004)
[9] Gang, L., Honggang, Z., Jun, G.: Application of HMM in Handwritten Digit OCR.
Journal of Computer Research and Development 40(8), 1252–1256 (2003)
[10] Keou, S., Fenggang, H., Xiaoting, L.: A fast algorithm for searching and tracking object
centroids in binary images. Pattern Recognition and Artificial Intelligence 11(2), 161–168
(1988)
 New Algorithms for Complex Fiber Image Recognition

Yan Ma and Shun-bao Li

Shanghai Normal University, Dept. of Computer Science, No. 100 GuiLin Road,
200234 Shanghai, China
{ma-yan,lsb}@shnu.edu.cn

Abstract. This paper presents an automatic recognition approach for complex

fiber images, including the binarization method based on fiber boundary conti-
nuity and variance, the corner detection algorithm based on chain codes, the
recognition method based on the curvature similarity in the same fiber. The ex-
perimental results show that the most fibers can be recognized by using the pro-
posed automatic recognition methods.

Keywords: Fiber image, Chain code, Binarization, Recognition.

1 Introduction
At present, many approaches have been presented for the fiber image recognition
using digital image processing technology. The processes of the fiber image analysis
and recognition are: image preprocessing, image analysis and image recognition. It
was presented to process data by the principal factor analysis and create experts li-
brary[1]. In [2], the method was designed to extract the image fiber edge, and use the
curve fitting to make a quantitative analysis of the fiber. The gray value morphologi-
cal method was given to process the fiber image[3].
The most above methods are designed for simple fiber image, that is, there is only
1-2 fibers and less cross-points in the image. In this paper, the fiber image being rec-
ognized is complex, that is, they are different in size, cross-cutting and adhesive in
different fibers. This paper presents the binarization method based on fiber boundaries
continuity and variance, thinning and pruning for binary fiber image, the corner detec-
tion algorithm based on chain codes, the recognition method based on the curvature
similarity in the same fiber. The experimental results show that the better results has
been obtained by using the proposed automatic recognition methods.

2 Image Preprocessing
2.1 Image Binarization

The input gray image is taken by melectron microscope (As shown in Fig. 1). The
fibers are mutual adhesive with different intensity from Fig.1. If we use the traditional
differential operators, such as Sobel operator, Canny operator for edge detection, the
results are sensitive to the threshold. When selecting a big threshold, high-contrast
edge can be detected, and low-contrast edge may be neglected. When selecting a small

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 306–312, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 New Algorithms for Complex Fiber Image Recognition 307

threshold, the small stripes may be detected and it is be not beneficial for the following
recognition.

Fig. 1. The original fiber image

The fiber image has the characteristics that the boundary variance is big and non-
boundary variance is small. We present the edge detection method based on the vari-
ance and reduce the threshold based on the continuity of the fiber boundary until all
edge has been detected. Fig.2 is the variance distribution along horizontal direction in
the period of Fig.1, x represents the pixel location, y respresents the variance. If we
encounter the valley point B while detecting, the variance of B is recorded as VarB ,
and the peak point belong to the neighborhood of B is further searched. If we encoun-
ter the peak point A in Fig.2, the variance of A is recorded as VarA . The initial
threshold is defined as Th0 . If VarA − VarB > Th0 , the point A is considered as
boundary point and recorded as 1, otherwise recorded as 0. While taking a big Th0 ,
other peak points with small variance, such as point C, D and E can not be detected.
It is found through out experiments that the distribution of variance of the same fi-
ber boundary has greater difference in the influence of outside light and the distance
between the fiber and electron microscopy. But the variance for the local adjacent
boundary is related and the image binarization method based on variance and bound-
ary continuity is presented:

A
Variance

C D
B

Fig. 2. The variance distribution

308 Y. Ma and S.-b. Li

Step 1: Define the initial threshold Th0 ; While the difference of variance diff var
between the peak point and the valley point is greater than Th0 , the peak point is
recorded as 1, otherwise recorded as 0, and diff var is recorded;
Step 2: Define step as step. Define Th0 = Th0 − step . The peak point, satisfied
that it is recorded as 0, diff var < Th0 , and at least one point in its 8-point neighhood
has been recorded as 1, is recorded as 1;
Step 3: Return to Step 2, until Th0 < Thmin , Thmin is the predefined minimum
value.
Through our experiments, Th0 , step and Thmin can be set to 20, 1 and 5 respec-
tively. The experimental results show that we can receive the better results by using
the proposed image binarization method. Fig.3 is the binarizaiton result for the image
shown in Fig.1.In some cases, it is the Contact Volume Editor that checks all the pdfs.
In such cases, the authors are not involved in the checking phase.

Fig. 3. Binary image

2.2 Thinning and Pruning

In this paper, the binary image is thinned using the look-up table thinning algorithm
proposed in the literature [8]. The 8-neighbor look-up table, in which the value of 8-
neighbor is corresponding with their value, is predefined. The binary image is
scanned from up to bottom, left to right, according to the value in the look-up table to
determine whether to delete the point.
The fiber image after thinning will have burr and the wrong recognized fiber may
occur due to the burr. The deburr method is to start from the endpoint, walk along the
connected pixels and calculate the walk length. The algorithm will continue until
encountering the cross point or another endpoint. When the walk length is lower than
the threshold (usually set to 10), it can be considered as burr and the values of the
walking through pixels are set to 0.
After binarization, thinning and deburr in the image process, it will occur that one
continuous fiber is segmented into two parts wrongly. So the filling algorithm is re-
quired to reduce measure error. In this paper, the angle between adjacent fiber end-
points is calculated and the filling algorithm proposed in the literature [9] is fulfilled.
 New Algorithms for Complex Fiber Image Recognition 309

2.3 Corner Detection

The thinned, pruned fiber image may be crossed and adhesive in different fibers, and
that will occur the wrong connecting point, corner, which should be deleted in order
to ensure the accuracy of recognition.
In this paper, a fiber from the starting point to terminal point is assigned 8 neighbor
chain codes. The chain code values are 0-7 shown in Fig.4.

2
3 1
4 0

5 7
6
Fig. 4. The value of chain codes

Define a chain code as the array c = {c1 , c2 , L , cn } (in this code only consider the
chain code values, regardless of the starting point coordinates). Define the difference
between the adjacent chain code values in the array as d i (that is, the relative chain
code), d i is calculated using the following equation:

⎧ ci +1 − ci ci +1 −c i ≤ 2
⎪
d i = ⎨ c i +1 − c i − 8 ci +1 − ci ≥ 6 (1)
⎪c −c +8 ci +1 − ci ≤ −6
⎩ i +1 i
Since that the largest difference between the directions represented by the adjacent
chain codes is ± 90o , the possible values for d i is 0, ±1, ±2. If the value is zero, the
direction of fiber is unchanging. If positive, it turns right. If negative, it turns left. The
point with d i of 2 may be considered as corner in the theory. But it may not be the
real corner due to the noise in the fiber image. To address the problem, the corner
detection algorithm based on chain code is given as follows:
(1) Calculate the difference d i between adjacent chain code values according to
formula (2), the difference between equation (2) and equation (1) is that the difference
d i between the values of the adjacent chain codes separated by length1 ( length1 can
be set to about 20 based on our experiments), which can avoid the impact of local
noise in the fiber image.

⎧ ci + length1 − ci ci + length1 −c i ≤ 2
⎪
d i = ⎨ci + length1 −c i −8 ci + length1 − ci ≥ 6 (2)
⎪c −c +8 ci + length1 − ci ≤ −6
⎩ i + length1 i
310 Y. Ma and S.-b. Li

(2) Once d i = 2 , define n1 = n d i+ j = 2

j = 1,2, L n in the predefined length n ,

that is, n1 is the number of d i of 2 among the following n points after i . Assume s
n1
be the scale value. While > s , the point corresponding to d i is considered as
n
corner. The features of the proposed corner detection algorithm is, while d i is 2, the
point is not seen as corner immediately, the point that satisfies that, the number of d i
of 2 in its following points exceeds the predefined value s ( s can be set to 0.4), is seen
as corner instead. That can avoid the impact of the local noise.

3 Fiber Recognition
The fiber boundaries are represented by chain codes. The chain codes must be
matched each other for the purpose of recognition. Since there exists many different
lengths of chain codes in an image, and two boundaries belonging to the same fiber
have the similarity in curvature, the fast recognition algorithm is given:
(1) Select a chain code ci = {ci1 , ci 2 , L , cin } , assume the cordination of the initial
point p0 is xi , yi , calculate the slope k of the point p0 . While k ≥ 1 , the direction
of point p0 is up and down, the search zone is [xi − h, xi + h] , While k < 1 , the
direction of point p0 is left and right, the search zone is [ yi − h, yi + h] , h can be
defined as the width of the thickness fiber and generally about 50.
(2) Assume a chain code c j = {c j1 , c j 2 , L , c jm } be found in the search zone, due
to unequal length for each chain codes, that is, n ≠ m . Suppose n ≥ m , select m
chain code values from the start of ci to compare with c j one by one. Once the
values of chain code are equal, the match value match increases 1. While one round
comparison stops, the chain codes of ci are shifted to right by one position and the
algorithm continues to compare it with c j until arriving the position of n − m + 1 .
The n − m + 1 number of match are obtained and we represent the largest match as
matchmax . Normalize matchmax , let matchmax = matchmax / m . Define the largest
{
match value of each chain codes as M = matchmax 1 , matchmax 2 , L , matchmax k . }
Find the maximum in M , once the maximum is greater than the predefined value, ci
is considered to be matched with c j , and they are recognized as the same fiber.

4 The Experimental Results

The above algorithms are achieved by MATLAB. Fig.5 shows the recognized fibers
using different gray value.
 New Algorithms for Complex Fiber Image Recognition 311

3
1
4
2

Fig. 5. The recognized fibers with same gray value

The most fibers are recognized correctly using the proposed algorithm as Fig.5. For
example, No.1, No.2 and No.3 are cross-cutting in Fig.5, while No.1 is recognized as
one fiber, No.2 and No.3 are recognized as the same fiber and the results are
consistent with the judge of human eyes. No.1 and No.4 are considered as the same
fiber in the position of No.5, since the bifurcation of the two fibers is not obvious in
the position of No.5.

5 Summary
This paper presents binarization, prunning, corner detection and recognition methods
for the recognition of cross-cutting and winding complex fibers. The experimental
results show that the most fibers can be recognized by using the proposed methods.
The diameter and length of fibers can be further measured with the help of the
recognized fibers.
Acknowledgments. Supported by Leading Academic Discipline Project of Shanghai
（
Normal University Project Number DZL805, DCL200802).

References
1. Jinhu, S., XiaoHong, W.: The Application of Image Processing in Specialty Animal Hair
Fibers Recognition. Journal of Textile Research 25(4), 26–27 (2004)
2. ZaiHua, Y., Yuhe, L.: Animal Fiber Length Measurement Based on Image Processing.
Computer Engineering and Applications 41(8), 180–181 (2005)
3. Jian, M., WenHua, Z., WeiGuo, C.: The Application of Gray Value Morphology in Animal
Fiber Image Processing. Computer Engineering and Applications 25(5), 42–44 (2004)
4. ShuaiJie, R., WenSheng, Z.: Measuring Diameter and Curvature of Fibers Based on Image
Analysis. Journal of Image and Graphics 13(6), 1153–1158 (2008)
5. SuPing, Y., PeiFeng, Z., JianPing, C.: Segmentation on Cotton Fiber Cross Secctions Based
on Mask. Computer Engineering 33(13), 188–191 (2007)
6. Bribiesca, E.: A Geometric Structure for Two-Dimensional Shapes and Three-Dimensional
Surfaces. Pattern Recognition 25(5), 483–496 (1992)
7. Canny, J.A.: Computational Approach to Edge Detection. IEEE Trans. on PAMI 8(6),
679–698 (1986)
312 Y. Ma and S.-b. Li

8. Wei, Y., Ke, G., YiKun, W.: An Efficient Index Thinning Algorithm of Fingerprint Image
Based on Eight Neighbourhood Points. Journal of SiChuan University of Science & Engi-
neering 21(2), 61–63 (2008)
9. XiaoFeng, H., WenYao, L., ShouDong, X.: The Computer Measurement for Cotton Fiber
Length. Cotton Science 15(6), 339–343 (2003)
 Laplacian Discriminant Projection Based on Affinity
Propagation

Xueping Chang and Zhonglong Zheng

Department of Computer Science, Zhejiang Normal University, Zhejiang, China

cxp0407@163.com, zhonglong@zjnu.cn

Abstract. The paper proposes a new algorithm for supervised dimensionality re-
duction, called Laplacian Discriminant Projection based on Affinity Propagation
(APLDP). APLDP defines three scatter matrices using similarities based on rep-
resentative exemplars which are found by Affinity Propagation Clustering. After
linear transformation, the considered pairwise samples within the same exemplar
subset and the same class are as close as possible, while those exemplars between
classes are as far as possible. The experiments on several data sets demonstrate
the competence of APLDP.

1 Introduction
Linear dimensionality reduction is widely spread for its simplicity and effectiveness
in dimensionality reduction. Principal component analysis(PCA), as a classic linear
method for unsupervised dimensionality reduction, aims at learning a kind of subspaces
where the maximum covariance of all training samples are preserved [2]. Locality Pre-
serving Projections(LPP), as another typical approach for unsupervised dimensionality
reduction, seeks projections to preserve the local structure of the sample space [3]. How-
ever, unsupervised learning algorithms can not properly model the underlying structures
and characteristics of different classes [5]. Discriminant features are often obtained by
supervised dimensionality reduction. Linear discriminant analysis(LDA) is one of the
most popular supervised techniques for classification [6,7]. LDA aims at learning dis-
criminant subspace where the within-class scatter is minimized and the between-class
scatter of samples is maximized at the same time. Many improved LDAs up to date
have demonstrated competitive performance in object classification [10,11,12,14,16].
The similarity measure of the scatter matrices of traditional LDA is based on the
distances between sample vectors and the corresponding center vectors. Frey proposed
a new clustering method, called affinity propagation clustering(APC), to identify a sub-
set of representative examples which is important for detecting patterns and processing
sensory signals in data[21]. Compared with center vector, the exemplars are much more
representative because the similarities between the samples and exemplars within the
same subsection are more closely than those of between the samples and the center
vectors. Motivated by APC, traditional LDA, Laplacian Eigenmaps(LE) and the near-
est neighborhood selection strategy [4,8,9], we propose a new dimensionality reduction
algorithm, Laplacian discriminant projection based on affinity propagation (APLDP),


Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 313–321, 2009.


c Springer-Verlag Berlin Heidelberg 2009
314 X. Chang and Z. Zheng

for discriminant feature extraction. In our algorithm, we play much emphasis on exem-
plar based scatter similarity which can be viewed as extensions of within-class scatter
and the between-class scatter similarity. We formulate the exemplar based scatter by
means of similarity criterions which were commonly used in LE and LPP. The ex-
tended exemplar scatter are governed by different Laplacian matrices. Generally, LDA
can be regarded as a special case of APLDP. Therefore, APLDP not only conquers the
non-Euclidean space problem[5], but also provides an alternative way to find potential
better discriminant subspaces.

2 Related Work
2.1 Linear Discriminant Analysis
Let X = [x1 , x2 , . . . , xn ] ∈ RD×n denote a data set matrix which consists of n samples
{xi }ni=1 ∈ RD . Linear dimensionality reduction algorithms focus on constructing a small
number, d, of features by applying a linear transformation W ∈ RD×d that maps each
sample data {xi } of X to the corresponding vector {yi } ∈ Rd in d−dimensional space as
follows:
W : xi ∈ RD → yi = W T xi ∈ Rd . (1)
Assume that the matrix X contains c classes, and is ordered such that samples appear
by class
X = [X1 , X2 , . . . , Xc ]
(2)
= [x11 , . . . , x1c1 , . . . , xc1 , . . . , xccc ].
In traditional LDA, two scatter matrices, i.e., within-class matrix and between-class
matrix are defined as follows [6]:
1


c
Sw = (x − m(i) )(x − m(i) )T (3)
n i=1 x∈X
i

1

c
Sb = ni (m(i) − m)(m(i) − m)T , (4)
n i=1
where ni is the number of samples in the i − th class Xi , m(i) is the mean vector of the
i − th class, and m is the mean vector of total samples. It follows from the definition that
trace(S w ) measures the within-class compactness, and trace(S b ) measures the between-
class separation.
The optimal transformation matrix W obtained by traditional LDA is computed as
follows [6]:
tr(W T S b W)
Wopt = arg max . (5)
w tr(W T S w W)
To solve the above optimization problem, the traditional LDA computes the following
generalized eigenvalue equation
S b wi = λS w wi , (6)
and takes the d eigenvectors that are associated with the d largest eigenvalues λi , i =
1, . . . , d.
 Laplacian Discriminant Projection Based on Affinity Propagation 315

2.2 Affinity Propagation Clustering

Affinity Propagation Clustering(APC), recently proposed by Frey in Science, is a kind
of clustering algorithm that works by finding a set of exemplars in the data and assigning
other data points to the exemplars[21]. The input of APC is pair-wise similarities s(n, m)
which are set to be negative squared Euclidean distance:
s(n, m) = −Vn − Cm 2 . (7)
In APC algorithm, all data points are simultaneously considered as potential exemplars.
Therefor, all s(m, m) can be initialized to be the same. In the process of clustering, two
kinds of message, the responsibility r(n, m) and the availability a(n, m), are exchanged
among data points by taking competition mechanism into account. To start with, the
availabilities are set to be zeroes, and the updating rule in the whole process is defined
as:
r(n, m) = s(n, m) − max {a(n, m ) + s(n, m )} (8)
m s.t.m
m
r(m, m) = s(m, m) −  max {a(m, m ) + s(m, m )} (9)
m s.t.m
m


a(n, m) = min{0, r(m, m) + max{0, r(n , m)}} (10)
n s.t.n
n,m


a(m, m) = max{0, r(n , m)} (11)
n s.t.n
m

Messages are updated on the above simple formula that search for minima of an ap-
propriately chosen energy function. The message-passing procedure may be terminated
under a fixed number of iterations, or after the local decisions keep constant for some
number of iterations.

3 Laplacian Discriminant Projection Based on Exemplars

3.1 Discriminant Within-Exemplar Scatter
Let xis denote the s − th sample in i − th class. We formulate Eq(1) as follows:

yis = W T xis (12)

Assume that each xis belongs to its exemplar neighborhood within i − th class.
Each X j ∈ X can be divided into k exemplar neighborhoods κis , s = 1, ..., k.
Let κis denote the center vector of exemplar neighborhoods κis . We define the within-
exemplar scatter of κis as a s as follows:


as = αis yis − γis 2 , (13)
yis ∈γis

where αis is the weight defined as

xis − κis 2
αis = exp(− ). (14)
t
316 X. Chang and Z. Zheng

To obtain the compact expression of Eq(13), let Ais = diag(α1i , ..., αni s ) be a diagonal
matrix and Yis = [y1i , ..., yni s ]. In addition, let ens denote the all one column vector of
length n s . Then γis = n1s Yis ens . Eq(13) can be reformulated as:


ais = αis tr{(yis − γis )(yis − γis )T }




= tr{ αis yis (yis )T } − 2tr{ αis yis (γis )T } + tr{ αis γis (γis )T }
2
= tr{Yis Ais (Yis )T } − tr{Yis Ais ens (ens )T (Yis )T } (15)
ns
(ens ) Ais ens
T
+ tr{Yis ens (ens )T (Yis )T }
n2s
= tr{Yis Lis (Yis )T }
where
2 s (ens )T Ais ens
Lis = Ais −
Ai ens (ens )T + ens (ens )T (16)
ns n2s
The within-exemplar scatter of class i:


ks

ks
ai = ais = tr{Yis Lis (Yis )T } (17)
s=1 s=1

There also exists a 0-1 indicator matrix Pis satisfying Yis = Yi Pis . Each column of Pis
records the exemplar information which is derived from the clustering process.


ks
ai = tr{Yi Pis Lis (Pis )T (Yis )T }
s=1 (18)
= tr{Yi Li YiT }
ks
where Li = s=1 Pis Lis (Pis )T . The total within-exemplar scatter of all classes:


c

c
A = ai = tr{Yi Li (Yi )T } (19)
i=1 i=1

There exists a 0-1 indicator matrix Qi satisfying Yi = YQi . Each column of Qi records
the class information which is known for supervised learning. Then Eq(19) can be re-
formulated as

c
A = tr{YQi Li (Qi )T (Yi )T }
i=1 (20)
= tr{YLExem Y T },

where LExem = ci=1 Qi Li (Qi )T can be called the within-exemplar Laplacian matrix.
Plugging the expression of Y = W T X into Eq(20), we obtain the final form of the
total within-exemplar scatter:
A = tr(W T DExem W) (21)
where DExem = XLExem X is the within-exemplar scatter matrix.
T
 Laplacian Discriminant Projection Based on Affinity Propagation 317

3.2 Discriminant Within-Class Scatter

Besides the within-exemplar scatter, APLDP also takes the within-class scatter into
account. The with-class scatter of i − th class, ηi , is defined as follows:


ki
ηi = αis yis − yi 2 . (22)
s=1

By the similar deduction, ηi can be formulated as

ηi = tr{Yi Lηi YiT }, (23)
where
2 (eni )T Aηs i eni
Lηi = Aηi − Aηi eni (eni )T + eni (eni )T . (24)
ni n2i
The total within-class scatter of all classes:


c
Y= ηi = tr{Yi Lηi YiT } = tr(W T Dw W), (25)
i=1

where Dw = XLηi X T is the within class scatter matrix.

3.3 Discriminant Between-Class Scatter

Let B denote the exemplar based between-class scatter of all classes. B is defined as


c

cs
B= αij yij − y2 , (26)
j=1 i=1

where yij implies its corresponding original i − th exemplar center of j − th class of X j .

αij is the weight, defined as

xij − x2
αij = exp(− ). (27)
t
Let Y = [yk11 , . . . , yk11 , . . . , ykcc , . . . , ykcc ] consists of exemplar center vectors of all classes.
By the similar deduction, B can be formulated as follows
T
B = tr{Y Lb Y }, (28)
where
2 eTexem Ab eexem
Lb = Ab − Ab eexem eTexem + eexem eTexem (29)
nexem n2exem
is the exemplar based between-class Laplacian matrix.
Taking Y = W T X into account, we re-write Eq.(28) as follows
B = tr(W T DB W). (30)
T
where DB = XLb X is the total exemplar based between-class scatter matrix.
318 X. Chang and Z. Zheng

3.4 Discriminant Projection

We construct the following Fisher criterion
B tr(W T DB W)
f (W) = arg max = . (31)
w A + Y tr(W T (DExem + DW )W)
Let DEW = DExem + DW , then
B
f (W) = arg max (32)
w DEW
To solve the above optimization problem, we take the similar approach used in the
traditional LDA. We take the d eigenvectors derived from the following generalized
eigenvalue analysis
DB wi = λi DEW wi (33)
that are associated with the d largest eigenvalues λi , i = 1, . . . , d.

4 Experiments
In this section, we investigate the use of APLDP on several data sets including UCI1
and PIE-CMU face data set [13]. The data sets used in the paper belong to different
fields in order to test the performance of APLDP algorithm. We compare our proposed
algorithm with PCA[2], LDA[7], LPP[3] and Marginal Fisher Analysis (MFA)[1].

4.1 On UCI Data Set

In this experiment, we perform on iris data taken from the UCI Machine Learning
Repository. There are 150 samples of 3 classes (50 samples per class) in iris data set.
We randomly select 20 samples per class for training and the remaining samples for
testing. The average results are obtained over 50 random splits. All algorithms reduce
the original samples to 2-dimensional space. The classification is based on k-nearest
neighbor classifier. The experimental results are shown in Table1. In terms of APLDP
algorithm, there are several parameters which should be set before the experiments. In
this experiment, the number of clusters in each class is set k = 2, and the time variable
t = 10.
Table 1. Recognition accuracy of different algorithms

Algorithm PCA LDA LPP MFA APLDP

Accuracy 95.112 95.391 95.391 95.383 95.913

To demonstrate the performance of APLDP algorithm, we randomly select one split

from the 50 splits. The embedding results of APLDP in 2D space, together with other
four algorithms, are shown in Fig.1. As illustrated in Table1, APLDP algorithm outper-
forms other methods with recognition rate of 95.913%. We can find that the within-class
embedding result of APLDP is more compact than those of others four methods, as il-
lustrated in Fig.1.
1
Available at http://www.ics.uci.edu/ mlearn/MLRepository.html
 Laplacian Discriminant Projection Based on Affinity Propagation 319

3 3 3 3 3 3

2 2 2 2 2 2

1 1 1 1 1 1

0 0 0 0 0 0

−1 −1 −1 −1 −1 −1

−2 −2 −2 −2 −2 −2

−3 −3 −3 −3 −3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(a) PCA (b) LDA (c) LPP (d) MFA (e) APLDP1 (f) APLDP2

Fig. 1. Embedding results in 2 − D space of PCA, LDA, LPP, MFA and APLDP1,2(with and
without AP)

4.2 On PIE-CMU Face Data Set

The PIE-CMU face data set consists of 68 subjects with 41, 368 face images [13]. In
this experiment, we select 40 subjects with 120 face images for each from CMU data
set, 60 images for training, and the other 60 images for testing. Before the experiment,
faces in the images are detected by the face detection system described in [17][18]. The
detected faces are converted to gray scale images and resized to 32 × 32. Some samples
are shown in Fig.2. Totally, there are 2, 400 images in the training set and the testing
set, respectively.
It should be mentioned that we take PCA as preprocessing step for LDP. The number
of principal components is a free parameter to choose. As pointed out in [19,20], the
dimension of principal subspaces significantly affects the performance of recognition
tasks. Besides, they confirmed that the optimal number lies in the interval [50, 200].

Fig. 2. Some samples of CMU-PIE face data set

Fig. 3. From the top row to the bottom row, the face-like images are Eigenfaces, Fisherfaces,
Laplacianfaces, MFAfaces and APLDPfaces, respectively

Table 2. Best recognition accuracy of different algorithms

Algorithm PCA LDA LPP MFA APLDP

Dims 180 39 91 58 112
Accuracy 69.786 79.510 79.533 83.328 84.04
320 X. Chang and Z. Zheng

Based on their work, we find the best dimension of PCA is 182. Therefore, we take 182
as the number of principal components in the following experiments.
For the sake of visualization, we illustrate algorithmic-faces derived from different
algorithms, such as Eigenfaces from PCA, Fisherfaces from LDA and Laplacianfaces
from LPP, in Fig.3. The special face-like images derived from MFA and APLDP can be
called MFAfaces and ELDPfaces, respectively.
The average results with corresponding reduced dimensions are obtained over 50
random splits. The classification is also based on k-nearest neighbor classifier. The ex-
perimental results are shown in Table2. In the experiment, the parameters of APLDP
algorithm are set as k = 3, and the time variable t = 10.

5 Conclusions
In this paper, based on affinity propagation and LDA algorithms, we propose a new
method, exemplar based Laplacian Discriminant Projection(APLDP) for supervised di-
mensionality reduction. Using similarity weighted discriminant criterions, we define the
exemplar based within-class Laplacian matrix and between-class Laplacian matrix. In
comparison with the traditional LDA, APLDP focuses more on the enhancement of the
discriminability while keeping local structures. Therefore, APLDP has the flexibility of
finding optimal discriminant subspaces.

Acknowledgement
The authors confirm that the research was supported by National Science Foundation
of China (No.60805001).

References
1. Yan, S., Xu, D., Zhang, B., Zhang, H., Yang, Q., Lin, S.: Graph Embedding and Extension: A
General Framework for Dimensionality Reduction. IEEE Transactions on Pattern Analysis
and Machine Intelligence 29(1), 40–51 (2007)
2. Turk, M., Pentland, A.: Eigenfaces for recognition. Journal of Cognitive Neuroscience 3(1),
71–86 (1991)
3. He, X., Yan, S., Hu, Y.X., Niyogi, P., Zhang, H.: Face recognition using Laplacianfaces.
IEEE Transactions on Pattern Analysis and Machine Intelligence 27(3), 328–340 (2005)
4. Belkin, M., Niyogi, P.: Laplacian Eigenmaps for dimensionality reduction and data represen-
tation. Neural Computation 15, 1373–1396 (2003)
5. Zhao, D., Lin, Z., Xiao, R., Tang, X.: Linear Laplacian Discrimination for Feature Extraction.
In: CVPR (2007)
6. Fukunaga, K.: Introduction to Statistical Pattern Recognition, 2nd edn. Academic Press,
Boston (1990)
7. Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs. fisherfaces: Recognition
using class specific linear projection. IEEE Transactions on Pattern Analysis and Machine
Intelligence 19(7), 711–720 (1997)
8. Weinberger, K., Blitzer, J., Saul, L.: Distance metric learning for large margin nearest neigh-
bor classification. In: NIPS, pp. 1475–1482 (2006)
 Laplacian Discriminant Projection Based on Affinity Propagation 321

9. Nie, F., Xiang, S., Zhang, C.: Neighborhood MinMax Projections. In: IJCAI, pp. 993–998
(2007)
10. Howland, P., Park, H.: Generalizing discriminant analysis using the generalized singular
value decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(8),
995–1006 (2004)
11. Liu, C.: Capitalize on dimensionality increasing techniques for improving face recognition
grand challenge performance. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 28(5), 725–737 (2007)
12. Martinez, A., Zhu, M.: Where are linear feature extraction methods applicable. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 27(12), 1934–1944 (2006)
13. Sim, T., Baker, S., Bsat, M.: The CMU Pose, illumination, and expression (PIE) database.
In: IEEE International Conference of Automatic Face and Gesture Recognition (2002)
14. Wang, X., Tang, X.: Dual-space linear discriminant analysis for face recognition. In: CVPR,
pp. 564–569 (2004)
15. Yan, S., Xu, D., Zhang, B., Zhang, H.: Graph embedding: A general framework for dimen-
sionality reduction. In: CVPR (2005)
16. Yang, J., Frangi, A., Yang, J., Zhang, D., Jin, Z.: KPCA plus LDA: a complete kernel Fisher
discriminant framework for feature extraction and recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence 27(2), 230–244 (2005)
17. Zheng, Z.L., Yang, J., Zhu, Y.: Face detection and recognition using colour sequential im-
ages. Journal of Research and Practice in Information Technology 38(2), 135–149 (2006)
18. Zheng, Z.L., Yang, J.: Supervised Locality Pursuit Embedding for Pattern Classification.
Image and Vision Computing 24, 819–826 (2006)
19. Wang, X., Tang, X.: A unified framework for subspace face recognition. IEEE Transactions
on Pattern Analysis and Machine Intelligence 26(9), 1222–1228 (2004)
20. Wang, X., Tang, X.: Random sampling for subspace face recognition. International Journal
of Computer Vision 70(1), 91–104 (2006)
21. Frey, B.J., Dueck, D.: Clustering by Passing Messages Between Data Points. Science 315,
972–994 (2007)
 An Improved Fast ICA Algorithm for IR Objects
Recognition

Jin Liu and Hong Bing Ji

School of Electronics Engineering, Xidian Univ., Xi’an, China

jinliu@xidian.edu.cn

Abstract. In this paper, an improved algorithm based on fast ICA and optimum
selection for IR objects recognition is proposed. Directed against the problem
that the Newton iteration is rather sensitive to the selection of initial value, this
paper presents a one dimension search to improve its optimum learning algo-
rithm in order to make the convergence of the results independent of the selec-
tion of the initial value. Meanwhile, we design a novel rule for the distance
function to retain the features of the independent component having major
contribution to object recognition. It overcomes the problem of declining of
recognition rate and robustness associated with the increasing of training image
samples. Compared with traditional methods the proposed algorithm can reach
a higher recognition rate with fewer IR objects features and is more robust in
different kinds of classes.

Keywords: IR objects recognition, fast ICA, one dimension search, distance

function, robustness.

1 Introduction
Objects recognition is one of the key techniques in the field of pattern recognition.
The key of objects recognition is feature extraction and dimension reduction. Directed
against the lower recognition rate of IR image due to the characteristics of IR image
itself, it is necessary to look for a new method for image information processing and
feature extraction which can remove redundant information in the object image data
and presents object image using feature vector with invariant features. Several charac-
teristic subspace methods have been developed in the field of objects recognition,
such as Principal Component Analysis (PCA) characteristic subspace algorithm[1][2],
Independent Component Analysis (ICA) characteristic subspace algorithm[3][4] and
so on. ICA is different from the PCA in that it does not focus on signals’ second-order
statistical correlation. Instead it is based on the higher-order statistics of signals, i.e., it
studies the independent relationship between signals. Therefore ICA method can
reveal the essential structure between image data. Hyvarinen A. [5] [6] etc proposed a
fast independent component analysis (Fast ICA) algorithm which is a fast iterative
optimization algorithm with fast convergence rate and exempts the need to determine
learning step. In recent years, fast ICA has obtained universal attention and has been
widely used in feature extraction [7], blind source separation [8], speech signal proc-
essing [9], target detection [10] and face recognition [11] [12], etc.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 322–329, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 An Improved Fast ICA Algorithm for IR Objects Recognition 323

Directed against the problem that the Newton iteration in the fast ICA algorithm is
rather sensitive to the selection of initial value and the problem of declining in recogni-
tion rate and robustness with the increasing of training image samples, this paper pre-
sents an improved fast ICA and features optimization algorithm which can be applied
to IR multi-object recognition. The paper is organized as follows. Section 2 introduces
ICA theory and fast fixed-point algorithm for ICA. Section 3 IR describes multi-object
classification recognition based on improved fast ICA and a feature optimization algo-
rithm and gives the rule for feature optimization. Section 4 gives experimental results
produced by using improved fast ICA algorithm and conducts performance evaluation.
Section 5 is conclusion.

2 Related Works

2.1 Review of ICA Theory

Suppose {x1 , x2 ,L , xn } are n observed random signals formed by linear mixing of

m unknown independent sources {s1 , s2 , L , sm } . X = [ x1 , x2 ,L , xn ]T is the matrix
representation of the observed signals. S = [ s1 , s2 ,L , sm ]T is the independent
source signals. Assuming the observed random signals and independent source signals
are zero mean, we have
m
xi = ∑ aij s j (1)
j =1

where i = 1, 2, L , n. j = 1, 2, L , m .
The linear model of ICA can be expressed as
m
X = AS = ∑ a j s j (2)
j =1

where A = [a1 , a2 ,L , am ] is a full rank n × m mixed matrix, a j ( j = 1, 2,L , m)

is the base vector of the mixed matrix. The observed signal X can be viewed as a linear
combination of the column vector a j of the mixing matrix A, and the corresponding
weights are just the independent source signals s j ( j = 1, 2, L , m) .

2.2 Fast and Fixed-Point Algorithm for ICA

For a random variable x, its entropy is defined as

H ( x) = − ∫ f ( x) log f ( x)dx (3)

where f ( x) is the probability density distribution function of the random variable x.

324 J. Liu and H.B. Ji

The negative entropy of a random variable x is defined as

J ( x) = H ( xgauss ) − H ( x) (4)

where xgauss is a random variable having Gaussian distribution and the same variance
with x.
From (4), we can see that the value of negative entropy is always non-negative, and
only when the random variable x is of Gaussian distribution, the negative entropy is
zero. The stronger the non-Gaussian random of variable x, the bigger J ( x ) is. The
more effective approximate negative entropy formula is

J ( wi ) ≈ [ E{G ( wiT X )} − E{G (v)}] (5)

where G (•) are some non-quadratic functions, and v is a random variable with stan-
dard normal distribution.
By finding the projection direction wi which makes J ( wi ) maximum, an inde-
pendent component can be extracted. If J ( wi ) takes the maximum value,
E{G ( wiT X )} will also have the maximum value, and the extreme point wi of
E{G ( wiT X )} is the solution of equation

E{ Xg ( wiT X )} = 0 (6)

where g (•) is the derivative of G (•) . According to Newton iteration theorem, we

have

E{ Xg ( wiT X )} E{ Xg ( wiT X )}
wi+ = wi − ≈ wi − (7)
E{ X T Xg '( wiT X )} E{g '( wiT X )}

where E{ X T Xg '( wiT X )} ≈ E{ X T X }E{g '( wiT X )} = E{g '( wiT X )} . Multiply the two sides
of (7) with − E { g '( w iT X ) } , and let w i∗ = − E { g '( w iT X ) } w i+ , we
have

⎧⎪ wi∗ = E{ Xg ( wiT X )} − E{g '( wiT X )}wi

⎨ (8)
⎪⎩ wi = wi∗ wi∗

T
By conducting iteration according to (8), the wi obtained through convergence corre-
sponds to the row vectors of the separation matrix W, and then an independent com-
ponent si can be extracted.
 An Improved Fast ICA Algorithm for IR Objects Recognition 325

3 IR Objects Recognition Based on Improved Fast ICA and

Features Optimization

3.1 Improved Iteration Learning Algorithm Based on One Dimension Search

Fast ICA algorithm utilized the principle of Newton iteration. However, since the
direction of Newton iteration is not necessarily in the descent direction, when the
initial value is far from the minimum point, a situation may arise that the object func-
tion obtained by iteration cannot reach the best point or even not convergent. Targeted
at this problem, this paper proposes a strategy that adds a one-dimensional search
along the direction of Newton iteration, and designs the new iteration formula

⎧ x ( k +1) = x ( k ) + λk d ( k )
⎪⎪ ( k ) ( k ) −1
⎨d = −∇ f ( x ) ∇f ( x )
2 (k )
(9)
⎪ f ( x ( k ) + λ d ( k ) ) = min f ( x ( k ) + λ d ( k ) )
⎪⎩ k
λ

where λ is the damping factor, ∇ is the first order partial derivative of the function,
∇ 2 is the second order partial derivative of the function. The optimum solution of the
function can be obtained by iteration.
In the case of wi = 1 , it is transformed into an unconstrained extreme value
problem, and the new cost function is

F ( wi , c) = E{G ( wiT X )} + c( wi − 1) (10)

where c is referred to as cost factor.

Therefore, the damping iteration of transformation matrix w is modified to

E{ Xg ( wiT X )} − cwi
wi+ = wi − λi (11)
E{g '( wiT X )} − cI

Multiply both sides of (11) by E{g '( wi

T
X )} − cI , let wi∗ = wi+ [ E{g '( wiT X )} − cI ] ,
and normalize the obtained parameters, we get the new iteration formula

⎧⎪ wi∗ = λi [ E{ Xg ( wiT X )}] − E{g '( wiT X )}wi − c(λi − I ) wi

⎨ (12)
⎪⎩ wi = wi∗ wi∗

where λi is the one-dimensional search factor, which makes the cost function enter
the convergence region of Newton iteration beginning with a certain wi in a certain
range, and ensures the convergence of the algorithm. When λi = 1 , it becomes the
original fast ICA algorithm.
326 J. Liu and H.B. Ji

3.2 Feature Optimization for IR Object Classification Recognition

The optimal selection of features is a key issue that affects the correctness of object
classification and recognition. In feature subspace, the concentrated the feature vectors
of the same class and the more far apart the feature vectors of different classes, the
easier the classification and recognition. Therefore, when extracting features from IR
object images, it is required to extract features that are far different between different
classes of objects and are almost the same for one class of objects. With the increasing
of the number of observation images, the number of independent components and the
corresponding features are also increasing. Among these added features, some will
have less contribution to classification recognition due to similarity to existing fea-
tures; others may even reduce recognition rate due to feature distortion. In order to
ensure correctness and efficiency of classification and recognition, we select a mini-
mum number of features from the original features that are most effective in classifica-
tion. For this purpose, we propose a new feature selection criteria based on intra-class
and inter-class distance function.
Suppose observation image matrix F is composed of M classes and there
are N i (i = 1, 2,L , M ) observations for each class. Column vector A j of mixed
matrix A corresponds to independent component s j , and aij is the jth feature of the
ith observation. U j is defined as the mean value of intra-class distance of the jth
feature in M classes. This can be expressed as

⎧ 1 M Ni Ni

⎪U j = ∑∑∑ aq + s , j − aqc + p, j
MN ( N − 1) c =1 s =1 p =1 c
⎪
⎨ p≠s (13)
⎪ c −1

⎪ qc = ∑ N i , where N 0 = 0; i = 1, 2, L, M
⎩ i =0

Similarly, V j is called the mean value of inter-class distance of the jth feature in M
classes, and can be expressed as

⎧ 1 M M
1 Nl 1 Nk
⎪ j
V = ∑∑ ∑ q + s, j N ∑a
M ( M − 1) l =1 k =1 Nl s =1 l
− aqk + p , j
⎪ k p =1
⎨ k ≠l (14)
⎪ l −1 k −1

⎪ql = ∑ Ni , qk = ∑ Ni , where N 0 = 0; i = 1, 2,L , M

⎩ i =0 i =0

Effective classification factor βj is defined as

Uj
βj = (15)
Vj
 An Improved Fast ICA Algorithm for IR Objects Recognition 327

which reflects the difficulty degree of classification of M classes by the jth feature.
The smaller β j is, the more sensitive the jth feature is, and the more easy and effec-
tive the classification of the M classes.
We also have the relationship in practical situations
0 < U j < Vj (16)
In accordance with the definition of effective classification factor, we can further
obtain
0 < β j <1 (17)

We choose the independent components corresponding to βj satisfying (17) to form

the optimal feature subspace, and carry out object recognition and classification in
this optimal characteristic subspace by minimum distance classifier.

4 Experiments and Analysis

Multi-object classification/recognition experiments have been made by using real
infrared object image database. This database contains nine classes of infrared ob-
jects, and the images for each class were taken each time when the targeted object has
rotated 10 degrees in front of the IR camera. So there are 36 photos for each class
(324 IR object images for 9 classes). The training set consists of 162 images (18 pho-
tos for each class), and image size is 200 × 100. Particularly, the 18 photos for each
class were taken each time when the object has rotated 20 degrees, such as at 5 de-
grees, 25 degrees, 45 degrees etc., and 9 photos are chosen randomly to add Gaussian
noise to. So the testing set is made up of 162 frames of images from 9 classes. The 18
images for each class were taken each time when the object has rotated 20 degrees,
such as at 15 degrees, 35 degrees, 55 degrees etc., and 9 photos are chosen randomly
to add Gaussian noise to. Part of the infrared object images in the training set are
shown in Fig. 1.

(a) hatchback (b) microbus (c) hummer

(d) armored car (e) jeep (f) helicopter

(g) tank (h) sedan (i) rocket bike

Fig. 1. Part of the IR object images in the training set

The experiment verifies the feasibility and effectiveness of the algorithm through
IR database built by ourselves. We use PCA feature extraction algorithm (Traditional
PCA) in [2], Fast ICA feature extraction algorithm (Traditional FastICA) in [6] and
our proposed object recognition algorithm based on improved fast ICA and feature
328 J. Liu and H.B. Ji

(a) Classification rate in different dimension (b) Classification rate in different classes

Fig. 2. Performance comparison with different algorithms

optimization (Improved FastICA) to conduct feature extraction of independent com-

ponents, and use minimum distance classifier (Mindist) in optimal feature subspace
for classification. The respective ratio of correct classification of the above three
methods is tested by using different classes. In order to test the performance of the
algorithms more accurately, the experimental results are obtained by running the
algorithms 10 times and using independent data sets. Each algorithm is measured by
correctness rate of classification (correct classification sample numbers divided by all
tested sample numbers). Experimental results are shown in Fig. 2.
In Fig.2.(a) we can see that the classification rate of the proposed Improved FastICA
+ Mindist algorithm can reach 75% under the condition when the dimension of charac-
teristics is 20, meanwhile the Traditional FastICA + Mindist and Traditional PCA +
Mindist algorithm is 60% and 70% respectively. When the dimension is 40, the recogni-
tion rate of our proposed Improved algorithm achieves maximum 81%, and with the
increasing of dimensions the classification rate has a little reduction, which shows that
some characteristics are not helpful to the classification; the recognition rate of Tradi-
tional FastICA + Mindist algorithm is only 66%; the recognition rate of Traditional
PCA + Mindist algorithm remains unchanged in 71%. Because of the incompleteness
and uncertainty of infrared images information, the improved method has more clear
superiority than the other two methods. It just retains the characteristic components
which are helpful to the classification and recognition, so the recognition rate is im-
proved obviously. In Fig.2.(b) we know that the proposed improved algorithm obtains
more robustness in terms of recognition rate compared with the other two methods.
The above two groups of experiment results show that our proposed object rec-
ognition algorithm based on improved fast ICA and feature optimization can obtain
a higher recognition rate with fewer features and dimensions of subspace. Our
proposed algorithm also has the advantage of great robustness in different classes,
so it lays a foundation for the engineering application of the recognition algorithm.

5 Conclusions
In this paper, we develop an IR object recognition technique employing an improved
fast ICA and feature optimization algorithm. Targeted at the problem that the Newton
 An Improved Fast ICA Algorithm for IR Objects Recognition 329

iteration in the fast ICA algorithm is sensitive to the selection of initial value, one
dimension search is imposed on the direction of Newton iterative in order to ensure
the convergence of the results and the robustness to initialization. Meanwhile, a novel
rule based on distance function is designed to select the optimal features favorable for
recognition according to the characteristics of infrared image. It overcomes the prob-
lem of declining in recognition rate and robustness along with the incensement of the
number of training image samples. Experimental comparison on IR object database
under similar conditions shows that our proposed method can obtain highest recogni-
tion rate. Theoretical derivation and experimental results show that the improved fast
ICA features optimization algorithm can improve recognition accuracy and effi-
ciency. In the future, we aim to further improve the recognition rate in combination
with other feature information.

Acknowledgement. This paper is supported by the National Natural Science

Foundation of China (NSFC) under Grant Number 60677040.

References
1. Oja, E.: Neural Networks, Principle Components and Subspaces. International Journal of
Neural Systems 1(1), 61–68 (1989)
2. Jolliffe, I.T.: Principal Component Analysis, 2nd edn. American Statistical Association,
American (2003)
3. Hyvarinen, A., Karhunen, J., Oja, E.: Independent Component Analysis. Wiley, New York
(2001)
4. Karhunen, J., Hyvarinen, A., Vigario, R., et al.: Applications of Neural Blind Separation to
Signal and Image Processing. In: IEEE International Conference on ICASSP, vol. 1(1), pp.
131–134 (1997)
5. Hyvarinen, A.: Blind Source Separation by Nonstationarity of Variance: a Cumulant-based
Approach. IEEE Transactions on Neural Networks 12(6), 1471–1474 (2001)
6. Hyvarinen, A.: Fast and Robust Fixed-point Algorithms for Independent Component
Analysis. IEEE Transactions on Neural Networks 10(3), 626–634 (1999)
7. Jouan, A.: FastICA (MNF) for Feature Generation in Hyperspectral Imagery. In: 10th In-
ternational Conference on Information Fusion, pp. 1–8 (2007)
8. Shyu, K.-K., Lee, M.-H., Wu, Y.-T., et al.: Implementation of Pipelined FastICA on FPGA
for Real-Time Blind Source Separation. IEEE Transactions on Neural Networks 19(6),
958–970 (2008)
9. Shen, H., Huper, K.: Generalised Fastica for Independent Subspace Analysis. In: IEEE In-
ternational Conference on Acoustics, Speech and Signal Processing, vol. 4, pp. 1409–1412
(2007)
10. Ming-xiang, W., Yu-long, M.: A New Method for The Detection of Moving Target Based
on Fast Independent Component Analysis. Computer Engineering 30(3), 58–60 (2004)
11. Mu-chun, Z.: Face Recognition Based on FastICA and RBF Neural Networks. In: 2008 In-
ternational Symposium on Information Science and Engineering, vol. 1, pp. 588–592
(2008)
12. Li, M., Wu, F., Liu, X.: Face Recognition Based on WT, FastICA and RBF Neural Net-
work. In: Third International Conference on Natural Computation, vol. 2, pp. 3–7 (2007)
 Facial Feature Extraction Based on Wavelet Transform

Nguyen Viet Hung

Department of Mathematics and Computer Science,

Ho Chi Minh City, University of Pedagogy, Viet Nam
nvhung@math.hcmup.edu.vn

Abstract. Facial feature extraction is one of the most important processes in

face recognition, expression recognition and face detection. The aims of facial
feature extraction are eye location, shape of eyes, eye brow, mouth, head
boundary, face boundary, chin and so on. The purpose of this paper is to
develop an automatic facial feature extraction system, which is able to identify
the eye location, the detailed shape of eyes and mouth, chin and inner boundary
from facial images. This system not only extracts the location information of
the eyes, but also estimates four important points in each eye, which helps us to
rebuild the eye shape. To model mouth shape, mouth extraction gives us both
mouth location and two corners of mouth, top and bottom lips. From inner
boundary we obtain and chin, we have face boundary. Based on wavelet
features, we can reduce the noise from the input image and detect edge
information. In order to extract eyes, mouth, inner boundary, we combine
wavelet features and facial character to design these algorithms for finding
midpoint, eye’s coordinates, four important eye’s points, mouth’s coordinates,
four important mouth’s points, chin coordinate and then inner boundary. The
developed system is tested on Yale Faces and Pedagogy student’s faces.

Keywords: Computer vision, face recognition, wavelet transform, facial feature

extraction.

1 Introduction
In recent years, computer vision has rain-storm development. Computer vision con-
cerns with developing systems that can interpret the content of natural scenes [6, 7].
Like the aims of all applications, to make computers more user-friendly and to in-
crease computer ability to identify learn and gain knowledge, computer vision sys-
tems begin with the simple process of detecting objects and locating objects, continue
with the higher process of detecting and locating features to gain more information,
and then proceed to the most advanced process of extracting meaningful information,
to apply to intelligent systems.
Face recognition and face expression are one part of computer vision. To recognize
one person, we must have three main processes. One of the main processes is facial
feature extraction. In a computer vision society, a feature is defined as a function of
one or more measurements, the values of some quantifiable property of an object,
computed so that it quantifies some significant characteristics of the object [6,7]. One

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 330–339, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Facial Feature Extraction Based on Wavelet Transform 331

of the biggest advantages of feature extraction lies in that it significantly reduces the
information (compared to the original image) to represent an image for understanding
the content of the image [12]. Basically, the extraction of facial feature points (eyes,
mouth, nose, chin, inner boundary) plays an important role in various applications
such as face detection, face recognition, model based image coding, and expression
recognition, and has become a popular area of research due to emerging applications
in human-computer interface, surveillance systems, secure access control, video con-
ferencing, financial transaction, forensic applications, pedestrian detection, driver
alertness monitoring systems, image database management system and so on.
In space problem of facial feature extraction, we have some obstacles given as fol-
low: input images are so blur and have a lot of noise, face sizes and orientations.
Sometimes facial features may be covered by other things, such as a hat, a pair of
glasses, a hand, etc.
Many approaches to facial feature extraction have been reported in literature over
the past few decades, ranging from the geometrical description of salient facial fea-
tures to the expansion of digitized images of the face on appropriate basis of images
[8]. Different techniques have been introduced recently, for example, principal com-
ponent analysis [1], geometric modeling [2], auto-correlation [13], deformable tem-
plate [14], neural networks [15], elastic bunch graph matching [16], color analysis
[17] and so on. Lam and Yan [3] have used snake model for detecting face boundary.
Although the snake provides good results in boundary detection, the main problem is
to find the initial position [4].
We can divide these approaches to four main parts: geometry-based, template-
based, color segmentation, appearance-based approach. Generally geometry-based
approaches extract features using geometric information such as relative positions
and sizes of the face components. Template-based approaches match facial compo-
nents to previously designed templates using appropriate energy functional. The
best match of a template in the facial image will yield the minimum energy. Color
segmentation techniques make use of skin color to isolate the face. Any non-skin
color region within the face is viewed as a candidate for eyes and mouth. The per-
formance of such techniques on facial image databases is rather limited, due to the
diversity of ethnical backgrounds [18]. In appearance-based approaches, the con-
cept of “feature” differs from simple facial features such as eyes and mouth. Any
extracted characteristic from the image is referred to as a feature [5]. Methods such
as principal component analysis (PCA), independent component analysis, and Ga-
bor wavelets [21] are used to extract the feature vector.
This paper explores facial feature extraction such as eyes, mouth, chin, and inner
boundary. Facial feature extraction is established by two approaches: (i) edge detec-
tion based on wavelet transform and (ii) geometry property. Based on wavelet prop-
erty, we can make the edge of an image bolder and reduce the noise from that image.
Appling geometry property of human face, eyes, and mouth, we give some thresholds.
Experimental results indicate that the algorithm can be applied to image having
noises, human having beard and human expression.
332 N.V. Hung

2 Background Wavelets and Edge Detection Based on Wavelets

The concept of wavelet analysis has been developed since the 1980s. It is a relatively
recent development of applied mathematics. Independent from its developments in
harmonic analysis, the wavelet transform has been studied in its continuous form and
initially applied to analyze geological data by A. Grossmann, J.Morlet and their co-
workers. However, at that time, the roughness of wavelets made mathematics suspect
the existence of a “good” wavelet basic until two great events took place in 1988,
namely:
Daubechies, a female mathematician, constructed a class of wavelet bases, which
are smooth, compactly supported and orthonormal. They are referred to as Daube-
chies bases, and are successfully applied to many fields today.
French signal analysis expert, S.Mallat, with mathematician, Y.Meyer, proposed a
general method to construct wavelet bases. It’s termed multiresolution analysis
(MRA) and is intrinsically consistent with sub-band coding in signal analysis [9].
In 1985, Y.Meyer [22,23], discovered that one could obtain orthonormal bases for
L2(R) of the type:
ψ j ,k ( x) = 2 j /2ψ (2 j x − k), j,k ∈ Z . (1)
and that the expression
f = ∑
j ,k∈Z
f ,ψ j ,k ψ j ,k . (2)
for decomposing a function into these orthonormal wavelet converged in many func-
tion spaces.
In 1987, Stéphane Mallat [22,23] made a giant step in the theory of wavelets when
he proposed a fast algorithm for the computation of wavelet coefficients. He proposed
the pyramidal schemes that decomposed signals into subbands. These algorithms and
theory of wavelets is the concept of multi-resolution analysis (MRA) [10].

2.1 Wavelets

Wave is an oscillation function in the time domain or frequency domain. What is a

wavelet? The simplest answer is a “short” wave (wave + let =>wavelet). The suffix
“let”, in the mathematical term “wavelet”, comes from “short”, which indicates that
the duration of function is very limited. In other words, the wavelet is a special signal,
in which, two conditions have to be satisfied:
1. First, the wavelet must be oscillatory (wave).
2. Second, its amplitudes are nonzero only during a short interval (short).
A signal or function f(t) can be analyzed, described or processed if we give linear
analysis as follows
f (t ) = ∑ alψ l (t ). (3)
l

where l is infinite or finite summary, al is real coefficient of expansion and ψ l (t ) is real

function, an expansion. If (1) is unique, the set of ψ l (t ) is called base of function class.
 Facial Feature Extraction Based on Wavelet Transform 333

If base is orthogonal, i.e.

ψ k ( t ) ,ψ l ( t ) = ∫ψ k ( t )ψ l ( t ) dt = 0, ( k ≠ l ). (4)
then coefficient can be calculated by product
ak = f ( t ) ,ψ k ( t ) = ∫ f ( t )ψ k ( t ) dt. (5)

In Fourier series, the orthogonal functions,ψ l (t ) , are sin(kw0t) and cos(kw0t)

In Wavelets, we have
f ( t ) = ∑∑ aj ,kψ j ,k ( t ) . (6)
k l

where

( )
j
ψ j ,k (t ) = 2 2
ψ 2 j t − k , k, j ∈ Z. (7)

In general, g(t) ∈ L2 (R) can be written as follows:

∞ ∞ ∞
g(t ) = ∑ c ( k )ϕ ( t ) + ∑ ∑ d ( j , k )ψ ( t ).
k j ,k (8)
k = −∞ j = 0 k = −∞

where
c ( k ) = c0 ( k ) = g ( t ) ,ϕk ( t ) = ∫ g ( t )ϕk ( t ) dt . (9)

d ( j , k) = d j ( k) = g( t ) ,ψ j ,k ( t ) = ∫ g( t )ψ j ,k ( t ) dt. (10)

ϕk ( t ) ,ψ j ,k ( t ) are scaling function, and wavelet function, respectively.

( ) ( )
j j
ψ j ,k ( t ) = 22ψ 2 j t − k , ϕ j ,k ( t ) = 22 ϕ 2 j t − k . (11)

2.2 Edge Detection Based on Wavelets

The edge of an image is a boundary which the brightness of image changes abruptly.
In image processing, an edge is often interpreted as a class of singularities. In a func-
tion, singularities can be presented easily as discontinuities where the gradient ap-
proaches infinity. However, image data is discrete, so edges in an image are often
defined as the local maxima of the gradient [10].
Edge detection is a technique related to gradient operators. Because an edge is
characterized by having a gradient of large magnitude, edge detectors are approxima-
tions of gradient operators. Noise influences the accuracy of the computation of gra-
dients. Usually an edge detector is a combination of a smooth filter and a gradient
operator. First, an image is smoothed by the smooth filter and then its gradient is
computed by the gradient operator.
334 N.V. Hung

The 2-norm of the gradient is

∇f = f 2x + f 2y . (12)
2

In oder to simplify computation in image processing, we often use the 1-norm instead.

∇f 2
= fx + fy . (13)

Their discrete forms are

( f (i − 1, j ) − f (i + 1, j ) ) + ( f (i , j − 1) − f (i , j + 1) )
2 2
∇f = . (14)
2

and
∇f 1
=
1
2
(
f (i − 1, j ) − f (i + 1, j ) + f (i , j − 1) − f (i , j + 1) . ) (15)

1. Local maxima definition: Let f (x, y) ∈H2 ( Ω) . A point ( x, y) ∈ Ω0 is called an

edge point of the image f(x,y), if ∇f has a local maximum at ( x, y)
i.e. ∇f ( x, y) > ∇f ( x, y) in a neighborhood of ( x, y) . An edge curve in an image
is a continuous curve on which all points are edge point. The set of all edge points
of f(x,y) is called an edge image of f(x,y).
2. Threshold definition: Let f ( x, y) ∈ H 2 ( Ω ) . Assume that

max ∇f ( x, y) = M. (16)
( x, y)∈Ω

Choose K, 0 < K < M, as the edge threshold. If ∇f ( x, y) ≥ K the ( x, y) is called an

edge point of f(x,y) [19].

Fig. 1. Lena image and Lena image with noise

 Facial Feature Extraction Based on Wavelet Transform 335

Edge detection is a very important process in image processing. Basing on this

process, we can do recognition pattern, image segmentation, scene analysis, etc. The
traditional methods to detect edge fail to deal with images having noise. Noise is
something we don’t expect in the original image.
Maxima of wavelet transform modulus proved by Mallat, Hwang, and Zhong [19,20]
can detect the location of the irregular structures. Edge of the image data f(x,y) is the
singularities of f(x,y) so they relate to local maximal modulus of wavelets.
Basing on wavelet theory, we use a filter J-steps including two-channels, lowpass
and highpass channel to filter low frequency and high frequency of the signal. After
we use analysis bank which has J steps, we separate image frequency to each parts of
frequency. Depending on goal requirement, we keep or remove parts of frequency,
then we inverse these by synthesis bank. In edge detection, we are interested in high
frequency (details of image) and omit low or medium frequency (the coarse of im-
age). Depending on frequency and J-step, we can get a different output image. For
example, if we keep many parts of high frequency, the output image will blur with
boundary, but if we keep few parts, the output image will lack information. Therefore,
the number of parts we keep and omit depends on each aims.

Fig. 2. Decomposition Steps

After analyzing an image by wavelet transform, we will omit some unnecessary

coarses. We have some solutions and predictable results of these solutions.
• If we do not omit any parts, the reconstruction image will be an original image.
• If we keep HH part and omit the others, the reconstruction image will be an image
lacking in information.
• If we keep HL part or LH part and omit the others, the reconstruction image will be
an image having a respectively thicker vertical or horizontal boundary than if we
keep LH part or HL part. And if we keep both HL and LH part, the output image
will be blurred.
• If we need more image information, we apply analysis bank to LL part to get more
information in J+1 step.

3 Algorithms for Extracting Facial Feature

The proposed facial feature extraction method consists of two main steps: preprocess,
and main process. In preprocess, we use wavelet properties to reduce noise and find-
ing space. Noise can make it difficult to find facial feature. The feature we need to
extract is always on the small part of the human face, so unnecessary information
336 N.V. Hung

must be removed in advance so as to decrease finding space, which increases accu-

racy and lessens running time.
In the main process, we have four mains steps:
Step 1 (Find eyes) In this step, we build some algorithms to get coordinates of two
eyes, coordinates of eye’s top and bottom, and coordinates of eye corners.
Step 2 (Find mouth) We have 3 algorithms to gain coordinate of mouth, and coor-
dinates of mouth corners.
Step 3 (Find chin) To obtain coordinate of bottom chin, we build one algorithm.
Step 4 (Find inner boundary) After using several algorithms, we get a set point of
inner boundary.
Step one is the most important in this paper. First, we apply MES algorithm. The
aim of this algorithm is to reduce eye space by finding 2 lines which are parallel with
vertical axis. The left line is a line closest to the left of left eye. The right line is simi-
lar. We use one algorithm to find the eye’s region. The next step, we find approximate
middle point of two eyes, left and right eyes, find top and bottom point of eyes and
then find coordinate of two eye corner. Step two includes: find coordinate of mouth,
two corners of mouth, top point of upper lip and bottom point of lower lip. Step three
has one algorithm, approximately chin coordinate. Step 4 includes algorithm of the
inner boundary.

Coordinate of two eyes,

coordinate of mouth, image
matrix, clipped by 3 lines.

Left_eyes = ptLeftEye, Right_eyes = ptRightEye, Mouth =

ptMouth
Equalize histogram of InputMatrix
Apply TransformCross.
Wavelet transform with filter D4, 1-scale, and get HL part

D = RIGHT_EYES.X – LEFT_EYE.X
Build matrix arWeight[] whose height is InputMatrix.width
Assign arWeight[i] = sum{InputMatrix(j,i), j = Left_eyes.y .. Left_eyes.y + D}.
Determine Average = sum(arWeigh[i],where i < Left_eyes.X or i >
Right_eye.x) / (intW-D)
Do normalized noise to matrix arWeight.

Find the first left point Find the first right point
cLeft is abciss of the first left point cRight is abciss of the first right point
Assign cLeft = i, the first i satisfied arWeight(i) > Assign cRight = i, the first i satisfied arWeight(i) >
Average where i = Left_eye.x –D/4 downto 0 Average where i = Right_eye.x + D/4 to intW.

d = D/(intCountDetect-1)
Buid two matrixs arLeft and arRight to save left inner poins and right inner points
Assign start = Left_eye.y + d*i;end = left_eye.y + d*(i+1)
for i = 0 to intCountDetect – 1
Apply SelectBorderSegment (matrix arW, start to end, cLeft, cRight)
This method will return the next coordinate left and right of inner boundary.
Do normalized data arLeft[] and arRight[] again.

Two matrix arLeft[] and arRight[] (left

inner boundary and right inner
boundary).

Fig. 3. Inner boundary

 Facial Feature Extraction Based on Wavelet Transform 337

Fig. 4. All features

4 Experimental Results
The facial feature extraction method was implemented in C#, Visual Studio 2007,
Windows XP and Dot.net framework 2.0.
To evaluate the accuracy in finding these facial features, we compare facial fea-
tures obtained by hand and facial feature gained by these algorithms. The mean of
1 i =n
error of point to point is calculate by this formula me = ∑ di
ns i =1
where di is an error of point to point, s is the distance between two eyes and n is the
number of features.

Fig. 5. Mean of error of point to point

The accurate scale of eyes is 93.54% with me2 = 0.0534 in 124 Yale images (front
human face) and 91.17% with me2 = 0.043 in 34 student’s images.
The accurate scale of mouth is 91.93% with me1 = 0.068 in 124 Yale images and
88.23% with me1 = 0.054 in 34 student’s images.
The accurate scale of chin is 91.93% with me1 = 0. 0017 in 124 Yale images and
88.23% with me1 = 0.0034 in 34 student’s images.
The rate between the inner boundary we obtain by these algorithms and by hand is
90.32 in 124 Yale images and 85.29% in 34 student’s images.
338 N.V. Hung

Fig. 6. Input and output of noise images

5 Conclusion and Future Work

Facial feature extraction is one of the most interesting research studies and there are many
approaches to it. In our paper, we proposed some algorithms based on wavelet transform
and geometry features to extract eye features, mouth features, chin and inner boundary.
However, this program may not always give accurate results due to several issues, includ-
ing some lossy information, concealing, and different orientations. In the future, we can
improve these algorithms so that they can work well with different orientations and build
face recognition system or facial feature analysis relying on these features.

References
1. Turk, M., Pentland, A.: Eigenfaces for Recognition. Journal of Cognitive Neurosci-
ence 3(1), 71–86 (1991)
2. Ho, Y.Y., Ling, H.: Facial Modeling from an Uncalibrated Face Image Using a Coarse-to-
Fine Genetic Algorithm. Pattern Recognition 34(8), 1015–1031 (2001)
 Facial Feature Extraction Based on Wavelet Transform 339

3. Lam, M.K., Hong, Y.: Locating and Extracting the Eye in Human Face Images. Pattern
Recognition 29(5), 771–779 (1996)
4. Bhuiyan, A., Ampornaramveth, V., Muto, S., Ueno, H.: Face Detection and Facial Feature
Localization for Human-Machine Interface. NII Journal 5 (2003)
5. Bagherian, E., Rahmat, R., Udzir, N.: Extract of Facial Feature Point. JCSNS 9(1) (Janu-
ary 2009)
6. Castleman, R.: Digital Image Processing. Prentice-Hall, Englewood Cliffs (1996)
7. Van Vliet, J.: Grey-Scale Measurements in Multi-Dimensional Digitized Images. Doctor
Thesis. Delft University Press (1993)
8. Hjelmas, E., Low, K.B.: Face detection: A Survey. Computer Vision and Image Under-
standing 83(3), 236–274 (2001)
9. Tang, Y., Yang, L., Liu, J., Ma, H.: Wavelet Theory and Its Application to Pattern Recog-
nition. World Science Publishing Co. (2000) ISBN 981-02-3819-3
10. Li, J.: A Wavelet Approach to Edge Detection. Master Thesis of Science in the Subject of
Mathematics Sam Houston State University, Huntsville, Texas (August 2003)
11. Nixon, S., Aguado, S.: Feature Extraction and Image Processing (2002) ISBN 0750650788
12. Lei, B., Hendriks, A., Reinders, M.: On Feature Extraction from Images. Technical Report
on Inventory Properties for MCCWS
13. Goudail, F., Lange, E., Iwamoto, T., Kyuma, K., Otsu, N.: Face Recognition System Using
Local Autocorrelations and Multiscale Integration. IEEE Trans. Pattern Anal. Machine In-
tell. 18(10), 1024–1028 (1996)
14. Yuille, A., Cohen, D., Hallinan, P.: Feature Extraction from Faces Using Deformable
Templates. In: Proc. IEEE Computer Soc. Conf. on computer Vision and Pattern Recogni-
tion, pp. 104–109 (1989)
15. Rowley, H., Beluga, S., Kanade, T.: Neural Network-Based Face Detection. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 20(1), 23–37 (1998)
16. Wiskott, L., Fellous, J.-M., Kruger, N., Von der Malsburg, C.: Face Recognition by Elastic
Bunch Graph Matching. IEEE Trans. Pattern Anal. & Machine Intelligence 19(7),
775–779 (1997)
17. Bhuiyan, M., Ampornaramveth, V., Muto, S., Ueno, H.: Face Detection and Facial Feature
Extraction. In: Int. Conf. on Computer and Information Technology, pp. 270–274 (2001)
18. Chang, T., Huang, T., Novak, C.: Facial Feature Extraction from Colour Images. In: Pro-
ceedings of the 12th IAPR International Conference on Pattern Recognition, vol. 2,
pp. 39–43 (1994)
19. Mallat, M., Zhong, S.: Characterization of Signals from Multiscale Edges. IEEE Trans.
Pattern Anal. Machine Intell. 14(7), 710–732 (1992)
20. Mallat, S., Hwang, W.L.: Singularity Detection and Processing with Wavelets. IEEE
Trans. Inform. Theory 38(2), 617–643 (1992)
21. Tian, Y., Kanade, T., Cohn, J.: Evaluation of Gabor Wavelet-Based Facial Action Unit
Recognition in Image Sequences of Increasing Complexity. In: Proceedings of the Fifth
IEEE International Conference on Automatic Face and Gesture Recognition, pp. 218–223
(2002)
22. Meyer, Y.: Ondelettes, fonctions splines et analyses graduées. Rapport Ceremade 8703
(1987)
23. Meyer, Y.: Ondelettes et Opérateurs. Hermann, Paris (1987)
24. The Yale database, http://cvc.yale.edu/
 Fast Iris Segmentation by Rotation Average Analysis
of Intensity-Inversed Image

Wei Li1,∗∗ and Lin-Hua Jiang2,∗,∗∗

1
Institute of Digital Lib., Zhejiang Sci-Tech University, Zhejiang, China
2
Department of BFSC, Leiden University, Leiden, The Netherlands
L.Jiang@science.leidenuniv.nl

Abstract. Iris recognition is a reliable and accurate biometric technique used in

modern personnel identification system. Segmentation of the effective iris re-
gion is the base of iris feature encoding and recognition. In this paper, a novel
method is presented for fast iris segmentation. There are two steps to finish the
iris segmentation. The first step is iris location, which is based on rotation aver-
age analysis of intensity-inversed image and non-linear circular regression. The
second step is eyelid detection. A new method to detect the eyelids utilizing a
simplified mathematical model of arc with three free parameters is implemented
for quick fitting. Comparatively, the conventional model with four parameters
is less optimal. Experiments were carried out on both self-collected images and
CASIA database. The results show that our method is fast and robust in seg-
menting the effective iris region with high tolerance of noise and scaling.

Keywords: Biometrics, Iris segmentation, Iris location, Boundary recognition,

Edge detection.

1 Introduction

Iris recognition is known as one of the most reliable and accurate biometric tech-
niques in recent years [1-5]. Iris, suspending between the cornea and lens and perfo-
rated by the pupil, is a round contractile membrane of the eye (A typical iris image
and related medical terminology are shown in Figure 1a). As a valuable candidate
which could be used in highly reliable identification systems, the iris owns many
advantages [4]: The patterns of iris are rather stable and not easy to change through-
out one’s whole life; Iris patterns are quite complex and can contain many distinctive
features; An image of iris is much easier to be captured, it can be collected within a
distance as far as one meter, and the process would be much less invasive. With all
these advantages, Iris recognition is more and more widely used in modern personnel
identification system.
The process of iris recognition mainly consists of six steps: image acquisition,
segmentation, normalization, encoding, matching, and evaluation.
∗
Correponding Author.
∗∗
Both authors equally contributed to this work.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 340–349, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Fast Iris Segmentation by Rotation Average Analysis 341

Iris segmentation is the first substantial step of image processing to subtract the ef-
fective texture region of iris from the acquired images. The latter steps, such as iris
feature encoding and evaluation, are directly based on the result of iris segmentation.
The main task of iris segmentation is to locate the inner and outer boundaries of
iris, which is so called iris location. Dozens of papers (e.g. [6-11]) can be found in
this active research field. Most of the cases, eyelids/eyelashes may strongly affect the
final accuracy of iris recognition. For a complete segmentation, eyelids and eyelashes
also need to be detected and removed. Besides the problem of iris location, eyelids
and eyelashes detection are also major challenges for effective iris segmentation.
Eyelid detection is a challenging task due to the irregular shape of eyelids and eye-
lashes. Normally, much computational cost is needed and higher chances to be failed
in eyelid detection. The corresponding research was addressed in many works (e.g.
[12-15]), however it is less discussed and resolved compared to the research of iris
location.
In the remainder of this paper, we will introduce the new method to do iris segmen-
tation by rotation average analysis of intensity-inversed image in a detail. For iris
location, in fact, circular edge detection, we first locate the inner boundary (as a cir-
cle) and then fitting the outer boundary (as a circle) by least-square non-linear circular
regression. A simplified mathematical model of arc for detecting the eyelids is created
to speed up the conventional process of eyelid detection. Other conventional methods
are compared and discussed.

2 Segmentation Methods
There are two steps to complete the iris segmentation. First, locating the circular
boundaries of the iris, which includes fitting the inner boundary of the iris (the pupil-
lary boundary) and fitting the outer boundary of the iris (the limbus boundary). The
fitted two circles are not necessarily to be concentric. According to the fact that the
distance between the centers of them is rather small, it will be much easier to locate
one after the other.
Normally the iris images in use are near infrared (NIR) images which contain
abundant texture features in dark and brown color iris. In the NIR iris images, the
sclera is not white, high contrast level can be found at the boundary between pupil
and iris (inner boundary), while on the other hand the contrast at the boundary be-
tween iris and sclera (limbus, outer boundary) is much less (figure 1a). So, we started
the whole process from locating the pupil.

2.1 Locate the Circular Boundaries of the Iris

Find the Pseudo Center and Radius of the Inner Boundary

An iris image consists of an array of pixels with positions (x,y). For an image, we
assume the top left pixel with the coordinate (1,1) and the bottom right pixel (n,n).
To find the center of the pupil (figure 1b), an image is scanned twice and the most
consecutive high-intensity pixels are individually counted both vertically and horizon-
tally. A pixel is taken as ‘high-intensity pixel’ only if its intensity is higher than a user
342 W. Li and L.-H. Jiang

Fig. 1. (a) A NIR Iris image and related medical terminology. (b) Locate the pupil in the inten-
sity-inversed iris image. Calculate the center coordinates by averaging Hx and Vx, as well as
Hy and Vy. Scan the image vertically to get the most consecutive high intensity pixels on thick
red line (Vx). Do the same horizontally showing as the thick blue line (Hy).

defined threshold, which is normally close to the global maximum of intensity, e.g.
85% of the maximum.
First, scan the image vertically, for each fixed x, scan along y axis. The x-
coordinate - for which the most consecutive high intensity pixels in the y-direction are
found - is stored (thick red line, Vx). This step is repeated for fixed y, and scan along
x axis, and a y-coordinate is stored (thick blue line, Hy).
It is not necessarily true that (Vx,Hy) is the correct center. Therefore the second
step is to calculate the middle point of the thick red and blue line. For example, if the
first high-intensity pixel on thick red line is found at (Vx,100) and the last high-
intensity pixel at (Vx,180), then the middle point (Vx, Vy) is (Vx, 140).
In a final step, Vx is averaged with Hx (thin blue line), and Hy is averaged with Vy
(thin red line),which result in the mediated center coordinate. We call it pseudo cen-
ter, because it may be affected by reflections and noise or irregular edge of the pupil.
It is necessary to finely locate the center in the following process. The average length
of the consecutive high intensity pixels is the diameter of the pupil.

Refine the Pupillary Center and Rotation Average Analysis

For an accurate refinement of the pseudo center, the center of mass of the pupil can be
quickly calculated within a narrow range to avoid the effect of eyelid/eyelash noise.
Formula to calculate the center of mass in a digital image is:

1 m, n
M= ∑ I (i, j ) ∗ Χ
N 1, 2
(1)

Where, M is the coordinate of centroid, N is the number of total pixels, I is the inten-
sity of each pixel, Χ is the coordinate of each pixel. The area of summation can be a
limited region near the estimation.
Normally the pseudo center of the pupil slightly runs off the real center, and this
refinement may finely calibrate it in several pixels. Figure 2a shows the inversed
image, centered with the detected coordinate of the center of the pupil.
 Fast Iris Segmentation by Rotation Average Analysis 343

Fig. 2. (a) The intensity-inversed iris image is centered in a square panel. (b) The rotation aver-
age image of (a).

The pupil is thus quickly located. The time-consuming procedure of ‘integro- dif-
ferential’ of intensities [2] and Canny edge detection [16] are not needed at all.
After an accurate center of the pupil and its radius are determined, we can next
center the iris image in a larger-size square panel so that a rotation average image and
curve can be calculated. Figure 2b shows the rotation average image according to the
refined center, and figure 3a shows the rotation average curve.
Different averaging method can be used to calculate the rotation average curve:
A) Mean average
It is the most common way to calculate the mean average of all the pixels in a cir-
cular shell. It runs rather quickly, and most of the time, the result is stable.
B) Median average
Median is the number in the middle, when all numbers are listed in order. The sta-
tistic calculation of median average makes it more stable and less sensitive to singular
points. Disturbance from the reflection spots and the spike noise can be reduced.
Mean and median are both types of averages, although mean is the most common
type of average and usually refers to the arithmetic mean. Here, we choose the median
average in calculating the rotation average curve.
From the rotation average curve (figure 3a), we clearly see an oscillation section
near the middle-peak, which corresponds to the iris region containing abundant tex-
ture information. The curve is smooth and continuous, so that it is easy to find the
middle-valley and middle-peak on the curve. The mean-value of the middle-peak and
maximum intensity on the curve just corresponds with the papillary boundary, while
the mean-value of the middle-peak and minimum intensity can be used to roughly
segment the limbus boundary.
This limbus boundary calculated by the rotation average analysis is a rough value,
assuming the pupil and limbus share the same center. Since the limbus boundary and
pupillary boundary are not necessarily concentric, the pupillary boundary center we
have found can not be used as the center of limbus boundary directly (though it’s very
close to the truth). A refinement step of fitting the outer boundary is necessary.
344 W. Li and L.-H. Jiang

Fig. 3. (a) Rotation average curve and analysis of the intensity-inversed image. (b) Intermediate
result of iris segmentation.

Fitting the Outer Boundary by Circular Regression

Based on this rotation average analysis of intensity-inversed image, rough center and
radius of the outer boundary of iris are known from the former step. Now we are try-
ing to refine the circle to fit the outer boundary.
In mathematics, “circle fitting” is a question of non-linear circular regression:
Given a set of measured (x, y) pairs that are supposed to reside on a circle, but with
some added noise. A circle to these points, i.e. find the center (xc, yc) and radius r,
such that:

( x − xc ) 2 + ( y − yc ) 2 = r 2 (2)

The threshold used for roughly segmenting the limbus boundary can be utilized to
binarize the iris image and generate a simple edge map for this circle fitting (figure 4a).
This process is much faster than classical Canny edge detection.
Direct mathematical calculation of circular regression requires a good initial guess
and less massive noise (e.g. eyelashes). Since we already had a very close estimation
of the outer boundary, it will be much easier to accurately locate the limbus boundary
by limiting the circle fitting in a certain range of estimation (for instance, +-10%).
The circle fitting is based on orthogonal distance regression. Most of the known fit-
ting routines rely on the Levenberg-Marquardt optimization routine. The Levenberg-
Marquardt algorithm requires an initial guess as well as the first derivatives of the
distance function. In practice it converges quickly and accurately even with a wide
range of initial guesses. For two-dimensional circle fitting, we give the distance, ob-
jective and derivatives functions:
Distance equation:
(3)
d ( x, y ) = ( x − xc ) 2 + ( y − yc ) 2 − r
Objective function:
(4)
E ( xc , yc , r ) = ∑ ( ( x − xc ) 2 + ( y − yc ) 2 − r )
2
 Fast Iris Segmentation by Rotation Average Analysis 345

Derivatives:

∂d ∂d ∂d
= −( x − xc ) /(d + r ) ; = −( y − yc ) /(d + r ) ; = −1 (5)
∂xc ∂yc ∂r
Good circular regression algorithms can readily be found to minimize the objective
function E. Usually such an algorithm requires an initial guess, along with partial
derivatives, either of E itself or of the distance function d. Of course, one can imple-
ment a least-squares algorithm which uses a different optimization algorithm.
We recommend and use Taubin’s method [17] for circle fitting (see figure 4 for the
result). It is a robust and accurate circle fitting method, which works well even if data
points are observed only within a small arc. It is more stable than the other simple
circle fitting methods, e.g. by Kasa [18].

Fig. 4. (a)Simple binarized edge map for circle fitting (b) Circular regression and result of
fitting the outer boundary

To be compared with, the known Hough transform algorithm [19] needs to calcu-
late a Canny edge map first, and then use edge points to vote for particular model.
The idea of both the Hough transform and non-linear circular regression is to find the
contours which can fit in a circle with its center as (xc, yc) and the radius as r. Hough
transform uses “voting” mechanism, while circular regression uses directly mathe-
matical calculation.
Another method is to use integro-differential method to search for the circle of
outer boundary, due to the limited searching range, this method is also fast. Here we
do not intend to go into the details.

2.2 A New Mathematical Model of Arc for Detecting the Eyelid

The integro-differential approach can be adapted to detect the eyelid, when the contour
model of circle is replaced by an arc model. This arc edge detector can be described as:

I ( x, y )
max Gσ (r ) * ∫ arc ( d , a ,θ ) (6)
d , a ,θ L(arc)
346 W. Li and L.-H. Jiang

G is a Gaussian low-pass filter, I is the intensity of pixels on arc, L is the length of arc. d,
a, and θ are the three parameters of our new arc model, which will be introduced latter.
The contours of upper/lower eyelid can be regarded as a piece of parabolic arc ap-
proximately. Both Hough transformation and integro-differential method could be
applied in this task. The methods used in finding the upper and lower eyelid arcs are
in fact pattern fitting algorithms. In our implementation, we chose integro-differential
method to avoid the calculation of edge maps. A standard mathematical model of arc
(equation 7 & figure 5a) has four parameters for the j’th image: the apex coordinate
(hj, kj), the curvature aj, and the rotation angle θj . To fit the data with the mathemati-
cal model of arc is a time-consuming searching problem in a 4D space.
(−( x − h j ) sinθ j + ( y − k j ) cosθ j )2 = a j (( x − h j ) cosθ j + ( y − k j ) sin θ j ) (7)

To speed up this process, we proposed an efficient adapted mathematical model of arc

with three free parameters to detect the eyelid/eyelash. In Hough transformation and
integro-differential methods, the model of circular iris boundaries has three parame-
ters, while here the model of arc has one more parameter. Consequently, the computa-
tion complexity would be much higher with the arc contours. In order to improve the
performance, we simplified the eyelid model in our implementation by assuming that
the axis of the arc goes through the center of pupillary boundary. In this way, the two
parameters of the apex coordinate (hj, kj) can be replaced by a single parameter dj,
which refers to the distance from the apex to the original point (center of pupillary
boundary). That is, hj= djcosθj ; kj = djsinθj, and equation 7 is simplified to equation 8.
Practically, the curvature aj can be replaced by the distance from a predefined point on
the arc to the tangent line that passes through the apex. Now we have only three easy-
to-get parameters (dj, aj, θj). This new model is illustrated in figure 5b.
(− x sin θ j + y cosθ j ) 2 = a j ( x cosθ j + y sin θ j − d j ) (8)

By simplifying the fitting problem from a 4D space to a 3D space, the processing

speed is greatly improved. In order to gain more improvement of speed, one can sim-
ply constrain the searching range of these parameters. (e.g. to constrain the ‘d’ as the
value between the inner and outer radius of the iris.)

Fig. 5. (a) Conventional mathematical model of parabolic arc with 4 parameters. (b) Novel
model of arc with 3 parameters and central axis passing through the center of pupil.
 Fast Iris Segmentation by Rotation Average Analysis 347

3 Results and Discussion

The new method for iris segmentation has been implemented in Matlab. To evaluate
their performances, both self-obtained iris images and the CASIA iris image database
are used. Figure 6 gives some examples of the segmentation results.
The time performance is listed in Table 1. Table 2 presents the comparison of accu-
racy and speed performance of iris location between our new method and other con-
ventional methods. These results showed significant improvements when applying
our new method.
Compared with other methods implemented based on Hough transformation or the
integro-differential method [6-11], the new method of locating the pupil is different
from conventional ‘model fitting’ methods in the sense that the time to locate the
pupil is a constant for images with the same size, no matter how big the actual areas
of pupils occupy in the images. If the pupils are quite different in size, conventional
‘model fitting’ methods usually consume much more time in searching a wider range,
thus slow down the entire process.
Why we inverse the intensity of the image in the analysis? By inversing the inten-
sity of the image, the circular pupil becomes the highest intensity area. It allows locat-
ing the pupil by checking the consecutive high intensities in the inversed image, and

Fig. 6. Examples of iris segmentation results of the proposed method. (a)-(c) Iris segmentation
of images from CASIA iris image database. (d) Segmentation of self-obtained iris image. A
dummy arc of lower eyelid is drawn on the edge, if the eyelid and eyelash do not cover the iris.

Table 1. Typical time cost in each steps of segmentation for the new method presented

Step Task Time cost (s)

1. Locate pupil and coarse limbus boundary. 0.16
2. Fitting limbus boundary by circular regression. 0.05
3. Detect upper or lower eyelid. 1.5

Table 2. Comparison of performance of iris boundaries location (not include eyelid detection)

Method Success rate Time cost (s)

Doughman’s 96.7% 2.83
Wildes’ 94.2% 5.74
Proposed 98.1% 0.21
348 W. Li and L.-H. Jiang

refining the pupillary center with the centroid (the center of mass) of the pupil. If not
inverse the intensity, the calculated center of mass will not be stable, since the sur-
rounding area has higher intensity than the center area, the center of mass will easily
move away from the real center.

4 Conclusions
In this paper, we present a novel iris segmentation method. The effective iris is sub-
tracted in two steps: Iris boundaries location and eyelid detection. Intensity analysis of
rotation average of inversed image is applied to quickly locate the pupil and give a
close initial guess in circular regression of the outer boundary. The outer boundary is
refined by least-square non-linear circular regression. A new simplified mathematical
model of arc is created and implemented, which also works fast and efficiently com-
pared to the conventional methods. Our method provides a new idea to do iris seg-
mentation based on statistic analysis of rotation average.
In the experiment based on both self-acquired images and the public database
(CASIA iris image database), the new method used in the iris segmentation shows
fast and robust performance. Prospectively, there are still spaces to be improved in
our method: automatic eyelashes detection and removal.
Acknowledgments. This work was mainly done in the Catholic University of Leuven
(K.U.Leuven), Belgium. Thanks very much for the instruction of Prof. Dr. Dirk Van-
dermeulen in the Center for Processing Speech and Images (PSI) at the Department of
Electrical Engineering (ESAT). Thanks for Ying Zhang in the Terry Fox Laboratory,
University of British Columbia, Canada for acquisition of iris images.Parts of the
experimental data were supplied by the Institute of Automation, Chinese Academy of
Sciences (CASIA) [20].

References
1. Daugman, J.: New Methods in Iris Recognition. IEEE Trans. System, Man, and Cybernet-
ics–Part B: Cybernetics 37(5), 1167–1175 (2007)
2. Daugman, J.: How Iris Recognition Works. IEEE Trans. Circuits and Systems for Video
Technology 14(1), 21–30 (2004)
3. Daugman, J.: The Importance of being Random: Statistical Principles of Iris Recognition.
Pattern Recognition 36(2), 279–291 (2003)
4. Daugman, J.: Statistical Richness of Visual Phase Information: Update on Recognizing
Persons by Iris Patterns. Int. J. Computer Vision 45(1), 25–38 (2001)
5. Wildes, R.: Iris Recognition: An Emerging Biometric Technology. Proc. IEEE 85(9),
1348–1365 (1997)
6. Trucco, E., Razeto, M.: Robust iris location in close-up images of the eye. Patter Anal.
Applic. 8, 247–255 (2005)
7. Tang, R., Han, J., Zhang, X.: An Effective Iris Location Method with High Robustness.
Optica Applicata. 37(3), 295–303 (2007)
8. He, Z., Tan, T., Sun, Z.: Iris Localization via Pulling and Pushing. In: ICPR, vol. 4,
pp. 366–369 (2006)
 Fast Iris Segmentation by Rotation Average Analysis 349

9. Yuan, W., Xu, L., Lin, Z.: An Accurate and Fast Iris Location Method Based on the Fea-
tures of Human Eyes. In: Wang, L., Jin, Y. (eds.) FSKD 2005. LNCS (LNAI), vol. 3614,
pp. 306–315. Springer, Heidelberg (2005)
10. Sun, C., Zhou, C., Liang, Y., Liu, X.: Study and Improvement of Iris Location Algorithm.
In: Zhang, D., Jain, A.K. (eds.) ICB 2005. LNCS, vol. 3832, pp. 436–442. Springer,
Heidelberg (2005)
11. Lee, J.C., Huang, P.S., Chang, C.P., Tu, T.M.: Novel and Fast Approach for Iris Location.
IIHMSP 1, 139–142 (2007)
12. He, X.F., Shi, P.F.: An efficient iris segmentation method for recognition. In: Singh, S.,
Singh, M., Apte, C., Perner, P. (eds.) ICAPR 2005. LNCS, vol. 3687, pp. 120–126.
Springer, Heidelberg (2005)
13. Arvacheh, E.M., Tizhoosh, H.R.: Iris Segmentation: Detecting Pupil, Limbus and Eyelids.
In: ICIP, pp. 2453–2456 (2006)
14. He, Z., Tan, T., Sun, Z., Qiu, X.: Towards Accurate and Fast Iris Segmentation for Iris
Biometrics. IEEE Trans. Pattern Analysis and Machine Intelligence 99(2008),
doi:10.1109/TPAMI.2008.183
15. Jang, Y.K., Kang, B.J., Park, K.R.: Study on eyelid localization considering image focus
for iris recognition. Pattern Recognition Letters 29(11), 1698–1704 (2008)
16. Canny, J.: A Computational Approach to Edge Detection. IEEE Trans. Pattern Analysis
and Machine Intelligence 8, 679–714 (1986)
17. Taubin, G.: Estimation Of Planar Curves, Surfaces And Nonplanar Space Curves Defined
By Implicit Equations, With Applications To Edge And Range Image Segmentation. IEEE
Trans. PAMI 13, 1115–1138 (1991)
18. Kasa, I.: A curve fitting procedure and its error analysis. IEEE Trans. Inst. Meas. 25, 8–14
(1976)
19. Duda, R.O., Hart, P.E.: Use of the Hough Transformation to Detect Lines and Curves in
Pictures. Comm. ACM 15, 11–15 (1972)
20. CASIA iris image database, http://www.sinobiometrics.com
 A New Criterion for Global Asymptotic
Stability of Multi-delayed Neural Networks∗

Kaiyu Liu1 and Hongqiang Zhang2

1
College of Mathematics and Econometrics Hunan University,
Changsha, Hunan 410082, China
2
College of Mathematics and Compution Science,
Changsha University of Science and Technology,
Changsha, Hunan 410076, China
liukyhnu@yahoo.com.cn, hqzhangcs@gmail.com

Abstract. This Letter presents some new sufficient conditions for the unique-
ness and global asymptotic stability (GAS) of the equilibrium point for a class
of neural networks with multiple constant time delays . It is shown that the use
of a more general type of Lyapunov-Krasovskii functional enables us to estab-
lish global asymptotic stability of a class of delayed neural networks than those
considered in some previous papers. Our results generalize or improve the pre-
vious results given in the literature.

Keywords: Delay neural Networks, Global asymptotic stability, Equilibrium

analysis, Lyapunov-Krasovskii functional.

1 Introduction
In the last few years, stability of different classes of neural networks with time delays,
such as Hopfield neural networks, cellular neural networks, bidirectional associative
neural networks, has been extensively studied[1-7,10-12], particularly regarding their
stability analysis. Recently, LMI-based techniques have been successfully used to
tackle various stability problems for neural networks,mainly with a single delay(see,
for example[8,9,15]) and a few with multi-delays[13,14].In the present paper, moti-
vated by [14],by employing a more general Lyapunov functional and the LMI
approach , we establish some sufficient conditions for the uniqueness and global
asymptotic stability for neural networks with multiple delays . The conditions given in
the literature are generalized or improved .
Consider the following neural networks with multiple delays
m
u& (t ) = − Au (t ) + Wg (u (t )) + ∑ W j g (u (t − τ j )) + I , (1)
j =1

where u (t ) = [u1 (t ),L , u n (t )]T is the neuron state vector, A = diag (a1 ,L , a n ) is
a positive diagonal matrix, W , W j = (ω ik( j ) ) n×n , j = 1,2,L, m are the interconnection
∗
Project supported Partially by NNSF of China (No:10571032).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 350–360, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 A New Criterion for Global Asymptotic Stability 351

matrices, g (u ) = [ g 1 (u1 ), L , g n (u n )]T denotes the neuron activation, and

I = [ I 1 ,L , I n ]T is a constant external input vector, while τj >0
( j = 1,2,L , m) being the delay parameters.
The usual assumptions on the activation functions are as follows

g j (ξ1 ) − g j (ξ 2 )
(H ) 0≤ ≤σ j, j = 1,2,L , n ,
ξ1 − ξ 2
for each ξ1 , ξ 2 ∈ R, ξ1 ≠ ξ 2 , where σ j are positive constants. In the following, we

will shift the equilibrium point u ∗ = [u1∗ ,L , u n∗ ]T of system (2) to the origin. The
transformation x(⋅) = u (⋅) − u ∗ puts system (3) into the following form
m
x& (t ) = − Ax (t ) + Wf ( x (t )) + ∑ W j f ( x (t − τ j )), (2)
j =1

where x = [ x1 ,L , x n ]T is the state vector of the transformed system, and

f j ( x j ) = g j ( x j + u ∗j ) − g j (u ∗j ) with f j (0) = 0, ∀j. Note that the functions
f j (⋅) satisfy the condition (H ) , that is

f j (ξ1 ) − f j (ξ 2 )
0≤ ≤σ j, j = 1,2,L , n (3)
ξ1 − ξ 2
for each ξ1 , ξ 2 ∈ R, ξ1 ≠ ξ 2 , where σ j are positive constants.

2 Main Stability Results

We first introduce the following Lemma 1, which will be used in the proof of our
main results.
Lemma 1. (Schur complement)The linear matrix inequality (LMI)

⎛S S12 ⎞
S = ⎜⎜ 11 ⎟>0
⎝ S 21 S 22 ⎟⎠

S11 > 0 (respectively, S 22 > 0 ) and

is equivalent to the following conditions
S 22 − S S S > 0, (respectively, S11 − S12 S 22−1 S 21 > 0 ) where S11T = S11 ,
−1
21 11 12
T
S 22 = S , S = S 21 .
22
T
12
352 K. Liu and H. Zhang

Theorem 1. The The origin of (2) is the unique equilibrium point and it is globally
asymptotically stable if there exists a positive diagonal matrix
P = diag ( pi > 0), and positive definite matrix Q j ( j = 1,2,L , m) such that
⎡ Ω
L − Wm ⎤ − W1
⎢− W T
L 0 ⎥⎥Q1
⎢ 1 > 0, (4)
⎢ M
L M ⎥ M
⎢ ⎥
⎣ − Wm
T
L Qm ⎦ 0
where Ω = 2 A ∑ P − WP − PW − P ∑ j =1 Q j P ( ∑ = diag (σ i > 0) ).
−1 T m

Proof. We will first prove the uniqueness of the equilibrium point. To this end, let us
consider the equilibrium equation of (2) as follows
m
Ax ∗ − Wf ( x ∗ ) − ∑ W j f ( x ∗ ) = 0, (5)
j =1

Where x ∗ is the equilibrium point. We have to prove x ∗ = 0 , for this purpose, mul-
T ∗ −1
tiplying both sides of (5) by 2 f ( x ) P , we obtain
m
2 f T ( x ∗ ) P −1 Ax ∗ − 2 f T ( x ∗ ) P −1Wf ( x ∗ ) − 2∑ f T ( x ∗ ) P −1W j f ( x ∗ ) = 0, (6)
j =1
Using (3) in (6) results in

f T ( x ∗ )(2 P −1 A ∑ −1 − P −1W − W T P −1 ) f ( x ∗ )
m
− 2∑ f T ( x ∗ ) P −1W j f ( x ∗ ) ≤ 0.
j =1

m
Adding and subtracting the term f T ( x ∗ )∑ Q j f ( x ∗ ) in the left side of the above
j =1
inequality yields

UBU T ≤ 0, (7)

where U = [ f T ( x ∗ ) P −1 , f T ( x ∗ ),L , f T ( x ∗ )] ,

⎡ Ω − W1 L − Wm ⎤
⎢− W T Q1 L 0 ⎥⎥
B=⎢ 1 . (8)
⎢ M M L M ⎥
⎢ ⎥
⎣ − Wm
T
0 L Qm ⎦
 A New Criterion for Global Asymptotic Stability 353

（）
Condition 4 in Theorem 1 is B > 0 . Thus it follows from (7) that f (x∗ ) = 0 ,
∗ ∗
substitutes this into (5), we get Ax = 0 which implies x = 0 . This has shown that
the origin is the unique equilibrium point for every I .
We will now prove the GAS of the origin of (2). To this end, let us consider the
following Lyapunov-Kraasovskii functional
n
V ( x(t )) = x T (t )Cx(t ) + 2∑ p i−1 ∫
xi ( t )
f i ( s )ds
0
i =1
m
+ ∑∫
t
f T ( x( s ))Q j f ( x( s ))ds ,
t −τ j
j =1

where C , P = diag ( pi > 0) and Q j ( j = 1,2,L , m) are positive definite matri-

ces with C being being chosen appropriately later on. Clearly, V ( x (t )) is positive
except at the origin and it is radially unbounded. Evaluating the time derivative of
V along the trajectory of (2), we have
V& ( x(t )) = −2 x T (t )CAx(t ) + 2 x T (t )CWf ( x(t ))
m
+ 2∑ x T (t )CW j f ( x(t − τ j )) − 2 f T ( x(t )) P −1 Ax(t )
j =1
m
+ 2 f T ( x(t )) P −1Wf ( x(t )) + ∑ f T ( x(t ))Q j f ( x(t ))
j =1
m
+ 2∑ f T ( x(t )) P −1W j f ( x(t − τ j ))
j =1
m
− ∑ f T ( x(t − τ j ))Q j f ( x(t − τ j )) .
j =1

Since f T ( x(t )) P −1 A ∑ −1 f ( x(t )) ≤ f T ( x(t )) P −1 Ax(t ) ,we can write

V& ( x(t )) ≤ − x T (t )(CA + AC ) x(t ) + 2 x T (t )CWf ( x(t ))
m m
+ 2∑ x T (t )CW j f ( x(t − τ j )) + f T ( x(t ))∑ Q j f ( x(t ))
j =1 j =1
−1 −1 −1 −1
− f ( x(t ))(2 P A ∑ − P W − W P ) f ( x(t ))
T T

m
+ 2∑ f T ( x(t )) P −1W j f ( x(t − τ j ))
j =1
m
− ∑ f T ( x(t − τ j ))Q j f ( x(t − τ j )) ,
j =1
354 K. Liu and H. Zhang

which can be written in the form

V& ( x(t )) ≤ U 1 SU 1T , (9)

where U 1 = [ x T (t )C , f T ( x(t ) P −1 , f T ( x(t − τ 1 )),L , f T ( x(t − τ m ))] ,

⎡ AC −1 + C −1 A − WP − W1 L − Wm ⎤
⎢ ⎥
⎢ − PW Ω − W1 L − Wm ⎥
T

S=⎢ − W1 T
− W1T Q1 L 0 ⎥.
⎢ ⎥
⎢ M M M M ⎥
⎢ − Wm T
− WmT L Qm ⎥⎦
⎣ 0
In the following , we will choose a positive-definite matrix C such that S > 0 . For
this purpose, let

⎛ AC −1 + C −1 A D ⎞
S = ⎜⎜ ⎟, (10)
⎝ DT B ⎟⎠
whereD = [−W1 ,L ,Wm ] and B is given in (8). In view of (4) and (8), we have
B > 0 . Now, choose a positive definite matrix G such that
G > DB −1 D T . (11)

Since − A = diag (− ai < 0) is negative definite, there must be a positive-definite

matrix H such that
H (− A) + (− A) H = −G . (12)

Let C = H −1 , then C is positive definite and we have by (11) and (12)

AC −1 + C −1 A − DB −1 D T > 0 ,
from which and Lemma 1 we conclude S > 0 , and it follows by (9) that
V& ( x(t )) ≤ 0 . Moreover V& ( x(t )) = 0 if and only if

[ x T (t )C , f T ( x(t ) P −1 , f T ( x(t − τ 1 )),L , f T ( x(t − τ m ))] = 0 ,

i.e., V& ( x(t )) = 0 if and only if

x(t ) = f ( x(t ) = f ( x(t − τ 1 )) = L = f ( x(t − τ m )) = 0 .
 A New Criterion for Global Asymptotic Stability 355

Hence, V& ( x(t )) is negative definite and therefore, the origin of (2) or equivalently
∗
the equilibrium x of (1) is globally asymptotically stable. This completes the proofs.
In view of Lemma 1, the following Theorem 1' and Theorem 1'' are both equiva-
lent descriptions of Theorem 1.
Theorem 1'. The origin of (2) is the unique equilibrium point and it is globally as-
ymptotically stable if there exists a positive diagonal matrix P and positive definite
matrices Q j ( j = 1,2,L , m) such that

m m
2 A ∑ −1 P − WP − PW T − P ∑ Q j P − ∑ WQ −j 1W jT > 0 .
j =1 j =1

Theorem 1''. The origin of (2) is the unique equilibrium point which is globally as-
ymptotically stable if there exists a positive diagonal matrix P and positive definite
matrices Q j ( j = 1,2,L , m) such that

m
B1 = 2 A ∑ −1 P − WP − PW T − P ∑ Q j P > 0
j =1
and
⎡Q1 − W1T B1−1W1 L − W1T B1T Wm ⎤
⎢ ⎥
⎢ M M M ⎥ > 0.
⎢ − Wm B1 W1
T T
L Qm − Wm B1 Wm ⎥⎦
T T
⎣
In the following, we will give further result for GAS of the origin of (2)(or equiva-
lently the equilibrium point of (1)). We need the following

Lemma 2. Let Q j , S , S j ∈ R n×n ( j = 1,2,L , m) be positive-definite matrices and

let W j ∈ R n×n . Then the following LMI

⎡Q1 − W1T SW1 L − W1T SWm ⎤

⎢ ⎥
⎢ M M M ⎥>0 (13)
⎢ − WmT SW1 L Qm − Wm SWm ⎥⎦
T
⎣
hold if the following LMIs hold

Q j − W jT S jW j > 0 , j = 1,2,L , m (14)

and
m
S −1 > ∑ S −j 1 . (15)
j =1
356 K. Liu and H. Zhang

Proof. Let A denote the matrix given in the left side of (14), i.e.,
⎛ Q1 ⎞ ⎛W1 SW1 L W1 SWm ⎞
T T
⎜ ⎟ ⎜ ⎟
A=⎜ O ⎟ − ⎜ M L M ⎟,
⎜
⎝ Qm ⎟⎠ ⎝Wm SW1 L Wm SWm ⎟⎠
⎜ T T

consider the quantic form associate with A as follows

m m m
f ( x) = x T Ax = ∑ x Tj Q j x j − ∑∑ x Tj W jT SWi xi ,
j =1 j =1 i =1

where x = ( x , x ,L , x ) ∈ R
T
1
T
2
T
m
nm
with x j ∈ R n , j = 1,2,L , m .Adding and
m m m
subtracting the term ∑ x Tj W jT S jW j x j + ∑∑ x Tj W jT SWi xi , we have
j =1 j =1 i =1
m
f ( x) = x T Ax = ∑ x Tj (Q j − W jT S jW j ) x j
j =1
m m m m
+ ∑ [ x Tj W jT S jW j x j - 2 x Tj W jT S ∑ Wi xi ] + ∑ x Tj W jT S ∑ Wi xi . (16)
j =1 i =1 j =1 i =1
Note that
m
x Tj W jT S jW j x j − 2 x Tj W jT S ∑ Wi xi
i =1
m m
= [ S jW j x j − ∑ SWi xi ]T ⋅ S −j 1 ⋅ [ S jW j x j − ∑ SWi xi ]
i =1 i =1
m
− ∑ xiT WiT SS −j 1 SWi xi
i =1
for i = 1,2,L , m . This implies that
m m

∑[x W
j =1
T
j
T
j S jW j x j - 2 x Tj W jT S ∑ Wi xi ]
i =1
m m
≥ −∑ xiT Wi T S ∑ S −j 1 SWi xi .
i =1 j =1
Substituting this into (16) result in
m
f ( x) = x T Ax ≥ ∑ x Tj (Q j − W jT S jW j ) x j
j =1
m m m
+ ∑ [ xiT WiT [ S (−∑ S −j 1 + S −1 ) S ]∑ Wi xi ,
i =1 i =1 i =1
 A New Criterion for Global Asymptotic Stability 357

Which, by using the condition (14) and (15), implies that f ( x) = x T Ax > 0 for all
x ≠ 0 . Therefore we have A > 0 and the proofs is completed.
Theorem 2. The origin of (2) is the unique equilibrium point which is globally as-
ymptotically stable if there exists a positive diagonal matrix P and positive definite
matrices S j ( j = 1,2,L , m) such that

m m
2 A ∑ −1 P − WP − PW T − P ∑ W jT S jW j P − ∑ S −j 1 > 0 . (17)
j =1 j =1

Proof. In view of (17), there exists a sufficiently small positive constant δ >0 such
that
m m
2 A ∑ −1 P − WP − PW T − P ∑ (W jT S jW j + δE ) P − ∑ S −j 1 > 0 , (18)
j =1 j =1

where E is the appropriate unit matrix. Let

Q j = W jT S jW j + δE , j = 1,2,L , m (19)

and
m
B2 = 2 A ∑ −1 P − WP − PW T − P ∑ Q j P . (20)
j =1

Then Q j > W jT S jW j ≥ 0 ( j = 1,2,L , m ) and by (18)-(20), we have

m
B2 > ∑ S −j 1 > 0. (21)
j =1

It follows from Lemma 2 that

⎡Q1 − W1T B2−1W1 L − W1T B2T Wm ⎤

⎢ ⎥
⎢ M M M ⎥>0 (22)
⎢ − Wm B2 W1
T T
L Qm − Wm B2 Wm ⎥⎦
T T
⎣
In view of (20)-(22), the conclusion of Theorem 2 follows from Theorem 1''. This
completes the proofs.
Remark 1. Theorem 1 and Theorem 2 generalize Theorem 1 in [14] .
The following Theorem 3 give explicit condition which ensure the GAS of the
origin of (2), or equivalently the equilibrium point of (1).
358 K. Liu and H. Zhang

Theorem 3. The origin of (2) is the unique equilibrium and it is globally asymptoti-
cally stable if there exists a positive constant α > 0 such that the following condi-
tions hold
(a) − (2 A ∑ −1 +W + W T + αE ) is positive definite;
m
α
(b) ∑W
j =1
j 2 ≤
2
.

Proof. In view of condition (a), there must be a positive constant δ > 0 such that
− (2 A ∑ +W + W + αE ) > 2δE .
−1 T
(23)
Take P = λE and S j = λ j E ( j = 1,2,L , m) ,where λ and λ j are positive constant
and observe that
m m
2 A ∑ −1 P − WP − PW T − ∑ S −j 1 − P ∑ (W jT S jW j + δ 2 E ) P
j =1 j =1
m m
= 2λA ∑ −1 −λW − λW T − ∑ λ−j1 E − λ2 ∑ (λ jW jT W j + δ 2 E )
j =1 j =1
m
≥ 2λA ∑ −1 −λW − λW T − λ2 ∑ (λ j W j
2
+ δ 2 + λ− 2 λ−j1 )E
2
j =1

= −λ (−2 A ∑ +W + W )−1 T

m
− λ2 ∑ [λ j ( W j + δ − λ−1λ−j1 ) 2 + 2λ ( W j + δ )]E .
2 2
j =1

Let λ λ−j 1 = W j + δ . Then

−1
2

m m
2 A ∑ −1 P − WP − PW T − ∑ S −j 1 − P ∑ (W jT S jW j + δ 2 E ) P
j =1 j =1
m
≥ −λ (−2 A ∑ −1 +W + W T ) − 2λ ∑ ( W j + δ )E
2
j =1

α m
= −λ[(−2 A ∑ −1 +W + W T + αE ) − 2δE ] + 2λ ( − ∑ W j )E
2 j =1
2

Using (23) and conditions (b) in Theorem 3 , we get

m m
2 A ∑ −1 P − WP − PW T − ∑ S −j 1 − P ∑ (W jT S jW j + δ 2 E ) P > 0 .
j =1 j =1
 A New Criterion for Global Asymptotic Stability 359

Moreover we have
m m
2 A ∑ −1 P − WP − PW T − ∑ S −j 1 − P ∑ W jT S jW j P > 0 .
j =1 j =1

Thus the proof is completed by Theorem 2.

Let A = E and α = 2 + β in Theorem 3, we have the following Theorem 4 for
DCNN
m
x& (t ) = − x (t ) + Wg ( x (t )) + ∑ W j g ( x (t − τ j )), (24)
j =1

Theorem 4. The origin of (24) is the unique equilibrium point which is globally and
asymptotically stable if there exists a real constant β > −2 such that following
conditions hold
(I) − (W + W T + βE ) is positive definite;
m
β
(II) ∑W
j =1
j 2 ≤ 1+
2
.

Remark 2. Theorem 4 improve the results in [15] which β > 0.

3 Conclusions
The equilibrium and global asymptotic stability properties of neural networks with
multiple constant delays have been studied. Some new stability criteria have been
derived by employing a more general type of Lyapunov-Krasovskii functional. The
conditions are expressed in terms of the linear matrix inequality, which can be veri-
fied efficiently with the Matlab LMI Control Toolbox.

References
1. Hopfield, J.: Neurons with Graded Response Have Collective Computational Properties
Like Those of Two-State Neurons. Proc. Natl. Acad. Sci. USA 81, 3088–3092 (1984)
2. Marcus, C.M., Westervelt, R.M.: Stability of Analog Neural Network with Delay. Phys.
Rev. A 39, 347–359 (1989)
3. Belair, J.: Stability in A Model of A Delayed Neural Network. J. Dynam. Differential
Equations 5, 607–623 (1993)
4. Gopalsamy, K., Leung, I.: Delay Iinduced Periodicity in A Neural Network of Excitation
and Inhibition. Physica D 89, 395–426 (1996)
5. Ye, H., Michel, N., Wang, A.: Global Stability of Local Stability of Hopfield Neural
Networks with Delays. Phys. Rev. E 50, 4206–4213 (1994)
6. Liao, X.F., Yu, J., Chen, G.: Novel Stability Criteria for Bi-directional Associative Mem-
ory Neural Networks with Time Delays. Int. J. Circuit Theory Appl. 30, 519–546 (2002)
7. Gopalsamy, K., He, X.Z.: Stability in A Asymmetric Hopfield Networks with Transmis-
sion Delays. Physica D 76, 344–358 (1994)
360 K. Liu and H. Zhang

8. Liao, X.F., Chen, G., Sanchez, E.N.: LMI-Based Approach for Asymptotically Stability
Analysis of Delayed Neural Networks. IEEE Trans. Circuits Syst. 49(7), 1033–1039
(2002)
9. Liao, X., Chen, G., Sanchez, N.: Delayed-Dependent Exponential Stability Analysis of De-
layed Neural Networks: an LMI Approach. Neural Networks 15, 855–866 (2002)
10. Liao, X.F., Wu, Z., Yu, J.: Stability Analysis of Cellular Neural Networks with Continuous
Time Dlays. J. Comput. Appl. Math. 143, 29–47 (2002)
11. Arik, S.: Stability Analysis of Delayed Neural Networks. IEEE Trans. Circuits Syst. I. 47,
1089–1092 (2000)
12. Liao, X.X., Wang, J.: Algebraic Criteria for Global Exponential Stability of Cellular Neu-
ral Networks with Multiple Time Delays. IEEE Trans. Circuits Syst. 50, 268–275 (2003)
13. Wu, W., Cui, B.T.: Global Robust Exponential Stability of Delayed Neural Networks.
Chaos, Solitons and Fractals 35, 747–754 (2008)
14. Liao, X.F., Li, C.: An LMI Approach to Asymptotical Stability of Multi-Delayed Neural
Networks. Physica D 200, 139–155 (2005)
15. Arik, S.: Global Asymptotic Stability of A Larger Class of Neural Networks with Constant
Time Delay. Phys. Lett. A 311, 504–511 (2003)
 An Improved Approach Combining Random PSO
with BP for Feedforward Neural Networks

Yu Cui, Shi-Guang Ju, Fei Han∗, and Tong-Yue Gu

School of Computer Science and Telecommunication Engineering, Jiangsu University,

Zhenjiang, Jiangsu, China
allen790336@hotmail.com, jushig@ujs.edu.cn,hanfei@ujs.edu.cn,
gtyglx@tom.com

Abstract. In this paper, an improved approach which combines random particle

swarm optimization with BP is proposed to obtain better generalization per-
formance and faster convergence rate. It is well known that the backpropagation
(BP) algorithm has good local search ability but it is easily trapped to local min-
ima. On the contrary, the particle swarm optimization algorithm (PSO), with
good global search ability, converges rapidly during the initial stages of a global
research. Since the PSO suffers from the disadvantage of losing diversity, it
converges more slow around the global minima. Hence, the global search is
combined with local search reasonably in the improved approach which is
called as RPSO-BP. Moreover, in order to improve the diversity of the swarm
in the PSO, a random PSO (RPSO) is proposed in the paper. Compared with the
traditional learning algorithms, the improved learning algorithm has much bet-
ter convergence accuracy and rate. Finally, the experimental results are given to
verify the efficiency and effectiveness of the proposed algorithm.

Keywords: Particle swarm optimization, backpropagation, feedforward neural

networks.

1 Introduction
Feedforward neural networks(FNNs) have been widely used to approximate arbitrary
continuous functions [1,2], since a neural network with a single nonliner hidden layer
is capable of forming an arbitrarily close approximation of any continous nonliner
mapping [3]. There have been many algorithms used to train the FNN, such as back-
propagation algorithm (BP), particle swarm optimization algorithm (PSO) [4], simulat-
ing annealing algorithm (SSA) [5], gentic algorithm(GA) [6,7] and so on. Regarding
the FNN training algorithms, the BP algorithm is easily trapped into a local minima
and converge slowly [8]. To solve this problem, many improved BP algorithm has
been proposed [9-11]. However these algorithms have not removed the disadvantages
of the BP algorithm in essence.
In 1995, inspired from complex social behavior shown by the natural species like
flock of birds, PSO was proposed by James Kennedy and Russell Eberhart. Different
∗
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 361–368, 2009.
© Springer-Verlag Berlin Heidelberg 2009
362 Y. Cui et al.

from the BP algorithm, the PSO algorithm has good ability of global search. Although
the PSO algorithm has shown a very good performance in solving many problems, it
suffers from the problem of premature convergence like most of the stochastic search
techniques.
In order to make use of the advantages of PSO and BP, many researchers have
concentrated on hybrid algorithm such as PSO-BP, CPSO-BP and so on [12-14]. In
these algorithms, the PSO is used to do global search first and then switch to gradient
descending searching to do local search around global optimum. It is proved to be
better in convergence rate and convergence accuracy. However, premature conver-
gence still exists in the PSO-BP and CPSO-BP. During the global search, every parti-
cle has to follow both it's best historic position (pbest) and the best position of all the
swarm (gbest). Therefore, pbest will get close to gbest, and then the current particle
swarm may lose its diversity.
In this paper an improved PSO-BP approach named RPSO-BP is proposed to solve
the problems mentioned above. In this algorithm, there are three methods to update
each particle and each particle selects one randomly. In another word, each particle
does not have only one update method that following its best historic position and the
best position of the swarm, any more. In this way the gbest can avoid getting closed to
pbest and the likelihood of premature convergence can be reduced. Moreover, the
experiment results show that the proposed algorithm has better generalization per-
formance and convergence performance than other traditional algorithms.

2 Particle Swarm Optimization

PSO can be stated as initializing a team of random particles and finding the optimal
solutions by iterations. Each particle will update itself by two optimal values pbest
and gbest which are mentioned above.
The original PSO algorithm is described as follows:
V I
(t + 1) = V I (t ) + c1 * r1 * ( P I (t ) − X I
(t )) + c 2 * r 2 * ( P g (t ) − X I
(t )) (1)

X i
( t + 1) = X i
( t ) + V i ( t + 1) (2)
where Vi is the velocity of the ith particle; Xi is the position of the ith particle; Pi is the
best position achieved by the particle so far ; Pg is the best position among all parti-
cles in the population; r1 and r2 are two independently and uniformly distributed
random variables with the range of [0,1]; c1 and c2 are positive constant parameters
called accelerated coefficients, which control the maximum step size.
The adaptive particle swarm optimization (APSO) [15,16] algorithm is proposed
by Shi & Eberhart in 1998. This algorithm can stated as follows:
V I
(t + 1) = w * V I (t ) + c1 * r1 * ( P I (t ) − X I
(t )) + c2 * r 2 * ( P g (t ) − X I
(t )) (3)

X i
( t + 1) = X i
( t ) + V i ( t + 1) (4)
where w is called the inertia weight that controls the impact of the previous velocity
of the particle on its current. Several selection strategies of inertial weight w have
 An Improved Approach Combining Random PSO with BP 363

been given. Generally, in the beginning stages of algorithm, the inertial weight w
should be reduced rapidly, and the inertial weight w should be reduced slowly when
around global optimum.
Another important variant of standard PSO is the CPSO, which was proposed by
Clerc and Kennedy [17]. The CPSO ensures the convergence of the search producers
and can generate higher-quality solutions than standard PSO with inertia weight on
some studied problems.

3 The Proposed Algorithm (RPSO-BP)

In this section, RPSO-BP is proposed to solve the problems mentioned above. In
order to reduce the likelihood of the premature convergence, a random PSO (RPSO)
is proposed. In the RPSO-BP, the FNN is trained by RPSO first and then trained by
BP algorithm.
The searching process of the RPSO is started from initializing a group of random
particles, and a random parameter f1. On the meantime, Each particle has a random
parameter f2. The update of each particle is determined by the following rules:
(1) If f2 is greater than f1, all particles will be updated according to Eqs.(3)- (4).
(2) If f2 is less than f1 and the pbest is nearest to the gbest, a particle j will be se-
lected randomly and the particle will be updated according to Eqs.(4)-(5).
V I
(t + 1) = w *V I (t ) + c1* r1* ( P j (t ) − X I (t )) + c 2 * r 2 * ( P g (t ) − X I (t )) (5)

(3) If f2 is less than f1 and the pbest of the particle h is nearest to the gbest, the par-
ticle will be updated according to Eqs.(6) and (4).
V I
(t + 1) = w * V I (t ) + c1 * r1 * ( P h (t ) − X I
(t )) + c 2 * r 2 * ( P g (t ) − X I
(t )) (6)

where Pj and Ph are the personal (local) best position achieved by the particle j and
particle h so far respectively.
Similar to the PSO, the parameter w in the above RPSO-BP algorithm reduces
gradually as the iterative generation increases as follows[16]:
(W max − W min ) (7)
W (iter ) = W max −
iterma ∗ iter

where Wmax is the initial inertial weight, Wmin is the inertial weight of liner section
ending, iterma is the used generations that inertial weight, is reduced linearly, iter is a
variable whose range is [1,iterma].

4 Experimental Results
In this paper, some experiments are conducted to verify the effectiveness of the pro-
posed learning approach. In the following experiments, the performances of the BP,
PSO, CPSO, PSO-BP and CPSO-BP are compared with that of the RPSO-BP. All the
experiments are carried out in MATLAB 6.5 environment running in a Pentium 4,
2.40 GHZ CPU.
364 Y. Cui et al.

4.1 Function Approximation Problem

The function f(x), y=(sin(x)*cos(x.^3)+(x.^2/9+x/3+cos(x.^3))*exp((-0.5)*(x.^2)))/2,

is approximated by FNN in this experiment. For all the six algorithms, the activation
functions of the hidden neurons are tansig function, while the one in the output layers
is purline function.
First, assume that the number of the hidden neurons is 8, and the number of the to-
tal training data is 126 which are selected at an identical spaced interval from
[0,3.138], while the 125 testing data are selected at an identical spaced interval from
[0.0125,3.1249].

Table 1. The approximation accuracies and iteration number with six learning algorithms for
approximating the function f(x)

Learning algorithms Training MSE Testing MSE Iteration number

BP 0.0660675 0.0661 30000
PSO 0.0565 0.0573 20000
CPSO 0.0978 0.2104 20000
PSO-BP 9.99206e-005 6.9930e-004 15000
CPSO-BP 9.9592e-005 6.9584e-004 15000
RPSO-BP 6.59849e-005 4.8331e-004 15000

Fig. 1. The approximating result of the RPSO-BP algorithm for approximating the function f(x)

Second, we set the maximal iteration number as 30000 for the BP algorithm, 20000
for the PSO and CPSO algorithms. In the later three algorithms, the maximal iteration
number as 300 for PSO, and the one as 15000 for BP algorithm. The corresponding
results are summarized in Table 1 and Figs. 1-2.
It can be found from Table 1 and Figs.1-2 that the values of MSE of the new algo-
rithm are always less than those of the other five algorithms for approximating the func-
tion. The results also support the conclusion that the generalization performance of the
 An Improved Approach Combining Random PSO with BP 365

BP
2.5 PSO
CPSO

2 PSO-BP
CPSO-BP
RPSO-BP
1.5

t h e t e s t i n g e r ro r
1

0.5

-0.5

-1
0 0.5 1 1.5 2 2.5 3 3.5
the function range

Fig. 2. The testing error curves of six learning algorithms for approximating the function f(x)

Fig. 3. The relationship between the number of hidden neurons and testing error for approxi-
mating the function f(x) with three algorithms

Fig. 4. The relationship between the number of PSO and testing error for approximating the
function f(x) with three learning algorithms
366 Y. Cui et al.

proposed algorithm is better than the other algorithms. Moreover, the hybrid algorithms
such as PSO-BP, CPSO-BP and RPSO-BP converge faster than other algorithms.
Fig.3. shows the relationship between the number of hidden neurons and testing er-
ror for approximating the function f(x) with PSO-BP, CPSO-BP and RPSO-BP. It can
be seen that the testing error has a downward trend as the number of hidden neurons
increases. Fig. 4 shows the relationship between the number of PSO and testing error
for approximating the function f(x) with PSO-BP, CPSO-BP and RPSO-BP. The most
suitable number of PSO can be selected as 60 for RPSO-BP.

4.2 Oil Refining Problem

In petrochemical industry, the true boiling point curve of crude oil reflects the composi-
tion of the distilled crude oil. To build a model that takes the mass percentage of distilled
component as the independent variable and the distilled temperature as the dependent
variable is an important problem in petrochemical industry. In this subsection, we use
FNN to build nonparametric models to resolve this problem. In this subsection, the num-
ber of hidden neurons and the size of PSO we chose are 10 and 50 respectively.
Regarding the three hybrid algorithms in this example, the maximal iteration num-
ber of the corresponding PSO is set as 500, and the one for BP is set as 5000. For the
BP, the maximal iteration is assumed as 10000. For the PSO and CPSO, the maximal
iteration number is assumed as 10000. The simulation results are summarized in the
table 2 and the comparison result of the three hybrid algorithms are showed in Fig. 5.
0.4
CPSO-BP
0.35 PSO-BP
RPSO-BP
0.3

0.25
the testing error

0.2

0.15

0.1

0.05

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
the normalized data

Fig. 5. The testing error curves of six learning algorithms for modeling the true boiling point
curve of crude oil

Table 2. The approximation accuracies and iteration number with six learning algorithms for
modeling the true boiling point curve of crude oil with six algorithms

Learning algorithms Training MSE Testing MSE Iteration number

BP 0.0865 0.2668 10000
PSO 0.0737 0.2644 10000
CPSO 0.0908 0.2685 10000
PSO-BP 9.8348e-005 0.0604 5000
CPSO-BP 9.85675e-005 0.0610 5000
RPSO-BP 6.02109e-005 0.0296 5000
 An Improved Approach Combining Random PSO with BP 367

From the Table 2 and Fig.5 the same conclusions as the experiment of the function
approximation problem can be drawn.

5 Conclusions
In this paper, an improved PSO-BP approach which combines RPSO with BP is pro-
posed to train FNN. PSO has good global search ability, but the swarm loses its diver-
sity easily. However, the gradient descent algorithm has good local search ability. In
this paper, global search algorithm—RPSO is combined with BP reasonably. More-
over, in order to improve the diversity of the swarm in the PSO, RPSO are proposed
in the paper. The RPSO-BP could not only improve the diversity of the swarm but
also reduce the likelihood of the particles being trapped into local minima on the error
surface. Finally, the experimental results were given to verify that the proposed algo-
rithm has better generalization performance and faster convergence rate than other
traditional algorithms. In future research works, we shall focus on how to apply this
improved algorithm to solve more practical problems.

Acknowledgements. This work was supported by the National Natural Science

Foundation of China (No.60702056), Natural Science Foundation of Jiangsu Province
(No.BK2009197) and the Initial Funding of Science Research of Jiangsu University
(No.07JDG033).

References
1. Homik, K.: Mulitilayer feedforward networks are universal approximators. Neural
Networks 2, 359–366 (1989)
2. Chen, D.S., Jain, R.C.: A robust Back-propagation Algorithm for Function Approximation.
IEEE Trans. on Neural Network 5, 467–479 (1994)
3. Meng, J.L., Sun, Z.Y.: Application of combined neural networks in nonlinear function
approximation. In: Proceedings of the Third World Congress on Intelligent Control and
Automation, Hefei, pp. 839–841 (2000)
4. Kennedy, J., Eberhart, R.: Particle Swarm Optimization. In: IEEE International Conference
on Neural Networks, Perth, IEEE Service Cente, Piscataway, NJ, pp. 1942–1948 (1995)
5. Li, X.P., Zhang, H.Y.: Improvement of Simulated Annealing Algorithm. Software
Guide 7(4), 47–48 (2000)
6. Angeline, P.J., Sauders, G.M., Pollack, J.B.: An evolutionary algorithm that constructs re-
current neural networks. IEEE Trans. Neural Networks 5(1), 54–65 (1994)
7. Yao, X.: A review of evolutionary artifical neural networks. Int. J. Intell. Syst. 8(4),
539–567 (1993)
8. Han, F., Ling, Q.H., Huang, D.S.: Modified Constrained Learning Algorithms Incorporat-
ing Additional Functional Constraints Into Neural Networks. Information Sciences 178(3),
907–919 (2008)
9. Wang, X.G., Tang, Z., Tamura, H., Ishii, M.: A modified error function for the backpropa-
gation algorithm. Neurocomputing 57, 477–484 (2004)
10. Liu, Y.H., Chen, R., Peng, W., Zhou, L.: Optimal Design for Learning Rate of BP Neutral
Network. Journal of Hubei University of Technology 22(2), 1–3 (2007)
368 Y. Cui et al.

11. Ma, Y.Q., Huo, Z.Y., Yang, Z.: The Implementation of the Improved BP Algorithm by
Adding the Item of the Momentum. Sci-Tech Information Development & Econ-
omy 16(8), 157–158 (2006)
12. Zhang, J.-R., Zhang, J., Lok, T.-M., Lyu, M.R.: A hybrid particle swarm optimization-
back-propagation algorithm for feedforward neural network training. Applied Mathematics
and Computation 185(2), 1026–1037 (2007)
13. Guo, W., Qiao, Y.Z., Hou, H.Y.: BP neural network optimized with PSO algorithm and its
application in forecasting. In: Proc. 2006 IEEE International Conference on Information
Acquisition, Weihai, Shandong,China, August 2006, pp. 617–621 (2006)
14. Han, F., Ling, Q.H.: A New Approach for Function Approximation Incorporating Adap-
tive Particle Swarm Optimization And A Priori Information. Applied Mathematics and
Computation 205(2), 792–798 (2008)
15. Shi, Y.H., Eberhat, R.C.: Parameter selection in particle swarm optimization. In: Proc. of
1998 Annual conference on Evolutionary Programming, San Diego, pp. 591–600 (1998)
16. Shi, Y.H., Eberhat, R.C.: A modified particle swarm optimizer. In: Proc.of IEEE World
Conf. on Computation Intelligence, pp. 69–73 (1998)
17. Clerc, M., Kennedy, J.: The particle swarm:explosion,stability,and convergence in a multi-
dimensional complex space. IEEE Trans. Evolut. Comput. 6(1), 58–73 (2002)
 Fuzzy Multiresolution Neural Networks

Li Ying1,4 , Shang Qigang2 , and Lei Na3,4

1
Symbol Computation and Knowledge Engineer Lab of Ministry of Education,
College of Computer Science and Technology
liying@jlu.edu.cn
2
Department of Mechanical and Engineering, Academy of Armored Force
Engineering, Beijing, Academy of Armored Force Technology, Changchun
3
College of Mathematics
4
Jilin University, Changchun 130021, P.R. China

Abstract. A fuzzy multi-resolution neural network (FMRANN) based

on particle swarm algorithm is proposed to approximate arbitrary non-
linear function. The active function of the FMRANN consists of not
only the wavelet functions, but also the scaling functions, whose trans-
lation parameters and dilation parameters are adjustable. A set of fuzzy
rules are involved in the FMRANN. Each rule either corresponding to
a subset consists of scaling functions, or corresponding to a sub-wavelet
neural network consists of wavelets with same dilation parameters. Incor-
porating the time-frequency localization and multi-resolution properties
of wavelets with the ability of self-learning of fuzzy neural network, the
approximation ability of FMRANN can be remarkable improved. A par-
ticle swarm algorithm is adopted to learn the translation and dilation
parameters of the wavelets and adjusting the shape of membership func-
tions. Simulation examples are presented to validate the effectiveness of
FMRANN.

Keywords: fuzzy multiresolution neural network, wavelet neural

network, particle swarm optimization.

1 Introduction
Artificial neural networks (ANN) and wavelet theory become popular tools in
various applications such as engineering problems, pattern recognition and non-
linear system control. Incorporating ANN with wavelet theory, wavelet neural
networks (WNN) was first proposed by Zhang and Benveniste [1] to approximate
nonlinear functions. WNNs are feedforward neural networks with one hidden
layer, where wavelets were introduced as activation functions of the hidden neu-
rons instead of the usual sigmoid functions. As a result of the excellent properties
of wavelet theory and the adaptive learning ability of ANN, WNN can make an
remarkable improvement for some complex nonlinear system control and identi-
fication. Consequently, WNN was received considerable attention [2,3,4,5,6,7].
Especially, inspired by the fuzzy model , Daniel [8] put forward a fuzzy wavelet
model. The FWN consists of a set of fuzzy rules. Each rule corresponding to a

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 369–378, 2009.


c Springer-Verlag Berlin Heidelberg 2009
370 L. Ying, S. Qigang, and L. Na

sub-wavelet neural network consists of single fixed scaling wavelets, where the
translation parameters are variable. Such sub-WNNs at different resolution level
can capture the different behaviors of the approximated function. The role of the
fuzzy set is to determine the contribution of different resolution of function to
the whole approximation. The fuzzy model can improve function approximation
accuracy given the number of wavelet bases. But such FWNN require a complex
initialization and training algorithm. During the initialization step, the dilation
values of such FWNN is fixed beforehand through experience and a wavelet
candidate library should be provided by analysis the training samples.
As we know, the scaling function corresponds to the global behavior (low
frequency information) of the original function, while the wavelets corresponds
to the local behavior (high frequency information) of the original function. The
FWNN in [8] loses sight of the low frequency, though which may be approximated
by wavelet functions. But due to the theory of MRA, it would need much more
wavelet nodes.
Incorporating the idea of FWNN with pre-wavelet, we propose a fuzzy pre-
wavelet neural network. The hidden of layer of FMRANN consist of not only the
scaling function nodes, but also wavelet function nodes. The structure of FM-
RANN includes two sub-neural networks: sub-scaling function neural networks
and sub wavelet neural networks. The contribution of such sub-neural networks
is determined fuzzy rules.
In the learning of artificial neural network (ANN), the back propagation algo-
rithm by Rumelhart, Hinton, and Williams has long been viewed as a landmark
event. With the increasing of complexity of the practical problems, the standard
back propagation algorithm based the single gradient has ability not equal to
its ambition, which motivated many researchers to develop enhanced training
procedures with exhibit superior capabilities in terms of training speed, map-
ping accuracy, generalization, and overall performance than the standard back
propagation algorithm. The training method based upon second-order derivative
information exhibit better efficient and promising. Primary second-order meth-
ods are the back propagation algorithm based on quasi-Newton, Levenburg-
Marquardt, and conjugate gradient techniques. Although these methods have
shown promise, they often lead to poor local optima partially attributed to the
lack of a stochastic component in training procedure. Particle swarm optimiza-
tion (PSO) in [9,13] is a stochastic population-based optimization technique.
The simplicity of implementation and weak dependence on the optimized model
of PSO make it a popular tool for a wide range of optimization problems.
A fuzzy multi-resolution neural network (FMRANN) based on modified par-
ticle swarm algorithm is proposed to approximate arbitrary nonlinear function
in this paper. The basic concepts of Multi-resolution are introduced in Section 1.
In section 2, the two different structure of fuzzy multi-resolution neural network
(FMRANN) is presented and Aiming at the features of FMRANN, a modified
particle swarm optimization is introduced to update the parameters. Two simu-
lation examples are utilized to illustrate the good performance of the FMRANN
compared with other methods in Section 4 .
 Fuzzy Multiresolution Neural Networks 371

2 Multi-Resolution Analysis (MRA)

The main characteristics of wavelets are their excellent properties of
time-frequency localization and multi-resolution. The wavelet transform can cap-
ture high frequency (local) behavior and can focus on any detail of the observed
object through modulating the scale parameters, while the scale function cap-
tures the low frequency (global behavior). In this sense wavelets can be referred
to as a mathematical microscope. Firstly, we give a simplified review of wavelets
and multiresolution analysis of L2 (R).
A series sequence Vj , j ∈ Z of closed subspaces in L2 (R) is called a multires-
olution analysis (MRA) if the following holds:
(1) · · · ⊂ Vj ⊂ Vj+1 · · · .
(2) f (·) ∈ Vj if and only if f (2·) ∈ Vj+1 .
 
(3) j∈Z Vj = {0}, j∈z Vj = L2 (R).
(4) There exists a function φ ∈ V0 , such that its integer shift {φ(· − k), k ∈ Z}
form a Riesz basis of V0 .
It is noted that the function φ is a scaling function, which satisfies the following
refinement equation: √
φ(x) = 2 hn φ(2x − n),
n
where {hn } is low frequency filter (scaling fiter). For j ∈ Z, k ∈ Z, denote
φj,k (·) = 2j/2 φ(2j · −k).
For j ∈ Z, {φj,k , k ∈ Z} forms a Riesz basis of Vj . The associated wavelet
function ψ(x) satisfies the following equation:
√
ψ(x) = 2 gn φ(2x − n),
n

where {gn } is a high frequency filter (wavelet filter). For j, k ∈ Z, denote

ψj,k (·) = 2j/2 ψ(2j · −k),
then {ψj,k , k ∈ Z} forms a Riesz basis of Wj , where Wj is the orthogonal com-
plement of Vj in Vj+1 , i.e.
Vj+1 = Vj ⊕ Wj ,
where Vj is scaling function space and Wj represents difference between spaces
spanned by the various scales of scaling function, referred to as detail space or
wavelet function space. Obviously,
∞
 ∞

L2 (R) = VJ ⊕Wj , L2 (R) = ⊕Wj .
j=J j=−∞

For any function f (x) ∈ L2 (R) could be written

∞
f (x) = cJ,k φJ,k (x) + dj,k ψj,k (x), (1)
k∈Z j≥J k∈Z
372 L. Ying, S. Qigang, and L. Na

as a series expansion in terms of the scaling function and wavelets, i.e. as a

combination of global information and local information at different resolution
levels. On the other hand, f (x) could also be represented as

f (x) = dj,k ψj,k (x), (2)

j∈Z k∈Z

as a linear combination of wavelets at different resolutions levels, i.e. the combi-

nation of all local information at different resolution levels.
The above equations indeed embody the concept of multiresolution analysis
and imply the ideal of hierarchical and successive approximation, which is the
main motivation of our work.
The common used wavelets are continuous wavelets, Daubecies wavelets and
spline wavelets and so on. In WNN, the frequently adopted active functions are
continuous wavelets such as Morlet, Mexican Hat, Gaussian derivatives wavelets,
which do not exist scaling function.
Daubechies compact support orthonormal wavelets constructed by Irigrid
Daubechies become a milestone of wavelet fields, which make wavelet rapidly
overspread almost all engineer fields from theory to application.
Daubecies wavelets is provided by filters. The birth of Daubecies wavelets
make a more extensive application. But unfortunately, they have not analytic
formula, which limited some application of demanding analytic formula such as
WNN.
Here the cascade algorithm i.e. the single-level inverse wavelet transform re-
peatedly [11], is employed to approximate the corresponding scale function and
wavelet function, which are adopted as active functions in our proposed fuzzy
multi-resolution neural networks.
In common WNN, taking Daubechies orthonormal wavelets as active func-
tions, due to the un-redundancy of orthonormal wavelet, such kind of WNN can
provide a unique and efficient representation for the given function, but which
highly relies on the selected wavelets and lead to poor robustness. In our work
we use not only Daubechies orthonormal wavelets but also the according scal-
ing function to construct our proposed Fuzzy multi-resolution neural network.
Obviously, the adoption of the fuzzy mechanism and the introduction of scaling
function can provide better robustness.

3 Fuzzy Multiresolution Neural Network

In this subsection, a fuzzy multiresolution neural network (FMRANN) is pro-
posed to approximate arbitrary nonlinear function. The active function of the
FMRANN consists of not only the wavelet functions, but also the scaling func-
tions, whose translation parameters and dilation parameters are adjustable. A
set of fuzzy rules are involved in the FMRANN. Each rule either corresponding
to a subset consists of scaling functions, or corresponding to a sub-wavelet neu-
ral network consists of wavelets with same dilation parameters. Incorporating
the time-frequency localization and multi-resolution properties of wavelets with
 Fuzzy Multiresolution Neural Networks 373

the ability of self-learning of fuzzy neural network, the approximation ability of

FMRANN can be remarkable improved.
Artificial neural networks mimic the structure and function of neural models
of human brain. The significant characteristics of ANN are known as the paral-
lel processing ability and learning ability through examples. ANN have shown
its outstanding performance on many complex nonlinear problems even though
ANN is far away from the elaborating structure and function mechanism of real
human brain.
Fuzzy logic is based on the attempt to emulate the human reasoning mecha-
nism, which is not numbers but rather indicators fuzzy sets.
The property of Multi-resolution analysis of wavelets conform with the nature
of audition and vision of human.
In this section, we mainly propose the models of Fuzzy Multi-resolution Neural
Network which contains two fuzzy boxes referred as scale fuzzy box and wavelet
fuzzy box.
Let x = (x1 , x2 , · · · , xN ) ∈ RN , y ∈ R. Φ(x) and Ψ (x) are respectively scaling
function and wavelet function of L2 (RN ). For simplification, denote
Φk = det(Dk0 )Φ(Dk0 x − t0k ),
and
Ψk = det(Dk1 )Ψ (Dk1 x − t1k ).
Note that for i=0,1, Dki = diag(aik,1 , · · · , aik,N ), aik,1 > 0, is dilation matrix.
tik = (bik,1 , · · · , bik,N ), bik,1 ∈ R, is translation vector.
Incorporate with the idea of multi-resolution analysis and fuzzy logic, we give
the description of FMRANN as follows:
c
y= µ̂i (x)ŷi (x), (3)
i=1

where
µi (x)
µ̂i (x) = 
c ,
µi (x)
i=0
0

M
ŷ 0 (x) = wk0 Φk (x),
k=1
1

M
ŷ 1 (x) = wk0 Ψk (x).
k=1
From figure 1 to figure 2, we respectively show the graphs of scaling functions
and wavelet functions with different support length and vanishing moment which
are used for our FMRANN.

3.1 Training Algorithm of Fuzzy Multi-Resolution Neural Network

Based on Modified Particle Swarm Optimization
Particle swarm optimization (PSO) is a stochastic and population-based opti-
mization technique, which first introduced by Kennedy and Eberhart [9,13] in
374 L. Ying, S. Qigang, and L. Na

1.4 2

1.2
1.5

1
1
0.8

0.6 0.5

0.4 0

0.2
−0.5
0

−1
−0.2

−0.4 −1.5
0 2 4 6 0 2 4 6

Fig. 1. The graph of scaling function φ and wavelet function ψ according to Daubecies
wavelet with support width 5 and vanishing moment 3

1.2 1.5

1
0.8

0.6
0.5

0.4

0
0.2

0
−0.5

−0.2

−0.4 −1
0 2 4 6 8 0 2 4 6 8

Fig. 2. The graph of scaling function φ and wavelet function ψ according to Daubecies
wavelet with support width 7 and vanishing moment 4

1995. Particle swarm optimization method mimics the development of this tech-
nique was inspired by the animal social behaviors such as school of fish, flock
of birds etc and mimics the way they find food sources. PSO has gained much
attention and wide applications in different fields. Compared with genetic algo-
rithm (GA), PSO employs a cooperative strategy while GA utilizes a competitive
strategy. In nature, cooperation is sometimes more benefit for survival than com-
petition. For example,the PSO has shown its superior ability than CA in some
problems such as optimizing the parameters of artificial neural network (ANN)
[14]. Gradient decent is another popular algorithm for the learning of ANN,
which is a good choice when the cost function is unimodal. But the majority of
the cost functions are multimodal, gradient decent might lead to convergence to
the nearest local minimum. But PSO has the ability to escape from local minima
traps due to its stochastic nature.
In PSO, each particle represents an alternative solution in the multi-dimensional
search space. PSO can find the global best solution by simply adjusting the trajec-
tory of each particle towards its own best location and towards the best location
of the swarm at each time step (generation).
 Fuzzy Multiresolution Neural Networks 375

The implement of PSO can be categorized into the following main steps:
(1). Define the Solution Space and a Fitness Function:
(2). Initialize Random Swarm Position and Velocities:
(3). Updating particle Velocities and Position
v(k + 1) = w ∗ v(k) + c1 ∗ rand()(pbest − x)
+c2 ∗ rand() ∗ (gbest − x);
x(k + 1) = x(k) + v(k + 1)
where wis inertial weight, c1 and c2 are constriction factors. Use of inertia weight
and constriction factors has made the original implementation of the technique
very efficient [9,13].
The inertial weight factor provides the necessary diversity to the swarm by
changing the momentum of particles and hence avoids the stagnation of particles
at local optima. The empirical investigations shows the importance of decreasing
the value of from a higher valve (usually 0.9 to 0.4) during the search.
The majority problem of PSO is premature convergence of particle swarm.
How to efficiently avoid such case is considerably important. Various mechanisms
have been designed to increase the diversity among the particles of a swarm.
Here we propose an mechanism called forgetfulness to improve the the diver-
sity of particle the premature convergence. The basic ideal is that we impose each
particle and the swarm slightly amnesia when the particle swarm trend to pre-
mature convergence. This mechanism give the particles restarting chance based
on the existing better condition. Firstly, define a merit to reflect the convergent
status of the particle swarm as follows:
σf2
τ=
max1≤j≤N {(fi − f¯)2 }
Where
N N
1 1
f¯ = fi , σf2 = (fi − f¯)2
N i=1
N i=1

Where fi is the fitness of the i-th particle, N is the size of the particle swarm, f¯ is
the average fitness of all particles, and σf2 is the covariance of fitness. For a given
small threshold, if τ is less than this threshold and the expected solution haven’t
reached, then we think that this particle swarm tends to premature convergence.
Under this condition, we impose each particle and the swarm slightly amnesia.
Each particle forget its historical best position and consider the current position
as its best position. Similarly, the swarm don’t remember its historical global best
position and choose the best position from the current positions of all particles.

4 Simulation
In the following, the concrete numerical examples are given to demonstrate the
validity of the presented FMRANN. In this section, for evaluating the perfor-
mance of the network, we take the following merits defined in [1] as a criterion
to compare the performance of various methods.
376 L. Ying, S. Qigang, and L. Na

10 1
original function error data
8 estimated result from FMRANN 0.8 zero line

6 0.6

4 0.4

2 0.2

0 0

−2 −0.2

−4 −0.4

−6 −0.6

−8 −0.8

−10 −1
−10 −8 −6 −4 −2 0 2 4 6 8 10 −10 −8 −6 −4 −2 0 2 4 6 8 10

Fig. 3. The left figure is the comparison between original function and estimated re-
sult from FMRANN with performance index 0.019912. The right figure is the error
between original function and estimated result from FMRANN with performance in-
dex 0.019912.

10 1
original function error data
8 estimated result from FMRANN 0.8 zero line

6 0.6

4 0.4

2 0.2

0 0

−2 −0.2

−4 −0.4

−6 −0.6

−8 −0.8

−10 −1
−10 −8 −6 −4 −2 0 2 4 6 8 10 −10 −8 −6 −4 −2 0 2 4 6 8 10

Fig. 4. The left figure is the comparison between original function and estimated re-
sult from FMRANN with performance index 0.015591. The right figure is the error
between original function and estimated result from FMRANN with performance in-
dex 0.015591.



 N N
1
N
J = ( yl − ŷl )/( yl − ȳ), ȳ = yl
N
l=1 l=1 l=1

Where yl is the desired output and ŷl is the estimated output from the con-
structed neural networks.
We respectively selected spline wavelets with order 3 and Daubechies wavelet
with support width 2N-1 and vanishing moment N. We compare our FMRANN
with other works in [1], [7] and [8], which take the Mexican Hat function as
wavelet function.
Example I - Approximation of univariate piecewise function:
⎧
⎨ −2.186x − 12.864, −10 ≤ x < −2
f (x) = 4.246x, −2 ≤ x < 0 . (4)
⎩
10e−0.05x−0.5sin[(0.03x + 0.7)x], 0 ≤ x ≤ 10
 Fuzzy Multiresolution Neural Networks 377

Table 1. Comparison of our FMRANN with others work for Example I

Method Number of unknown parameters Performance index J

used for training in different networks
FMRANN-prewavlets 23 0.016547
FWN[8] 28 0.021
WNN [1] 22 0.05057
WNN [7] 23 0.0480

Table 2. Comparison of our FMRANN with others work for Example II

Method Number of unknown parameters Performance index J

used for training in different networks
FMRANN-prewavlets 45 0.018023
FWN [8] 107 0.0147
WNN [1] 442 0.03395

Example II - Approximation of bivariate function:

y = (x21 − x22 ) ∗ sin(0.5 ∗ x1 ), x1 , x2 ∈ [−10, 10]. (5)

In the simulation, we sampled 200 points distributed uniformly over [0 , 10] and
[-10,10] respectively as training data. In Table I, we compare the performance of
our FMRANN with other WNNs, here the number of rules used in FWN is 6. In
Table 1. and Table 2. , the compared results with others work are given. From the
Fig.3 to Fig.6., the approximation results and error for example I are provided.
It should be noticed that the performance of FMRANN shown in the Table 1
and Table 2 is the average of the nine simulations. Each simulation is computed
on uniformly sampled test data of 200 points. Among the results of FMRANN
for example I, the maximal Error value is 0.019912 and the minimal Error value
0.015591. Among the results of FMRANN for example II, the maximal Error
value is 0.0326 and the minimal Error value 0.01320.
Figure. 3 shows the excellent performance of our FMRANN, which corresponds
to the worst performance value among the nine simulations for example I.
Figure. 4 shows the excellent performance of our FMRANN, which corresponds to
the best performance value among the nine simulations for example I. Obviously,
the performance of our FMRANN with considerable smaller parameters is superior
to that of other WNNs. Our constructed FMRANN is significant and effective.

References
1. Zhang, Q.H., Benveniste, A.: Wavelet networks. IEEE Trans. Neural Netw. 3(6),
889–898 (1992)
2. Pati, Y.C., Krishnaprasad, P.S.: Analysis and synthesis of feedforward neural net-
works using discrete affine wavelet transformation. IEEE Trans. Neural Networks 4,
73–85 (1993)
378 L. Ying, S. Qigang, and L. Na

3. Zhang, J., Walter, G.G., Lee, W.N.W.: Wavelet neural networks for function learn-
ing. IEEE Trans. Signal Processing 43, 1485–1497 (1995)
4. Zhang, Q.: Using wavelet networks in nonparametric estimation. IEEE Trans. Neu-
ral Networks 8, 227–236 (1997)
5. Alonge, F., Dippolito, F., Mantione, S., Raimondi, F.M.: A new method for optimal
synthesis of wavelet-based neural networks suitable for identification purposes. In:
Proc. 14th IFAC, Beijing, P.R. China, June 1999, pp. 445–450 (1999)
6. Li, X., Wang, Z., Xu, L., Liu, J.: Combined construction of wavelet neural networks
for nonlinear system modeling. In: Proc. 14th IFAC, Beijing, P.R. China, June 1999,
pp. 451–456 (1999)
7. Chen, J., Bruns, D.D.: WaveARX neural network development for system identi-
fication using a systematic design synthesis. Ind. Eng. Chem. Res. 34, 4420–4435
(1995)
8. Daniel, J., Ho, W.C., et al.: Fuzzy wavelet networks for function learning. IEEE
Trans. on Fuzzy Systems 9(1), 200–211 (2001)
9. Eberhart, R.C., Kennedy, J.: A new optimizer using particle swarm theory.
In: Proc. 6th Int. Symp. Micro Machine and Human Science, Nagoya, Japan,
pp. 39–43 (1995)
10. Daubechies, I.: Ten lectures on wavelets CBMS. SIAM 61, 194–202 (1994)
11. Strang, G., Nguyen, T.: Wavelets and Filter Banks. Wellesley-Cambridge Press
(1996)
12. Chui, C., Wang, J.: A general framework for compactly supported splines and
wavelets. Proceedings of Americal Mathematical Soceity 113, 785–793
13. Eberhart, R.C., Simpson, P., Dobbins, R.: Computational Intelligence PC Tools:
Academic, ch. 6, pp. 212–226 (1996)
14. Eberhart, R.C., Shi, Y.: Comparison between genetic algorithms and particle
swarm optimization. In: Porto, V.W., Saravanan, N., Waagen, D., Eiben, A.E.
(eds.) EP 1998. LNCS, vol. 1447, pp. 611–618. Springer, Heidelberg (1998)
 Research on Nonlinear Time Series Forecasting of
Time-Delay NN Embedded with Bayesian Regularization

Weijin Jiang1, Yusheng Xu2, Yuhui Xu1, and Jianmin Wang1

1
School of computer, Hunan University of Commerce, Changsha 410205, China
jwjnudt@163.com
2
China north optical-electrical technology, Beijing 100000, China
yshxu520@163.com

Abstract. Based on the idea of nonlinear prediction of phase space reconstruc-

tion, this paper presented a time delay BP neural network model, whose gener-
alization capability was improved by Bayesian regularization. Furthermore, the
model is applied to forecast the imp&exp trades in one industry. The results
showed that the improved model has excellent generalization capabilities, which
not only learned the historical curve, but efficiently predicted the trend of busi-
ness. Comparing with common evaluation of forecasts, we put on a conclusion
that nonlinear forecast can not only focus on data combination and precision im-
provement, it also can vividly reflect the nonlinear characteristic of the forecast-
ing system. While analyzing the forecasting precision of the model, we give a
model judgment by calculating the nonlinear characteristic value of the com-
bined serial and original serial, proved that the forecasting model can reasonably
‘catch’ the dynamic characteristic of the nonlinear system which produced the
origin serial.

1 Introduction
Basically, the traditional time serial forecasting model was built on the foundation of
statistic technology, they are mainly subjected to linear model, which has advantage
of simple nature and strong explanation ability, but as for the forecasting of tremen-
dous system’s evolution serial, it fails, especially when the change of real economic
system is strong, high level nonlinear characteristic and classical chaos appears in the
system, the common forecasting method can hardly handle it. Chaos theory is an
important discovery in nonlinear dynamic, which come into being from 1960’s, until
1980’s, it have developed to be a new study with special concept system and method
frame. Because of this, people have some new acknowledgement on the complexity
of time serial. According to utilize the phase space reconstruction technology of cha-
otic theory, economic time serial is embedded in the reconstructed phase space, and
with the help of fractal theory and symbol dynamics, the complex moving characteris-
tic of economic dynamic system can be described, furthermore, the inner rule can be
found and the forecasting report is produced. Therefore, it has a realistic meaning and
scientific value[1-6] to find the evolution rule of economic time serial on the stand of
nonlinear dynamics and study the chaotic characteristic and nonlinear forecasting of
complex economic dynamic system.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 379–388, 2009.
© Springer-Verlag Berlin Heidelberg 2009
380 W. Jiang et al.

Artificial neural network is a kind of nonlinear dynamics system, which possesses

strong robustness and error-tolerance. The powerful ability to draw on nonlinear re-
flection guarantee it has special advance in solving nonlinear problems, and this make
it suitable for the modeling and forecasting of changeable complex economic system
and economic serial. According to some certain structure setting, the study has proved
that neural network forecasting had obvious advantage when the system possessed
nonlinear characteristic. Reference [1] firstly gives a try on using neural network to
study and forecast the tome serial data which the computer produced, then neural
network forecasting application like Matsuba is used in stock market, after that, there
were a lot of product of neural network application appeared in neural network inter-
national meeting. Up to this day, the scholars has applied neural network in every
occupation of economic, such as securities, exchange ratio, GDP forecasting, finan-
cial alarm and market demand forecasting[2]. Paper [3,4] gives a demonstration about
the relation between phase space reconstruction and forecasting, theoretically, it is
reasonable, but the calculation of embedded dimension is usually be fake nearest spot
method, and the way of embedded time-delay calculation is average information
method. Paper [5,6] discusses respectively about both their advantages and disadvan-
tages, and comes a conclusion that there maybe some certain error by using these
methods, which has been mentioned before, to calculate embedded dimension and
time-delay because of short actual data and noise, as a results, reconstruction state
vector based on these method may receive a bad forecasting report. In paper [7]’ se-
lection about embedded dimension follows the rule that the dimension possess mini-
mum error, whose appearance come along with tremendous times of forecasting, can
be promoted to embedded dimension, however, the real dimension may not show up
because the times of common is limited, and this method can only cut down the error
to a local minimum, furthermore, many aimless experiment will waste time, finally[8],
the data is always divided into two parts (training set and forecasting set), determining
embedded dimension directly by the forecasting error in forecasting set is not useful,
because in actual forecasting, forecasting error can not be gained.
These studies usually use BP neural network model—error reverse transform neural
network, which contains a great many parameters, besides, these parameters always
judged by experience, as a result, the model is hard to establish. The large repetitious
training and experiment is a kind of ‘Over Fitting’ which may give the training sample
a fine combination effect, but as for new input sample, the output may have big differ-
ence with the objective value, which has no popularization capability. Conversely
Bayesian Regularization can overcome this problem, and it can improve the generali-
zation capability of network. That why this paper founded a new model for nonlinear
time serial forecasting, which is called Time Delay Neural Network (acronym: TDBP)
forecasting model by some kinds of tricks, such as following phase space reconstruc-
tion theory, utilizing G-P algorithm to calculate the saturation embedded dimension—
the number of input-level’s spot, combining Bayesian Regularization to certain the
number of the hidden-lever’s spot and improving the generalization capability of net-
work. In this model, when the forecasting error of certain set is relatively small, the
forecasting error of forecasting error will be accordingly tiny. The results of experi-
ment proved that chaos time serial forecasting and control system based on this method
possessed fine precision, furthermore, it can track the original time serial quickly,
 Research on Nonlinear Time Series Forecasting of Time-Delay NN 381

besides, it has faster response capability and smaller number of iterate than other simi-
lar methods.
As for the nonlinear characteristic of trade serial, this paper put the established
TDBP model into some certain occupation’s trade forecasting, according to adopt the
actual amount of imp&exp of 1/1989-5/2003 as the ample of training network, a new
model was founded[9,10], which successfully forecast and combine this occupation’s
international trade affairs[8]. Meanwhile, on the foundation of analyzing forecasting
precision, this paper gave a capability evaluation for the nonlinear forecasting model
according the maximum Lyapunov index and the relation dimension between com-
bined serial and original data.

2 Phase Space Reconstruction and BP Model Based on Bayesian

Regularization
The intention of forecasting is to deduce object’s future by its own develop rule. Ac-
cording to Kolmogorov theorem, all time serial can be regarded as a substitute of
input-output system which is certain by nonlinear mechanism. Accompany with the
development and maturity of neural network theory and phase space reconstruction
technology, a new road is provided for nonlinear forecasting of economic time serial.
Suppose disperse time serial is {X(t)}(t=1, 2, …, T), select proper embedded di-
mension m, time-delay τ and reconstruction phase space {Y(j)}(j=1, 2, …, N, N=T-
(m-1)), and:
(1)
Y(j)=[X(j),X(j- τ ),X(j-2 τ ),…,X(j-(m-1) τ )]
According to Takens embedding theorem, there is reflection F ：R →Rm m
makes Y
（ ） （ （ ））
j+ τ =F Y j , it is:

⎡X ( j +τ ) ⎤ ⎛ ⎡ X ( j) ⎤⎞
⎢ X ( j) ⎥ ⎜⎢ ⎥⎟
⎢ ⎜
⎥=F ⎢ X ( j − τ ) ⎥⎟
(2)

⎢M ⎥ ⎜ ⎢M ⎥⎟
⎢ ⎥ ⎜ ⎟
⎜ ⎢ X ( j − (m − 1)τ ⎥ ⎟
⎣ X ( j − (m − 2)τ ⎦ ⎝ ⎣ ⎦ ⎠
Then forecasting reflection F : Rm→Rm can be denoted as:
X(j+ τ )= F (X(j),X(j- τ )…,X(j-(m-1) τ )) (3)

Here, it is a nonlinear needed to evaluate, definitely, to find a concrete function ex-

pression is difficult, But the function can be obtained by utilizing neural network’s
capability of approaching nonlinear reflection.
Usually, when the size of training set is certain, the generalization capability of
network directly concern with its scale, if the scale of neural network is far smaller
than the size of training sample set, then the change of ‘over-training’ is small. Based
on this idea, reference [8] change neural network’s object function to:
F( w ) = βsse + αssw (4)
382 W. Jiang et al.

And sse denotes the error’s square sum, ssw denotes the square sum of all power
coefficient’s valve value in the network, α , β are the coefficient of objective func-
tion. Suppose that network’s power value is random variable, after the output data is
figured out, we can use Bayesian formula to revamp the ratio density function of
power value:
P ( D | w, β , M ) P ( D | w | α , M )
P ( w | D, α , β , M ) = (5)
P( D | α , β , M )
And w denotes network power value’s vector, D denotes data set, M denotes network
model. P ( D | α , β , M ) is full probability, P ( w | α , M ) denotes the first check
density function of power vector, P ( D | w, β , M ) is similar function of timing
output, commonly supposed that noise and power vector exist in data follow the
Gauss distribution, it means:

P ( D | w, β , M ) = e − βsse / Z D ( β ) (6)

P( w | α , M ) = e −αssw / Z w (α ) (7)

And Z D ( β ) = (π / β ) n / 2 , Z w (α ) = (π / α ) N / 2 , n denotes the number of sample,

N denotes the number of power value, put (6), (7) in formula (5), noticed that the
standard factor P ( D | α , β , M ) had no relation with power vector w, so:

e − (αssw+ βsse) /[ Z w (α ) Z D ( β )]
P ( w | D, α , β , M ) = (8)
P( D | α , β , M )
= e − F ( w) / Z F (α , β )
From formula (8), we can see that the optimal power value posses the biggest last
check probability, and maximum last validation probability equals to the minimum
regularization objective function F( w ) = βsse + αssw . If α << β , network train-
ing can reduce error to minimum; and if α >> β , network training will automati-
cally reduce the efficient network parameters, which is to make up to big network
error. On the condition that the error of network training is as small as possible, it is
evident that new capability objective function is efficient, which can minimize the
efficient parameter and lessen the scale of network.
Here, we adopt Bayesian method to certain the optimal regularization parame-
ter α , β , suppose α , β ‘s first validation function is P (α , β , M ) , then the last is:

P(D | α, β, M)P(α, β, M)
P(α, β | D, M) = (9)
P(D | M)
 Research on Nonlinear Time Series Forecasting of Time-Delay NN 383

Formula (9) illustrate that the last validation of maximum regularization parame-
ter α , β is equal to maximum similar function P ( D | α , β , M ) . Combining (5) and
(8), formula (10) can be concluded:
Z F (α , β )
P( D | α , β , M ) = (10)
Z w (α ) Z D ( β )
（）
In order to figure Z F (α , β ) out, unfolding F w on the minimum point w*, because
gradient is 0, then:
F(w)=F(w*)+(w-w*)TH(w*)(w-w*)/2 (11)

At least, Hison matrix H ( w*) = ∇ 2 F ( w*) is semi-symmetry, so it can be written

1 1 1
as H = ( H 2 )T H , suppose that u = H
2 2 (w − w*) . Putting formula (11) into
formula (8), and integrating the two sides:
1
1= e − F (w*)∫ N e −uTu / 2[det(H (w*)−1)]1/ 2 du
Z F (α , β ) R (12)

(2π ) N / 2 − F ( w*)
= e [det( H ( w*)−1 )]1/ 2
Z F (α , β )
Then:

ZF(α, β ) = (2π ) N / 2 e−F (w*)[det(H(w*)−1 )]1/ 2 (13)

Putting formula (13) into formula(10), obtaining logarithm from both sides, besides,
utilizing the one-power condition of optimal value to obtain the optimal regularization
parameter:
γ n−γ
α* = , β* = (14)
2 E w ( w*) 2 E D ( w*)

And γ = N − 2atr ( H ) denotes the number of efficient network parameter, which

−1

reflect the actual scale of network. Paper [8] pointed out that: as for Hison matrix H,
when using quick algorithm Levenberg-Marquardt to train the network, it was easy to
approach by Gauss-Network method. In the process of training, sse and ssw can be
the factor to figure out the number of nerve cell of hidden level (marked as Nhl) by
following the efficient parameter γ , after certain times of iterate, when these three
parameters is in a forever status or smaller, it gives a sign that this network is in con-
vergence status, and it is the time to stop training; selecting a smaller Nhl, continu-
ously increasing the number of hidden nerve cell until current Nhl makes network
convergence, then the state of sse and ssw will be unchangeable, as a result, this Nhl
can be regarded as the spot number of hidden-level.
This paper adopts single-output three-level BP neural network, separately, the
stimulating function of these three levels is logsig, logsig, pureline. The nerve cell
384 W. Jiang et al.

number of input level equals to the embedded dimension number of nonlinear time
serial, besides, many study has proved that the number of maturity embedded dimen-
sion of time serial can be the nerve cell number of network input level. Here, we ob-
tained the maturity embedded dimension number by utilizing G-P algorithm. we use
MatLab language to build the concrete model and on the aid of tool box function
trainbr() Bayesian Regularization algorithm is realized. Fig.1 gives out the basic
frame of the model.

Nhl=Nhl
Figure +1
the Find
Num- N
The
ber Build Result
Of Y num- Ana-
Giv- ber. net-
Input
Level
spot
Study
Sam-
ple
ing
train-
ing
，
doesγ
sseey
Of
Hid-
work
Fore-
cast
Emu-
lator
And
lyze
And
Fore-
And den
Origi- com- ssw is level Mode Fore- cast
nal Tutor bine-d steady l cast evalua-
Data signal Nl/L- ? tion
Selec- M
tion And
BR
Algo-
Data rithm
uni-
tary

Application
Data dealing in Founding the forecasting /evaluation
advance model

Fig. 1. TDBP forecasting model based on Bayes Regularization

3 The Selection of Trade Data and the Build of Forecasting Model

Considering that the scale of international trade in our country is small before 1990,
because the opening strategy for world was still in the tentative stage. But there were
lots changes in the international trade environment and trade system after 1990, every
kinds of nonlinear factor came out off water, therefore, we selected the month data of
one corporation’s imp&exp business from January 1st 1989 to June 2003 as the train-
ing sample. Before the network study, we did unitary operation for the original data.

X ' (t ) − Min{ X ' (t )}

X (t ) = (15)
Max{ X ' (t )} − Min( X ' (t )}
And X’(t) denotes original serial, X(t) denotes unitary serial.
We utilized G-P algorithm to confirm the relation dimension of imp&exp serial,
and then obtained the maturity embedded dimension: when the embedded dimension
of imp&exp trade is 13, relation dimension is mature; when it embedded dimension
 Research on Nonlinear Time Series Forecasting of Time-Delay NN 385

Table 1. Import trade

Nhl sse ssw N γ

2 0.2773 38.9909 213 27.3
4 0.2780 35.8799 243 27.0
5 0.2780 36.8654 258 27.0
6 0.2780 35.7243 273 26.9
8 0.2780 35.6423 303 26.9
10 0.2780 35.7176 333 26.9

Table 2. Export trade

Nhl sse ssw N γ

4 0.2164 22.1673 345 17. 7
6 0.2160 22.1450 381 17. 6
7 0.2160 22.1971 399 17. 6
8 0.2159 22.2076 417 17. 6
10 0.2159 22.2452 453 17. 6
12 0.2159 22.2283 489 17. 6

comes to 16, relation dimension is still mature. The time-delay τ is 1. When training
the network, we adopted unitary data of imp&exp trade’s month data from 1/1989 to
5/2003 to do the phase space reconstruction, and then respectively obtained 13-
dimension matrix or 16-dimension matrix which is be the study sample; the tutor
signal of network model is the unitary data of imp&exp trade’s month data from
2/1990 to 6/2003 and from 5/1990 to 6/2003 separately. As for different Nhl, utilizing
these study samples and tutor signals to do network training until the network is in the
status of convergence, the obtained parameter is illustrated by table.1 and table 2.
The network training parameter in table.1 demonstrates that: when network model
of imp&exp trade is under the condition that N hl ≥ 6 or N hl ≥ 8 , γ , sse and ssw
is steady, as a results, the hidden level nerve cell number of nonlinear forecasting
model of imp&exp trade are 6 and 8. On the foundation of these parameters, TDBP
network model of imp&exp trade can be established, which can give nonlinear
forecast for future development.

4 Example Verification
This paper used multi-steps forecasting method, which feedback the forecasting value
to the input end, reconstruct input matrix and do next step forecasting. Here the 12-
step of 7/2003-6/2004 is given out: as for the data of 6/2003-12/2003, we can com-
pare it directly from the original data, from which we can judge the generalization
capability of neural network and the forecasting precision of non-linear model; the
forecasting results of 1/2004-6/2004 shows the develop trend of our country’s
386 W. Jiang et al.

Table 3. The actual month data of imp&exp trade and forecasting results from 7/2003 to
12/2003 (10 thousand dollar)

Export trade data import trade data

Actual Forecasting Relative Forecasting Relative
Actual value
value value (error%) value error%
July 380.9 377.02 -1.0 365.1 367.92 0.8
August 374 382.69 2.3 346.2 379.89 9.7
September 419.3 405.26 -3.3 416.5 389.23 -6.5
October 409.1 387.8 -5.2 351.9 377.42 7.3
November 417.6 397.62 -4.8 368.9 404.14 9.6
December 480.6 411.95 -14.3 423.4 429.32 1.4

Table 4. Imp&exp trade forecasting results from 1/2004 to 6/2004 (10 thousand dollar)

January February March April May June sum

export 395.3 347.64 417.67 458.92 441.58 444.94 2506.05
import 448.37 405.84 469.43 489.43 486.12 486.04 2785.23

Table 5. Nonlinear characteristic calculation result of original month serial and forecast serial

import export
characteristic sample
Original serial Forecast serial Original serial Forecast serial
Relation dimension 1.4576 1.4206 1.6819 1.9697
Maximum Lyapunov index 0.0061 0.0060 0.0115 0.0167

imp&exp trade. The actual month data of imp&exp trade and forecasting results is
generalized in table.3 and table.4.
The results of table.3 shows that forecasting value is very close to the actual value
except the data of 1/2003’s export trade, and the relative error is under 10 percent,
single step forecasting relative error is below 1 percent. Meanwhile combining the
figure and forecasting data, apparently, the seasonal change trend of imp&exp trade
can be predicted, and the forecasting point it out that the amount of imp&exp trade of
the second half year of 2003 is larger than the first half one, when it come to Decem-
ber, the volume steps to the maximum number, then imp&exp volume of January and
February suffer a slow down, but January is little bigger than February, all the change
is in accordance with the actual development. Usually, the corporation like to finish the
good exchange at the end of one year, and the custom’s statistic is balanced before the
new year, because China’s spring festival is always in January or February, as a results,
the amount of December is usually the largest one, January and February the least,
owing to the lag factor of last year, amount of January will bigger than February. The
 Research on Nonlinear Time Series Forecasting of Time-Delay NN 387

forecasting demonstrated that: there will be a trade deficit partly because of model
factor; actually, it is also the results of environment change. On the stand of interna-
tional, the value of dollar may rise, which can hoist the value of RMB, unfortunately,
this can sharpen down the capability of China’s export trade and prompts import deals.
After the Cancun meeting, EU and USA will give more attention on business coordina-
tion of bilateral and distribute economic, which can prompt the coordination of distrib-
ute economic, strengthen the international trade protectionism and descent a blow for
our country’s strategy of broadening export volume and absorbing foreign investment.
As for the domestic aspect, the drop of custom will promote China’s import volume;
every blade has two curves, proper unfavorable balance of trade will buffer the pres-
sure of RMB appreciation. Base on these factors, the forecasting that the amount of
export trade in 2004 will change more, and the import trade may increase faster than
export, In a word, emulation and forecasting basically combined the training data and
develop trend.
Table.5 shows that the nonlinear characteristic value of the two import serials al-
most identical; as for export trade, the forecasting serial and original serial have tiny
difference, but it is not a big deal, besides, when the embedded dimension is 16, these
two serial’s relation dimension reach the point of maturity. Actually, many scientific
studies had proved that there are more factors are influencing the export trade than
import, which produces more noise in the export trade serial and fades the nonlinear
characteristic of actual system. Fortunately, the model of this paper is insensitive of
noise, and it has a good capability of generalization, which can enable the model
obtain a direct nonlinear expression form from the dynamic system, produce the com-
bination serial and efficiently ‘catch’ the dynamics characteristic of nonlinear system
that produced the original serial.

5 Conclusions
As a kind of nonlinear dynamics method, neural network has a strong capability to
close nonlinear reflection, and that is why it is so suitable in modeling and forecasting
of nonlinear economic system. Up to this day, the scholar has reaped enormous actual
fruit from applying neural network in the problems of economic management. Based
on nonlinear forecasting theory of phase reconstruction, this paper has built a time-
delay BP neural network model. Considering the weak generalization capability of
common BP neural network, this paper combined the Bayesian Regularization to
improve the objective function and the generalization capability of neural network. In
order to explain more vividly, this paper then adopted some certain occupation’s ac-
tual monthly imp&exp data of 1/1989-5/2003 as training sample, finally, it built a
multi-steps nonlinear forecasting model. The results of imp&exp trade’s multi-step
forecast showed that the model could not only reasonably forecast the develop trend
of actual serial, but also combined the actual data, besides, the relative error of single
forecast is in the scope of 1 percent. In addition, nowadays, the forecast evaluation
only focus on the improvement of the combination of comparing data point and the
forecasting precision, which lead to the overlook of the difference between combined
serial and original serial. When the system is nonlinear, it got a problem, because in
nonlinear system, even two identical serial may come from two different sub-systems.
388 W. Jiang et al.

For this problem, this paper made an improvement: utilizing the relation dimension of
combined serial and original data and the maximum Lyapunov index, comparing the
memory structure of combined serial and original serial, and then finding that the
model can efficiently catch up the inner structure and rule of the nonlinear system that
produce the original serial. The model of this paper has something really special from
others, its free parameter is fewer, it has a strong capability of robustness and gener-
alization, which make it very suitable for the forecasting of nonlinear chaos time
serial.

Acknowledgement
This paper is supported by the Society Science Foundation of Hunan Province of
China No. 07YBB239.

References
1. Ott, E., Grebogi, C., Yorke, J.A.: Controlling chaos. Physical Review Letters 64(11),
1196–1199 (1990)
2. Islam, M.N., Sivakumar, B.: Characterization and prediction of runoff dynamics: a nonlin-
ear dynamical view. Advances in Water Resources (25), 179–190 (2002)
3. Ramirez, T.A., Puebla, H.: Control of a nonlinear system with time-delayed dynamics.
Phys. Lett. A 262, 166–173 (1999)
4. Dongmei, L., Zhengou, W.: Small parameter perturbations control algorithm of chaotic
systems based on predictive control. Information and Control 32(5), 426–430 (2003)
5. Weijin, J.: Non-Linear chaos time series prediction & control algorithm based on NN.
Journal of Nanjing University (NaturalSciences) 38(Computer issue), 33–36 (2002)
6. Weijin, J.: Research of ware house management chaotic control based on alterable. Micro
Electronics & Computer 19(11), 55–57 (2002)
7. Sivakumar, B., Jayawardena, A.W., Fernando, T.M.K.G.: River flow forecasting: use of
phase-space reconstruction and artificial neural networks approaches. Journal of Hydrol-
ogy (265), 225–245 (2002)
8. Islam, M.N., Sivakumar, B.: Characterization and prediction of runoff dynamics: a nonlin-
ear dynamical view. Advances in Water Resources (25), 179–190 (2002)
9. Jayawardena, A.W., Li, W.K., Xu, P.: Neighborhood selection for local modeling and pre-
diction of hydrological time series. Journal of Hydrology (258), 40–57 (2002)
10. Kim, H.S., Eykholt, R., Salas, J.D.: Nonlinear dynamics, delay times, and embedding win-
dows. Phy. D (127), 48–60 (1999)
 An Adaptive Learning Algorithm for Supervised Neural
Network with Contour Preserving Classification

Piyabute Fuangkhon and Thitipong Tanprasert

Distributed and Parallel Computing Research Laboratory

Faculty of Science and Technology, Assumption University
592 Soi Ramkamhang 24, Ramkamhang Road, Huamak, Bangkapi, Bangkok, Thailand
piyabute@yahoo.com, t_tanprasert@yahoo.com

Abstract. A study of noise tolerance characteristics of an adaptive learning al-

gorithm for supervised neural network is presented in this paper. The algorithm
allows the existing knowledge to age out in slow rate as a supervised neural
network is gradually retrained with consecutive sets of new samples, resem-
bling the change of application locality under a consistent environment. The al-
gorithm utilizes the contour preserving classification algorithm to pre-process
the training data to improve the classification and the noise tolerance. The ex-
perimental results convincingly confirm the effectiveness of the algorithm and
the improvement of noise tolerance.

Keywords: supervised neural network, contour preserving classification, noise

tolerance, outpost vector.

1 Introduction
It is known that repetitive feeding of training samples is required for allowing a super-
vised learning algorithm to converge. If the training samples effectively represent the
population of the targeted data, the classifier can be approximated as being generalized.
However, there are many times when it is impractical to obtain such a truly representa-
tive training set. Many classifying applications are acceptable with convergence to a
local optimum. For example, a voice recognition system may be customized for effec-
tively recognizing voices of a group of limited number of users, so the system may not
be practical for recognizing any speaker. As a consequence, this kind of application
needs occasional retraining when there is sliding of actual context locality.
Our focus is when only part of the context is changed; thereby establishing some
new cases, while inhibiting some old cases, assuming a constant system complexity.
The classifier will be required to effectively handle some old cases as well as new
cases. Assuming that this kind of situation will occur occasionally, it is expected that
the old cases will age out, the medium-old cases are accurately handled to a certain
degree, and new cases are most accurately handled. Since the existing knowledge is
lost while retraining new samples, an approach to maintain old knowledge is required.
While the typical solution uses both prior samples and new samples on retraining, the
major drawback of this approach is that all the prior samples for training must be
maintained.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 389–398, 2009.
© Springer-Verlag Berlin Heidelberg 2009
390 P. Fuangkhon and T. Tanprasert

Research works related to the proposed algorithm are in the field of adaptive
learning [1], [2], incremental learning [16], and contour preserving classification [3].
The prior one can be categorized into three strategies [3]. The first strategy (increas-
ing neurons) [4], [5], [6], [7] will increase number of hidden nodes when the error is
excessively high. These algorithms will adapt weights of neuron that is closest with
input sample and neighbors only. However, the increasing size of neural networks is a
cause of accuracy tradeoff. The second strategy (rule extraction) [8], [9], [10], [11]
will interchange between rules and weight of neuron. The accuracy of networks de-
pends on discovered rules. The translation of weight vector into rules also partly sup-
presses certain inherited statistical information. The last strategy (aggregation) [12],
[13], [14], [15] will allow existing weights to change in bounded range and will al-
ways add new neurons for learning samples of new context. This method requires two
contexts be similar and network’s size is incrementally larger.
In this paper, an alternative algorithm is proposed for solving adaptive learning
problem for supervised neural network. The algorithm improved from [17], [18] is
able to learn new knowledge while maintaining an old knowledge through the decay
rate while allowing the adjustment of the number of new samples. In addition, the
improvement of the classification and the noise tolerance is archived by utilizing the
contour preserving classification algorithm which helps expanding the territory of
both classes while maintaining the shape of both classes.
Following this section, section 2 summarizes the outpost vector model. Section 3
describes the methodology. Section 4 demonstrates the experimental results of a 2-
dimension partition problem. Section 5 discusses the conclusion of the paper.

2 The Outpost Vector Model

This section summarizes the outpost vector method as originally published in [3]. Fig.
1 illustrates the concept of outpost vector. Each input vector is modeled to span its
territory as a circle (sphere in case of 3-dimensional space or hyper sphere in case of
more-dimensional space) until the territories collide against one another.
In Fig. 1, the territory of input vector k of class A (denoted by Ak) is founded by
locating the input vector in class B which is nearest to Ak (denoted by B*(Ak)) and
declaring the territory at half way between Ak and B*(Ak). Accordingly, the radius of
Ak’s territory is set at half of the distance between Ak and B*(Ak). This is to guarantee
that if B*(Ak) sets its territory using the same radius, then the distance from the hyper
plane to either Ak or B*(Ak) will be at maximum.
For B*(Ak), the nearest input vector of class A to B*(Ak) is not necessarily Ak.
*
B (Ak) may not place its outpost vector against Ak although Ak has placed an outpost
vector against it. However, B*(Ak) places its outpost vector against Aj because it is
nearest to B*(Ak). Since it is attempted to carve the channel between the two classes in
such a way that its concavity or convexity can influent the learning process, it is de-
sirable to generate an additional set of outpost vectors. As illustrated in Fig. 1, an
additional outpost vector of B*(Ak) (designated as “class B”) is placed on B*(Ak)’s
territory in the direction of Ak in response to the existence of an outpost vector of Ak
in that direction. In [3], a theorem was provided for proving that the territory created
will not overlap across different classes. All the outpost vectors are combined with the
original input vectors for training the network.
 An Adaptive Learning Algorithm for Supervised Neural Network 391

Outpost Vector of
Ak
taking B*(Ak) as
Ak

Ak’s
territory

B*(Ak)
Optional outpost
vector B*(Ak)’s
territory
A
Outpost Vector of B*(Ak)
taking Aj as the nearest

Fig. 1. Territory and Outpost Vectors

3 Methodology
The algorithm [17], [18] utilizes the concept of adaptive learning and outpost vectors
in modeling the new training samples. The adaptive learning algorithm will maintain
limited number of training samples from the previous training session (decayed prior
samples) to be used in the next training session while the outpost vectors will help
expanding the territory of both classes and maintaining the shape of the boundary
between both classes.
There are three parameters in the algorithm: new sample rate, outpost vector rate,
and decay rate. Firstly, the new sample rate is the ratio of the number of selected new
samples over the number of new samples. It determines the number of selected new
samples to be included in the final training set. The larger new sample rate will cause
the network to learn new knowledge more accurately. The number of selected new
samples is calculated by formula:

nss = nw × ns (1)

where nss is the number of selected new samples,

nw is the new sample rate [0, ∞),
ns is the number of new samples
Secondly, the outpost vector rate is the ratio of the number of generated outpost vec-
tors over the number of new samples. It determines the number of outpost vectors to
be included in the final training set. Using larger outpost vector rate will generate
more outpost vectors at the boundary between both classes. When outpost vector rate
is equal to 1.0, the number of outpost vectors will be at maximum. The number of
outpost vectors is calculated by formula:

nov = ov × ns (2)
392 P. Fuangkhon and T. Tanprasert

where nov is the number of outpost vectors,

ov is the outpost vector rate [0,1],
ns is the number of new samples
Lastly, the decay rate is the ratio of the number of decayed prior samples over the
number of new samples. It determines the number of decayed prior samples to be
included in the final training set. The larger decay rate will cause the network to for-
get old knowledge slower. When decay rate is greater than 1.0, more than one in-
stance of some prior samples will be included in the decayed prior sample set. There
is an exception for the first training session where the new samples will also be used
as the decayed prior samples. The number of decayed prior samples is calculated by
formula:

ndc = dc × ps (3)

where ndc is the number of selected prior samples,

dc is the decay rate [0, ∞),
ps is the number of prior samples
After the final training set is available, it will be trained with a supervised neural
network. The whole process will be repeated when new samples have arrived.

Algorithm. Sub_Trainning _with_Outpost_Vectors

1 Set new sample set as prior sample set for the first training session
2 for each training session
3 Construct selected new sample set by
3 If new sample rate is equal to 1.0
4 Select all new samples from new sample set
5 Else
6 Calculate the number of select new samples (1)
7 Randomly select samples from new sample set
8 End if
9 Construct outpost vector set by
10 Calculate the number of outpost vectors (2)
11 Generate outpost vectors from new sample set
12 Construct decayed prior sample set by
13 Calculate the number of decayed prior samples (3)
14 Randomly select samples from prior sample set
15 Construct final sample set by
16 UNION ( selected new sample set, outpost vector set ,decayed prior sample set )
17 Train the network with the final sample set
18 Set final sample set as prior sample set for the
next training session
19 end for
In order to study the noise tolerance of the algorithm, a random noise radius is
used to move the samples to their nearby location. Samples at the boundary can even
 An Adaptive Learning Algorithm for Supervised Neural Network 393

be moved across their boundary. However, the new location must also be outside the
territory of the other class.

4 Experiment
The experiment was conducted on a machine using Intel® Pentium D™ 820 2.4 GHz
with 2.0 GB main memory running Microsoft® Windows™ XP SP3. Feed-forward
backpropagation neural network running under MATLab 2007a is used as the classifier.
The proposed algorithm was tested on the 2-dimension partition problem. The dis-
tribution of samples was created in limited location of 2-dimension donut ring as
shown in Fig. 2. This partition had three parameters: Inner Radius (R1), Middle Ra-
dius (R2) and Outer Radius (R3). The class of samples depended on geometric posi-
tion. There were two classes which were designed as one and zero.

Inner Radius 0 < Radius ≤ R1 Class 0

Middle Radius R1 < Radius ≤ R2 Class 1
Outer Radius R2 < Radius ≤ R3 Class 0

The context of the problem was assumed to shift from an angular location to another
while maintaining some overlapped area between consecutive contexts as shown in
Fig. 3. The set numbers shown in Fig. 3 identify the sequence of training and testing
sessions.
In the experiment, each training and testing set consisted of eight sub sets of sam-
ples generated from eight problem contexts. Each sub set of samples consisted of 400
new samples (200 samples from class 0 and 200 samples from class 1). The samples
from class 0 were placed in the outer radius and inner radius. The samples from class
1 were placed in the middle radius.
In the sample generation process, the radius of the donut ring was set to 100.
Placement of the samples from both classes was restricted by the gap or empty space
introduced between both classes to test the noise tolerance of the network.
There were two training sets having the gap of size 5 (Set A) and 10 (Set B) as
shown in Fig. 4. The gap of size 0 was not used to generate the training set because
there was no available space at the boundary between both classes to generate outpost
vectors.

0 1 0

R1 R2 R3

Θ = 0°

Fig. 2. Shape of Partition Area Fig. 3. Shifting of Problem Context

394 P. Fuangkhon and T. Tanprasert

100 100
Class 0 Class 1 Class 0 Class 0 Class 1 Class 0
90 90

80 80

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 20 40 60 80 100 0 20 40 60 80 100

(a) Set A, Gap 5 (b) Set B, Gap 10

Fig. 4. Sample Sub Training Sets in S8

There were six main testing sets having the gap and noise radius 5:0 (Set M), 5:5
(Set N), 5:10 (Set O), 10:0 (Set P), 10:10 (Set Q), and 10:20 (Set R) as shown in
Fig. 5. To test the noise tolerance of the training sets, four testing sets having the gap
and noise radius 5:5 (Set S), 5:10 (Set T), 10:10 (Set U), and 10:20 (Set V) were gen-
erated by adding noise radius to all samples in the two training sets. The noise sam-
ples introduced into the testing set (Set N, O, Q, R, S, T, U, V) were intended to test
the noise tolerance of the network when outpost vector was applied.

100 100 100

Class 0 Class 1 Class 0 Class 0 Class 1 Class 0 Class 0 Class 1 Class 0
90 90 90

80 80 80

70 70 70

60 60 60

50 50 50

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100

(a) Set M, Gap 5, Noise 0 (b) Set N, Gap 5, Noise 5 (c) Set O, Gap 5, Noise 10
100 100 100
Class 0 Class 1 Class 0 Class 0 Class 1 Class 0 Class 0 Class 1 Class 0
90 90 90

80 80 80

70 70 70

60 60 60

50 50 50

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100

(d) Set P, Gap 10, Noise 0 (e) Set Q, Gap 10, Noise 10 (f) Set R, Gap 10, Noise 0
Fig. 5. Sample Sub Testing Sets in S8
 An Adaptive Learning Algorithm for Supervised Neural Network 395

In the training process, the sub training (training on sub set of training set) was
conducted eight times with eight sub training sets to cover eight problem contexts in a
training set. Each final sub training set was composed of three components:
1. Selected new samples taken from sub training set
2. Outpost vectors generated from sub training set
3. Decay prior samples randomly selected from final sub training set from the
previous training session
The numbers of vectors in each part of the final sub training set were determined by
the new sample rate, the outpost vector rate, and the decay rate. Table 1 shows the
number of vectors in a sample final sub training set when new sample set consisted of
200 samples, new sample rate was equal to 1.0, outpost vector rate was equal to 0.5
and decay rate was equal to 1.0. Fig. 6 shows samples of final training set for the last
training session when decay rate is equal to 1.0 and 2.0 respectively.

100 100
Class0 Class1 Class0 Class1
80 80

60 60

40 40

20 20

0 0
-100 -50 0 50 100 -100 -50 0 50 100
-20 -20

-40 -40

-60 -60

-80 -80

-100 -100

(a) 1200 Samples, Gap 5, Decay 1 (b) 1600 Samples, Gap 5, Decay 2
Fig. 6. Sample Last Sub Training Set

Because the prior sample set was constructed at the end of the algorithm, there
was no prior sample set to make decayed prior sample set for the first sub training
session. An additional step for solving this problem is to also use the new sample set
as the prior sample set.
The procedure for the experiment started from the feed forward back-propagation
neural network being trained with the following parameters:
1. network size = [10 1]
2. transfer function for hidden layer = “logsig”
3. transfer function for output layer = “logsig”
4. max epochs = 500
5. goal = 0.001
After the first training session (S1), seven sub training sessions followed. At the end
of the eighth sub training session (S8), the supervised neural network was tested with
the sub testing samples from every context to evaluate the performance of testing data
from each context.
396 P. Fuangkhon and T. Tanprasert

The testing results are shown in Table 2, 3, 4, 5, 6, 7, and 8. For the testing sets
without noise samples (Set M, P), applying outpost vector can lower the mean square
error (MSE) effectively. For the testing sets with noise samples and small gap (Set N,
O), medium outpost vector rate (OV 0.5) give better results. For the testing sets with
noise samples and large gap (Set Q, R), large outpost vector rate (OV 1.0) give better
results. For the testing sets with noise samples generated from the training sets (Set S,
T, U, V), large outpost vector rate (OV 1.0) generally give better results.
The testing results show that the proposed algorithm can classify samples in the
newer contexts (S1, S7, S8) accurately while the accuracy of classifying samples from
the older contexts (S2, S3, S4, S5, S6) is lower because its old knowledge is decaying.
The proposed algorithm presents some level of noise tolerance because the difference
between the mean square errors (MSEs) of the classification of testing sets with and
without noise samples is insignificant.

Table 1. The Number of Vectors in a Final Sub Training Set

Type Rate Vectors

Selected New Samples 1.0 200 ( 200 × 1.0 )
Outpost Vectors 0.5 100 ( 200 × 0.5 )
Decayed Prior Samples 1.0 200 ( 200 × 1.0 )

Table 2. MSEs from Training Set A with Testing Set M, N, O and Decay Rate 1

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
M 0.0 00 0.13 0.21 0.17 0.16 0.25 0.14 0.01 0.03
N 0.0 05 0.14 0.19 0.12 0.17 0.26 0.14 0.01 0.04
O 0.0 10 0.14 0.21 0.16 0.19 0.25 0.14 0.03 0.04
M 0.5 00 0.02 0.14 0.15 0.02 0.00 0.00 0.00 0.00
N 0.5 05 0.02 0.14 0.15 0.03 0.00 0.00 0.00 0.00
O 0.5 10 0.02 0.15 0.14 0.03 0.01 0.00 0.00 0.00
M 1.0 00 0.20 0.38 0.27 0.14 0.06 0.02 0.01 0.01
N 1.0 05 0.21 0.38 0.27 0.13 0.06 0.03 0.01 0.01
O 1.0 10 0.21 0.37 0.27 0.14 0.07 0.03 0.02 0.01

Table 3. MSEs from Training Set A with Testing Set M, N, O and Decay Rate 2

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
M 0.0 00 0.27 0.30 0.28 0.28 0.28 0.23 0.21 0.21
N 0.0 05 0.28 0.30 0.28 0.27 0.28 0.25 0.20 0.22
O 0.0 10 0.27 0.30 0.27 0.28 0.27 0.23 0.21 0.21
M 0.5 00 0.12 0.23 0.13 0.02 0.01 0.00 0.00 0.00
N 0.5 05 0.12 0.23 0.13 0.03 0.01 0.00 0.00 0.00
O 0.5 10 0.12 0.24 0.13 0.03 0.02 0.01 0.01 0.01
M 1.0 00 0.18 0.32 0.31 0.20 0.11 0.08 0.07 0.05
N 1.0 05 0.18 0.32 0.31 0.20 0.10 0.09 0.07 0.06
O 1.0 10 0.18 0.32 0.31 0.20 0.11 0.09 0.08 0.06
 An Adaptive Learning Algorithm for Supervised Neural Network 397

Table 4. MSEs from Training Set B with Testing Set P, Q, R and Decay Rate 1
SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
P 0.0 00 0.17 0.30 0.26 0.19 0.21 0.16 0.08 0.06
Q 0.0 10 0.18 0.31 0.27 0.20 0.21 0.16 0.08 0.07
R 0.0 20 0.17 0.30 0.27 0.20 0.21 0.17 0.10 0.09
P 0.5 00 0.27 0.45 0.38 0.39 0.35 0.19 0.08 0.04
Q 0.5 10 0.27 0.45 0.38 0.39 0.34 0.20 0.08 0.04
R 0.5 20 0.27 0.45 0.37 0.38 0.33 0.21 0.09 0.07
P 1.0 00 0.22 0.31 0.22 0.26 0.25 0.23 0.12 0.01
Q 1.0 10 0.25 0.34 0.22 0.27 0.25 0.23 0.13 0.01
R 1.0 20 0.24 0.34 0.25 0.29 0.24 0.25 0.14 0.05

Table 5. MSEs from Training Set B with Testing Set P, Q, R and Decay Rate 2

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
P 0.0 00 0.22 0.32 0.35 0.33 0.28 0.25 0.20 0.17
Q 0.0 10 0.24 0.32 0.35 0.33 0.29 0.25 0.21 0.18
R 0.0 20 0.25 0.32 0.35 0.34 0.29 0.25 0.21 0.20
P 0.5 00 0.16 0.31 0.34 0.29 0.19 0.08 0.03 0.03
Q 0.5 10 0.16 0.31 0.34 0.29 0.19 0.08 0.03 0.04
R 0.5 20 0.16 0.31 0.33 0.29 0.20 0.10 0.05 0.06
P 1.0 00 0.17 0.28 0.15 0.03 0.09 0.10 0.01 0.00
Q 1.0 10 0.17 0.30 0.16 0.04 0.09 0.09 0.01 0.01
R 1.0 20 0.18 0.30 0.17 0.08 0.11 0.11 0.02 0.02

Table 6. MSEs from Training Set A with Testing Set S, T and Decay Rate 1

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
S 0.0 05 0.13 0.19 0.14 0.15 0.24 0.14 0.01 0.02
T 0.0 10 0.13 0.25 0.18 0.14 0.24 0.14 0.08 0.02
S 0.5 05 0.01 0.14 0.15 0.02 0.00 0.00 0.00 0.00
T 0.5 10 0.02 0.15 0.14 0.02 0.00 0.00 0.02 0.00
S 1.0 05 0.19 0.37 0.27 0.14 0.06 0.03 0.01 0.01
T 1.0 10 0.18 0.39 0.26 0.14 0.07 0.02 0.02 0.01

Table 7. MSEs from Training Set B with Testing Set U, V and Decay Rate 1

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
U 0.0 10 0.18 0.30 0.27 0.20 0.21 0.16 0.07 0.08
V 0.0 20 0.20 0.30 0.26 0.21 0.21 0.16 0.09 0.07
U 0.5 10 0.29 0.43 0.41 0.39 0.35 0.17 0.07 0.05
V 0.5 20 0.31 0.44 0.37 0.42 0.35 0.17 0.08 0.06
U 1.0 10 0.24 0.32 0.22 0.27 0.24 0.21 0.11 0.01
V 1.0 20 0.28 0.29 0.23 0.27 0.25 0.21 0.14 0.03

Table 8. MSEs from Training Set B with Testing Set U, V and Decay Rate 2

SET OV NS S1 S2 S3 S4 S5 S6 S7 S8
U 0.0 10 0.23 0.37 0.30 0.33 0.29 0.24 0.20 0.19
V 0.0 20 0.28 0.26 0.37 0.29 0.28 0.24 0.21 0.18
U 0.5 10 0.18 0.33 0.34 0.29 0.19 0.07 0.03 0.04
V 0.5 20 0.20 0.30 0.33 0.29 0.19 0.07 0.03 0.05
U 1.0 10 0.18 0.31 0.17 0.03 0.08 0.10 0.01 0.01
V 1.0 20 0.21 0.26 0.15 0.08 0.09 0.10 0.02 0.01
398 P. Fuangkhon and T. Tanprasert

5 Conclusion
A study of noise tolerance characteristics of an adaptive learning algorithm for super-
vised neural network is presented. The noise samples are used to test the noise
tolerance of the algorithm. Overall result shows that combining adaptive learning
algorithm with contour preserving classification yields effective noise tolerance, bet-
ter learning capability, and higher accuracy.

References
[1] Tanprasert, T., Kripruksawan, T.: An approach to control aging rate of neural networks
under adaptation to gradually changing context. In: ICONIP 2002 (2002)
[2] Tanprasert, T., Kaitikunkajorn, S.: Improving synthesis process of decayed prior sam-
pling technique. In: Tech 2005 (2005)
[3] Tanprasert, T., Tanprasert, C., Lursinsap, C.: Contour preserving classification for maxi-
mal reliability. In: IJCNN 1998 (1998)
[4] Burzevski, V., Mohan, C.K.: Hierarchical growing cell structures. In: ICNN 1996 (1996)
[5] Fritzke, B.: Vector quantization with a growing and splitting elastic net. In: ICANN 1993
(1993)
[6] Fritzke, B.: Incremental learning of local linear mappings. In: ICANN 1995 (1995)
[7] Martinez, T.M., Berkovich, S.G., Schulten, K.J.: Neural-gas network for vector quantiza-
tion and it application to time-series prediction. IEEE Transactions on Neural Networks
(1993)
[8] Chalup, S., Hayward, R., Joachi, D.: Rule extraction from artificial neural networks
trained on elementary number classification tasks. In: Proceedings of the 9th Australian
Conference on Neural Networks (1998)
[9] Craven, M.W., Shavlik, J.W.: Using sampling and queries to extract rules from trained
neural networks. In: ICML 1994 (1994)
[10] Setiono, R.: Extracting rules from neural networks by pruning and hidden-unit splitting.
Neural Computation (1997)
[11] Sun, R.: Beyond simple rule extraction: Acquiring planning knowledge from neural net-
works. In: ICONIP 2001 (2001)
[12] Thrun, S., Mitchell, T.M.: Integrating inductive neural network learning and explanation
based learning. In: IJCAI 1993 (1993)
[13] Towell, G.G., Shavlik, J.W.: Knowledge based artificial neural networks. Artificial Intel-
ligence (1994)
[14] Mitchell, T., Thrun, S.B.: Learning analytically and inductively. Mind Matters: A Tribute
to Allen Newell (1996)
[15] Fasconi, P., Gori, M., Maggini, M., Soda, G.: Unified integration of explicit knowledge
and learning by example in recurrent networks. IEEE Transactions on Knowledge and
Data Engineering (1995)
[16] Polikar, R., Udpa, L., Udpa, S.S., Honavar, V.: Learn++: An incremental learning algo-
rithm for supervised neural networks. IEEE Transactions on Systems, Man, and
Cybernetics (2001)
[17] Tanprasert, T., Fuangkhon, P., Tanprasert, C.: An Improved Technique for Retraining
Neural Networks In Adaptive Environment. In: INTECH 2008 (2008)
[18] Fuangkhon, P., Tanprasert, T.: An Incremental Learning Algorithm for Supervised Neu-
ral Network with Contour Preserving Classification. In: ECTI-CON 2009 (2009)
 Application Study of Hidden Markov Model and
Maximum Entropy in Text Information Extraction

Rong Li1,∗, Li-ying Liu2, He-fang Fu1, and Jia-heng Zheng3

1
Department of computer, Xinzhou Teachers’ College,
Xinzhou, Shanxi 034000, China
2
Weather bureau of Xinzhou City, Xinzhou, Shanxi 034000, China
3
School of Computer and Information Technology, Shanxi University,
Taiyuan, Shanxi 030006, China
{Rong Li,Li-ying Liu,He-fang Fu,
Jia-heng Zheng,lirong_1217}@126.com

Abstract. Text information extraction is an important approach to process large

quantity of text. Since the traditional training method of hidden Markov model
for text information extraction is sensitive to initial model parameters and easy
to converge to a local optimal model in practice, a novel algorithm using hidden
Markov model based on maximal entropy for text information extraction is pre-
sented. The new algorithm combines the advantage of maximum entropy
model, which can integrate and process rules and knowledge efficiently, with
that of hidden Markov model, which has powerful technique foundations to
solve sequence representation and statistical problem. And the algorithm uses
the sum of all features with weights to adjust the transition parameters in hidden
Markov model for text information extraction. Experimental results show that
compared with the simple hidden Markov model, the new algorithm improves
the performance in precision and recall.

Keywords: text information extraction, maximum entropy, hidden markov

model, generalized iterative scaling algorithm.

1 Introduction
The universal application of WWW causes on-line text quantity to increase exponen-
tially, therefore how to deal with these huge amounts of on-line text information
becomes the present important research subject. The automatic text information ex-
traction is an important link of text information processing [1].Text information ex-
traction refers to extract automatically the related or specific type information from
text. At present, the text information extraction model mainly has three kinds: the
dictionary-based extraction model[2], the rule-based extraction model[3] and the
extraction model based on Hidden Markov Model (HMM)[4-8].
The text information extraction using HMM is one kind of information extraction
method based on statistics machine learning. HMM is easy to establish, does not need

∗
This work is supported by the Natural Science Foundation of China (Grant #60775041).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 399–407, 2009.
© Springer-Verlag Berlin Heidelberg 2009
400 R. Li et al.

the large-scale dictionary collection and rule set, its compatibility is good and the
extraction precision is high, thus HMM obtains researchers’ attention. For example,
In References [4], Kristie Seymore et al. used HMM to extract forehead information
of computer science research paper such as title, author and abstract. In References
[5], Dayne Frietag and Andrew McCallum adopted one kind of “shrinkage” technol-
ogy to improve the probability estimate of the HMM information extraction model. In
References[6], Freitag D and McCallum A used the stochastic optimization technique
is to select automatically the most suitable HMM model structure for information
extraction. Souyma Ray and Mark Craven[7] applied phrase structure analysis tech-
nology in the natural language processing to the HMM text information extraction. T
Scheffer et al.[8] used the active learning technology to reduce the flag data which is
needed when training HMM information extraction model. The HMM method doesn’t
consider the characteristic information of text context and the characteristic informa-
tion contained in text word itself, but these information are very useful to realize the
correct text information extraction. Freitag D et al.[9]proposed a Maximum Entropy
Markov Model(MEMM) for the segmentation of questions and answers in the
FAQs(Frequently Asked Questions) text. MEMM is one kind of exponential model. It
mainly takes the text abstract characteristic as the input and selects the next state in
the base of the Markov state transition, and on this point, it is near to finite state
automaton. Since MEMM incorporates the text context characteristic information and
the characteristic information contained in text word itself into Markov model, it can
improve the performance of information extraction. But it hasn’t concrete text word
count and only considers the abstract characteristic, which causes its performance to
be inferior to that of HMM in some certain circumstances.
In the paper, we combine the advantage of maximum entropy model, which can in-
tegrate and process rules and knowledge efficiently, with that of hidden Markov
model, which has powerful technique foundations to solve sequence representation
and statistical problem, and present a Maximal Entropy-based Hidden Markov
Model(ME-HMM) for text information extraction. The algorithm uses the sum of all
features with weights to adjust the transition parameters in hidden Markov model.
Experimental results show that compared with the simple hidden Markov model, the
new algorithm improves the performance in precision and recall.

2 Text Information Extraction Based on HMM

2.1 Hidden Markov Model

An HMM includes two layers: one observation layer and one hidden layer. The ob-
servation layer is the observation sequence for recognition and the hidden layer is a
Markov process, (i.e. a limited state machine), in which each state transition all has
transition probability.
An HMM is specified by a five-tuple(S,V,A,B,∏):
S={S1,S2,…,SN}
V={V1,V2,…VM}
A={aij=p(qt+1=Sj|qt=Si),1≤i,j≤N}
 Application Study of Hidden Markov Model and Maximum Entropy 401

B={bj(Vk)=P(Vk at t|qt=Sj), 1≤i,j≤N , 1≤k≤M}

∏={πi=P(q1=Si), 1≤i≤N}
Among the notation, S denotes the state set of N states, V represents the word set of M
possible output words, and A ,B, ∏ are the probabilities for the state transition, the
observation emissions, and the initial state, respectively.

2.2 Text Information Extraction Based on HMM

Hidden Markov model is mainly used to solve the following three fundamental prob-
lem: evaluation problem, study problem and decoding problem. For the commonly
used algorithm, see also Reference [10-12]. Text information extraction needs to
solve the study problem and the decoding problem of HMM. The purpose of text
information extraction is to extract those specific and compelling information from
large quantity of information, namely to extract some information such as author,
publishing press, publication time, affiliation and so on from the “word string ”array
composed of different text information, which is similar to part-of-speech tagging in
the field of Chinese information processing. When extracting text information using
HMM, Maximum Likelihood(ML) algorithm for marked training sample set or
Baum-Welch algorithm for unmarked training sample set is adopted generally to
obtain HMM parameters. Then Viterbi algorithm is used to find out the state label
sequence of the maximum probability from input text information and the state label
is the content label to be extracted that is we define beforehand. In the paper, the
manually marked training sample is adopted to train model parameters in the text
information extraction model based on HMM. Information extraction is a two-stage
process:
1) Obtain HMM parameters by using the statistical method from the training sam-
ple. ML algorithm is adopted to construct HMM model and obtain model parameters
aij ,bj(Vk),πi by the methods of statistic. ML algorithm demands to have sufficient
marked training array, and each word in each array uses corresponding class label.
Therefore, the transition number of times from state Si to Sj may be calculated and
noted as Cij. The number of times of outputting word Vk can be calculated in the spe-
cific state, noted as Ej(K).Similarly, the number of times of which a sequence starts
from a certain specific state may be calculated, noted as Init(i).Therefore, the prob-
ability can be described as follows:
Init (i)
πi = N
,1 ≤ i ≤ N (1)
∑
j =1
Init ( j )

cij
aij = N ,1 ≤ i, j ≤ N (2)
∑ cik
k =1

E j (k )
bj (k ) = M
,1 ≤ i ≤ N , 1 ≤ j ≤ M (3)
∑
i =1
E j (i )
402 R. Li et al.

）
2 Apply the established HMM for text information extraction.The text informa-
tion extraction process based on HMM is that state sequence Q* generating the
maximum probability symbol sequence should be sought if given HMM and a symbol
sequence and then the observation text for marked target state label is the content of
information extraction. Viterbi algorithm is the classical approach to solve the HMM
decoding problem. To avoid data underflow problem, an improved Viterbi algorithm
is brought forwarded in the paper. The place of improvement is that all the probabili-
ties in Viterbi formulas are multiplied by proportion factor 102, then the logarithm is
taken to both sides of formulas so as to obtain improved Viterbi formulas.

3 Text Information Extraction Using ME-HMM

3.1 Maximum Entropy Principle

Maximum entropy principle, as a very important principle in thermodynamics, has

widespread application in many other domains, and it is also one main processing
method in the natural language processing aspect [13-15].
If the natural language is regarded as a stochastic process, we will construct a sto-
∈ ∈
chastic process model p, p P. The set of output value is Y, y Y. The set of N sam-
ples is S={(x1, y1),(x2, y2)…,(xn, yn)}, where (xi, yi) is an observed event, the event
space is X*Y; The language knowledge is represented with the characteristic which is
a two-valued function f: X*Y→{0,1}. Entropy describes the uncertainty of the random
variable. The greater the random variable's uncertainty, the bigger is its entropy.
The entropy of model p may be described as :
H( p ) ≡ −∑ p( x, y )log p( x, y ) (4)
x,y

The maximum entropy model is represented as:

p* = arg max H( p ) (5)
p∈C

where C is the model set which satisfies the restraint. The remaining problem is to
seek p* in C. p* can be described as the following form:
1
p* ( y | x ) = exp[ ∑ λi f i ( x, y )] (6)
Z( X ) i

where Z(x) is a normalized constant.

Z( X ) = ∑ exp[ ∑ λi f i ( x, y )] (7)
y i

Where λi is the model parameter and also can be seen as the characteristic weight.
In 1972, Darroch and Ratcliff proposed a called GIS algorithm(generalized iterative
scaling algorithm), the algorithm is a general iterative algorithm [16]. An advantage of
the ME method is that the experimenter only needs to concentrate on what characteris-
tic should be selected, but does not need to consider how to use these characteristics.
 Application Study of Hidden Markov Model and Maximum Entropy 403

Each characteristic's contribution is decided by the corresponding weight, and these

weights may be obtained automatically by the GIS learning algorithm.

3.2 Text Information Extraction Based on ME-HMM

In the maximum entropy method, we call a rule as a characteristic. The idea of the
maximum entropy method is to find a characteristic set, and determine the important
degree of each characteristic. The maximum entropy model can integrate each kind of
characteristic and rule into a unified frame, and hidden Markov model has the very
strong superiority in sequence representation problem and the statistical learning
aspect. So, if the maximum entropy method is incorporated into the hidden Markov
model for text information extraction, both can we solve the knowledge expression
problem, and may momentarily add the newly gained language knowledge to the
model, and the combination of the two methods is that of the regular method and the
statistical method union. In accordance with the idea, the paper adopts hidden Markov
model based on maximum entropy principle for text information extraction. For ease
of comparison, three models including HMM, MEMM and ME-HMM proposed in
the paper are shown in Figure 1, respectively.

state

observation characteristic
(a) state
state
observation
characteristic
(c)
(b)

Fig. 1. (a)HMM; (b)MEMM; (c)ME-HMM

Given the characteristic set, we define a two-valued function:

⎧ 1, if the observation ot has charastic i,
f i, j ( ot ,st ) = ⎪ (8)
⎨ and the marked state st = s j
⎪ 0,
⎩ otherwise

The ME-HMM model constructs a characteristic - state transition probability matrix

M= {Mi,j}, where the matrix unit Mi,j is the transition probability from the characteris-
tic i to the state j , and Mi,j satisfies the following condition:
∑ M i, j = 1, 1 ≤ i ≤ N F , 1 ≤ j ≤ N S (9)
j

In the above formula, NF is the chosen characteristic number, NS is the state number of
the model. In the training stage, each observation is performed for feature extraction.
And each observation in the training data set implicitly corresponds to a flag state,
therefore we count the relation between the state and the characteristic. By using the
GIS algorithm [16], the probability matrix of the characteristic–state transition can be
obtained. The GIS algorithm is as follows:
404 R. Li et al.

1) Calculate the average value of each characteristic-state in the training data.

When the total length of the observation sequence is ms, for the ith characteristic,
the jth state, its average value is
1
Fi, j = t ∑ m f ( o ,s )
s
(10)
ms k =1 i, j t t k k

2) Take as the 0th iteration of the GIS algorithm by the beginning of the random
parameter and make M i,( 0j ) = 1 .
3) In the nth iteration, use the current M i,( nj ) value and calculate the the expected
value of each characteristic-state:

1 m
Ei,( nj ) = ∑ ∑ P ( s | ot )f i, j ( ot ,s )
s (n)
(11)
ms k =1 k∈S s' k k

where s’ is the previous state of s, and

1
Ps'( n )( s | ot ) = exp[ ∑ M i,( nj ) f i, j ( ot ,s )] (12)
k
Z( o,s') i k

Z(o,s’) is the normalization constant which can guarantee that the value is one by
method of summation for all next state s of all s’ state.
4) In the situation of satisfying the constraints limit, combine the expected value
with the average value of the training data to decide how to adjust the parame-
ter. Designate a constant C, the adjustment formula is as follows.

1 Fi, j
M i(,nj+1) = M i(,nj) + log[ (n) ] (13)
C Ei, j

5) If parameter values converge, then terminate the algorithm, else go to Step (3).
In order to reduce the iterative number of times, we use the statistical method to as-
sign the initial value for M i,( 0j ) . When utilizing the Viterbi algorithm, the state at time t
is determined together by the probability of the state at time t -1 and the observation
characteristic at time t.

1
p(st = s j | st −1, ot ) = (λ, αt −1, j + (1 − λ) * ∑ (M i, j * f i, j (ot , st ))) (14)
γ i

Where α t −1, j is the transition probability from the known state at time t-1 to state j.
γ is a normalized parameter.

γ = ∑ (λαt −1, j + (1 − λ) * ∑ (M i, j * f i, j (ot , st ))) (15)

j i

here λ is the weight of adjusting the relatively important degree of the characteristic-
state transition probability and state transition probability.
 Application Study of Hidden Markov Model and Maximum Entropy 405

4 Experiments

For ease of comparison, we conduct the experiment by using the standard data set
which is provided by Carnegie Mellon University for computer scientific research
paper forehead information extraction. In order to reduce the time complexity of the
characteristic selection process in the maximum entropy method, we artificially select
some useful characteristics. The concrete state can be divided into the positive charac-
teristic and negative characteristic according to the contribution of the characteristic
attribute relative to the state.
The positive characteristic of the state indicates that when observation presents this
characteristic, the characteristic tends to shift to this state. The negative characteristic
of the state indicates that when observation presents this characteristic, the probability
of shifting to the state can decrease. In the analysis whether to contain the personal
name characteristic, we use the American personal name dictionary downloaded from
on-line to match.
When extracting information from the computer scientific research paper forehead,
owing to considering the useful information such as the typeset format, newline char-
acter, separating character, the text preprocessing first needs to be conducted by
adopting the HMM text information method based on text blocks. In the training
stage, the initial probability and the transition probability take the block as the Fun-
damental unit, and formula (1) and (2) are used. The output probability takes the word
as the Fundamental unit, and formula (3) is used. In the extraction stage, after the text
to be extracted is partitioned into small blocks, we calculate the output probability of
each block. The text sequence to be extracted is converted to the block sequence and
the block emission probability is the sum of the emission probability of each word in
the block. Suppose the block observation sequence is O = O1 O2…OT, if the length of
tth block is K, namely including K words, noted as Ot1 Ot2…OtK, the probability that
state j outputting the tth block is:
b j (ot ) = ∑ kK=1 b j (otk ) (16)

We only use a few characteristics such as personal name, all digits, month or its abbre-
viation, email symbol “@”(the block without containing the symbol can’t be email.).
In the experiment, by changing the value of parameter λ , the relatively important
degree of the characteristic-state transition probability and state transition probability.
When λ =0.6 and train-set=400, experimental results are shown in Table 1.
When using MEMM for text information extraction without considering the statis-
tical probability of words, the performance in the experimental data set in the paper is
poor and experimental results are not convenient for combination, therefore we don’t
list the related experimental results. As shown in Table 1, the text information extrac-
tion algorithm based on ME-HMM improves in some degree in the precision and
recall rate. When using the person name dictionary to match, Table 1 shows that the
recall rate of state “author” increases approximately 8%, which illustrates that the
joined related knowledge plays a large role.
406 R. Li et al.

Table 1. Precision and recall comparison of information extraction concrete domaion using
HMM and ME-HMM

HMM ME-HMM
domain
Precision(%) Recall(%) Precision(%) Recall(%)
title 0.820347 0.830480 0.838308 0.864868
author 0.823954 0.916533 0.898551 0.955185
affiliation 0.872078 0.914193 0.887361 0.916536
address 0.902686 0.840000 0.925010 0.853007
email 0.888176 1.000000 0.988436 1.000000
note 0.880784 0.715228 0.921323 0.692785
web 1.000000 0.546508 1.000000 0.546508
phone 0.986667 0.888878 0.986301 0.912470
date 0.660505 0.990808 0.697607 0.990808
abstract 0.953434 1.000000 0.910071 1.000000
Intro 0.872307 1.000000 0.872307 1.000000
keyword 0.824168 0.932762 0.824168 0.931662
degree 0.512413 0.838224 0.616536 0.812032
pubnum 0.884603 0.642828 0.954045 0.703434
page 1.000000 1.000000 1.000000 1.000000

5 Conclusions and Further Work

The maximum entropy model provides one natural language processing method, and
can integrate each kind of characteristic and rule into a unified frame. Incorporating
the maximum entropy method into hidden Markov model for text information extrac-
tion can solve the problem of knowledge representation problem and may momentar-
ily add the newly gained language knowledge to the model. Experimental results
show that this method is effective.
In the follow-up work, we plan that the system can select automatically characteris-
tic from a lot of candidate sets. Moreover, how to reduce the algorithm complexity is
also worth the further research.
Acknowledgments. We would like to thank to the reviewers for their helpful com-
ments. This work has been supported by the Natural Science Foundation of China
(Grant #60775041).

References
1. Lawrence, S., Giles, L., Bollacker, K.: Digital libraries andautonomous citation indexing.
Computer 32(6), 67–71 (1999)
2. Riloff, E., Jones, R.: Learning dictionaries for information extraction by multi-level boot-
strapping. In: Proceedings of the Sixteenth National Conference on Artificial Intelligence,
pp. 811–816. AAAI Press, Orlando (1999)
3. Kushmerick, N.: Wrapper induction: Efficiency and expressiveness. Artificial Intelli-
gence 118(12), 15–68 (2000)
 Application Study of Hidden Markov Model and Maximum Entropy 407

4. Seymore, K., McCallum, A., Rosenfel, R.: Learning hidden Markov model structure for in-
formation extraction. In: Proceedings of the AAAI 1999 Workshop on Machine Learning
for Information Extraction, pp. 37–42. AAAI Press, Orlando (1999)
5. Frietag, D., McCallum, A.: Information extraction with HMMs and shrinkage. In: Proceed-
ings of the AAAI 1999 Workshop on Machine Learning for Information Extraction,
pp. 31–36. AAAI Press, Orlando (1999)
6. Freitag, D., McCallum, A.: Information extraction with HMM structures learned by sto-
chastic optimization. In: Proceedings of the Eighteenth Conference on Artificial Intelli-
gence, pp. 584–589. AAAI Press, Edmonton (2002)
7. Ray, S., Craven, M.: Representing sentence structure in hidden Markov models for infor-
mation extraction. In: Proceedings of the Eighteenth Conference on Artificial Intelligence,
pp. 584–589. AAAI Press, Edmonton (2002)
8. Scheffer, T., Decomain, C., Wrobel, S.: Active Hidden Markov Models for Information
Extraction. In: Hoffmann, F., Adams, N., Fisher, D., Guimarães, G., Hand, D.J. (eds.) IDA
2001. LNCS, vol. 2189, p. 309. Springer, Heidelberg (2001)
9. Freitag, D., McCallum, A., Pereira, F.: Maximum entropy Markov models for information
extraction and segmentation. In: Proceedings of The Seventeenth International Conference
on Machine Learning, pp. 591–598. Morgan Kaufmann, San Francisco (2000)
10. Rabiner, L.E.: A tutorial on hidden Markov models and selected application in speech rec-
ognition. Proceedings of the IEEE 77(2), 257–286 (1989)
11. Kwong, S., Chan, C.W., Man, K.F.: Optimization of HMM topology and its model
parameters by genetic algorithms. Pattern Recognition, 509–522 (2001)
12. Hong, Q.Y., Kwong, S.: A Training Method for Hidden Markov Model with Maximum
Model Distance and Genetic Algorithm. In: IEEE International Conference on Neural
Network & Signal Processing (ICNNSP 2003), Nanjing,P.R.China, pp. 465–468 (2003)
13. Berger, A.L., Stephen, A., Della Pietra, V.J.: A maximum entropy approach to natural lan-
guage processing. Computational Linguistics 22(1), 39–71 (1996)
14. Chieu, H.L., Ng, W.T.: Named entity recognition with a maximum entropy approach. In:
Daelemans, W., Osborne, M. (eds.) Proc. of the CoNLL 2003, pp. 160–163. ACL, Edmon-
ton (2003)
15. Xiao, J.-y., Zhu, D.-h., Zou, L.-m.: Web information extraction based on hybrid condi-
tional model. Journal of Zhenzhou University 40(3), 52–55 (2008)
16. Darroch, J., Ratcli, D.: Generalized iterative scaling for log-linear models. Annals of
Mathematical Statistics 43(5), 1470–1480 (1972)
 Automatic Expansion of Chinese Abbreviations
by Web Mining

Hui Liu1 , Yuquan Chen2 , and Lei Liu2

1
Shanghai Institute of Foreign Trade,
1900 Wenxiang Rd., Shanghai, China
liuh@shift.edu.cn
2
Shanghai Jiao Tong University,
800 Dongchuan Rd., Shanghai, China
{yqchen,liu-lei}@sjtu.edu.cn

Abstract. Abbreviations are common in everyday Chinese. For applica-

tions like information retrieval, we want not only to recognize the abbre-
viations, but also to know what they stand for. To tackle the emergence
of all kinds of new abbreviations, this paper proposes a novel method
that expands an abbreviation to its full name employing the Web as the
main information source. Snippets containing full names of an abbrevia-
tion are obtained through a search engine by learned ”help words”. Then
the snippets are examined using linguistic heuristics to generate a list of
candidates. We select the optimal candidate according to a kNN-based
ranking mechanism. Experiment shows that this method achieves satis-
factory results.

Keywords: Abbreviation Expansion, Web Mining.

1 Introduction
Abbreviations take an important part in modern Chinese language. It is so com-
mon that native Chinese speakers will not even notice them in a newspaper or
a book. But for natural language processing, the wide spread of abbreviations
is not ideal. For computers, abbreviations are unrecognizable strings until they
are further processed. Even worse, the formation of Chinese abbreviations is not
quite like English ones[1]. Therefore, we cannot simply port an English analyzer
into Chinese.
A lot of works[2][3][4] treat abbreviations like out-of-vocabulary words, which
means they put an emphasis on how to tag a string as an abbreviation, just like
the recognition of a location name, a person name, etc. But for practical applica-
tions like information retrieval and machine translation[5], we should learn what
an abbreviation stands for: what is their original word form (called full name).
The relationship is crucial for better relevance evaluation of two documents.


This paper is supported in part by Chinese 863 project No. 2009AA01Z334 and
the Shanghai Municipal Education Commission Foundation for Excellent Young
University Teachers.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 408–416, 2009.


c Springer-Verlag Berlin Heidelberg 2009
 Automatic Expansion of Chinese Abbreviations by Web Mining 409

Chang and Lai[6] regard abbreviation expansion as an ”error recovery” prob-

lem in which the suspect full names are the errors to be recovered from a set
of candidates. They apply an HMM-based generation model for identification
and recovery of abbreviations. In subsequent works, Chang and Teng[1] use a
similar model to mine atomic Chinese abbreviation pairs. Fu et al.[7] suggest
a hybrid approach to the abbreviation expansion of Chinese. They generate
candidates using a mapping table and an abbreviation dictionary. Then the
candidates are selected also using an HMM model. Li and Yarowsky[5] com-
bine abbreviation expansion into a machine translation system to achieve bet-
ter translation results. Their unsupervised expansion strategy is based on data
occurrence.
One problem of these current approaches is data sparseness. Chang and Lai
note that in the expansion experiment 42% of test cases cannot be recovered
due to the sparseness of the training corpus. For statistical methods, sparseness
is always the problem. With the emergence of new abbreviations, using a static
corpus of texts from the last century is not ideal at all.
Therefore, in this paper we focus on expanding abbreviations to their full
names by a novel approach using World Wide Web as the main source of in-
formation. We resort to the Web to find co-occurrence of abbreviations and
their full names in returned snippets. The candidates are extracted and selected
based on a combination of linguistic heuristics and Web-based features. The ad-
vantage of using the Web as a corpus is that it is huge and full of redundancy,
which is nice for mining tasks and thus it can alleviate sparseness. With Web
mining method, we can keep up with new abbreviations by relatively simple
methods.
This paper is structured as following. In the next section we will talk about
some main ideas of our approach. We will introduce the construction of Chinese
abbreviations briefly and show the motivation and outline of our approach. In
section 3, we will probe into our method in detail. We will first discuss the
generation of candidates and then the selection of them. Section 4 will be devoted
to the experiments. The last section is the conclusion.

2 The Main Idea

2.1 The Construction of Chinese Abbreviations

There are three means of constructing abbreviations in Chinese, namely

1. Truncation;
2. Condensation;
3. Summarization.

In truncation, a integral part of a full name is taken as the abbreviation, like ”d

d” is the abbreviation of ”dddd”(Tsinghua University). It is the simplest ap-
proach but also the rarest. By contrast, condensation is the most common. Con-
densation takes out one morpheme for every selected word in the phrase to form
410 H. Liu, Y. Chen, and L. Liu

the abbreviation. For example, ”dddd”(Beijing University) is abbreviated as

”dd”. Summarization is also very common in Chinese, especially in slogans. In
summarization, the full name is a phrase or a sentence which contains several
words or clauses sharing a same constituent. The full name is abbreviated by a
number denoting the number of words/clauses plus the constituent. For instance,
”dd”(literally: Two Guang) is the abbreviation of ”dd”(Guangdong) and ”d
d”(Guangxi).
According to linguists[8], the formation of abbreviation conforms to some
weak rules. Abbreviations are as short as possible, in case that there will be
no ambiguities. The morphemes are taken sequentially to form the abbrevia-
tion. The first morpheme is favored in selection for abbreviation over other mor-
phemes in a word. These rules show light on further selection of candidates of full
names.

2.2 Outline of Our Approach

The crux of the expansion problem is to find possible candidates of full names. In
scientific writing, a full name, together with the proposed abbreviation, must be
presented before the use of the abbreviation. In this case, if having caught the re-
lated context of both full name and abbreviation, we can decipher the full name.
Unfortunately, it is not the case in daily usage like newspaper articles, since some
abbreviations are far more common than full names. For example, we will hardly
find ”dddddd”(The United States of America) in news except for govern-
ment documents, but the use of ”dd”(America) is very extensive. Full names
and abbreviations co-occur in some contexts after all, though the co-occurrence
is hard to find in articles or static corpus, for it is scarce. Web, on the contrary,
contains thousands of terabytes of information, chances are that we can find the
co-occurrence in some web documents. Moreover, with the wider and wider usage
of the Web, new abbreviations will definitely be recorded in the Web.
Our approach then consists of two parts naturally (Fig. 1). First, we extract
the possible contexts from the Web. Second, we recognize the full name from the
context. Since we have only a limited access to the Web, that is to say, through
a search engine like Google or Baidu1 , which will give us less than 1000 results
per query, in most times even less, we cannot happily expect a full name to
emerge in a single query like ”dd”(America). In contrast, we have to tweak
the queries to have the results containing possible full names. After the snippets
are extracted, we move on to an extraction module in which we use linguistic
heuristics to generate a list of possible candidates. The list is further pruned
using co-occurrence data (local Web data) in the snippets for efficiency’s sake.
Finally, we rank the candidates in the pruned list to find the optimal result by
features acquired from the Web (global Web data) using the k nearest neighbor
method[9].

1
http://www.google.cn, http://www.baidu.com
 Automatic Expansion of Chinese Abbreviations by Web Mining 411

Fig. 1. The overview of this paper’s system

3 The Web-Based Full Name Expansion

3.1 Query Learning and Candidate Extraction
First, we have to find text segments containing abbreviations together with their
full names through a search engine. If we construct the query using the target ab-
breviation only, like ”dd”, the returned snippets will not possibly contain any
full names in a single search, because abbreviations appear far more frequently
than full names, as mentioned before. Therefore we shall tweak the queries to find
needed snippets. In other words, we will learn some types of queries containing
specific ”help words” that will lead to snippets containing full names.
The learning method is as following. Having a list of training examples of
pairs of <abbreviation, full name>, which will also be used in later stages of
processing, we first search both the abbreviation and the full name in a search
engine using the following query.
”abbreviation” ”full name”
The resulting snippets are segmented and some words are taken as potential
help words. The helpfulness of a word w for an abbreviation a are calculated as:

0 w ∈ ST OP
h(w) = (1)
|Q| else.
In Equation 1, Q is a set of queries whose results sq contain both w, a and its
full name f . Moreover, we hope that the occurrence of (w, a, f ) is above some
threshold γ.

Q = {q|sq ⊃ w ∧ sq ⊃ a ∧ sq ⊃ f ∧ Occ(w, a, f ) > γ} (2)

Note that if w occurs in some query’s results for over γ times, we add h(w) for
1, rather than count the occurrences of w directly from all returned snippets,
412 H. Liu, Y. Chen, and L. Liu

because for a specific abbreviation there may be some special word that will
appear again and again in the results. For generality’s sake, we will consider the
count of related queries only.
With all the help words H, the queries are constructed as

”abbreviation” ”full name” ”help word”

Putting these queries through a search engine, we will have a list of returned
snippets that may contain full names. We will then extract all the possible can-
didates of full names. The strategy is simple: according to linguistic observations
in section 2.1, we assume that a full name will always include all characters in
the abbreviations.2 Therefore, if locating all the characters from the abbrevia-
tions in the snippets, we can find the first candidate by simply extracting the
shortest string containing them. For example, for ”dd” we will first find ”d
ddddd”. For ”dd” we will first extract ”dddd”. Starting from the
first candidate, we scan in both directions to construct more candidates until we
meet a ”boundary word”. A boundary word is either a word in a given stop word
list, or a word having appeared next to the abbreviation in question. Namely,
the set of candidate C is a set of word sequences defined as following:

C = {s|∀s ⊂ s s ∈ ST OP  ∧ ∀ch ∈ a, ch ∈ s} (3)

In the above equation, s and s are strings, ST OP  is a list of boundary words,

ch a character.

3.2 Candidate Selection

After the candidates are extracted, we have to decide which one is the full name.
Before selection, we will first prune the list to exclude less likely candidates, since
selection approach, which is based on Web features, is quite time intensive.
In a straight-forward manner we can just count the occurrences of the candi-
date strings in the snippets and delete ones with the lowest count. This approach
is somewhat clumsy in the sense that c1 which is a substring of c2 will always
has higher counts than c2 , because each occurrence of c2 contains a c1 .
Though the occurrence count is not applicable directly as the sole evidence,
we can use it as a basis for further judgment. We can model the scenario as a
comparison problem. If two strings do not have a subsumption relation between
them, the above count is the measurement. Otherwise, we will subtract the co-
occurrence count of c2 , the super-string from that of the substring, c1 .

Definition 1. Let  be a relation on the Cartesian product of two string sets,

S × S, in which ∀s1 , s2 ∈ S, s1  s2 iff s1 is more likely to be a full name.

For any s1 and s2 we can draw the relation  between them according to the
following:
2
The assumption is not always true as we show in section 4.
 Automatic Expansion of Chinese Abbreviations by Web Mining 413
⎧
⎨ O(s1 ) > 2O(s2 ) s1 ⊆ s2
Proposition 1. ∀s1 , s2 ∈ S, s1  s2 iff O(s1 ) > O(s 2)
s2 ⊆ s1
⎩ 2
O(s1 ) > O(s2 ) Otherwise.
With  defined on any string pairs of S, we know that  is a total order.
Therefore, we can have some top candidates for further selection.
The selection of an optimal full name from the pruned list is seen as a ranking
problem. The ranking values are obtained using kNN. The training examples
are transformed into a real value vector − →
x of features. The ranking value of a
candidate c for an abbreviation a is the average Euclid distance from k nearest
neighbors in the examples. Candidates with the lowest value (the nearest) is
taken as the final result.
Two kinds of real value features are used here, the structural features and
the Web-based features. The structural features represent how the abbreviation
is constructed, i.e. the distribution of abbreviation characters (characters in the
abbreviation word) in full names. The Web-based features show the occurrences
of the candidates in Web context. Table 1 shows the features and some formulae
for computing them. In the table Len is a function for a string’s length, Ch is
the set of abbreviation characters, I(t) is a function that maps an abbreviation
character t to its index in the full name, and E is the set of characters that are
expanded into 2-character words in the full name.

4 Experiments and Discussions

The test set of our experiment consists of 480 abbreviations. The data are from
two sources. One half of the abbreviations (Data Set 2) are extracted randomly

Table 1. Features used

Type Name Formula and Description

Len(a)
density Len(c)
Length
 of abbreviation characters versus the candidate
t∈Ch I(t)/Len(c)
Structural
center
Position
 of characters in the candidate.
0<i<Len(a) (I(ti ) − I(t( i − 1)))
2
diversity
How abbreviations are scattered in the candidate;
maxt,k∈Ch I(t)−I(k)
span Len(c)
The length of span occupied by abbreviation characters.
|E|
expand |Ch|
Number of characters that are expanded into words in the candidate.
/
wordcount
Occurrence of candidate in Web; Query: ”c”
Web
/
titlecount
Occurrences of candidate in titles of Web pages; Query: intitle:”c”
1
 Len(c)
titledensity Nsnippet Len(T itle)
Density of candidate in titles; Query: intitle:”c”
/
indcount
Occurrence of the candidate in whole if search has no quotation; Query: c
414 H. Liu, Y. Chen, and L. Liu

from the PFR People’s Daily Corpus3 , which contains articles from People’s
Daily in 1998. We take 2-character words tagged as ”j” for abbreviations. We
do not include single character words because they are often covered in an ab-
breviation dictionary. Another half of the data (Data Set 1) are hand selected
from the Web by an editor who is specialized in neither computer science nor
linguistics. The part of information represents the usage of abbreviations in our
daily life.
The training examples contain 50 abbreviations from the PFR corpus, to-
gether with their full names. In each query throughout the expansion we retrieve
100 snippets from Baidu. Since Web-based expansion is time intensive, we only
use 5 help words from learning. We evaluate the accuracy of expansion of Data
set 1, Data set 2 and the two sets in general. The overall results are shown in
Table 2. Apart from our final system, we present several other baseline systems,
B DICT, B NOKNN and B NOHELPWORD. B DICT relies on an abbreviation
dictionary for expanding. The dictionary is generated from the Modern Chi-
nese Standardized Dictionary[10]. The other baseline systems B NOKNN and
B NOHELPWORD only utilize parts of the final system. B NOKNN uses only
local Web count (occurrence counts from snippets of candidate generation) for
selecting the optimal result without the kNN ranker, while B NOHELPWORD
do not use help words but only the abbreviation itself as a query to generate
snippets.
Our approach does not perform so well on data set 2 as on data set 1. It
is partly due to the fact that the PRF corpus contains texts ten years ago,
so the abbreviations are somewhat too ”old-fashioned” to appear on the Web.
Moreover, some abbreviations tagged in the corpus are used as common words
now. From the Table 2 we can also see that B DICT has the lowest performance,
since the dictionary does not cover many abbreviations in use. Both B NOKNN
and B NOHELPWORD have similar performance, about 0.3 lower than our final
system. This inferiority shows the indispensability of the two main components:
help words and the kNN based selector.

Table 2. Results of Evaluation: Comparison of baseline systems

Systems Data set 1 Data set 2 All

B DICT 0.1625 0.1 0.1313
B NOKNN 0.5125 0.3458 0.4292
B NOHELPWORD 0.5833 0.4042 0.4938
Final 0.8792 0.8033 0.8412

We also look into the contribution of each feature used in the kNN-based
selector. We evaluate the contribution in a ”negative” way, i.e. excluding one
feature each time to see the impact on the evaluation result. The results are
shown in Table 3. We can see that all the features contribute to the final result,
3
Available at http://icl.pku.edu.cn/icl groups/corpus/dwldform1.asp
 Automatic Expansion of Chinese Abbreviations by Web Mining 415

Table 3. Contribution of Features

Feature Excluded Data set 1 Data set 2 All

None 0.8792 0.8033 0.8412
-titlecount -0.0833 -0.0875 -0.0854
-titledensity -0.075 -0.06667 -0.0708
-wordcount -0.2917 -0.3167 -0.3042
-idcount -0.125 -0.1667 -0.1458
-density -0.2583 -0.3167 -0.2875
-span -0.0958 -0.125 -0.1104
-diversity -0.15 -0.175 -0.1625
-center -0.2333 -0.3208 -0.2771
-expand -0.1583 -0.2125 -0.18542

with the contribution ranging from 8% to more than 30%. For Web generated
features, one interesting thing is that features about the title of snippets (title-
count and titledensity) have less contribution than others. This may be due to
the fact that for some long full names, people tend to use their abbreviations
in titles. Another interesting finding is that all the features except for one con-
tribute more to data set 2 than data set 1. This is because it is more difficult to
generate full names for abbreviation from Data set 2 from the Web than those
in data set 1. Therefore, the candidate list of full names is more error prone,
which leads to the larger contribution of the kNN-based selector together with
involving features.
One surprise for us is that our approach can handle abbreviations for very
complex and long slogans, for example, ”dddd”’s full name is ”ddddd
dddddddddddddddddddddddddddddddddd
ddddd”. This shows the merit of help words, for if we search ”dddd”
only, we will not find the complete slogan in the snippet.
The errors of expansion can be sorted in several categories. First, some ab-
breviations coincide with some common expressions. For example, ”dd” is
the abbreviation of ”ddddd” (Shanghai Library), while it also means ”the
above graph”. Therefore, we are not able to get the desired full name through
Web search due to the overflow of the more common usage. Second, for some
long full names the later parts of them do not appear in the abbreviations, thus
often neglected in both snippets and our processing. For example, ”dd” is ex-
panded into ”dddddd” instead of ”dddddddd” (Chinese Football
League A). Third, some abbreviations, especially those are not named entities,
are always used as a single word. Consequently, it is hard to find its original
form, even for a man. For example, ”ddd” which is labeled as an abbrevi-
ation in PRF corpus, is the abbreviation of ”ddddd” (the young and the
middle aged). However, ”ddd” is so widely used that we can hardly find the
expression of ”ddddd”.
416 H. Liu, Y. Chen, and L. Liu

5 Conclusion
In this paper we suggest a novel method to expand Chinese abbreviations into
their full names. Unlike previous works, we use World Wide Web as the main
source of information. The approach consists of two parts: first, we conjure
queries using learned help words to extract related Web contents containing both
the abbreviation and full name through a search engine. Second, we extract can-
didates of full names from the contents and select the best result. The extraction
is based on linguistic heuristics. The final selection consists of two parts: a pruner
based on occurrence information within obtained snippets and a kNN-based se-
lector utilizing a set of Web-based and structural features. Experiment shows
that our approach enjoys satisfactory results. We also show the indispensability
of help words and kNN-based ranker through experimental results. Moreover,
the contributions of different features are discussed and compared.
For further work, we are to compare the role of search engines in our task.
We are curious that whether different search engine will have different impact
on our task, or other Web mining tasks. Another interesting topic is to apply
content information for refined expansion. In this paper we have not considered
polysemy, which also exists in abbreviations. Disambiguation is a tough topic to
deal with in the future.

References
1. Chang, J., Teng, W.: Mining atomic Chinese abbreviation pairs: A probabilis-
tic model for single character word recovery. Language Resources and Evalua-
tion 40(3/4), 367–374 (2007)
2. Chen, K., Bai, M.: Unknown word detection for Chinese by a corpus-based learning
method. Computational Linguistics 3(1), 27–44 (1998)
3. Sun, J., Gao, J., Zhang, L., Zhou, M., Huang, C.: Chinese named entity identifi-
cation using class-based language model. In: COLING 2002, pp. 24–25 (2002)
4. Sun, X., Wang, H.: Chinese abbreviation identification using abbreviation-template
features and context information. In: Matsumoto, Y., Sproat, R.W., Wong, K.-F.,
Zhang, M. (eds.) ICCPOL 2006. LNCS (LNAI), vol. 4285, pp. 245–255. Springer,
Heidelberg (2006)
5. Li, Z., Yarowsky, D.: Unsupervised Translation Induction for Chinese Abbrevia-
tions using Monolingual Corpora. In: Proceedings of ACL, pp. 425–433 (2008)
6. Chang, J., Lai, Y.: A preliminary study on probabilistic models for Chinese abbre-
viations. In: Proceedings of the Third SIGHAN Workshop on Chinese Language
Learning, pp. 9–16 (2004)
7. Fu, G., Luke, K., Zhang, M., Zhou, G.: A hybrid approach to Chinese abbrevia-
tion expansion. In: Matsumoto, Y., Sproat, R.W., Wong, K.-F., Zhang, M. (eds.)
ICCPOL 2006. LNCS (LNAI), vol. 4285, pp. 277–287. Springer, Heidelberg (2006)
8. Huang, L.: More on the construction of modern Chinese abbreviations. Journal of
Suihua University (004) (2008)
9. Mitchel, T.: Machine Learning 48(1) (1997)
10. Li, X.: Modern Chinese Standardized Dictionary. Foreign Language Teaching and
Researching Press, Language and Literature Press, Beijing (2004)
 Adaptive Maximum Marginal Relevance Based
Multi-email Summarization

Baoxun Wang, Bingquan Liu, Chengjie Sun, Xiaolong Wang, and Bo Li

School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001,
China
{bxwang,liubq,cjsun,wangxl,lib}@insun.hit.edu.cn

Abstract. By analyzing the inherent relationship between the maximum mar-

ginal relevance (MMR) model and the content cohesion of emails with the same
subject, this paper presents an adaptive maximum marginal relevance based
multi-email summarization method. Due to the adoption of approximate com-
puting of email content cohesion, the adaptive MMR is able to automatically
adjust the parameters according to the changing of the email sets. The experi-
mental results have shown that the email summarizing system based on this
technique can increase the precision while reducing the redundancy of the
automatic summary results, consequently improve the average quality of email
summaries.

Keywords: multi-email summarization, maximum marginal relevance, content

cohesion, adaptive model.

1 Introduction
With the popularization of email service, multi-email summarization is highly de-
manded. As an information fusion technique based on email contents, multi-email
summarization can supply users with direct and compact summaries. Since users are
used to talking about a specific subject via emails [1], a set of emails of the same
subject can be clustered together in the email client, and users from client side usu-
ally would like to obtain a panoramic view of these emails for further analysis and
decision.
Multi-email summarization is a kind of operation on email contents, which have
some characteristics that web page contents do not possess. We take the following
two facts as examples: firstly, email contents are usually very short, so there is much
less information in email contents for language understanding. Secondly, when writ-
ing emails, people tend to use a casual tone, which is more closed to their oral habits,
so parsing the contents with a set of rules becomes very tough. The two aspects above
introduce the main difficulty to multi-email summarization. Besides, some other tech-
niques should be considered in order to build a multi-email summary system, such as
Base64 and Quoted Printable decoding, time and title extracting, and so on.
In this paper, we present an adaptive maximum marginal relevance (AMMR) based
approach to generate multi-email summaries, which takes the extracting summary way
and is able to automatically adjust the parameters according to the content cohesion of

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 417–424, 2009.
© Springer-Verlag Berlin Heidelberg 2009
418 B. Wang et al.

the email collections. Experimental results show that our method can improve the
average quality of multi-email summaries and meet the need of system application.
The rest of this paper is organized as follows: Section 2 discusses the related work.
Section 3 presents the proposed adaptive MMR model and its application in multi-
email summarization. We evaluate our work in section 4. Section 5 concludes this
paper and discusses the future work.

2 Related Work
The researches on multi-email summarization have just started now, so most people
take the extracting approach. Wan et al. [2] extract question-answer pairs to form
email summaries, and a similar method is also discussed in [3], in which Shrestha
et al. generate email summaries by joining extracted question-answer sentences. Since
QA pairs can reflect the semantic clue of a given email thread, such simple strategies
improve the performance of summarization systems to a certain extent, however the
disadvantage of this kind of approach is obvious, that is, email summaries based on
this method are lack of consistency. Moreover, not all the QA sentences are related to
the topic of the email thread.
Rambow et al. [4] introduce machine learning techniques to multi-email summari-
zation. They first extract features for every sentence, and then decide whether a given
sentence should be chosen as a part of the summary. In fact, this strategy has been
applied in traditional multi document summarization. Ntt [5] employs SVM to extract
summary sentences, and their basic idea is to find the hyperplane to separate summary
sentences from the sentence set. A summary achieved by classify methods has a
higher precision, however its redundancy is also higher, because the sentences ex-
tracted by classifier are similar to each other.
Comparing with multi-email summarization, summarization on multi documents is
developing better, and many approaches are used to solve the problem. On summari-
zation based on the extraction strategy, besides the method in [4], O’Leary et al. [6]
compute the probability of the sentence becoming a member of the summary by using
HMM. Kraaij et al. [7] extract summary sentences based on the term frequency. By
introducing the deep mining of document relation, people have explored information
fusion based summarization. Radev [8] presents a cross-document structure, so as to
fuse information of multi documents on every semantic level. Xu [9] proposes a
multi-document rhetorical structure (MRS), and has done many researches on sum-
mary sentence extracting, ranking, redundancy reducing, and summary generating.
Although multi document summarization is different from multi-email summarization
on the object to summarize, there is still some experience for the latter to learn.
Our approach employs the improved extracting strategy to generate multi-email
summaries. By adopting the maximum marginal relevance (MMR) model, the
sentences extracted to form a summary are more relative with the subject, and the
redundancy can be reduced in our summary system. However MMR has an inherent
disadvantage that it is not able to adjust its parameters automatically for different
email sets, which makes the average quality of email summaries decline. In this pa-
per, we analyze the inner relationship between MMR and the content cohesion of
emails of the same subject, and propose an AMMR model, which improves the aver-
age quality of email summaries.
 Adaptive Maximum Marginal Relevance Based Multi-email Summarization 419

3 AMMR Based Multi-email Summarization

3.1 MMR Model

Maximum marginal relevance (MMR) model is proposed by Golestein et al. [10],

which was employed to improve the performance of search engine by reducing the
redundancy of documents retrieved. This thought of lowering redundancy while main-
taining relative degree is also the objective of summary sentence selection. Obviously,
the sentences related to the subject of emails are repeated in semantic, so they are not
suitable to be selected as summary sentences. The sentences that are not related to the
subject have lower redundancy, but they can not be chosen to form a summary.
To some extent, subject relevance and redundancy form a pair of contradictory
concepts, which can be integrated by MMR model. The model is defined as:

MMR ( R , S ) = Arg max ⎡ λ Sim1 ( si , q ) − (1 − λ ) max Sim2 ( si , s j ) ⎤ .

def
(1)
s i ∈R \ S ⎢⎣ s j ∈S ⎥⎦
Where R stands for the set of all the sentences belonging to a given subject, S denotes
the set of sentences in the summary, and si indicates a sentence. Sim1 and Sim2 are
two functions for describing, which can be either the same or different. Sim1 is used to
depict the similarity of a given sentence and the subject, while Sim2 is used to describe
the similarity of two sentences. λ is a factor of linear interpolation.
From equation 1 we can see that the key of MMR based summarization is to bal-
ance between subject relevance and redundancy by tuning factor λ . Obviously, the
sentence which makes the result of Sim1 larger and that of Sim2 smaller is likely to be
selected as the summary sentence.
The subject of emails is another factor for consideration, which is normally pre-
sented by a set of terms and plays an important role in sentence selection. We
discover that users’ queries can reflect the subjects of the emails retrieved. Conse-
quently, we take users’ queries as subjects in the use of our AMMR model.

3.2 AMMR Based Multi-email Summarization

As mentioned above, λ is an important changeable parameter in MMR model, which is

the key factor to balance between subject relevance and redundancy. In traditional
MMR, this parameter is usually a human set value ranging from 0 to 1. Our experimen-
tal results show that, however, with λ changing in [ 0,1] , the quality of the summary
fluctuates significantly, and for different email sets, the values assigned to λ corre-
sponding the maximum values of summary quality are different. Adjusting λ for dif-
ferent email sets automatically to make the summary quality to approach to the maximal
is an effective method to improve the performance of multi-email summarization.
To explain AMMR, we first propose the idea of content cohesion, which is able to
reflect the cohesion degree of sentences belonging to a subject. Given a collection of
emails and their subject, we assume that all the sentences revolve around this subject,
420 B. Wang et al.

with different contributions. Assuming that there is a way to model the subject cor-
rectly and the reasonable features of sentences can be extracted to convert them to
vectors, we can regard content cohesion as the variance of the sentence vectors.
However, the content cohesion is difficult to compute since the variance is hard to
compute. Thus, we present an approximate method and the idea is as follows. The
content cohesion is closed to sentence selection during summary generation. Suppos-
ing that the content cohesion of a given email selection is large, we can draw the con-
clusion that the percentage of sentences related to their subject is large, so we only
have to select one or two most relative sentences as representative, and extract proper
sentences which are not very relative as the supplement of the summary content. Oth-
erwise we should choose more sentences that are related to the subject to generate a
summary, and reduce the number of irrelative sentences. In this paper, we introduce
content cohesion into MMR to build an adaptive model, which is defined as follows:

AMMR ( R , S ) = Arg max ⎡(1 − N q N ) Sim1 ( si , q ) − N q N max Sim2 ( si , s j ) ⎤ .

def

(2)
s i ∈R \ S ⎢⎣ s j ∈S ⎥⎦

Where N q stands for the number of sentences containing the words in the query, and
N denotes the number of all the sentences in the email collection. We estimate the
content cohesion by N q N , for the more query words a sentence containing, the
more relative it is to the subject. And the estimated content cohesion is adopted to
replace λ , in order to adjust the parameter of the model according to different email
collections automatically.
The proposed AMMR based multi-email summarization algorithm is depicted in
Figure 1.

Input: query q and a set of email with the same subject

Output: the email summary
Algorithm:
1. divide the contents of the emails into sentences to form a candidate
sentence set R;
2. count the number of sentences containing the words in query q, and
compute N q N ;
3. compute the similarity of sentences in R and q, take the most similar
sentence as the first member of the summary, remove it from R and put
it to the summary sentence set S;
4. use equation 2 to compute the AMMR value of the remaining sen-
tences, choose the sentence with the highest value and put it into S;
5. loop step 4 until the length of the summary is enough;
6. return the sentences in S ranked by time feature;

Fig. 1. The algorithm based on AMMR

 Adaptive Maximum Marginal Relevance Based Multi-email Summarization 421

3.3 Sentence Similarity Computing Based on HowNet

Sentence similarity computing plays an important role in both AMMR and the evalua-
tion of summary results. VSM is widely employed to compute similarity in NLP,
however it does not perform very well in our experiments. According to our investi-
gation, this problem is caused by the individuation and informality of people while
writing emails, which directly lead to the existence of word variants. As we know, the
fundamental idea of VSM is the words’ cooccurrence, but in email contents there are
many words with the same meaning but totally different morphology.
In this paper, we adopt the method based on HowNet to compute semantic similar-
ity of sentences, which is built on the computation of word semantic similarity [11].
Our strategy is to convert the sentence similarity to the weighted mean of word simi-
larities. A greedy algorithm is employed to match the most similar words in two sen-
tences to collect word pairs, and take inverse document frequency as the weight for
computing the weighted mean of the words. Experimental results have shown that the
average quality of summary has increased by 4% to 6% after introducing this strategy,
which will be given in detail in section 4.

4 Experiments

4.1 Evaluation

In this paper we adopt direct evaluation [12] to evaluate the performance of an email
summarization system. We first choose several typical user queries as the subjects,
and then acquire the top k emails respectively from the email retrieval system to form
email collections. For every group of emails, 5 summaries are generated manually by
different people as benchmark, with which we can evaluate the machine summaries.
We compute the precision and the redundancy for each email summary and get the
quality of the summary according to them. Precision can be computed as follows:

precision = Sim ( SP, SA ) . (3)

Where SP stands for the summary generated by human, and SA stands for the sum-
mary generated automatically. Sim is a similarity computing function.
We consider redundancy as the semantic similarity of the sentences within the
summary, so it can be obtained by computing the average similarity of the sentence
pairs within the machine summary, which is defined by:

max_ sim ( si , s j ) .
1
redundancy = ∑
n i , j =1,...n ,i ≠ j
(4)

And the quality of the summary is defined as follows:

quality = precision − redundancy . (5)
422 B. Wang et al.

4.2 Experimental Results

We choose 6 groups of real emails as test corpus, and each group contains 15 to 20
members. Since the emails are obtained from our email retrieval system, the subject
of every group can be expressed by the user’s query. The experimental results are
shown in figure 2, where the horizontal axis represents the value of linear interpola-
tion factor λ , and vertical axis stands for the quality of summaries defined in (5).

Fig. 2. The results of multi-email summary

`
 Adaptive Maximum Marginal Relevance Based Multi-email Summarization 423

From the results we can see the general trend is that the quality can reach its
maximum value while λ ranging from 0 to 1. But for different subjects, λ is not at a
fixed value or interval when the maximum value appears, which is reasonable accord-
ing to our discussion on the content cohesion of emails, so the summarization system
based on traditional MMR model is unable to achieve a high average quality.
In figure 2, the position of λ marked by the dashed line represents the value of
linear interpolation introduced by AMMR, and the corresponding value on vertical
axis is the quality AMMR can reach. We can see that our method is able to approach
the maximum value of the summary quality on all the 6 groups, especially for the last
group, the maximum quality appears when λ is equal to about 0.4, which signifi-
cantly deviates from the rest 5 groups, and the AMMR model performs well despite
of this situation.
In Figure 2, the red dotted line represents the results of the model taking HowNet
based sentence similarity computing method, and the black squared line stands for
that using VSM, we can see the quality has increased by 4% to 6%, which means that
the summary sentences selected tend to be more reasonable after introducing this
semantic based strategy.

5 Conclusion
Aiming at increasing the average quality of multi-email summaries, we propose a
novel multi-email summarization technique based on adaptive MMR model. We ana-
lyze the relationship between content cohesion and the parameters of MMR and find
out that the content cohesion is one of the main factors that lead to the decline of
summary quality. This motivates us to present an adaptive summarization model to
improve the quality of email summaries. The experimental results show that our
model can effectively improve the average quality of email summaries. Our work can
be applied to build an automatic summary module on the email client.
In the future, we will focus on the following problems: 1) to model email content
cohesion in an exact way; 2) to adopt more linguistic features to make the computa-
tion of sentence relevance reasonable; 3) to develop algorithms for improving the
readability of summary generated.
Acknowledgments. This investigation was supported by the project of the High
Technology Research and Development Program of China (grants No.
2006AA01Z197 and 2007AA01Z172), the project of the National Natural Science
Foundation of China (grant No. 60673037) and the project of the Natural Science
Foundation of Heilongjiang Province (grant No. E200635).

References
1. Fisher, D., Moody, P.: Studies of automated collection of email records. University of
Irvine ISR Technical Report UCI-ISR-02-4 (2002)
2. Wan, S., McKeown, K.: Generating overview summaries of ongoing email thread discus-
sions. In: Proceedings of COLING 2004, the 20th International Conference on Computa-
tional Linguistics, pp. 745–751 (2004)
424 B. Wang et al.

3. Shrestha, L., McKeown, K.: Detection of question-answer pairs in email conversations. In:
Proceedings of COLING 2004, pp. 889–895 (2004)
4. Rambow, O., Shrestha, L., Chen, J., et al.: Summarizing email threads. In: Proceedings of
HLT/NAACL, Boston, USA (2004)
5. Hirao, T., Sasaki, Y., Isozaki, H., et al.: NTT’s Text Summarization System for DUC
2002. In: Proceedings of the workshop on automatic summarization, Philadelphia, Penn-
sylvania, USA, pp. 104–107 (2002)
6. Zajic, D.M., O’Leary, D.P.: Sentence Trimming and Selection: Mixing and Matching. In:
Proceedings of the 2006 Document Understanding, New York (2006)
7. Kraaij, W., Spitters, M., van der Heijden, M.: Combining a Mixture Language Model and
Naive Bayes for Multi-document Summarization. In: Proceedings of the DUC 2001 work-
shop (SIGIR 2001), New Orleans (2001)
8. Radev, D.R.: A Common Theory of Information Fusion from Multiple Text Sources Step
One: Cross-Document Structure. In: Proceedings of the 1st ACL SIGDIAL Workshop on
Discourse and Dialogue, Hong Kong, pp. 74–83 (2000)
9. Xu, Y.: Research of Multi Document Automatic Summarization. Dissertation of the Doc-
toral Degree in Engineering, Harbin Institute of Technology (2007)
10. Carbonell, J., Goldstein, J.: The use of MMR, diversity-based reranking for reordering
documents and producing summaries. In: Proceedings of the 21st annual international
ACM SIGIR conference on research and development in information retrieval, Melbourne,
Australia, pp. 335–336 (1998)
11. Li, S.J., et al.: Semantic computation in a Chinese question-answering system. Journal of
Computer Science & Technology 17(6), 933–939 (2002)
12. Jones, K.S., et al.: Automatic Summarizing Factors and Directions. In: Advance in Auto-
matic Text Summarization. MIT Press, Cambridge (1998)
 Semantic Relation Extraction by Automatically
Constructed Rules

Rongfeng Huang

Shanghai Jiaotong University, Shanghai, China

silverfox@sjtu.edu.cn

Abstract. This paper presents an approach to construct rules auto-

matically to identify semantic relations, such as Hypernym, Part Of and
Material from Modern Chinese Standard Dictionary (MCSD)1 . This ap-
proach combines all the useful information such as part of speeches, syn-
tactic information, and extracted semantic relations together in order
to identify as many semantic relations as possible. However, not all the
information provided, like syntactic information, is reliable. A method
is required to ensure that the ”right” information up to a point can be
picked. And then all possible semantic relations are constructed as a net
after being identified. Finally, we evaluate the net by determining word
similarity.

Keywords: semantic relation,machine-readable dictionary, word

similarity.

1 Introduction

Semantic resource is very useful in many Natural Language Processing (NLP)

tasks like Information Retrieval, Machine Translation and so on. Some semantic
resources have been constructed manually like WordNet [1] and HowNet [2]. And
they have showed their immense values. However, manual construction of such
resources is time-consuming, high cost and susceptible to arbitrary decision of
human. Therefore, lots of automatic approaches to extract semantic relations
have been developed. One method is to identify a semantic relation between two
words in a sentence. Usually, words between two destination words are key in-
formation. For example, Dmitry Zelenko et al [3] use a kernel method to identify
relations. Chinatsu Aone et al [4] use a pattern-based method.
Another method, usually taking dictionary as source, is to identify semantic
relations between a head word and words in its definition. The main difference
between these two methods is head word does not exist in sentences of definition
in the latter one. Therefore, we have to consider whole context of a word in
definition. In English, the most impressive achievement is MindNet, which are
generated by applying rules on results proviede by a broad-coverage parser [5].
1
Xingjian Li, Modern Chinese Standard Dictionary, Foreign Language Teaching and
Research Press, Beijing, 2004.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 425–434, 2009.


c Springer-Verlag Berlin Heidelberg 2009
426 R. Huang

However, an excellent syntactic parser is usually inaccessible. As a way, a special

one may be designed. Barriere [6] implements a chart parser to aid his extraction
task from a children’s dictionary. Others use patterns, most of which are at string
level. Markowitz [7] gives several general patterns.
From recent works, string pattern matching seems an alternative way to ex-
tract semantic relations from Chinese dictionary, as a syntactic parser for Chi-
nese usually has a much lower precision than one for English. However, complex
structural information cannot be expressed adequately in string patterns [8]. In
other words, we should make use of more information besides string patterns.
This paper describes an approach which exploits MCSD to identify semantic
relations by automatically constructed rules. We use various types of information
to capture as much information as possible. And we try to uniform representa-
tions of different types’ information to make sure they work well with each other.
We also try to use a way to avoid abusing unreliable information. One hypothesis
used here is ”wrong” information occurs occasionally while ”right” information
doesn’t.
In the rest of this paper, we describe types of information used and a way of
using them to construct rules (Sect. 2). And then we introduce types of identified
semantic relations and a way of constructing all identified semantic relations as a
net (Sect. 3). We evaluate the result by hand-checking a scale of random samples
(Sect. 4). However, in order to get a more objective evaluation, we evaluate the
net by determining word similarity (Sect. 5). Also, we discuss which aspects of
the approach should be improved (Sect. 6). Finally, we give a conclusion (Sect. 7).

2 Identification Method
We use various types of information called features and we uniform their repre-
sentation to make them work well with each others. And then we construct rules
with a list of features.

2.1 Features
There are five types of features, including words, part of speeches, syntactic
information, positions, and extracted semantic relations. And we use a ”Feature-
Value” structure to uniform their representations.

Types. In order to be referred easily, a word in definition is called a ”candidate

word” if its part of speech matches the requirement of a target relation. For
example, if and only if a word is a noun, it can be a candidate word that holds a
Hypernym relation with the headword. Except for a candidate word, rest words
of a sentence are called ”definition words” of this candidate word.
Five feature types are described as below.
Firstly, definition words and part of speeches of them are considered just like
most other ways do, no matter they are rule-based or pattern-based. However,
we do not plan to encode any definition word and its part of speech as only
one unit feature. In fact, there is a transition sacrificing precision for recall from
 Semantic Relation Extraction by Automatically Constructed Rules 427

words with part of speeches to only part of speeches. Therefore, it’s necessary to
decide which words should be kept in their forms while others are better to be
kept as part of speeches only. In this paper, all decisions about these are made
depending on the distribution of words in corpus.
Secondly, result of syntactic analysis is another kind of important useful in-
formation. But unfortunately, a reliable syntactic parser like the broad-coverage
parser used in MindNet is inaccessible at most of the time, especially when we
deal with Chinese. So, we should decide which syntactic information provided by
a not very reliable parser is useful for us. All relevant decisions about these are
also made depending on the distribution of result of syntactic analysis in corpus.
In this paper, a dependency parser is used. Based on the syntactic results, a
further analysis is applied. For example, ”Is X a subject of a sentence? (X is a
definition word)”.
Thirdly, the positions of definition words are also taken into account. This
is an assistant strategy in fact. One reason is the unreliability of the syntactic
parser. And our suspicions will be raised when the number of syntactic relation-
ships of the path linking from a definition word to a candidate word is getting
bigger. Another reason is that a certain word may play different syntactic roles
in different sentences and its roles are related to its positions more or less. For
example, a subject tends to be in a position of the first several ones of a sentence
while an object prefers to appear in the tail of a sentence.
Finally, as a special resource, we use extracted semantic relations to aid in the
identification of other semantic relations. In this paper, about half of extracted
relations are type of Hypernym and the precision of it is 95% with a margin
error of +/-5% with 99.5% confidence. So a double-step pass strategy is applied.
Hypernym relations are identified first and then they are encoded as a type of
feature to be used in a second pass. Making use of extracted semantic relations
to improve precisions of other semantic relations’ identification is also mentioned
in [9].
Uniform Representation. In this paper, any feature is attached with a value
derived from a definition to form a ”Feature-Value” structure. And then, any
candidate word can be attached with a list of such structures. In fact, if we take
”Feature-Value” structure as a type of ”Feature”, value range of which is true,
false, then regardless of the types and value ranges of features in a ”Feature-
Value” structures, we can treat them as the same. We refer ”Feature-Value”
structures as features in the rest of this paper.

2.2 Construction of Rules

A rule consists of a list of features. If a candidate word owns all the features of
a rule, then we say that this candidate word satisfies this rule.
Feature Selection. To balance the precision and the recall, we tend to make
sure that each rule provides high precision and acceptable recall and all rules
working as a whole provide a good recall. The procedure similar to the one of
constructing a decision tree is showed below:
428 R. Huang

Firstly, select the best feature from features of candidate words by means of
a kind of evaluation way which is described later. And then candidate words
will be split into two sets, words in one of which own this feature and words in
another don’t.
Secondly, if the precision of the former is higher than a predefined value, then
a new rule is constructed. Otherwise, select more features and split candidate
words into more sets like the first step until the precision is acceptable, or give
up if the number of selected features exceeds a predefined value.
Thirdly, after a new rule is constructed, any candidate words satisfying this
new rule will be removed. And then repeat the procedure until no more new
rules can have a precision higher than the predefined value.
Feature Evaluation. There has been a lot of effective ways to choose features.
However, different evaluation ways have different trends on precision and recall.
And we notice that very few candidate words have relevant semantic relation
to a head word. For example, a candidate word has a probability of only 1.5%
, estimated from a random small set, to hold Part Of relation with the head
word. And this fact causes severer different trends of different evaluation ways.
In experiments of this paper, Information Gain and X 2 Statistic usually give
good recalls while perform poorly in precisions. In contrast, Odds Ratio shows
an excellent precision but gives an unsatisfactory recall. Therefore, it’s necessary
to combine them to get a good precision as well as an acceptable recall. In order
to make it, we choose Information Gain (see ( 1),( 2)) as the evaluation way for
selection of the first feature of a rule.
Entropy(C) = −p+ log2 p+ − p− log2 p− (1)
Where C is a set of candidate words, p+ is the proportion of words holding target
relation with head words and p− is the proportion of the rest words.
|Cv |
Inf oGain(C, F ) = Entropy(C) − Entropy(Cv ) (2)
|C|
v∈{true,false}

Where F is a feature, Ctrue is the set of words with feature F in C and Cfalse
is the set of the rest words in C .
And we choose odds ratio (see ( 3)) to evaluate the rest features of a rule.
P (F |C+ )(1 − P (F |C− ))
OddsRatio(C, F ) = log (3)
P (F |C− )(1 − P (F |C+ ))
Where C+ is the set of words holding target relation with head words, C− is the
set of the rest words in C, P (F |C+ ) is the proportion of words holding Feature
F in C+ and P (F |C− ) is similar to P (F |C+ ).

3 Semantic Relation
We apply our approach to identify ten types of semantic relations ( see Table 1.
in Sect. 4.2). And we construct the identified relations as a net to make them
more useful.
 Semantic Relation Extraction by Automatically Constructed Rules 429

Three examples, covering from abstract concept to concrete concept and from
substance to creature, are showed in Fig. 1, Fig. 2 and Fig. 3.

Ÿ
Head word:
¯Ô¤kŠá5 §é¯Ô5Ÿ G¹ÚuÐåû
(essence)

½Š^ ‹ y– ƒ«O
Definition: , ,
( ” ” ) (The inherent nature of a thing or a class of things,
determines the character, state and development of that thing or that class of
things. (opposite to phenomena))

Ÿ(essence) –<Hypernym>→ á5(nature)

| –<Means>→ û½(determine)
| –<Antonym>→ y–(phenomena)

Fig. 1. Essence’s Definition and Semantic Relations

Head word: €Ð
j§/G”&|§‹fõãÚ§¶ õkxÚ½‘ãڀ
(culver)

:"
Definition:
(A bird, looking like homing pigeon, usually has taupe feathers and a neck
with white or filemot spots.)

€Ð(culver) –<Hypernym>→ j(bird)

| –<Part>→ ‹f(feather)
| –<Part>→ ¶(neck)

Fig. 2. Culver’s Definition and Semantic Relations

v(titanium)
Head word:
7ყÎÒTi"՝ڧŸM ”§òÐ5r§L:p§
F@¡"Œ^5›EA«Ü7g"(A kind of metal element, signed with ”Ti”,
Definition:

is silver gray, light, hard, ductile and corrosion-resistant. It can be use to make
special alloy steel.)

v(titanium) –<Hypernym>→ ƒ(element)

| –<Color>→ ՝Ú(silver gray)
| –<Material>→ Ü7g(alloy steel)

Fig. 3. Titanium’s Definition and Semantic Relations

3.1 Constructing Semantic Relations as a Net

It won’t be very useful if we isolate every small set of semantic relations extracted
from a definition of a word. Therefore, we need to construct all semantic relations
430 R. Huang

as a net. Every word with its part of speech has and only has one node in the
net. In other words, every word has not only semantic relations to some words
in its definition, but can have additional semantic relations to other words once
it is used to explain them. That is because when we look up a word, we may
not only consult the definition of this word, but also the definitions of any word
which mentions it [9]. Chodorow et al. [10] exploit such insight in developing a
tool for helping human users disambiguate hyper/hyponym links among pairs of
lexical items. MindNet implements it with full inversion of structures [5].

4 Evaluation
4.1 Experimental Settings
Preprocessing. Before extracting semantic relations, it’s necessary to process
raw text first, including word segment, part of speech tagging and dependency
parsing. All these are done by LTP v2.02 . Precision of dependency parsing is
78.23% without labeling type of dependency and 73.91% with labeling.

Dictionary. Rounding number of entries to the nearest thousand, 44000 nouns

with their definitions in MCSD are parsed with a set of rules automatically
constructed from a small corpus which was also made from MCSD.

4.2 Results
After a double-step pass, 43000 relations are identified. For each type of semantic
relations, a random sample of 400 relations is hand-checked and the precisions of
all types are listed in Table 1. Using common statistical techniques, we estimate
that these rates are representatives of each type with a margin error of +/-5%
with 99.5% confidence. From this result, not only a large number of relations
have been extracted but also high precisions of each type have been achieved.
However, such a ”hand-check” evaluation way may not be objective enough as
some of the identified relations are somewhat fuzzy to discriminate. Also, the
recall is hard to estimate. Further, it’s necessary to see the relations and the
net they form are really useful. Therefore, we suggest a more objective way of
evaluation.

Table 1. Set of semantic relation types

Type Precision Type Precision Type Precision Type Precision

Antonym 0.96 Color 0.96 Domain 0.88 Hypernym 0.95
Hyponym 0.81 Material 0.84 Means 0.92 Member 0.93
Part 0.88 Synonym 0.99

2
LTP is a language platform based on XML presentation, http://ir.hit.edu.cn/
demo/ltp/Sharing Plan.htm
 Semantic Relation Extraction by Automatically Constructed Rules 431

5 Evaluating by Determining Word Similarity

To get a more objective evaluation, we evaluate the net by determining word

similarity. We generate automatically word pairs labeled with similarity levels
as corpus. Then, we use one half to identify path patterns and use another half
to test them.

5.1 Generation of Word Pairs

A thesaurus3 is used to generate similar word pairs and dissimilar word pairs.
Similar strategy is taken by Stephen [11]. And here, we generate six similarity
levels of word pairs.
Words in the thesaurus are classified in five levels, including Big Class, Middle
Class, Tiny Class, Word Group and Atom Group. Every item in this thesaurus
consists of a class code and one or more words. For example:

Aa01A08= z<,ˆ<, z (All mean everyone)

Ba01B10# N,ŒN,‡N(conductor,semiconductor,superconductor)
Ad02F01@ (<(alien)
Fig. 4. Items in Thesaurus

The class code in the head of the first item in Fig. 4 means that these three
words are classified in A [Big Class] a [Middle Class] 01[Tiny Class] A [Word
Group] 08 [Atom Group]. And ”=” means that these three words are the same.
If they are not completely the same, then they are denoted by ”#”. If only one
word contained in an item, then a tag ”@” is used. The procedure of generating
word pairs is showed below:
Firstly, randomly select word pairs from the thesaurus.
Secondly, look up the already selected word pairs to ensure the newly selected
pair doesn’t exist in them.
Thirdly, look up the net to find out if they exist and connect to each other
within 18 arcs.
Fourthly, look up the thesaurus to see if they belong to the same atom group,
then they are labeled with ”Atom Group”, if they belong to the same word group
but not the same atom group, then they are labeled with ”Word Group”, or they
are labeled with Tiny Class, Middle Class or Big Class by analogy. And finally
if they are from different Big Classes then they are labeled with ”General”.
Fifthly, repeat the above procedure until 20000 word pairs for each level are
available. And then they are split into two sets equally. One set is used for
training and another is used for testing.
3
HIT-IR Tongyici Cilin (Extended), http://ir.hit.edu.cn/demo/ltp/Sharing Plan.htm
432 R. Huang

5.2 Using Path Patterns to Determine Similarity

Paths that one word travels through to another word are identified. For exam-
ple, given a word pair showed in Fig. 5, there are several paths exist, and two
of them are showed in Fig. 5. From these two paths, we can get two path pat-
terns <Synonym, Hypernym, Hypernym Of> and <Synonym, Hypernym Of,
Hypernym>.

Word Pairs: < Æf ä3

Æf
(student), (apprentice)>
(student) – <Synonym> → Æ)
<
Path pattern1:
ä3
(student) – <Hypernym>
→ (person) ← <Hypernym> –
Æf
(apprentice)
(student) – <Synonym> → Æ)(student) ← <Hypernym>
pä
Path Pattern2:
– (excellent student) – <Hypernym> → ä3 (apprentice)

Fig. 5. Two Paths of Word pair <Student,Apprentice>

Stephen [11] also uses path patterns to determine similarity. However, the
method adopted here is different. In this paper, if the precision and frequency of
a pattern are both high, then it will be used to identify similar word pairs later.

5.3 Result
In experiments of this paper, there are three different definitions of similarity.
In the first experiment, only word pairs labeled with ”Atom Group” are consid-
ered as similar pairs. In the second experiment, word pairs labeled with ”Word
Group” are also considered. In the third experiment, word pairs labeled with
”Tiny Class” are also considered. And the result is showed in Table 2.
The result shows high precision and low recall. The possible reasons may be:
Firstly, the high precision is ensured by the high precision of extraction.
Secondly, the low recall of determining word similarity reflects the low recall
of the semantic relations extraction. And there is no doubt that the insufficient
semantic relations cause two similar words can’t find a stable and short path to
reach each other although they may be still connected by a long path. And the
longer the path is, the more instable and unreliable the path will be. Also it will
become sparser when the length of path grows up.
Thirdly, the way of determining similarity may still need to be improved.
Finally, the knowledge contained in the MCSD may not consist very well with
the one contained in the thesaurus.

Table 2. Result of Similarity Experiment

No Similar Word Pairs Dissimilar Word Pairs Precision Recall

1 10000 50000 95.02% 14.14%
2 20000 40000 91.76% 16.6%
3 30000 30000 87.26% 25.66%
 Semantic Relation Extraction by Automatically Constructed Rules 433

6 Discussion

Although a large number of semantic relations have been identified, the recall
still may be low, reflected by the result of determining word similarity. And
lots of other knowledge in definitions is still unexploited. Take the definition of
culver (see Fig. 2) for example. In Fig. 6, only top level semantic relations linking
from head words to words in definitions are identified, and lower level semantic
relations are still unexploited.

Top Level: Ϫᬵ(culver) – <Hypernym> → ୔(bird)

Lower Level: ვર(feather) – <Color> → ߧ‫ލ‬೤(taupe)

Fig. 6. Two different level Semantic Relations

7 Conclusion

This paper describes a strategy to construct rules automatically to identify

semantic relations from machine readable dictionary, and a conceptual net is
constructed from the extracted relations. In order to estimate the precision, a
random sample of 400 relations for each type has been hand-checked. However,
it is still possibly subjective. Therefore, a more objective way, that is determin-
ing word similarity, is applied. And the result shows again that the precision of
extracted semantic relations is high. But what frustrates us is the result also
reflects indirectly the possible low recall of semantic relation extraction. Our
future work is to improve the ability of the approach to identify more seman-
tic relations including top level ones and lower level ones. Also, we will try
our approach to extract semantic relations from cyclopedias such as Chinese
Wikipedia.

Acknowledgement

Supported by National Natural Science Foundation of China (NSFC) (No.

60873135).

References
1. Miller, G.A., Beckwith, R., Fellbaum, C., Gross, D., Miller, K.J.: Introduction to
wordnet: An on-line lexical database. Int. J. Lexicography 3(4), 235–244 (1990)
2. Zhendong, D., Qiang, D.: Hownet, http://www.keenage.com
3. Zelenko, D., Aone, C., Richardella, A.: Kernel methods for relation extraction. J.
Mach. Learn. Res. 3, 1083–1106 (2003)
434 R. Huang

4. Aone, C., Ramos-Santacruz, M.: Rees: a large-scale relation and event extraction
system. In: Proceedings of the sixth conference on Applied natural language pro-
cessing, Morristown, NJ, USA, pp. 76–83. Association for Computational Linguis-
tics (2000)
5. Richardson, S.D., Dolan, W.B., Vanderwende, L.: Mindnet: acquiring and struc-
turing semantic information from text. In: Proceedings of the 17th international
conference on Computational linguistics, Morristown, NJ, USA, pp. 1098–1102.
Association for Computational Linguistics (1998)
6. Barriere, C.: From a children’s first dictionary to a lexical knowledge base of con-
ceptual graphs. PhD thesis, Burnaby, BC, Canada, Adviser-Popowich, Fred. (1997)
7. Markowitz, J., Ahlswede, T., Evens, M.: Semantically significant patterns in dic-
tionary definitions. In: Proceedings of the 24th annual meeting on Association for
Computational Linguistics, Morristown, NJ, USA, pp. 112–119. Association for
Computational Linguistics (1986)
8. Vanderwende, L.H.: The analysis of noun sequences using semantic information
extracted from on-line dictionaries. PhD thesis, Washington, DC, USA, Mentor-
Loritz, Donald (1996)
9. Dolan, W., Vanderwende, L., Richardson, S.D.: Automatically deriving structured
knowledge bases from on-line dictionaries. In: Proceedings of the First Conference
of the Pacific Association for Computational Linguistics, pp. 5–14 (1993)
10. Chodorow, M.S., Byrd, R.J., Heidorn, G.E.: Extracting semantic hierarchies from a
large on-line dictionary. In: Proceedings of the 23rd annual meeting on Association
for Computational Linguistics, Morristown, NJ, USA, pp. 299–304. Association for
Computational Linguistics (1985)
11. Richardson, S.D.: Determining similarity and inferring relations in a lexical knowl-
edge base. PhD thesis, New York, USA (1997)
 Object Recognition Based on Efficient Sub-window
Search

Qing Nie1, Shouyi Zhan2, and Weiming Li1

1
School of Information and Electronic, Beijing Institute of Technology, Beijing 100081, China
2
School of Computer Science, Beijing Institute of Technology, Beijing 100081, China
qingnie@bit.edu.cn

Abstract. We propose a new method for object recognition in natural images.

This method integrates bag of features model with efficient sub-window search
technology. sPACT is introduces as local feature descriptor for recognition task.
It can capture both local structures and global structures of an image patch effi-
ciently by histogram of Census Transform. An efficient sub-window search
method is adapted to perform localization. This method relies on a branch-and-
bound scheme to find the global optimum of the quality function over all possi-
ble sub-windows. It requires much fewer classifier evaluations than the usually
way does. The evaluation on PASCAL 2007 VOC dataset shows that this object
recognition method has many advantages. It uses weakly supervised training
method, yet has comparable localization performance to state-of-the-art algo-
rithms. The feature descriptor can efficiently encode image patches, and local-
ization method is fast without losing precision.

Keywords: object recognition, sPACT, object detection, bag of features.

1 Introduction
Object class recognition is a key issue in computer vision. In the last few years,
Object class recognition has received a lot of attentions. It has been approached in
many ways in literatures. These approaches can typically be divided into two types.
One type is training a part-based shape model, the model describes objects by simpler,
robust and stable features connected in a deformable configuration. These features are
usually generic, which can be edge points, lines or even visual words, and features
must be grouped to describe a given object. Some typical part-based shape models
include constellation model [1], implicit shape model (ISM) [2], k-fans model [3],
pair-wise relationships model [4] and star model [5,6]. When detect an object, it first
detect individual object parts, which be group together to reason about the position of
the entire objects. This is accomplished via generalized Hough transform, where the
detections of individual object parts cast probabilistic votes for possible locations of
the whole object. These votes are summed up into a Hough image, the peaks of it
being considered as detection hypotheses.
A shortcoming of part-based shape model is the need for clean training shapes.
Most of this type approaches need pre-segmentation training images or label the ob-

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 435–443, 2009.
© Springer-Verlag Berlin Heidelberg 2009
436 Q. Nie, S. Zhan, and W. Li

ject boundary box. This quickly becomes intractable when hundreds of categories are
considered. Some weakly supervised training methods exist, but with high computa-
tion cost.
Another type for object localization is commonly performed using sliding window
classifiers [7]. This approach trains a quality function (classifier) and then scans over
the image and predicts that the object is present in sub-windows with high score.
Siding window object localization has been shown to be very effective in many situa-
tions, but suffers a main disadvantage: it is computationally inefficient to scan over
the entire image and test every possible object location. To overcome this shortcom-
ing, researchers typically use heuristics to speed up the search, which introduces the
risk of missing target object. Lampert C.H. et al.[8] proposes an Efficient Sub-
window Search (ESS) to solve this problem. It relies on a branch-and-bound scheme
to find the global optimum of the quality function over all possible sub-images, thus
returning the same object locations that an exhaustive sliding window approach
would. At the same time it requires less classifier evaluations and typically runs in
linear time or faster.
Motivated by Lampert C.H.’ work, we proposed a multi-object recognition
method. This method integrates bag of features model with efficient sub-window
search technology. Bag of features model is a very powerful and effective classifica-
tion model. It allows weakly supervised training method. And ESS is applied in our
recognition frame for fast localization. In order to improve the recognition perform-
ance, we also designed a new local feature descriptor suitable for our recognition task.
The rest of the paper is organized as follows. We first introduce a modified spatial
PACT for our local feature descriptor in Section 2. Then in Section 3 we explain how
to use an efficient sub-window search method to perform localization. In Section 4 we
present our experiments results on PASCAL 2007 dataset. And come to the conclu-
sions in section 5.

2 Local Feature Descriptor

Bag of features model contains five major steps: feature extraction and description,
codebook creation, image feature description, model training and classification. In order
to learn and detect object classes in cluttered images of real-world scenes, some kind of
efficient local feature descriptors are necessary. Many local feature descriptors have
been proposed in the last decade. Some descriptors are appearance-based, such as SIFT,
GLOH, Spin image, PCA-SIFT. Some descriptors are structure-based, such as PCBR
[9], KAS[10] and so on. Structure-based descriptors are preferred in object detection
task. But most of structure-based descriptors need pre-segment images and segmenta-
tion itself is an open question in computer vision field. To address this problem, Wu J.
[11] proposed to use spatial histogram of Census Transform values (sPACT) to repre-
sent a whole image. Census Transform (CT) compares the intensity values of a pixel
with its eight neighboring pixels. The eight bits generated from intensity comparisons
can be put together and converted to a base-10 number. Census Transform can capture
local structures of an image, while large-scale structures are captured by the strong
correlation between neighboring CT values and the histogram. The PCA operation
results in a compact representation, and Spatial PACT further incorporates global struc-
 Object Recognition Based on Efficient Sub-window Search 437

tures in the image. SPACT demonstrates high performance on commonly used datasets.
Besides, it need not pre-segment image and evaluates extremely fast.

(a) (b) (c)

Fig. 1. Examples of CT images. (a) Original images (b) CT images(c) Modified CT images.

Our local feature descriptor is motivated by Wu J.’s work. But we have modified it
to fit our situation. And trim it to fit well in our bag of feature model. According to
Wu J.’s analysis, the powerful representation capacity of CT histogram is mostly
depend on its successful encoding shapes information. But from some examples of
CT images shown in fig.1(b), we find that Census Transform is sensitive to neighbor
pixels variation. The CT image is unnecessary noisy even in consistent region. In
order to remove these noises and keep shape information more clearly, we modified
the CT by introducing a threshold CTVALUE. Only when the difference between two
neighbor pixels great than CTVALUE, the CT value is 1, otherwise the CT value is 0.
The modified CT images are shown in fig.1(c). They are obviously clearer than stan-
dard CT images and keep the shape structures better.
In order to remove pixels correlation effects and get a more compact representa-
tion, Principal Component Analysis (PCA) operation is used to CT histograms data.
①
The key steps for PCA operation are: Random select some training images, con-
②
vert them to gray images, and calculate the modified CT values. Use regular grids
，
sample patches. The sampling interval set to 32 the size of image patches set to
③
32X32. Calculate CT histograms of these patches and normalize the CT histo-
grams. When calculate CT histogram, remove two bins with CT = 0, 255. Perform ④
PCA operation on these CT histograms. The PCA results show that the first 40 ei-
genvectors are enough to cover 92% information. So we get the first 40 eigenvectors
as main components.
To add spatial information, an image patch is split into 2x2 blocks, add the
center block, the normalized PACT histograms of these blocks are concatenated to
form a feature vector which has 40x5= 200 dimensions. Together with the gray
438 Q. Nie, S. Zhan, and W. Li

histogram of image patch, which result in a feature vector with an overall 200+256=
456 dimension.
The above contents are how to represent local image patches. In order to use bag of
feature model, the second step need create a codebook. There are various methods for
creating visual codebooks. K-means clustering, mean-shift and hierarchical k-means
clustering are the most common methods. Yet Extremely Random clustering forests
(ERC forests)[12,13] have recently attracted a lot of attention. ERC Forests are very
efficient at runtime, since matching a sample against a tree is logarithmic in the
number of leaves. And they can be trained on large, very high-dimensional datasets
without significant over fitting and within a reasonable amount of time (hours). In
practice, the average-case time complexity for building ERC forest is O( DN logk) ,
where D is the feature dimension, N is the number of patches and k is the number of
leaf nodes. In contrast, k-means has a complexity of O(DNk) . In our recognition
system, ERC forest is applied to create visual codebook.
After creating a codebook, we can represent an image as histogram of visual
words. And the training images’ histograms of visual words are fed to SVM to train a
classifier model. LIBSVM[14] is selected for training classifier model. Linear kernels
with C=2−5 is used to ensure real-time classification. The trained classifier model can
be used for later localization.

3 Using ESS to Localize Objects

To perform localization, one can take a sliding window approach, but this strongly
increases the computational cost, because the classifier function f has to be evaluated
over a large set of candidate sub-windows. In order to overcome this shortcoming,
several approaches have been proposed to speed up the search. One way is reducing
the number of evaluations by searching only over a coarse grid and by allowing only
rectangles of certain fixed sizes. Another way is applying local optimization methods
instead of global ones. These techniques sacrifice localization robustness to achieve
acceptable speed. It can lead to missing of the objects.
In order to overcome this shorting, we adopt an Efficient Sub-window Search
(ESS) technical proposed by Lampert C.H. et al.[8]. ESS is guaranteed to find the
locally maximal region. At the same time ESS is very fast, because it relies on a
branch-and-bound search instead of an exhaustive search.
The principle of ESS is the following: when searching for object regions, only very
few of candidate regions actually contain object instances. It is wasteful to evaluate
the quality function for all candidate regions. One should target the search directly to
identify the regions of highest score and ignore the rest of the search space where
possible. The branch-and-bound framework plits the parameter space into disjoint
subsets, while keeping bounds of the maximal quality on all of the subsets. By this
way, large parts of the parameter space can be discarded early if their upper bounds
are lower than a guaranteed score.
In our localization case, the parameter space is the set of all possible rectangles in
an image. These rectangles are parameterize by their top, bottom, left and right coor-
dinates (t, b, l, r). We represent set of rectangles as tuples [T,B,L,R], where T = [tlow,
thigh ] etc.. For each rectangle set, the highest score that the quality function could
 Object Recognition Based on Efficient Sub-window Search 439

take on any of the rectangles in the set are calculated. And this highest score is the
upper bound of the rectangle set.
Before giving the algorithm of Efficient Sub-window Search, we first discuss how
to construct a bound function fˆ . Denoting rectangles by R and sets of rectangles
by ℜ , the bound function fˆ has to fulfill the following two conditions:

1) fˆ ( ℜ ) ≥ max f (R )
R ∈ℜ

2) fˆ (ℜ) = f (R), if R is the only element in ℜ

Condition 1) ensures that fˆ is an upper bound to f, and condition 2) guarantees the

optimality of the solution to which the algorithm converges. The possible
bounds fˆ are not unique. A good bound fˆ should be fast to evaluate but also tight
enough to ensure fast convergence. In our experiments, we adopt bag of features
model for object classification, and we use SVM with a linear kernel as classifier. The
corresponding SVM decision function is as formula (1)
f (I ) = β + ∑α i i 〈h, h i 〉 (1)

where 〈 ., 〉 denotes the scalar product. hi is the histogram of the training examples i,
and α i and β are the weight vectors and bias that were learned during SVM training.
Because of the linearity of the scalar product, we can rewrite this expression as a sum
over per-point contribution with weights w j = ∑ α i h ij
i

f (I ) = β + ∑
n
w cj (2)
j =1

Here cj is the cluster index belonging to the feature point xj and n is the total number
⊂
of feature points in I. This form allows us to evaluate f over sub-images R I by sum-
ming over the feature points that lie within R. Since we are only interested in the
⊂
argmax of f over all R I, we can ignore the bias term β . Set f = f + + f − , where f +
contains only the positive summands of Equation 2 and f − only the negative ones. If
we denote by Rmax the largest rectangle and by Rmin the smallest rectangle contained
in a parameter region ℜ , then formula (3)
+ −
fˆ ( ℜ ) = f ( R max ) + f ( R min ) (3)

has the desired properties 1) and 2). Using integral images we can evaluate f + and
f − at constant time complexity, thus making the evaluation of fˆ a constant time
operation.
Suppose we use a priority queue P to hold the search states, the pseudo-code which
use ESS to locate an object class instance is like follows.
① Initialize P as empty priority queue
set [T,B,L,R] = [0, n] × [0, n] × [0,m] × [0,m] (m, n is the size of image I)
440 Q. Nie, S. Zhan, and W. Li

② repeat
split [T,B,L,R]→[T1,B1,L1,R1] ∪[T2,B2,L2,R2]
push([T1,B1,L1,R1], fˆ ([T1,B1,L1,R1] ) into P
push([T2,B2,L2,R2], fˆ ([T2,B2,L2,R2] ) into P
retrieve top state [T,B,L,R] from P
until [T,B,L,R] consists of only one rectangle
③ set (t max , bmax , lmax , rmax ) = [T,B,L,R]

This algorithm always examines the most promising(with highest upper bound) rec-
tangle set. The candidate set is split along its largest coordinate interval into halves,
thus forming two smaller disjoint candidate sets. The search is stopped if the most
promising set contains only a single rectangle with the guarantee that this is the
rectangle of globally maximal score.

4 Experiments
We experiments our methods on PASCAL VOC 2007 dataset [15]. The images in this
dataset contain objects at a variety of scales and in varying context. Multiple objects
from multiple classes may be present in the same image. The complexity of these
images is very high. The data has been split into 50% for training/validation and 50%
for testing. We choose ten object classes for our recognition task. We use approximate
1000 training images train classification models, and use approximate 2500 test
images test detection performance.
The experiment has three major steps. The first step is codebook creation. ERC
forest is applied to create visual codebook. Regular grids sampling method is used to
，
extract features from training images. The sampling interval is set to 32 the size of
image patches is set to 32. After extract local features in all training images, describe
them using modified sPCAT descriptors as discussed in section 2. These feature vec-
tors are used to train an ERC Forest. We use 4 assembling trees and they contain 4000
leaf nodes in total. The leave nodes are numbered and each leaf node is assigned a
distinct visual word index.
The second step is training a classifier. Regular grids sampling method also be se-
lected to extract local features. But in this phase, more patches need be sampled than
，
cookbook creation phase. The sampling interval is set to 8 the size of image patches
set to different scales. After getting the feature descriptors, query the corresponding
visual words. During a query, for each descriptor tested, each tree is traversed from
the root down to a leaf and the returned leaf index is the visual word index. After
getting all visual words in an image, calculate the 4000-D histogram of visual words.
These histograms are fed to SVM to train classifier model. After get SVM decision
function, calculate the weight of each word w j = ∑ α i h ij , rewrite decision function
i

as a sum over per-point contribution according to formula (2).

The third step is object localization. Only test images that contain object instances
are chosen in this step. For each image, image features are extracted as the way in
training phase. After getting the corresponding visual words, visual words integral
image are calculated. The integral image can be used to speedup the calculation of
 Object Recognition Based on Efficient Sub-window Search 441

f and f − . Then we can localize object instance using the algorithm discussed in
+

section 3. To find multiple object locations in an image, the best-first search can be
performed repeatedly. Whenever an object is found, the corresponding region is re-
moved from the image and the search is restarted until a threshold is not satisfied.
In our experiments, we adapt average precision (AP) scores to compare the per-
formance of detection methods. Detections are considered true or false positives
based on the area of overlap with ground truth bounding boxes. To be considered a
correct detection, the area of overlap a0 between the predicted bounding box BB p and
ground truth bounding box BB g must exceed 50% by the formula:

area ( BB ∩ BB g )
a0 =
p

area ( BB ∪ BB g )
(4)
p

Fig. 2. shows the AP scores on the test data for 10 of the categories..
For illustration, we also compare our AP scores with the maximum results and
median results in the PASCAL VOC 2007 challenge. From the table we can see that
although our results are not prior than best results, our results are better than median
results of PASCAL VOC 2007 in most of 10 categories. We believe that our method
is general for object categories with different features.

Average Precision
sofa
sheep sheep
person
aeroplane aeroplane
tvmonitor AP
train train
horse
motorbike motorbike
bicycle
car car
0 10 20 30 40 50
AP(%)

Fig. 2. AP scores on the 10 categories of PASCAL VOC 2007

Table 1. Average Precision(AP) of our method and PASCAL2007 results

Object classes Our Max Median

car 0.39 0.43 0.29
bicycle 0.37 0.41 0.27
motorbike 0.23 0.38 0.22
horse 0.29 0.34 0.20
train 0.29 0.34 0.16
Tv/monitor 0.26 0.29 0.24
aeroplane 0.23 0.27 0.15
person 0.11 0.23 0.12
sheep 0.14 0.18 0.05
sofa 0.10 0.15 0.10
442 Q. Nie, S. Zhan, and W. Li

5 Conclusions
We have proposed a new method for object localization in natural images. It uses a
weakly supervised training method, yet can get state-of-the-art localization result. It
achieves this in several ways. First, we designed a modified sPACT patch descriptor
which is suitable for representing local features. Then, ERC forest is applied to
speedup both training and testing process. And we further improve the localization
performance by employing an Efficient Sub-window Search technical. In future work,
we will explore more powerful local feature descriptors, and adapt our method to
different SVM kernels.
Acknowledgments. The authors also would like to thank the researchers who kindly
published their datasets and software packages online.

References
1. Fergus, R., Perona, P., Zisserman, A.: A visual category filter for google images. In:
Pajdla, T., Matas, J.G. (eds.) ECCV 2004. LNCS, vol. 3021, pp. 242–256. Springer,
Heidelberg (2004)
2. Leibe, B., Leonardis, A., Schiele, B.: Combined object categorization and segmentation
with an implicit shape model. In: 8th European Conference on Computer Vision,
pp. 17–32. Springer, Heidelberg (2004)
3. Crandall, D., Felzenszwalb, P., Huttenlocher, D.: Spatial priors for part-based recognition
using statistical models. In: IEEE Computer Vision and Pattern Recognition 2005,
pp. 10–17. IEEE Press, San Diego (2005)
4. Leordeanu, M., Heber, M., Sukthankar, R.: Beyond Local Appearance: Category Recogni-
tion from Pairwise Interactions of Simple Features. In: IEEE Computer Vision and Pattern
Recognition 2007, pp. 1–8. IEEE Press, Minnesota (2007)
5. Shotton, J., Blake, A., Cipolla, R.: Contour-Based Learning for Object Detection. In: 10th
International Conference on Computer Vision, pp. 503–510. IEEE Press, Beijing (2005)
6. Opelt, A., Pinz, A., Zisserman, A.: A Boundary-Fragment-Model for Object Detection. In:
Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3952, pp. 575–588.
Springer, Heidelberg (2006)
7. Dalal, N., Triggs, B.: Histograms of Oriented Gradients for Human Detection. In: IEEE
Computer Vision and Pattern Recognition 2005, pp. 886–893. IEEE Press, San Diego
(2005)
8. Lampert, C.H., Blaschko, M.B., Hofmann, T.: Beyond Sliding Windows: Object Localiza-
tion by Efficient Subwindow Search. In: IEEE Computer Vision and Pattern Recognition
2008, pp. 1–8. IEEE Press, Anchorage (2008)
9. Deng, H.L., Zhang, W., Mortensen, E.: Principal Curvature-Based Region Detector for
Object Recognition. In: IEEE Computer Vision and Pattern Recognition 2007, pp. 1–8.
IEEE Press, Minnesota (2007)
10. Ferrari, V., Fevrier, L., Jurie, F., Schmid, C.: Groups of Adjacent Contour Segments for
Object Detection. IEEE Trans. Pattern Anal. Machine Intell. 30, 36–51 (2008)
11. Wu, J., James, M.R.: Where am I: Place instance and category recognition using spatial
PACT. In: IEEE Computer Vision and Pattern Recognition 2008, pp. 1–8. IEEE Press,
Anchorage (2008)
 Object Recognition Based on Efficient Sub-window Search 443

12. Geurts, P., Ernst, D., Wehenkel, L.: Extremely randomized trees. Machine Learning Jour-
nal 63, 3–42 (2006)
13. Moosmann, F., Triggs, B., Jurie, F.: Fast Discriminative Visual Codebooks using Random-
ized Clustering Forests. In: Advances in Neural Information Processing Systems, vol. 19,
pp. 985–992 (2006)
14. LIBSVM: a library for support vector machines,
http://www.csie.ntu.edu.tw/cjlin/libsvm
15. PASCAL 2007 VOC dataset, The PASCAL Visual Object Classes Challenge (2007),
http://www.pascal-network.org/challenges/VOC/voc2007/
 A Multi-Scale Algorithm for Graffito Advertisement
Detection from Images of Real Estate

Jun Yang and Shi-jiao Zhu

School of Computer and Information Engineering, Shanghai University of Electric Power,

Shanghai 200090, China
{mail_yangjun,zhusj707}@hotmail.com

Abstract. There is a significant need to detect and extract the graffito adver-
tisement embedded in the housing images automatically. However, it is a hard
job to separate the advertisement region well since housing images generally
have complex background. In this paper, a detecting algorithm which uses
multi-scale Gabor filters to identify graffito regions is proposed. Firstly, multi-
scale Gabor filters with different directions are applied to housing images, then
the approach uses these frequency data to find likely graffito regions using the
relationship of different channels, it exploits the ability of different filters tech-
nique to solve the detection problem with low computational efforts. Lastly, the
method is tested on several real estate images which are embedded graffito ad-
vertisement to verify its robustness and efficiency. The experiments demon-
strate graffito regions can be detected quite well.

Keywords: Multi-Scale, Graffito Advertisement Detection, Gabor Filter.

1 Introduction
With the development of network and multimedia technology, people are used to
retrieval housing images from Internet to get real estate information instead of text
description. However, a large number of housing images are embedded graffito adver-
tisement by real estate agents, for instance, an agent often write a contact telephone
number or embed a logo in the images, it is a tedious job for people to select the pol-
luted images one by one from a large scale web image database, therefore, there is
great demand for detecting and extracting the graffito advertisement from the images
automatically by computer.
Since the background of housing images are very complex, it is difficult to detect
and extract these polluted information well. Fortunately, from consistency and ro-
bustness of observations from these images, we find that images include graffito ad-
vertisement are more relevant than other images and the graffito advertisement is
composed of some connected region that we can make good use of these traits to
detect and extract the polluted information.
Multi-scale frequency filter is a popular method in image analysis. This technique was inspired
by a biological system [1] which has been studied in the past decades. The results of physiological
studies have supported that visual processing in the biological visual system at the cortical level

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 444–452, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 A Multi-Scale Algorithm for Graffito Advertisement Detection 445

involves a set of different scale filtering mechanisms [2][3]. Some frequency mathematic models
have been studied, such as Differences of Gaussian[4] , Gabor functions[5] and Wavelet func-
tions[6].In the frequency domain, it is generally observed that a clutter field in an image has higher
frequency content compared to other scatter regions. Therefore, filtering with frequency analysis
is an attractive approach for the task of image region analysis.
In this paper, firstly terms used in the proposed multi-scale frequency analysis are
defined and invariant operators are introduced to represent a housing image. Since
multi-scale method has been studied for its localized frequency filters alike human’s
visual cortical neurons [3] and used for representation of an image’s opponent fea-
tures [7], a multi-scale algorithm using frequency domain filters to detect polluted
regions in a housing image is then proposed. In this approach, because frequency
domain features of an image are divided into different subgroups by analyzing re-
gions’ distribution, it is more suitable and robust than other feature detection models
and it can be used for detecting words, numbers and company’s logo which embedded
by a real estate agent without any explicit knowledge. The experiments results indi-
cate that images have graffito advertisement with complex background can be
detected quite well.
The remainder of this paper is organized as follows: necessary definitions and
operators for images which have graffito advertisement are presented in section 2,
then a multi-scale algorithm is proposed and some important issues are discussed in
section 3. Section 4 provides the experimental comparison results between the pro-
posed algorithm with others algorithm and the concluding remarks are given in
Section 5.

2 Graffito Region Detection Using Multi-Scale Filtering

2.1 Definitions for Advertisement Region

Since the advertisement embedded in an image usually is not an isolated line or an

edge but a specific text region or a logo, as shown in Figure 1, in order to describe
the issues briefly, the necessary symbols and definitions are given as follows.

(a) broker’s telephone number (b) company’s logo.

Fig. 1. Examples of images embedded advertisement

446 J. Yang and S.-j. Zhu

Definition 1. For an image distributed in a discrete domain I ⊂ Z 2 , a binary image is

defined as
⎧1 I ( x, y ) ≥ Thr (1)
Bin( x, y ) = ⎨
⎩0 otherwise

Here I(x, y) denotes a pixel point position, L is gray level and Thr presents a specific
threshold between 0 and L.
Definition 2. The region which excludes advertisement is presented as B, so the pos-
sible advertisement region in an image is defined as

∑S
i =1.. n
i =I −B (2)

Definition 3. For a subset Λ ⊆ Si , a candidate advertisement region is a region has two

properties as follows:
(1) For a point in the region’s neighborhood area I Λ ( x, y ) ∈ Si of Λ , it can be
combined into Λ .
(2) For a subarea AreaΛ of Λ , it is define as
_
AreaΛ =∪{I(x, y)| || I(x, y) −IΛ ||<δ} (3)

The ratio of sub region can be defined as

Ratio A = AreaΛ / SiΛ (4)

S
where T0 is a threshold and Ratio A > T0 , we can regard Λ ⊆ Si as a candidate
S
region.
Definition 4. For multi-scale frequency, a multi-scale operator for the region Λ is
defined as

Fm = I Λ ⊗ Gm ,θ , m = m1 ...mn , θ = θ1 ...θ N (5)

Where Gm ,θ is the m scale with θ direction operator which is described as Gabor

filter as below.
Definition 5. In frequency domain, the different covariance of same m is defined as
θM
1
σF = m
M
∑ | (F
θ θ =
m ,θ − Fm )2 | (6)
1

When σ F < σ o , Fm has isotropy we can get σ Fθ at the same θvalue with different
m

frequency.
 A Multi-Scale Algorithm for Graffito Advertisement Detection 447

Definition 6. For the different scale m or σ , a region has advertisement is defined

as

⎧ ∆
⎪1 1/ M | σ Fm|θ − σ Fm|θ |< σ m|θ
P=⎨ (7)
⎪⎩0 otherwise

∆
Where σ m|θ is an assigned threshold value.
The other property of the polluted region is that the region is continuous not only
in the image but also in its corresponding frequency domain. We can verify the region
by comparing it with the x-y axes after getting the candidate regions by multi-scale
transformation in frequency domain.

2.2 Multi-filters Bank

For implement, Gabor filter is chosen for transforming an image to a frequency do-
main. The transformation of a 2-D image is modulated by a 2-D Gabor function. A
Gabor filter is consisted of two functions having phase by 90 degree, conveniently
located in the real and imaginary parts in complex field. Gabor is a function modu-
θ
lated by a complex sinusoid which can be rotated in the x-y plan by , it can be sym-
bolized as x' and y' .

x' = x*cos(θ )+y*sin(θ ) (8)

y' = x*cos(θ )-y*sin(θ ) (9)

The even symmetric Gabor is describe as

1 x' y'
）
G(x,y,θ ,f ) = exp(− [( )2 + ( )2 ])cos(2*pi *f *x'
2 sx ' sy '
(10)

sx and sy are variances along x and y-axes respectively, f is the frequency of the sinu-
soidal function, θg ives the orientation of Gabor filter. By changing the radial fre-
quency and the orientation, the filter makes different target features. The result of
spatial domain by Gabor functions is shown in Fig. 2.
After filtering the input image with filters under various frequency and orientation
of a channel, these paralleling channels can be analyzed. Since the polluted regions
have some common properties under different channels, we can extract these special
regions by analyzing the relationship of the channels. The proposed paradigm is
shown in Fig.3.
The analysis can be performed in each filter channel using different scale and theta,
so it is possible to mark polluted regions in an output image.
448 J. Yang and S.-j. Zhu

(a) Original image (b) Embedded advertisement

(c) f=0.3730, θ = 0o (d) f=0.3730, θ = 45o

(e) f=0.3730, θ = 90o (f) f=0.3730, θ = 135o

Fig. 2. The output filtered housing image by Gabor Filters

Filter(f1 ,T1 )

Filter(f1 ,T 2 )

Filter(f1 , T N )

Image Image
Filter(fM ,T1 ) Region
Procession Combining
Filter(f M ,T 2 )
Various Filters

Filter(f M ,T N )

Fig. 3. Filtering bank process paradigm

 A Multi-Scale Algorithm for Graffito Advertisement Detection 449

3 The Proposed Method

From large amounts of observation, we find a housing image has some direction traits
even if it has a complex background, for instance it has many straight lines and right
angles. These properties can help us separate embedded advertisement from the image
using Gabor function. Here we combine the filtering results from original image
with each region’s location to find the candidate polluted regions. The detection proc-
ess is described as bellow.
Step 1. Operation of Multi-Scale Gabor Filters.
For the properties of housing image, it adopts different orientations with variant
frequencies. For theta, there are six angles which is 0o,30o,45o,90o,120o,135o respectively.
In order to find regions contain straight lines in resulting image, we set θ = 0o ,90o . For
the adaptive of scale, we use the relative size defined as below

f = (2w - 0.5) / W , w=[0...log(2W/8) ] (11)

Where W is the width of an image. Therefore the total number of channel is

6*sizeof(f).
Step 2. Channels Combination computation.
After getting the filter results, we can combine the different channels to mark the
polluted regions as follows:
(1) Computing the set of Λ in different regions of an original image.
(2) Using the frequency results to find the σ Fm to mark and erase some dynamic
value regions.
(3) Using the frequency results to find the different frequency at the same
theta σ Fθ .
(4) Getting the transformed gray image under σF m
and σ Fθ , and then using K-
Mean method to cluster the gray image. After that, histogram equalization is
performed to scale values of processed gray images.
(5) Calculating by formula (7), a region is marked by Pm∩θ = Pm I Pθ .
(6) Calculating Pm ∩θ and Λ to mark the region.

⎧1 ( Pm∩θ = 1) ∧ (Λ = 1)
Region = ⎨ (12)
⎩0 otherwise
By formula (12), we can get a 0-1 array which can describe the region is polluted
region or not well.
In addition, many other features can be identified using transformed. In our imple-
mentation, we used the orientation and frequency dimensions. In order to prove the
effectiveness of the method, in the next section, we detect graffito regions from
housing images downloaded from Internet.
450 J. Yang and S.-j. Zhu

4 Experimental Results
In order to verify the correctness and effectiveness of our proposed solution, a prototype
system is developed by MATLAB. The scheme is tested on the real housing images
included advertisement from Web page. The theta equals 0o,30o,45o,90o,120o,135o respec-
tively, the frequency is selected by f = (2 w - 0.5) / W , w=[0...log(2W/8) ] , where the
test W of image is scaled the range of 0 to 255. The comparison results between the
proposed method and Summing Squared Response of filters (SSR) (where threshold is
0.5 of range of total summing squared value) are shown in Figure.4. The method of
summing squared response is to find a threshold of filters to separate the original image
[8]. From Fig.4, we can find that the different level of multi-scale has its own character-
istic. When performed by the algorithm, image (d) is obtained which contains graffito
features. From the comparison shown in Fig.5, we can find that multi-scale method can
gives more graffito regions than the summing squared responses, which can gives more
supports for the next step of classification. The result indicates that the detecting of
regions from an image having embedded advertisement based on multi-scale algorithm
is more suitable for housing images and it is robust for these types of images with more
complex background. In the point of image procession, the method also has the ability
of eliminating light shadows. Therefore, the result of filtered image cannot give more
details of buildings which often have more lines in an image. Additionally, the computa-
tion complexity of this method can performed using fast FFT technique which is a
critical factor for some real-time applications.

(a) Original image (b) θ = 45o , f=0.2537

(c) θ = 135o , f=0.3278 (d) The result

Fig. 4. The filtering results
 A Multi-Scale Algorithm for Graffito Advertisement Detection 451

(a) Original image (b) Result of proposed method (c) Result of SSR

Fig. 5. The comparison result between proposed method and SSR

5 Conclusions
In this paper, we present a method for detecting graffito polluted regions from an
image based on multi-scale algorithm. The main contribution of this paper is that the
proposed algorithm can find the relationship of channels in different frequency, there-
fore, more effective method can be proposed based on its characteristics. On the other
hand, the method using fast FFT transform can make algorithm more efficient which
can solve problems in the real application of housing images embedded advertise-
ment. Experimental results indicate our method is superior in terms of accuracy, ro-
bustness and stability in detection of advertisement in housing images. More studies
in the aspect of basic theory and its application should be studied in the near future.

Acknowledgement
This work is supported by Shanghai Education Commission Research Project under
Grant 09YZ344, and by Shanghai Special Research Fund of Young College Teacher
under Grant sdl08026. The authors are grateful for the anonymous reviewers who
made constructive comments.

References
1. De Valois, R.J., Albracht, D.G., Thorell, L.G.: Spatial-frequency selectivity of cells in ma-
caque visual cortex. Vision Research 22, 545–559 (1982)
2. Wilson, H.R.: Psychological evidence for spatial channels. In: Braddick, O.J., Sleigh, A.C.
(eds.) Physical and Biological Processing of Images. Springer, Berlin (1983)
3. Watt, R.J., Morgan, M.J.: Spatial lters and the localisation of luminance changes in human
vision. Vision Res. 24(24), 1387–1397 (1984)
4. Petrosino, A., Ceccarelli, M.: A scale-space approach to preattentive texture discrimination.
In: International Conference on Image Analysis and Processing, 1999. Proceedings,
pp. 162–167 (1999)
5. Kamarainen, J.-K., Kyrki, V., Kalviainen, H.: Invariance properties of Gabor filter-based
features-overview and applications. IEEE Trans. on Image Processing 15(5), 1088–1099
(2006)
6. Jain, A., Healey, G.: A multiscale rep resentation including opponent color features for tex-
ture recognition. IEEE Transactions on Image Processing 7(1), 124–128 (1998)
452 J. Yang and S.-j. Zhu

7. Pollen, D.A., Ronner, S.E.: Visual cortical neurons as localized spatial frequency filters.
IEEE Transactions on System, Man and Cybernetics 13(15), 907–916 (1983)
8. Jung, K.: Neural network-based text location in color images. Pattern Recognition
Letter 22(14), 1503–1515 (2001)
An Improved Path-Based Transductive Support Vector
Machines Algorithm for Blind Steganalysis Classification

Xue Zhang1 and ShangPing Zhong1,2

1
College of Mathematics and Computer Science, Fuzhou University,
Fuzhou, 350108, China
2
Fujian Supercomputing Center, Fuzhou, 350108, China
fjzx33@163.com, spzhong@fzu.edu.cn

Abstract. With the technologies of blind steganalysis becoming increasingly

popular, a growing number of researchers concern in this domain. Supervised
learning for classification is widely used, but this method is often time consum-
ing and effort costing to obtain the labeled data. In this paper, an improved
semi-supervised learning method: path-based transductive support vector ma-
chines (TSVM) algorithm with Mahalanobis distance is proposed for blind
steganalysis classification, by using modified connectivity kernel matrix to im-
prove the classification accuracy. Experimental results show that our proposed
algorithm achieves the highest accuracy among all examined semi-supervised
TSVM methods, especially for a small labeled data set.

Keywords: Path-based TSVM, Semi-supervised learning, Blind steganalysis.

1 Introduction
Techniques for semi-supervised learning (SSL) are becoming increasingly sophisti-
cated and widespread. Semi-supervised learning is halfway between supervised and
unsupervised learning, using a relatively small labeled data set and a large unlabeled
data set to obtain the classification. With the rising popular application of support
vector machines (SVM), one of the semi-supervised method, transductive SVM
[1](TSVM), becomes more popular nowadays. TSVM, maximizes the margin using
both labeled data set and unlabeled data set, which is not used in standard SVM.
However, the drawback of TSVM is that the objective function is non-convex and
thus difficult to minimize. Joachims designs a heuristically optimization algorithm for
TSVM called SVMlight [2], which has a wide range of applications in diverse do-
mains. Chapelle and Zien in [3] represent a path-based gradient descent method on
the primal formulation of TSVM objective function, called Low Density Separation
algorithm (LDS algorithm for short). A novel large margin methodology SSVM and
SPSI is proposed by Wang and Shen [4].
Information hiding techniques have recently received a lot of attention. With digi-
tal images as carriers, the goal of blind steganalysis is to classify testing images by
any steganographic algorithm into two categories: cover images and stego images.
Nowadays, most blind steganalysis methods use supervised learning technologies,

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 453–462, 2009.
© Springer-Verlag Berlin Heidelberg 2009
454 X. Zhang and S. Zhong

which require a large amount of labeled data for training. In fact, labeled samples are
often difficult, expensive, or time consuming to obtain, while unlabeled data may be
relatively easy to collect: we can download a large amount of unlabeled images from
the internet. It is of great interest both in theory and in practice because semi-
supervised learning requires less human effort and gives higher accuracy [5]. Re-
searches for semi-supervised blind steganalysis classification are still on the initial
stage. Our research group once proposed semi-supervised blind steganalysis classifi-
cation [6] based on SVMlight, but it has high complexity. Thus we consider another
semi-supervised algorithm, path-based TSVM algorithm (LDS algorithm [3]) for
blind steganalysis. As far as we know, this algorithm is not yet applied to blind stega-
nalysis classification.
In this paper, we build up on the work on LDS algorithm for blind steganalysis
classification. But its classification accuracy is not high enough, partly because of the
special distribution of steganalysis feature data set, and the lack of utilizing labeled
data set information to build the graph. Thus, an improved LDS algorithm for blind
steganalysis classification is proposed. It uses Mahalanobis distance instead of
Euclidean distance to measure the distance between samples, and modifies the edges
between labeled samples when modeling initial weight-graph.
The rest of this paper is organized as follows. Section 2 reviews LDS algorithm
and shows its drawbacks. In Section 3, we propose an improved LDS algorithm to
overcome those drawbacks. Experimental results on blind steganalysis classification
are presented in Section 4. Final conclusions are in Section 5.

2 Review of Path-Based TSVM Algorithm

Path-Based TSVM algorithm is a way to enforce the decision boundary lying in low
density regions, and thus not crossing high density clusters, that is why it is equally
called Low Density Separation (LDS) algorithm. LDS algorithm first builds a weight-
graph derived from the data, using path-based dissimilarity measure to compose con-
nectivity kernel matrix; and then takes manifold mapping method for dimensionality
reduction; finally, minimizes TSVM objective function by gradient descent. The first
part is the most significant process, and we will show some details.

2.1 Connectivity Kernel on Path-Based Dissimilarity Measure

Connectivity kernel is firstly proposed by Fischer, whose main idea is to transform

elongated structures into compact ones [7]. Let a labeled data set
Tl = {( x1 , y1 ) ,L , ( xl , yl )} ∈ ( X × Y ) , where xk ∈ X = R , yk ∈ Y = {1, −1} , k = 1,..., l , an
l m

unlabeled data set Tu = { xl +1 ,L , xl +u } ∈ X u . Given the undirected graph G = (V , E , D ) ,

where V = { xi } is the set of nodes (samples), E denotes the edges between nodes.
Each edge e ( i, j ) ∈ E is weighted by Euclidean distance xi − x j . G can be a full-
2

connected graph or k-nearest neighbors (k-NN) graph.

 An Improved Path-Based Transductive Support Vector Machines Algorithm 455

Path-based dissimilarity measure matrix D is a matrix to measure the dissimilarity

of two nodes based on graph-path. Let us denote Pij by all paths from node xi to
node x j , the dissimilarity measure d ij ∈ D from node xi to node x j is defined as [3],

2
⎛ p −1
⎞
ρ

ρ p∈Pij
(
d ( i, j ) = 2 ln ⎜⎜1 + min ∑ exp ( ρ e ( pk , pk +1 ) ) − 1
1
) ⎟⎟ . (1)
⎝ k =1 ⎠
dij denotes the minimum of all path-specific maximum weight. If two nodes are in the
same cluster, their pairwise dissimilarity will be small; by contrast, if two nodes are in
different clusters, at least one large edge of all paths from node xi to node x j crosses
low density regions, which will cause large dissimilarity value. Then we can get con-
nectivity kernel K as [3],

⎛ d ρ ( i, j ) ⎞
K ( i, j ) = exp ⎜ − ⎟ . (2)
⎝ 2 σ 2
⎠

The linear case corresponds to σ = ∞ , that is K ij = dijρ .

2.2 Gradient Descent on TSVM

Chapelle and Zien [3] represent a path-based gradient descent on the primal formula-
tion of TSVM objective function, which can be rewritten as follows,
l l +u
w + C ∑ L ( yi ( w ⋅ xi + b ) ) + C * ∑ L* ( w ⋅ xi + b )
1 2
min
2 i =1 i = l +1

⎧
⎪ L(t ) = max ( 0,1 − t ) . (3)
⎪
⎪
s.t. ⎨ L* (t ) = exp ( −3t 2 )
⎪ l +u
⎪1 1 l
∑
⎪⎩ u i =l +1
w ⋅ xi + b = ∑ yi
l i =1
The last constraint is to enforce that all unlabeled data are not put in the same class.
Unlike the traditional SVM learning algorithms, Chapelle and Zien solve the problem
by gradient descent on equation (4), and this primal method has obvious lower time
complexity than SVMlight algorithm. We can see the contrast experimental results in
Section 4. The implemental details of original LDS algorithm can be found in [3].

2.3 Drawbacks of LDS Algorithm

Although LDS algorithm has the advantages of low time complexity, easy implemen-
tation, its classification accuracy is not high enough to some extent. As we mentioned
456 X. Zhang and S. Zhong

(a) (b)

Fig. 1. An example of showing the drawback of LDS algorithm: (a) a synthetic data set is con-
structed by Delta data set and its mirror points; (b) the classification result done by LDS
algorithm

in Section 2.1, the graph is built as all nodes are equal, no matter labeled or unlabeled.
Here is a synthetic data set constructed by Delta data set and its mirror points, with
four labeled samples, see fig.1(a). We can see that there is a wide gap between Delta
data set and its mirror points, thus, if we don’t use labeled sample information to build
the graph, the classification will be done like fig.1(b).
Fig.1 shows that although two labeled samples are in the same cluster, they some-
times produce a large dissimilarity calculated by Euclidean distance. Chang and Ye-
ung in [8] shows a way of robust path-based semi-supervised clustering, thus we can
take it in LDS algorithm. In the synthetic Delta data set, the same cluster samples can
be linked together by shrinking the distance of purple red dotted lines in fig.1(a), and
the different cluster samples distance (green dotted line in fig.1(a)).can be enlarged to
distinguish obviously.
Furthermore, LDS algorithm uses the most common Euclidean distance to build
original graph, but the results obtained from Euclidean distance often comes with a
deviation from the actual distance, because of the various weights of different fea-
tures and the mutual influences among the features [9]. On the contrary, Mahalano-
bis distance can overcome these shortcomings. Mahalanobis distance utilizes the
covariance matrix to eliminate the mutual influences among the features, thus can
capture the distance more correctly than Euclicean distance. We can use the Within-
and-Between-Class Distribution Graph (B-W Graph) [10] to compare the difference
between Euclidean distance and Mahalanobis distance for steganalysis features. In
B-W Graph, the horizontal axis denotes the between-class distribution B, which
provides the main information for classification; and the vertical axis denotes the
with-in distribution W. B≥0 means correct classified samples, B<0 means error
classified samples.
Fig.2(a) shows the B-W Graph of steganalysis feature set (extracted from 300
original samples and 300 stego samples with Shi’s method [11]) by Euclidean dis-
tance. In this figure, a lot of stego and original samples are set on the region of B<0,
which affects the accuracy of classifier. But if Mahalanobis distance is used instead of
Euclidean distance, the result will be better, see Fig.2(b).
 An Improved Path-Based Transductive Support Vector Machines Algorithm 457

(a) (b)

Fig. 2. The steganalysis feature set B-W Graph: (a) B-W Graph is drawn using Euclidean dis-
tance; (b) B-W Graph is drawn using Mahalanobis distance

3 A Novel Path-Based TSVM Algorithm for Blind Steganalysis

Classification

3.1 Improved Connectivity Kernel on Path-Based Dissimilarity Measure

Now we define an improved path-based dissimilarity measure to construct connec-

tivity kernel. All symbols are the same with that mentioned in Section 2.1. We replace
Euclidean distance e in Equation (1) with Mahalanobis distance em ,

em ( i, j ) = (x − x ) S −1 ( xi − x j ) .
T
(4)
i j

Where S denotes the sample covariance matrix for all samples x ∈ X l U X u . Then
we add the information of labeled samples by shrinking the edges of two samples in
the same clusters to the minimum em , and expanding the edges of two samples in
different clusters to the maximum em ,

⎧ min ( em ) if xi , x j ∈ X l AND yi × y j >0

⎪⎪
em′ ( i, j ) = ⎨ max ( em ) if xi , x j ∈ X l AND yi × y j <0 . (5)
⎪
⎪⎩em ( i, j ) otherwise

Then we use novel em′ to calculate the linear case connectivity kernel ( σ = ∞ ),

( ( ) )
2
1 ⎛ p −1
⎞
d m ( i, j ) = 2 ln ⎜⎜1 + min ∑ exp ρ em′ ( pk , pk +1 ) − 1 ⎟⎟ .
ρ

ρ ⎝
p∈Pi, j
k =1 ⎠
(6)

K m ( i, j ) = d m ( i, j )
ρ
458 X. Zhang and S. Zhong

3.2 Blind Steganalysis Classification Framework and Path-Based TSVM

Algorithm

The process of blind steganalysis is actually the process of pattern recognition. It

includes three steps: First, do image preprocessing, such as transforming the images
into gray scale. Second, extract features and select important features to construct the
classifier. Finally, classify the testing data set. For blind steganalysis, testing images
can be classified into two categories: cover images and stego images. See Fig.3.
In this paper, we focus on path-based TSVM algorithm in step 3 and want to find a
suitable semi-supervised algorithm for blind steganalysis classification.

Fig. 3. Blind steganalysis framework based on Path-based TSVM

So combining with the characteristics of blind steganalysis data set, we proposed

an improved Path-Based TSVM algorithm.
Algorithm 1. A novel Path-based TSVM algorithm for blind steganalysis
Step 1: Form ( l + u ) × ( l + u ) connectivity kernel matrix by equation (4)-(6).
a) Build a fully connected graph with standard Mahalanobis distance (equation (4));
b) Modify the edge lengths between labeled samples (equation (5));
c) Calculate kernel matrix K m (equation (6)).

Step 2: Perform multidimensional scaling (MDS) of ( l + u ) × ( l + u ) matrix K m , and

obtain ( l + u ) × p matrix K m′ .
 An Improved Path-Based Transductive Support Vector Machines Algorithm 459

Step 3: Train TSVM using gradient descent.

index − nofIter
a) Set original C = 2
*
C , index=0.
b) Minimize by gradient descent on the objective function (3).
c) Increase index by one, if index equals nofIter, goto (d), else goto (b).
d) Print the classification result.
Actually, algorithm 1 can be used in other domains for its flexibility and good per-
formance. We redo the synthetic data set experiment mentioned in Section 2.3 with
the same four labeled samples, and our proposed path-based TSVM algorithm gets
satisfied result, see fig.4.

Fig. 4. The classification result done by proposed path-based TSVM algorithm

3.3 Analysis of Algorithmic Time Complexity

The analysis of time complexity for our proposed algorithm is shown as follows: It
starts with the l × m matrix X l and u × m matrix X u , and then builds a fully con-
nected graph with standard Mahalanobis distance. The computation of calculating
( )
covariance matrix is O ( l + u )2 m . For each two samples, the Mahalanobis distance
( )
costs O m2 + ( l + u )2 m . Next, we modify the edge lengths that costs O ( l 2 ) . And
Dijkstra’s algorithm with a priority queue based on a binary heap allows equation (6)
( )
to have O ( l + u )2 ⎣⎡( l + u ) + log ( l + u ) ⎦⎤ time complexity [3]. That is, step 1 costs
(
O m2 + (l + u ) m + (l + u ) .
2 3
)
( )
The time complexity of MDS is approximately equal to O ( l + u ) in step 2. The
3

kernel matrix K m′ calculated by MDS sharply reduces the time complexity of gradient
( )
descent algorithm, which is approximately equal to O ( l + u ) . Thus, the time com-
3

( )
plexity of whole algorithm is O m 2 + ( l + u ) m + ( l + u ) . It nearly equals to the
2 3

time complexity of primal LDS algorithm.

4 Experimental Result
The original image database comes from all the 1096 sample images contained in the
CorelDraw Version 10.0 software CD#3. The stego image database comes from those
460 X. Zhang and S. Zhong

1096 images embedded by five typical data embedding methods [11]: the non-blind
spread spectrum (SS) method (α=0.1) by Cox et al., the blind SS method by Piva et
al., the 8×8 block based SS method by Huang and Shi, a generic quantization index
modulation (QIM, 0.1 bpp (bit per pixel)), and a generic LSB (0.3 bpp). The hidden
data are a random number sequence obeying Gaussian distribution with zero mean
and unit variance.
We randomly choose 300 original images as positive samples and 300 stego im-
ages as negative samples (totally 600) for testing in our experiment. They also com-
pose the unlabeled data set X u . And we randomly choose some other samples from
original image database and stego image database respectively for the labeled data set
X l . In experiments, we set the default values of parameters as recommended in [3],
while setting soft margin parameter C=50 by cross-validation.

4.1 Comparative Experiment for Supervised and Semi-supervised Learning

In this experiment, we randomly choose 10 to 100 labeled samples and combine with
600 samples unlabeled data set for training. The experiment is done for 20 times and
then calculates the mean accuracy and variance. The results of semi-supervised learn-
ing algorithm (LDS algorithm and our proposed algorithm) and the result of super-
vised learning algorithm (SVM algorithm) are shown below.

(a) (b)

Fig. 5. Comparative Experiment for supervised and semi-supervised learning: (a) the accuracy
of three algorithms; (b) the variance of three algorithms

Fig.5(a) illuminates that our proposed algorithm achieves the highest accuracy,
even with 10 labeled samples; our proposed algorithm can reach almost 90% accu-
racy. Fig.5(b) shows the variance of three algorithms. Algorithm with larger variance
is usually more unstable. Variance becomes much larger when the number of labeled
samples is small, especially for LDS algorithm. This phenomenon might be caused by
the biased selection of labeled samples, but we can see that our proposed algorithm
still has the best stability.
 An Improved Path-Based Transductive Support Vector Machines Algorithm 461

4.2 Comparative Experiment for Path-Based TSVM Algorithm and SVMlight

Algorithm

SVMlight algorithm has a wide range of applications despite its high time complexity,
while our proposed algorithm has lower time complexity. SVMlight software tool can
be obtained at http://www.cs.cornell.edu/People/tj/. In this experiment, we also ran-
domly choose 10 to 100 labeled samples and combine with 600 samples unlabeled
data set for training. The experiment is done for 10 times and then calculates the mean
accuracy and cost time. It is worth mentioning that the cost time of building graph in
our proposed method is included. The testing results are shown in Fig. 6 and table 1.
From the results, we can make the conclusion that our proposed algorithm has much
lower time complexity and higher accuracy than SVMlight algorithm.

Fig. 6. The accuracy of SVMlight and proposed algorithm with 10 to 100 labeled samples

Table 1. Average time of proposed algorithm and SVMlight

No. of No. of
Proposed (s) SVMlight(s) Proposed (s) SVMlight(s)
labeled labeled
10 4.50 277 60 7.12 601
20 5.25 357 70 7.39 532
30 5.51 520 80 7.54 553
40 6.06 541 90 7.74 408
50 6.52 593 100 8.12 436
average 5.568 457.6 average 7.582 506

5 Conclusions
In this paper, an improved path-based TSVM algorithm is proposed for blind stegana-
lysis classification. It uses Mahalanobis distance instead of Euclidean distance to
measure the distance between samples, and adds the information of labeled samples
by shrinking the edges of two samples in the same clusters, and expanding the edges
462 X. Zhang and S. Zhong

of two samples in different clusters. It improves the accuracy and time complexity.
Experiments show that our proposed path-based TSVM algorithm performs well even
in a small labeled data set. But it still has some drawbacks. With the increasing size of
labeled data set, the result is not significantly increasing accordingly. It might be
caused by the gradient descent method which will converge to a local minimal value
sometimes. Another drawback is, when the size of unlabeled samples becomes very
large, the time complexity will be extremely high.
In the future, we will try to do more comparative experiment for other semi-
supervised algorithms and find a low time complexity algorithm to deal with the large
size of unlabeled samples.
Acknowledgments. This paper is supported by the National Natural Science Founda-
tion of China under Grant NO.10871221, Fujian Provincial Young Scientists Founda-
tion under Grant No.2006F3076.

References
1. Chapelle, O., Scholkopf, B., Zien, A. (eds.): Semi-Supervised Learning. MIT Press,
Cambridge (2006)
2. Joachims, T.: Transductive inference for text classification using support vector machines.
In: Proceeding of ICML, Bled, pp. 200–209 (1999)
3. Chapelle, O., Zien, A.: Semi-Supervised Classification by Low Density Separation. In:
Proceeding of the AISTAT, Barbados, pp. 57–64 (2005)
4. Wang, J., Shen, X.: Large margin semi-supervised learning. J. Mach. Learn. Res. 8,
1867–1891 (2007)
5. Zhu, X.: Semi-Supervised Learning Literature Survey. Technical report, University of
Wisconsin-Madison,
http://pages.cs.wisc.edu/~jerryzhu/pub/ssl_survey.pdf
6. Zhong, S.P., Lin, J.: An Universal Steganalysis Method for GIF Image Based On TSVM.
J. Beijing Jiaotong University 33, 122–126 (2009)
7. Fischer, B., Roth, V., Buhmann, J.M.: Clustering with the connectivity kernel. In: Thrun,
S., Saul, L., Scholkopf, B. (eds.) Advances in Neural Information Processing Systems,
vol. 16. MIT Press, Cambridge (2004)
8. Chang, H., Yeung, D.Y.: Robust Path-Based Spectral Clustering. J. Pattern Recogni-
tion 41, 191–203 (2008)
9. Chen, Y.L., Wang, S.T.: Application of WMD Gaussian kernel in spectral partitioning. J.
Computer Applications 28, 1738–1741 (2008)
10. Tong, X.F., Teng, J.Z., Xuan, G.R., Cui, X.: JPEG Image Steganalysis Based on Markov
Model. J. Computer Engineering 34, 217–219 (2008)
11. Shi, Y.Q., Xuan, G.R., Zou, D., Gao, J.J.: Steganalysis Based on Moments of Characteris-
tic Functions Using Wavelet Decomposition, Prediction-Error Image, and Neural Network.
In: Proceeding of ICME 2005, Netherlands (2005)
12. Chen, Y.W., Lin, C.J.: Combining SVMs with various feature selection strategies,
http://www.csie.ntu.edu.tw/~cjlin/
 Resolution with Limited Factoring

Dafa Li

Dept of mathematical sciences

Tsinghua University, Beijing 100084, China

Abstract. The resolution principle was originally proposed by

J.A. Robinson. Resolution with factoring rule is complete for the first-
order logic. However, unlimited applications of factoring rule may gen-
erate many irrelevant and redundant clauses. Noll presented resolution
rule with half-factoring. In this paper, we demonstrate how to eliminate
the half-factoring.

Keywords: factoring rule, linear resolution, resolution.

1 Introduction

Resolution with factoring rule is complete[1]-[7]. There are four kinds of

resolvents[1]:
(1). a binary resolvent of clause C1 and clause C2 ,
(2). a binary resolvent of clause C1 and a factor of C2 ,
(3). a binary resolvent of a factor of C1 and clause C2 ,
(4). a binary resolvent of a factor of C1 and a factor of C2 .
For the definition of the binary resolvent, see [1]. However, unlimited applica-
tions of factoring rule may generate many irrelevant and redundant clauses. For
example, if a clause contains m literals with the same sign, then we need to test
2m − (m + 1) cases to decide which literals of the m literals have a mgu (most
general unifier) to find all the factors of the clause. This will require exponential
time.
Noll [5] showed that it is sufficient to factorize only one of two parent clauses of
a resolvent. See the Basic Lemma in [5] and also see [4]. He called the refinement
of the resolution as the resolution rule with half-factoring. Note that lemma 4
in this paper also reports the similar result. Also, in this paper, furthermore, we
investigate how to eliminate the half-factoring. In which cases can factoring be
eliminated? In lemmas 1, 2, 3 and 4, we discuss all the cases, where factoring
is applied. We show that factoring in lemmas 1, 2 and 3 can be eliminated
and factoring in lemma 4 is inevitable. Conclusively, we show that if one can
binary resolve two clauses on literals that might also both factor in their clauses,
then also allow factoring of only one of the clauses on those literals. Otherwise,
factoring can be ignored.


The paper was supported by NSFC(Grants: No. 60673034 and 10875061).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 463–468, 2009.


c Springer-Verlag Berlin Heidelberg 2009
464 D. Li

We need the following notation and definitions.

Notation: Let θ be a substitution and by the notation in [1], we can write θ =
{t1 /x1 , t2 /x2 , ..., tn /xn }. V AR{θ} = {x1 , .., xn }.
Definition
If literals L(1) , L(2) , ..., and L(m1 ) of clause C have a mgu σ, then Cσ is called a
factor of C[1] and L(1) σ the factored literal.
Definition
Linear binary resolution is a linear deduction, of which each resolvent is a binary
resolvent.

2 Factoring Can Be Eliminated in the Following Cases

2.1 Case 1. A Resolvent Has Only One Factored Parent
Lemma 1.
Let C1 = L1 ∨ L1 ∨ ... ∨ L1 1 ∨ C1 and C2 be parent clauses. Assume that
(1) (2) (m )

(1) (2) (m )
L1 , L1 , ..., and L1 1 have a mgu α, and C1 and C2 have no common variables.
Then the binary resolvent of the factor C1 α of clause C1 and clause C2 , where the
(1)
factored literal L1 α is the literal resolved upon, is an instance of the resolvent
obtained by linear binary resolution with C1 as a top clause and C2 as side
clauses. See Fig. 1.
Proof.
(1) (2) (m )
Let α be a mgu of L1 , L1 , ..., and L1 1 . Then,
(1) (2) (m1 )
L1 α = L1 α = ... = L1 α. (1)

Then L1 α ∨C1 α is a factor of C1 . Assume that C2 = L2 ∨ C2 , where L2 is the

(1)

(1)
literal resolved upon. Let σ be a mgu of L1 α and ¬L2 . Then
(1)
L1 ασ = ¬L2 σ. (2)

Let R be the resolvent of C2 and factor C1 α of C1 . Then

R = C1 ασ ∨ C2 σ. (3)

Since C1 and C2 have no common variables, we can choose α such that no

variables in V AR(α) occur in C2 . It means C2 α = C2 . Especially L2 α = L2 .
Thus, by Eq. (2)
(1)
L1 ασ = ¬L2 ασ. (4)
(1) (1)
This says that ασ is a unifier of L1 and ¬L2 . Let β be a mgu of L1 and ¬L2 .
Then there is a substitution γ such that

ασ = βγ. (5)
 Resolution with Limited Factoring 465

 


  


 




 


  

Fig. 1.

We take C1 as a top clause and C2 as a side clause and let R1 be the binary
(1)
resolvent of C1 and C2 , where L1 and L2 are the literals resolved upon. Then
β ∨ C1 β ∨ C2 β.
(2) (m1 )
R1 = L1 β ∨ ... ∨ L1 (6)

Now R1 is considered as a center clause and C2 is taken a side clause again, let
us compute a binary resolvent of R1 and C2 . Notice that from Eq. (1) and Eq.
(1) (2) (m )
(4), L1 ασ = L1 ασ = ... = L1 1 ασ = ¬L2 ασ. From Eq. (5),
(2)
L1 βγ = ¬L2 βγ. (7)

By notation in [1], we write C2 as L2 [x1 , ..., xl ] ∨ C2 [x1 , ..., xl ], where x1 , ..., xl
are all the variables which occur in C2 . After renaming all the variables in C2 ,
we obtain C2∗ = L∗2 ∨ C2∗ = L2 [y1 , ..., yl ] ∨ C2 [y1 , ..., yl ], where y1 , ... and yl are
new variables which do not occur in C1 , C2 , β, γ, α or σ. Clearly, L2 βγ =
L2 [x1 , ..., xl ]βγ = L2 [x1 βγ, ..., xl βγ]. Let λ = {x1 βγ/y1 , ..., xl βγ/yl }. Then,

L∗2 λ = L2 βγ = L2 βγλ, (8)

C2∗ λ = C2 βγ = C2 ασ. (9)

So λ is a unifier of L∗2 and L2 βγ, and λ is also a mgu of L∗2 and L2 βγ. From
Eqs. (7) and (8), L1 βγλ = ¬L2 βγλ = ¬L∗2 λ = ¬L∗2 γλ since y1 , ..., and yl do
(2)

not occur in V AR(γ). Therefore γλ is a unifier of L1 β and ¬L∗2 . Let η be a

(2)

mgu of L1 β and ¬L∗2 . Then for some substitution ρ,

(2)

γλ = ηρ. (10)

Let R2 be the binary resolvent of R1 and C2∗ , where L1 β and L∗2 are the literals
(2)

resolved upon. Then R2 = L1 βη ∨ ... ∨ L1 1 βη ∨ C1 βη ∨ C2 βη ∨ C2∗ η. For

(3) (m )
466 D. Li

simplicity, here let m1 = 2. Then R2 becomes R2 = C1 βη ∨ C2 βη ∨ C2∗ η. By Eq.

(10), R2 ρ = C1 βγλ∨C2 βγλ∨C2∗ γλ. By Eq. (5), R2 ρ = C1 ασλ∨C2 ασλ∨C2∗ γλ.
Thus, R2 ρ = C1 ασ ∨ C2 ασ ∨ C2∗ λ. By Eqs. (9) and (3), R2 ρ = C1 ασ ∨ C2 ασ =
C1 ασ ∨ C2 σ = R since C1 and C2 have no common variables. This implies that
R is an instance of R2 .
Example 1. The deduction of
can be obtained from P (a) and ¬P (x) ∨ ¬P (y)
by binary resolution.

2.2 Case 2. A Resolvent Has Two Factored Parents, Where the

Two Factored Literals Are Not Resolved Upon

Lemma 2. Let C1 = L1 ∨ L1 ∨ L1 ∨ ... ∨ L1 1 ∨ C1 and C2 = L2 ∨ L2 ∨ L2 ∨

(1) (2) (m ) (1) (2)

...∨L2 2 ∨C2 . Assume that L1 , L1 , ..., and L1 1 have a mgu α, L2 , L2 , ...,

(m ) (1) (2) (m ) (1) (2)

(m2 )
and L2 have a mgu β, and C1 and C2 have no common variables. Then the
binary resolvent of the factor C1 α of C1 and the factor C2 β of C2 , where L1 α
and L2 β are the literals resolved upon, is an instance of the binary resolvent of
C1 and C2 , where L1 and L2 are the literals resolved upon.
Proof.
Let µ = α ∪ β. Then factor C1 α = C1 µ = L1 µ ∨ L1 µ ∨C1 µ and factor C2 β =
(1)

C2 µ = L2 µ ∨ L2 µ ∨C2 µ. Let σ be a mgu of L1 α (= L1 µ) and ¬L2 β (= ¬L2 µ)

(1)

and Rf be the binary resolvent of factor C1 α of C1 and factor C2 β of C2 . Then

Rf = L1 µσ ∨ C1 µσ ∨ L2 µσ ∨C2 µσ.
(1) (1)

Since L1 µσ = ¬L2 µσ, µσ is a unifier of L1 and ¬L2 . Let υ be a mgu of

L1 and ¬L2 . Then µσ = νγ for some substitution γ. Let Rb be the binary
resolvent of C1 and C2 , where L1 and L2 are the literals resolved upon. Then
Rb = L1 ν ∨ L1 ν ∨ ... ∨ L1 1 ν ∨ C1 ν ∨ L2 ν ∨ L2 ν ∨ ... ∨ L2 2 ν ∨ C2 ν. Since
(1) (2) (m ) (1) (2) (m )

µσ = νγ, Rb γ = L1 µσ ∨ C1 µσ ∨ L2 µσ ∨ C2 µσ = Rf . It means that Rf is an

(1) (1)

instance of Rb .
Example 2. Let C1 = P (x) ∨ ¬Q(x) ∨ ¬Q(y) and C2 = ¬P (a) ∨ R(u) ∨ R(v).
Then instead of computing the resolvent of factors of the two clauses, we just
compute the binary resolvent of the two clauses.

2.3 Case 3. A Resolvent Has Two Factored Parents, Where Only

One Factored Literal Is the Literal Resolved Upon

Lemma 3. Let C1 = L1 ∨ L1 ∨ ... ∨ L1 1 ∨ C1 and C2 = L2 ∨ L2 ∨ L2 ∨ ... ∨

(1) (2) (m ) (1) (2)

L2 2 ∨ C2 . Assume that L1 , L1 , ..., and L1 1 have a mgu α1 , L2 , L2 , ...,

(m ) (1) (2) (m ) (1) (2)

(m2 )
and L2 have a mgu α2 , and C1 and C2 have no common variables. Then the
binary resolvent of the factor C1 α1 of C1 and the factor C2 α2 of C2 , where the
(1)
literal L2 α2 and the factored literal L1 α1 are the literals resolved upon, is an
instance of the resolvent obtained by linear binary resolution with C1 as a top
clause and C2 as side clauses. See Fig. 1.
 Resolution with Limited Factoring 467

Proof
Let α = α1 ∪ α2 . Then L1 α ∨C1 α is a factor of C1 and L2 α ∨ L2 α ∨ C2 α
(1) (1)

(1)
is a factor of C2 . Let σ be a mgu of L1 α and ¬L2 α, and R be the resolvent
of the factors of C1 and C2 . Then R = C1 ασ ∨ L2 ασ ∨ C2 ασ. For simplicity,
(1)

here let m1 = 2. Then, by using the method used in Lemma 1, we can obtain
a deduction of R2 by linear binary resolution with C1 as a top clause and C2
as side clauses, and we can show that for some substitution ρ, R2 ρ = C1 ασ ∨
(L2 ∨ L2 ∨ ... ∨ L2 2 ∨ C2 )ασ = C1 ασ ∨L2 ασ ∨ C2 ασ = R. It means that
(1) (2) (m ) (1)

R is an instance of R2 .
Example 3. Let C1 = ¬P (a) ∨ ¬P (u) ∨ R(v) and C2 = P (x) ∨ ¬Q(x) ∨ ¬Q(y).
Then instead of computing the resolvent of factors of the two clauses, we derive
the deduction by linear binary resolution with C1 as a top clause and C2 as side
clauses.

3 When a Resolvent Has Two Factored Parents, Where

Two Factored Literals Are the Literals Resolved Upon,
Also Allow Factoring Only One Parent

Lemma 4. Let C1 = L1 ∨ L1 ∨ ... ∨ L1 1 ∨ C1 and C2 = L2 ∨ L2 ∨ ... ∨

(1) (2) (m ) (1) (2)

L2 2 ∨ C2 . Assume that L1 , L1 , ..., and L1 1 have a mgu α1 , L2 , L2 , ...,

(m ) (1) (2) (m ) (1) (2)

(m )
and L2 2 have a mgu α2 , and C1 and C2 have no common variables. Then the
(1) (1)
binary resolvent of the two factors, where the factored literals L1 α1 and L2 α2
are the literals resolved upon, is an instance of the resolvent obtained by linear
binary resolution with C1 as a top clause and the factor C2 α2 of C2 as side
clauses. See Fig. 1.
Proof
let α = α1 ∪ α2 . Then C1 α = L1 α ∨C1 α and C2 α = L2 α ∨ C2 α. Let σ be a
(1) (1)

(1) (1)
mgu of L1 α and ¬L2 α and R be the binary resolvent of the factors C1 α and
(1) (1)
C2 α, where the factored literals L1 α1 and L2 α2 are the literals resolved upon.
 
Then R = C1 ασ ∨ C2 ασ. Now let us consider C1 as a top clause and the factor
C2 α of C2 as side clauses. For simplicity, here let m1 = 2. By lemma 1, we can
derive R2 ρ = C1 ασ ∨ C2 ασ = R for some substitution ρ, where the deduction of
R2 is obtained by linear binary resolution with C1 as a top clause and the factor
C2 α2 of C2 as side clauses. See Fig. 1. It implies that R is an instance of R2 .
We can find proofs of the following examples by using the limited factoring.
Example 4. S = {(1). P (x) ∨ P (y), (2). ¬P (x) ∨ ¬P (y)}.
Proof
(3). ¬P (z) a factor of (2)
(4). P (y) a binary resolvent of (1) and (3)
(5).
a binary resolvent of (3) and (4)
468 D. Li

Example 5. S = {(1). P (z, y)∨P (x, g(x)),(2). Q(b)∨¬P (u, v), (3). ¬P (w, g(b))∨
¬Q(w)}[7].
Proof
(4). ¬P (u, v) ∨ ¬P (b, g(b)) a binary resolvent of (2) and (3)
(5). ¬P (b, g(b)) a factor of (4)
(6). P (z, y) a binary resolvent of (1) and (5)
(7).
a binary resolvent of (5) and (6)

4 Summary
Noll [5] showed that it is sufficient to factorize only one of two parent clauses of
a resolvent. He called the refinement of the resolution as the resolution rule with
half-factoring. In this paper, furthermore, we demonstrate how to eliminate the
half-factoring. We show that if one can binary resolve two clauses on literals that
might also both factor in their clauses, then also allow factoring of only one of
the clauses on those literals. Otherwise, factoring can be ignored. For example,
lemmas 2 and 3 declare that it is not necessary to factorize the literals which
are not resolved upon.

References
[1] Chang, C.C., Lee, R.C.T.: Symbolic logic and mechanical theorem proving. Aca-
demic Press, San Diego (1973)
[2] Duffy, D.: Principles of automated theorem proving. John Wiley & Sons, Chichester
(1991)
[3] Fitting, M.: First-order logic and automated theorem proving. Springer, New York
(1990)
[4] Kowalski, R.: Studies in completeness and efficiency of theorem proving by resolu-
tion. Ph.D. thesis, University of Edingburgh (1970)
[5] Noll, H.: A Note on Resolution: How to get rid of factoring without loosing com-
pleteness. In: Bibel, W. (ed.) CADE 1980. LNCS, vol. 87, pp. 250–263. Springer,
Heidelberg (1980)
[6] Robinson, J.A.: A machine-oriented logic based on the rosolution principle. J.
ACM 12(1), 23–41 (1965)
[7] Socher-Ambrosius, R., Johann, P.: Deduction systems. Springer, New York (1997)
 Formal Analysis of an Airplane Accident in
N Σ-Labeled Calculus

Tetsuya Mizutani1 , Shigeru Igarashi1, Yasuwo Ikeda2 , and Masayuki Shio3

1
Department of Computer Science, University of Tsukuba, Tsukuba, Japan
{mizutani,igarashi}@cs.tsukuba.ac.jp
2
Mejiro University, Tokyo, Japan
yikeda@krd.biglobe.ne.jp
3
College of Community Development, Tokiwa University, Mito, Japan
shio@tokiwa.ac.jp

Abstract. N Σ-labeled calculus is a formal system for representation,

verification and analysis of time-concerned recognition, knowledge, belief
and decision of humans or computer programs together with related ex-
ternal physical or logical phenomena. In this paper, a formal verification
and analysis of the JAL near miss accident is presented as an example of
cooperating systems controlling continuously changing objects including
human factor with misunderstanding or incorrect recognition.

Keywords: Formal representation and analysis, Time-concerned

program systems, Belief and decision, Near miss accident.

1 Introduction

N Σ-labeled calculus [9], [12] is a formal system, which is a kind of “logic of

knowledge and thought [11]”, in order to describe time-concerned recognition,
knowledge, belief and decision of humans or computer programs together with
related external physical or logical phenomena. Formal verification and analysis
of the JAL near miss accident [1], [10] in this formalism is presented as an exam-
ple of cooperating systems controlling continuously changing objects including
human factor with misunderstanding or incorrect recognition. In this analysis,
some sufficient conditions that the accident would not occur are also shown.
Through this example, relationship among artificial intelligence, external en-
vironment and human factors are investigated.
There are many related works in this area. There are many reasoning systems
for temporal or time-related logic of knowledge and belief of human, e.g. [3], [5],
[7], etc. And also, airplane controlling systems as well as vehicle controlling sys-
tems are formally treated in [4]. All of these systems use model checking method,
while N Σ-labeled calculus does the proof-theoretic method on the natural num-
ber theory, which is one of the remarkable feature of this calculus.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 469–478, 2009.


c Springer-Verlag Berlin Heidelberg 2009
470 T. Mizutani et al.

2 N Σ-Labeled Calculus
A brief explanation of N Σ-labeled calculus is introduced along with [9] and [12],
whose semantics is described in [9] with its soundness. We will partially follow
[14] in which PA stands for Peano arithmetic, while an occurrence of the variable
x in a formula A will be indicated explicitly as A[x], and µxA[x] will designate
the least number satisfying A, in this paper.
Peano arithmetic (PA) is extended to PA(∞) called pseudo-arithmetic [8],
including the infinity (∞) and the minimalization (µ).
Infinite number of special constants, J, J1 , J2 , . . . and labels, , 1 , 2 , . . .
∈ Σ are added, where Σ is the set of labels. The special constants correspond to
program variables taking natural number values and possibly ∞, and the change
of their values along the natural number time is expressed by the change of local
models, or worlds, and vice versa. It must be noticed that a program variable is
guarded against quantification in the calculus, since each J is not a variable but
a constant. A label indicates a personality that is a generalization of an agent of
multi-agent systems, or an observer in physics, including notion of subjectivity.
A tense means the time relative to a reference ‘observation time’ called now,
which is taken to be 0 throughout the calculus, where ∞ indicates the tense
when false holds. The value of a special constant may change along with tense.
@-mark as a logical symbol is called the coincidental operator. A formula
of the form A@a,  intuitively means that the personality designated by the
label believes at the tense designated by a the fact that “A holds now”, while
A@a and A@ mean that A holds at tense a, and that believes that A holds
now, respectively. The logical symbol “ ; ” called the futurity operator moves
the observation time toward a future time-point. a; b is the tense of b observed
at the tense designated by a. a; A means the least, i.e. the earliest, time when
A comes to hold, or rises, after or exactly at the tense a, which will be called
the ascent of A at a.

3 Representation of Cooperative Continuous Programs

Actions or changes of states of programs for verification are represented by pro-
gram axioms, used as the nonlogical axioms in the proofs of theorems for the
particular programs. Some primitive notions for program axioms and for repre-
senting cooperative continuous programs are introduced.

3.1 Spur and Program Labels

Definition 1. A spur α is defined as a term by

α = µy(0 < y&0 < J@y),

for an arbitrarily chosen and fixed special constant J.

Spurs serve as the generalization of schedulers and the next operator in temporal
logic. α, β, γ, . . . , κ are used as the metasymbols for spurs. Each component
 Formal Analysis of an Airplane Accident in N Σ-Labeled Calculus 471

process, or even certain external object, of a multi-CPU parallel program system

is assigned a distinct spur usually, and the whole program is described in accor-
dance with the causality, i.e. the relationships between the spurs as motives and
the changes of values.
Besides the spur, to each process is assigned a distinct special constant Li ,
the formula Li = j expressing that the control of the i-th process in the parallel
program is at its j-th block, statement, etc.
Definition 2. ai,j (i, j = 0, 1, . . .) is a program label, expressed by mutually
exclusive special boolean constants defined as:
ai,j ≡ Li = j (i, j = 0, 1, . . .).
where Li is an arbitrary special constant.

3.2 Conservation Axioms

We adopt the following two axiom schemata [8].
(CA1)J = z ⊃ x < α & x < β & . . . & x < κ ⊃ J = z@x,
for each special constant J and all spurs α, β, . . . , κ in the program system.
This axiom means that the value of J does not change until any spur rises, i.e.
the next step of any process rises, where each spur α, β, . . . , κ indicates the
tense when the next step of the process corresponding to the spur rises.
(CA2)J = z@a ⊃ α < β & . . . & α < κ ⊃ a ≤ x ≤ α ⊃ J = z@x,
for each α, a and J, such that J does not occur in the ‘action’ part of a corre-
sponding program axiom in the axiom tableau (section 4), where the action part
means the part of the program axiom denoting the change of program variables.
This axiom says that the value of the special constant J does not change until
the next step α rises, when the control of the corresponding process is on the
program block labeled by a. It means that the value of J does not change within
the block corresponding to a, i.e. it does not change if the program axiom does
not denote the change of its value.

3.3 Approximation of Continuous System

For representation, analysis and verification of cooperating systems with continu-
ity, the notion of differentiation must be dealt with. In this paper, the first-order
derivatives, i.e. speed, are treated as special constants. The distance are defined
by the integral using Euler’s approximation.

Definition 3. For a term a[x], 0≤x<y a[x] is inductively defined as follows.

a[x] = 0, a[x] = a[x] + a[y],

0≤x<0 0≤x<y+1 0≤x<y
⎧
⎨ 0≤x<µy(∀z(y≤z⊃a[z]=0)) a[x],
a[x] = if ∃y(∀z(y ≤ z ⊃ a[z] = 0)),
⎩
0≤x<∞ ∞, otherwise.

472 T. Mizutani et al.

 b+t  b+t
Definition 4. For a term a[x], x=b a[x], that is, an approximation of x=b a[x]dx,
is defined as follows.
 b+t
def
a[x] = h · a[b + yh],
x=b 0≤y<n+1

where t = nh.

It must be noted that the error of this approximation can be made arbitrary
small by choosing a sufficiently small h.

Definition 5. Let Ȧ be a special constant. A is defined as follows.

 t
def
A[t] = µy(y = Ȧ[x]@x).

x=0

4 Representation by Axiom Tableaux

By axiom tableaux [12], the program axioms representing actions or changes of
states of program systems to be verified are easily and readably expressed. The
following abbreviations are used in the program tableaux.

Abbreviation 1. 1. Each axiom is numbered and divided into ‘condition/

prefix’, ‘action’, ‘tense’ and ‘label’ parts.
2. “&” is replaced by comma “ , ” whenever there is no confusion.
3. Double line, in contrast to single ones, delimit the scope of free variables,
which are quantifiable by “∀”. If “forsome” occurs in the condition/prefix
cell, the variables in the row is bound by “∃”.
4. ‘∗’ mark in the label cell means the corresponding axiom is not only a belief
by the personalities indicating the label cell but also a fact, i.e. it holds
without the belief by the personalities.

5 Analysis of JAL Near Miss Accident

5.1 Outline of the Accident
The outline of the JAL near miss accident is as follows [1], [10].
On Jan. 31, 2001, JAL flight 907, called A in this paper, as well as in [10],
had departed Tokyo-Haneda for a flight with destination Naha. JAL Flight 958,
called B, was en route from Pusan to Tokyo-Narita. A trainee controller at
Tokyo Area Control Center (ACC) cleared A to climb to FL3901 at 15:46, Two
minutes later B reported at FL370. Both flights were on an intersecting course
near the Yaizu NDB. At 15:54 the controller noticed this, but instead of ordering
B to descend, he ordered A to descend. Immediately after this instruction, the
1
FL stands for Flight Level, and FL1, written as 1[FL] in formulas in this paper, is
100 [ft] in a standard nominal altitude.
 Formal Analysis of an Airplane Accident in N Σ-Labeled Calculus 473

Table 1. Program Axioms

index condition/prefix action tense label

↑ anticipateNM(A, B)
1 α = CNF = D0 ∗, Monitor
@Monitor
2 anticipateNM(A, B) ≡ CNF ACC
3 β = (CNF@ACC) = D1 ↑ CNF ∗, ACC
β = (cntl(A, X) ∨ cntl(B, Y ),
↑ anticipateNM(A, B)
4 forsome X, Y cntl(A, X) ≡ Ecntl@A, ∗, ACC
@ACC
cntl(B, X) ≡ Ecntl@B) = D2
cntl(A, X) ∨ cntl(B, Y ) ⊃
5 forsome X, Y ∗, ACC

180 Separation7,5
6 γP = ¬cntl(P, X)@P = 15sec ↑ cntl(P, X)@P ∗, P, ACC
↑ anticipateNM(A, B)
7 forsome X ρA = RA(A, X) = D3 ∗, A, TCASA
@TCASA
↑ anticipateNM(A, B)
8 forsome X ρB = RA(B, X) = D4 ∗, B, TCASB
@TCASB
9 γP = ope(P, X) = D5 ↑ cntl(P, X)@P ∗, P, ACC
10 ¬Ecntl γP = ope(P, X) = D6 ↑ RA(P, X)@P ∗, P, ACC
γA =
(|fv (A, v, h, s)| ≤ |ȧv | ,
11 ↑ ope(A, v, h) ∗, A, B, ACC
0 ≤ ȧv · f (A, v, h, s))
= D7

crew of A were given an aural TCAS Resolution Advisory (RA) to climb in

order to avoid a collision. At the same time the crew of B were given an aural
RA to descend. The captain of A followed the instructions of the controller by
descending. A now approaching close to B, because captain of B descended as
well, following the advisory of his TCAS. A collision was averted when the pilot
of A then put his airplane into a nosedive. A missed B by 105 to 165 meters in
lateral distance and 20 to 60 meters in altitude difference. About 100 crew and
passengers on A sustained injuries due to emergency moreover, while no one was
injured on B.

5.2 Formalization
In this paper, the decision and order of the controller at Tokyo ACC, the de-
cision of the crews of the 2 airplanes, and the order from TCASs installed in
the airplanes caused the near miss are formulated, verified and analyzed in the
calculus.
Only the changes of the vertical positions of the airplanes are dealt with in
the analysis, for the simplicity. To do it, the projection of the positions of the
airplanes to a vertical plain is considered.

Definition 6. The following schemata are defined, where Pi (i = 1, 2) is a meta-

expression representing the airplane A and B, respectively, and pi = piv , pih 
is a pair of the special constants piv representing the vertical position and pih
representing the horizontal position of Pi .

(t, A) ≡ ∃u ≤ t(A@u), (1)

hold(P) ≡ ∀xv xh (ṗv = xv &ṗh = xh ⊃ ∀t((ṗv = xv &ṗh = xh )@t)), (2)
474 T. Mizutani et al.

def
Separation(P1 , P2 , v, h) = v[FL] < |p1v − p2v | ∨ h [nm] < |p1h − p2h |,(3)
180 A ≡ (180[s], A), (4)
def
Separation7,5 = Separation(P1 , P2 , 7, 5), (5)
def
anticipateNM(P1 , P2 ) = hold(P1 )&hold(P2 ) ⊃ 180 ¬Separation7,5 , (6)

where the unit nm is the initials of nautical mile.

Separation is a minimal distance between two airplanes served by the controller

to keep safe and ordered traffic, defined as the vertical distance is greater than
v[FL] or the horizontal distance is greater than h[nm]. Separation7,5 expresses
that the vertical distance is greater than 7[FL] or the horizontal distance is
greater than 5[nm]. The near miss condition is defined when it does not hold.
anticipateNM(A, B) expresses that a near miss will occur within 180 seconds if
the speed or the direction of the airplane A or B does not change.
The program axioms representing the actions of ACC, A, B, TCASA and
TCASB by the axiom tableau are listed in Table 1, together with the facts what
and when they actually acted, or ordered, when the accident occurred, in Table
2. In these tables, Di (0 ≤ i ≤ 5) indicate delays. α, β, γA , γB are spurs for the
monitor, the controller, airplanes A and B, respectively. TCASA and TCASB
are TCASs installed in A and B, respectively. ρA and ρB are spurs of TCASA
and TCASB , respectively.
The intended meaning of these program axioms and facts are as follows.

– Axiom 1 (the row indicated by number 1 in Table 1) represents the recog-

nition of the monitor, where CNF is a special constant over Boolean values
indicating a conflict alert occurs or not. If the monitor anticipates a near
miss, a conflict alert will occur with D0 second delay.
– Axiom 2-5 indicate the recognition of the controller.

Table 2. Facts
index condition/prefix action tense label
↑ anticipateNM(A, B)
12 S ∗, Monitor
= 15:54
15


↑ cntl(A, ⊥, 350[FL])
13 S ∗, A
= 15:54
27


↑ (cntl(A, ⊥, 350[FL]) ≡ Ecntl)
14 S ∗, A
= 15:54
27


↑ RA(A, 1500ft/min, ⊥)
15 S ∗, A
= 15:54
35


↑ RA(B, −1500ft/min, ⊥)
16 S ∗, B
= 15:54
34


17 ¬Ecntl, ¬CNF S ∗, A, B, ACC
↑ cntl(B, ⊥, 350[FL])
18 S ACC
= 15:54
27


ope(B, ⊥, 350[FL]) @D6 ,
19
180 Separation7,5 S; 15:54
27

∗, ACC
hold(A)
20 hold(A) ACC
 Formal Analysis of an Airplane Accident in N Σ-Labeled Calculus 475

• Axiom 2: The controller recognizes that anticipating the near miss and
the corresponding alert will occur simultaneously.
• Axiom 3: The controller recognizes the alert D1 seconds after it actually
occurs.
• Axiom 4: The controller operates some emergency controlling order in-
dicated by Ecntl to each airplane A or B with D2 second delay after he
notices the near miss danger.
• Axiom 5: There exist some suitable orders that the two airplanes will
not in the near miss state, namely, they will keep a safety separation for
180 seconds.
– Axiom 6: The order will be voided after 15 seconds.
– Axiom 7: TCAS of the airplane A outputs resolution advisory, and orders
to climb or descend some value X, after D3 seconds when it recognizes the
near miss danger.
– Axiom 8 is similar. It is for the airplane B.
– Axiom 9: Each airplane operates to climb or descend some value X with D5
seconds delay after the controller at the ACC sends the control order.
– Axiom 10: The airplane must follow the order from its own TCAS if there
is no emergency control from ACC.
– Axiom 11 is the differential equation of the actual movement when the air-
plane operates the value X, i.e. the pair of values, v and h, where v is the
vertical speed to climb or descend and h is vertical position to go.
– Fact 12: The monitor on the panel of the controller anticipated the danger
of near miss at 15:54 15 , where S indicates the actual starting time of the
whole system.
– Fact 13: The pilot or the crew of the airplane A recognized at 15:54 27 that
the controller ordered to descend to FL350.
– Fact 14: The crew of the airplane recognized that the order was emergent.
– Fact 15: The crew, on the other hand, recognized at 15:54 35 that TCAS
of A output resolution advisory to climb at 1500ft per minute. Of course
these two operations contradicts, and the crew followed that of ACC for the
axioms 9 and 10.
– Fact 16: At the same time, the crew of B recognized the order from its own
TCAS to descend at 1500 ft per minute.
– Fact 17 is the initial condition. At the start time of the system, there are no
emergency control and no conflict alert.
– Facts 18-20 are recognition by the controller.
• Fact 18: Though the Fact 13 (and 14), the controller believed, or mis-
recognized, that he ordered B, instead of A, to descend to FL350.
• Fact 19: The controller believed that if he has ordered as Fact 18, the
near miss accident would occur.
• Fact 20: These two airplane held their vertical and horizontal speed.

In these axioms and facts, the program labels exist implicitly and do not repre-
sent explicitly.
476 T. Mizutani et al.

5.3 Verification and Analysis

Lemma 1.
∃XY (↑ (cntl(A, X) ∨ cntl(B, Y )) = 15:54 15 + D0 + D1 + D2 @S). (7)
Intuitively, it means that there existed some operation from the controller to A
or B at 15:54’15” with D0 + D1 + D2 [s] delay.
[Proof.] From Axiom 1 and Fact 12,
↑ CNF = 15:54 15 + D0 @S. (8)
From Axiom 3,
↑ CNF@ACC = 15:54 15 + D0 + D1 @S. (9)
From Axiom 2,
↑ anticipateNM(A, B)@ACC = 15:54 15 + D0 + D1 @S.
Therefore (7) is obtained from Axiom 4.

Actually, putting D0 + D1 + D2 = 8[s], Fact 13 satisfies Lemma 1.
Lemma 2.
ope(A, ⊥, 350[FL])@15:54 27 + D5 @S (10)
&ope(B, −1500ft/min., ⊥)@15:54 34 + D6 @S.
It means that the airplane A operated to go (descend) to 350[FL] at 15:54’27”
with D5 [s] delay, and B did to descend at 1500ft per minute at 15:54’34” with
D6 [s] delay.
[Proof.] From Axiom 9 and Fact 13,
↑ ope(A, ⊥, 350[FL]) = D5 @15:54 27 @S.
On the other hand, from the conservation axioms,
¬Ecntl@B@15:54 34 @S. (11)
Thus from Axiom 10 and Fact 16,
↑ ope(B, −1500ft/min., ⊥) = D6 @ 15:54 34 @S. (12)


Let us analyze the accident for applying the actual values to Lemma 2 and
Axiom 11. The actual values of the vertical positions av and bv of these planes
at 15:54 34 are both FL370. From Axiom 2, both airplanes descended between
15:54 27 and 15:54 34 . And also, when they started to descend, the horizontal
positions and speed of these planes were ah = 5[nm], bh = −5[nm], ȧh = 500[kt]
and ḃh = −500[kt], where the origin is the position that the near miss occurred.
Applying these values to Axiom 11 and integrating it, the following theorem
representing that the near miss really occurred is obtained.
Theorem 1.
180 ¬Separation7,5 . (13)
 Formal Analysis of an Airplane Accident in N Σ-Labeled Calculus 477

Table 3. Axiom when TCAS are given priority over ACC

index condition/prefix action tense label

9’ γP = ope(P, X) = D5 ↑ RA(P, X)@P ∗, P, ACC
10’ ¬RA(P, Y ) γP = ope(P, X) = D6 ↑ cntl(P, X)@P ∗, P, ACC

5.4 Some Other Cases

1. Defending the controller:
As Facts 13 and 18, the trainee controller wanted to order B to descend, but
he actually did to A. If he ordered correctly, the near miss did not occur,
i.e.

180 Separation7,5 , (14)

which is the negation of (13), is derived.

On the other hand, the defense of the controller argues in court that the
order by the controller is still safe, since no one know the order by TCASs,
and if they ordered nothing, then the near miss would not occur. Namely,
Facts 15-16 do not hold. From the conservation axiom [12], ḃv = 0 holds,
therefore (14) is also obtained.
2. A carried out RA instead of ACC:
If the crew of A carried out RA by TCASA instead of ACC, i.e. axioms 9’
and 10’ (Table 3) held instead of 9 and 10, if then A climbed while B carried
out the order of TCASB to descend, thus (14) also holds.

6 Conclusion
This paper has demonstrated an analysis of a concrete serious near miss acci-
dent of airplanes by N Σ-labeled calculus that is a formal system for complicated
control systems involving human factor, especially misunderstanding and inap-
propriate decision.
If the order from the controller at ACC contradicts that from TCAS, the
report from the Investigating Committee for Avion and Traffic Accident [10]
recommends that the crew must follow the latter, as mentioned in the Section
5.4. And also, on Jul. 1, 2002, a similar airplane accident occurred at Überlingen
[2]. A crew of one airplane followed the descending order from the controller,
though there was a climbing order from its own TCAS. On the other hand,
the crew of another airplane descended to follow the order from its own TCAS.
Consequently, these two airplanes collided.
From the investigations of these accidents, every crew now must follow orders
from TCAS if orders from it and from controllers contradicts. Namely, axioms 9’
and 10’ in Table 3 are substituted for axioms 9 and 10. Following to these new
rules, the fact that there is no near miss accident by such a conflict is proved.
One of our future studies is to construct some safety designs to control air-
planes, trains, vehicles, etc., against misunderstanding and incorrect decision.
478 T. Mizutani et al.

References
1. Avion Safety Network (ed.): ASN Aircraft accident McDonnell Douglas DC-10-40
JA8546 off Shizuoka Prefecture (2005),
http://aviation-safety.net/database/record.php?id=20010131-2
2. Avion Safety Network (ed.): ASN Aircraft accident Tupolev 154M RA-85816
Überlingen (2007),
http://aviation-safety.net/database/record.php?id=20020701-0
3. Curzon, P., Ruksenas, R., Blandford, A.: An Approach to Formal Verification of
Human-Computer Interaction. Formal Aspect of Computing 19, 513–550 (2007)
4. Damm, W., Hungar, H., Olderog, E.-R.: Verification of Cooperating Traffic Agents.
International Journal of Control 79, 395–421 (2006)
5. Fagin, R., Halpern, J.Y., Moses, Y., Vardi, M.Y.: Reasoning About Knowledge.
The MIT Press, Cambridge (1995)
6. Gentzen, G.: Untersuchungen über das logische Schließen. Mathematische
Zeitschrift 39, 176–210, 405–431 (1935); Investigations into Logical Deduction. In:
Szabo, M.E. (ed.) The Collected Papers of Gerhard Gentzen, Series of Studies in
Logic and the Foundations of Mathematics, pp. 68–131. North-Holland Publ. Co.,
Amsterdam (1969)
7. Halpern, J.Y., Vardi, M.Y.: The Complexity of Reasoning about Knowledge and
Time. I. Lower Bounds. Journal of Computer and System Sciences 38, 195–237
(1989)
8. Igarashi, S., Mizutani, T., Ikeda, Y., Shio, M.: Tense Arithmetic II: @-Calculus as
an Adaptation for Formal Number Theory. Tensor, N. S. 64, 12–33 (2003)
9. Mizutani, T., Igarashi, S., Ikeda, Y., Shio, M.: Formal Analysis of an Airplane
Accident in N Σ-labeled Calculus. In: Deng, H., Wang, L., Wang, F.L., Lei, J. (eds.)
AICI 2009. LNCS (LNAI), vol. 5855, pp. 469–478. Springer, Heidelberg (2009)
10. Investigating Committee for Avion and Traffic Accident (ed.): Report of the Near
Miss Accident by JA8904 belonging to Japan Airlines with JA8546 belonging to
the Same Company (2002) (in Japanese)
11. McCarthy, J., Sato, M., Hayashi, T., Igarashi, S.: On the Model Theory of Knowl-
edge. Stanford University Technical Report, STN-CS-78-657 (1979)
12. Mizutani, T., Igarashi, S., Shio, M., Ikeda, Y.: Human Factors in Continuous Time-
Concerned Cooperative Systems Represented by N Σ-labeled Calculus. Frontiers
of Computer Science in China 2, 22–28 (2008)
13. Presburger, M.: Über die Vollständigkit eines gewissen Systems der Arithmetik
ganzer Zahlen in welchem die Additon als einzige Operation hervortritt. In:
Comptes-Rendus du Congres des Mathematicians des Pays Slaves, pp. 92–101
(1930)
14. Shoenfield, J.R.: Mathematical Logic. Addison-Wesley Publishing Company, Read-
ing (1967)
 Using Concept Space to Verify Hyponymy in Building a
Hyponymy Lexicon

Lei Liu1, Sen Zhang1, Lu Hong Diao1, Shu Ying Yan2, and Cun Gen Cao2
1
College of Applied Sciences, Beijing University of Technology
2
Institute of Computing Technology, Chinese Academy of Sciences
{liuliu_leilei,zhansen,diaoluhong}@bjut.edu.cn,
{yanshuying,caocungen}@ict.edu.cn

Abstract. Verification of hyponymy relations is a basic problem in knowledge

acquisition. We present a method of hyponymy verification based on concept
space. Firstly, we give the definition of concept space about a group of candi-
date hyponymy relations. Secondly we analyze the concept space and define a
set of hyponymy features based on the space structure. Then we use them to
verify candidate hyponymy relations. Experimental results show that the
method can provide adequate verification of hyponymy.

Keywords: hyponymy, relation acquisition, knowledge acquisition, hyponymy

verification.

1 Introduction
Automatic acquisition of semantic relations from text has received much attention in
the last ten years. Especially, hyponymy relations are important in accuracy verifica-
tion of ontologies, knowledge bases and lexicons [1][2].
Hyponymy is a semantic relation between concepts. Given two concepts X and Y,
there is the hyponymy between X and Y if the sentence “X is a (kind of) Y” is accept-
able. X is a hyponym of Y, and Y is a hypernym of X. We denote a hyponymy rela-
tion as hr(X, Y), as in the following example:
中国是一个发展中国家 ---hr(中国,发展中国家)
(China is a developing country ---hr(China, developing country) )
Human knowledge is mainly presented in the format of free text at present, so
processing free text has become a crucial yet challenging research problem. The error
hyponymy relations in the phase of acquiring hyponymy from free text will affect the
building of hyponymy lexicon.
In our research, the problem of hyponymy verification is described as follows:
Given a set of candidate hyponymy relations acquired based on pattern or statistics,
we denoted these relations as CHR= {(c1, c2), (c3, c4), (c5, c6), …}, where ci is the
concept of constituting candidate hyponymy relation. The problem of hyponymy
verification is how to identify correct hyponymy relations from CHR using some
specific verify methods.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 479–486, 2009.
© Springer-Verlag Berlin Heidelberg 2009
480 L. Liu et al.

In this paper, we present a method of hyponymy verification based on concept

space. The rest of the paper is organized as follows. Section 2 describes related work
in the area of hyponymy acquisition, section 3 gives the definition of concept space
and analyzes the structure of concept space, section 4 gives a group of hyponymy
features and presents how to verify candidate hyponymy relations, section 5 conducts
a performance evaluation, and finally section 6 concludes the paper.

2 Related Work
There are two main approaches for automatic/ semi-automatic hyponymy acquisition.
One is pattern-based (also called rule-based), and the other is statistics-based. The
former uses the linguistics and natural language processing techniques (such as lexical
and parsing analysis) to obtain hyponymy patterns, and then makes use of pattern
matching to acquire hyponymy, and the latter is based on corpus and statistical lan-
guage model, and uses clustering algorithm to acquire hyponymy.
At present the pattern-based approach is dominant, and its main idea is the hy-
ponymy can be extracted from text as they occur in detectable syntactic patterns. The
so-called patterns include special idiomatic expressions, lexical features, phrasing
features, and semantic features of sentences. Patterns are acquired by using the lin-
guistics and natural language processing techniques.
There have been many attempts to develop automatic methods to acquire hy-
ponymy from text corpora. One of the first studies was done by Hearst [3]. Hearst
proposed a method for retrieving concept relations from unannotated text (Grolier’s
Encyclopedia) by using predefined lexico-syntactic patterns, such as
…NP1 is a NP2… ---hr(NP1, NP2)
…NP1 such as NP2… ---hr(NP2, NP1)
…NP1 {, NP2}*{,} or other NP3 … ---hr (NP1, NP3), hr (NP2, NP3)
Other researchers also developed other ways to obtain hyponymy. Most of these tech-
niques are based on particular linguistic patterns.
Morin and Jacquemin produced partial hyponymy hierarchies guided by transitiv-
ity in the relation, but the method works on a domain-specific corpus [4].
Llorens and Astudillo presented a technique based on linguistic algorithms, to con-
struct hierarchical taxonomies from free text. These hierarchies, as well as other rela-
tionships, are extracted from free text by identifying verbal structures with semantic
meaning [5].
Sánchez presented a novel approach that adapted to the Web environment, for
composing taxonomies in an automatic and unsupervised way. It uses a combination
of different types of linguistic patterns for hyponymy extraction and carefully de-
signed statistical measures to infer information relevance [6].
Elghamry showed how a corpus-based hyponymy lexicon with partial hierarchical
structure for Arabic can be created directly from the Web with minimal human super-
vision. His method bootstraps the acquisition process by searching the Web for the
lexico-syntactic patterns [7].
 Using Concept Space to Verify Hyponymy in Building a Hyponymy Lexicon 481

3 Concept Space
In Chinese, one may find several hundreds of different hyponymy relations patterns
based on different quantifiers and synonymous words, which is equivalent to the sin-
gle hyponymy pattern (i.e. (<?C1> is a <?C2>), (<?C3> such as <?C1>,<?C2>)) in
English. Fig. 1 depicts a few typical Chinese hyponymy relation patterns.

Pattern 1: <?C1>< 是一 为一 种 个|名|篇|片|块|部|颗|本|…><?C2>

| >< |
(Pattern: <?C1> is a <?C2>)

Pattern 2: <?C3>< 如 |象 | 包括 |包含 |囊括 |涵盖 ><?C1>{< 或 |及 |以

及|和|与|、><?C2>}*<等>
(Pattern: <?C3> such as <?C1>,<?C2>…)

Fig. 1. Defining Chinese hyponymy patterns

In Fig. 2, Pattern 1 means “Pattern: <?C1> is a <?C2>”. Pattern 2 means “Pattern:

<?C3> such as <?C1>,<?C2>…”. These items of the pattern are divided into constant
item and variable item. Constant item is composed of one or more Chinese words or
punctuations. Variable item is a non-null string variable. “<?C1>” is a variable item in
是一 为一
the pattern. “|” expresses logical “or”. In pattern1,“ | 种个名
”means “is a”; “ | |
篇片块部颗本 如象
| | | | | | …” is a group of quantifiers. In pattern 2, “ | ” means “such as”;
包括 包含 囊括 涵盖
“ | | | 和与 或
” means “include”; “ | ” means “and”; “ ” means “or”; “ 等
及 以及
” means “etc.”; “ | 、
” denotes “as well as”; Chinese dunhao “ ” is a special kind
of Chinese comma used to set off items in a series.
Chinese hyponymy relation patterns will be used to capture concrete sentences
from Chinese free corpus. In this process, variables <?C> will be instantiated with
words or phrases in a sentence, in which real concepts may be located. Let c and c’ be
the real concept in <?C>. If hr(c, c′) is true, then we tag c by cL, and c′ by cH, as
shown below.
众所周知， 中国
{ 是一个 社会主义 国家
{ }cL}<?C1>/ /{ { }cH}<?C2>
({It is well-known that {China}cL}<?C1>/is a/ { socialist nation}cH }<?C2> )
农作物 主要
{{ 有 水稻
}cH 、 玉米
}<?C5>/ /{{ }cL}<?C1> {{ 、 红
}cL}<?C2> {{
薯 、 烟叶
}cL}<?C3> {{等 }cL}<?C4>/ /
(The farm crop mainly includes paddy rice, corn, sweet potato, tobacco leaves etc.)
中国 国家 水稻 农作物 玉米 农作物
We can acquire hr( , ), hr( , ), hr( , 红薯 农作
), hr( ,
物 烟叶 农作物
) and hr( , ) from the above example.
As we know, Chinese is a language different from any western language. A
Chinese sentence consists of a string of characters which do not have any space or
delimiter in between [8]. So there are still many error relations in the acquired hy-
ponymy relations from free text. They must be verified further for the building of
hyponymy lexicon.
482 L. Liu et al.

Firstly we initially acquire a set of candidate hyponymy relation from large Chi-
nese free text based on Chinese lexico-syntactic patterns. Then we build concept
space using those candidate hyponymy relations.
Definition 1: The concept space is a directed graph G = (V, E, W) where nodes in V
represent concepts of the hyponymy and edges in E represent relationships between
concepts. A directed edge (c1, c2) from c1 to c2 corresponds to a hyponymy from con-
cept c1 to concept c2. Edge weights in E are used to represent varying degrees of cer-
tainty.
For a node c in a graph, we denote by I(c) and O(c) the set of in-neighbors and out-
neighbors of v, respectively. Individual in-neighbors are denoted as Ii(c), for <=i<=
|I(c)|, and individual out-neighbors are denoted as Oi(c), for 1 <= i <= |O(c)|.
The basic control process about building concept space is shown in Algorithm 1.
------------------------------------------------------------------------------------------------
Algorithm 1. The basic process of building concept space
Input: the set of candidate hyponymy relations CHR from large Chinese free text
based on Chinese lexico-syntactic patterns;
Output: the concept space G.
Step1: Initialize G = (V, E, W), let V=∅, E=∅, W=∅;
Step2: For each (c1,c2)∈CHR, continue Step3-Step4 ；
Step3: If c1∉V,c2∉V, then V=VU{c1,c2}; E=EU{(c1,c2)}; If c1∉V, c2∈V, then V=VU{c1};
E=EU{(c1,c2)}; If c1∈V, c2∉V, then V=VU{c2}; E=EU{(c1,c2)};
Step4: CHR= CHR-{(c1,c2)};
Step5: For each r∈E, set its w(r) ∈W to be 0.
Step6: return G ；
----------------------------------------------------------------------------------------------------------

4 Space Structure Features

When a group of candidate hyponymy relations are correct or error, they often satisfy
some space structure feature. If a candidate hyponymy satisfies a certain threshold
with matching those features, we think that it is a real hyponymy.
We analysed the influence of the space structure on verification of hyponymy rela-
tions. In space structure analysis, we use the coordinate relation between concepts.
The coordinate relations are acquired using a set of coordinate relation patterns in-
、 、
cluding “ ”. Chinese dunhao “ ” is a special kind of Chinese comma used to set off
items in a series. For example:
In a sentence of matching a coordinate pattern, if there exists concept c1 and con-
、
cept c2 divided by “ ”, then c1 and c2 are coordinate, denoted as cr(c1, c2). An exam-
ple is as shown below.
农作物主要有 水稻 、 玉米 、 红薯 、 烟叶 等
{ }c1 { }c2 { }c3 { } c4
(The farm crop mainly includes paddy rice, corn, sweet potato, tobacco leaves
etc..)
cr(水稻 玉米 红薯 烟叶
, , , ) (cr(paddy rice, corn, sweet potato, tobacco leaves) ) is
acquired from the above example.
 Using Concept Space to Verify Hyponymy in Building a Hyponymy Lexicon 483

Fig. 2 depicts a few typical structure features of hyponymy.

(a) (b) (c) (d)

(e) (f)

Fig. 2. Typical structure features of hyponymy

Structure (a): (c1, c2), (c2, c3).

For example:
( 苹果 水果 水果, 食品) ((apple, fruit), (fruit, foodstuff))
, ), (
Structure (b): (c1, c2), (c2, c3), (c3, c1).
For example:
(游戏 生活 生活 童话 童话 游戏
, ), ( , ), ( , )
((game, life), (life, fairy tale), (fairy tale, game))
Structure (c): (c1, c2), (c2, c3), (c1, c3).
For example:
( 番茄 蔬菜 番茄 食品 蔬菜 食品
, ), ( , ), ( , )
((tomato, vegetable), (tomato, foodstuff), (vegetable, foodstuff))
Structure (d): (c1, c), (c2, c), …, (cm, c), cr(c1, c2, …, cm),
(c’1, c), (c’2, c), …, (c’n, c), cr(c’1, c’2, …, c’n),
{c1, c2, …, cm}∩{c’1, c’2, …, c’n}≠∅.
For example:
职业 时装 休闲装 礼服 裤装 时装 休闲服
c= , cr( , , ), cr( , , ) {c1,…, cm}∩{c’1,…, c’n}
={ 时装 休闲装
, }
(c= clothing, cr(fashionable dress, sportswear, full dress), cr(trousers,
fashionable dress, sportswear), {c1,…, cm}∩{c’1, …, c’n} = {fashionable dress,
sportswear})
Structure (e): (c1, c), (c2, c), …, (cm, c), (c’1, c’), (c’2, c’), …, (c’n, c’), {c1, c2, …,
cm}∩{c’1, c’2, …, c’n}≠∅.
For example:
484 L. Liu et al.

c=食品, {c ,…, c }={牛肉饼, 蛋糕, 面包, 奶油}, c’=产品, {c’ ,…, c’ }={牛肉
饼, 牛肉干, 牛肉汤}, {c ,…, c }∩{c’ ,…, c’ } = {牛肉饼}
1 m 1 n

1 m 1 n
(c= foodstuff, {c1,…, cm}={hamburger, cake, bread, butter}, c’=product, {c’1,…,
c’n}={hamburger, beef jerky, brewis}, {c1,…, cm}∩{c’1, …, c’n} = {hamburger})
Structure (f): (c, c1,), (c, c2), …, (c, cm), (c, c’1), (c, c’2), …, (c, c’n) {c1, c2, …,
cm}∩{c’1, c’2, …, c’n}≠∅.
For example:
西 红柿
c= , {c1,…, cm}={ 植物 蔬菜 食品 果 实 茄子
, , , }, c’= , {c’1,…,
c’n}={ 蔬菜 食品 食材
, , 蔬菜 食品
}, {c1,…, cm}∩{c’1,…, c’n} = { , }
(c=tomato, {c1,…, cm}={plant, vegetable, foodstuff, fruit}, c’= aubergine,
{c’1,…, c’n}={vegetable, foodstuff, food for cooking}, {c1,…, cm}∩{c’1, …, c’n} =
{vegetable, foodstuff})
The structure features of hyponymy are converted into a set of production rules
used in uncertainty reasoning. We use CF (certainty factors) that is the most common
approach in rule-based expert system. The CF formula is as follows:
⎧ P ( C H R |f ) -P ( C H R )
⎪ , P ( C H R |f ) ≥ P ( C H R )
⎪ 1 − P (C R )
C F (C H R , f ) = ⎨ (1 )
⎪ P ( C H R |f ) -P ( C H R ) , P ( C H R |f ) < P ( C H R )
⎪⎩ P (C H R )

Where CHR is a set of candidate hyponymy, which has a precision P(CHR).

P(CHR|f) is the precision of a subset of CHR satisfying feature f. CF is a number in
the range from -1 to 1. If there exists CF(CHR, f)≥0, then we denote f as positive
feature and CF(CHR, f) denotes the support degree of feature f; if there exists
CF(CHR, f)<0, then we denote f as negative feature and CF(CHR, f) denotes the no
support degree of feature f.
For example, P(CHR)=0.71, the precision of candidate hyponymy relations satisfy-
ing the structure feature (b) is 23%, namely P(CHR|f)=0.23, the result of CF is (0.23-
0.74)/ 0.74= -0.69. The f is a negative feature.
The precision of candidate hyponymy relations satisfying the structure feature (c)
is 96%, namely P(CHR|fb)=0.94, the result of CF is (0.96-0.71)/(1- 0.71)=0.86. The f
is a positive feature.
After those structrue features are converted into a set of production rules, we can
carry uncertainty reasoning in concept space.

5 Evaluation
5.1 Evaluation Method

We adopt three kinds of measures: R (Recall), P (Precision), and F (F-measure). They

are typically used in information retrieval.
Let h be the total number of correct hyponymy relations in the CHR.
Let h1 be the total number of hyponymy relations in the classify set.
Let h2 be the total number of correct hyponymy relations in the classified set.
 Using Concept Space to Verify Hyponymy in Building a Hyponymy Lexicon 485

(1) Recall is the ratio of h2 to h, i.e. R = h2/h

(2) Precision is the ratio of h2 to h1, i.e. P = h2/h1
(3) F-measure is the harmonic mean of precision and recall. It is high when both
precision and recall are high. F = 2RP/(R+P)

5.2 Experimental Results

We used about 8GB of raw corpus from the Chinese Web pages. Raw corpus is
preprocessed in a few steps, including word segmentation, part of speech tagging, and
splitting sentences according to periods. Then we acquired candidate hyponymy
relations CHR from processed corpus by matching Chinese hyponymy patterns. For
analyzing the influence of threshold a, we choose several different values. We
manually evaluated 10% initial set CHR and 10% final result. The detailed result is
shown in Table 1.

Table 1. The result of Verification

Concept space: 62,265 hyponymy relations P(CHR)=71%

The result of verified hyponymy
α The number P R F
α=0 56,155(90.2%) 76% 97% 0.85
α=0.2 46,574(74.8%) 82% 86% 0.84
α=0.4 34,365(55.2%) 87% 68% 0.76
α=0.6 20,102(32.3%) 92% 42% 0.58
α=0.8 12,510(20.1%) 95% 27% 0.42

As we can see from table 1, there are 62,265 hyponymy relations in concept space
initially. With the increase of threshold α, the precision is also increase. If we want to
increase the precision, we can augment γ value. For example, when α=0.8, the preci-
sion is up to 95%, and but its recall decreases to 27%. That is to say, when threshold
α is a small value, our methods can throw away many error hyponymy relations under
the condition of skipping a few correct relations. But when threshold α is a large
value, our methods can throw away many error hyponymy relations and also skip
many correct relations at same time.

6 Conclusion

For the verification of candidate hyponymy relations, we present a method of hy-

ponymy verification based on concept space. Experimental results demonstrate good
performance of the method. It will raise the precision of hyponymy relations and
benefit to the building of ontologies, knowledge bases and lexicons.
There are still some inaccurate relations in the result. In future, we will combine
some methods (such as web page tag etc.) to the further verification of hyponymy.
486 L. Liu et al.

Acknowledgments. This work is supported by the National Natural Science Founda-

tion of China under Grant No.60573064, and 60773059; the National 863 Program
under Grant No. 2007AA01Z325, and the Beijing University of Technology Science
Foundation (grant nos. X0006014200803, 97006017200701).

References
1. Beeferman, D.: Lexical discovery with an enriched semantic network. In: Proceedings of
the Workshop on Applications of WordNet in Natural Language Processing Systems,
ACL/COLING, pp. 358–364 (1998)
2. Cao, C., Shi, Q.: Acquiring Chinese Historical Knowledge from Encyclopedic Texts. In:
Proceedings of the International Conference for Young Computer Scientists, pp. 1194–1198
(2001)
3. Hearst, M.A.: Automated Discovery of WordNet Relations. In: Fellbaum, C. (ed.) To Ap-
pear in WordNet: An Electronic Lexical Database and Some of its Applications, pp. 131–
153. MIT Press, Cambridge (1998)
4. Morin, E., Jacquemin, C.: Projecting corpus-based semantic links on a thesaurus. In: Pro-
ceedings of the 37th Annual Meeting of the Association for Computational Linguistics, pp.
389–396 (1999)
5. Lloréns, J., Astudillo, H.: Automatic generation of hierarchical taxonomies from free text
using linguistic algorithms. In: Advances in Object-Oriented Information Systems, OOIS
2002 Workshops, Montpellier, France, pp. 74–83 (2002)
6. Sánchez, D., Moreno, A.: Pattern-ed automatic taxonomy learning from the Web. AI Com-
munications 21(3), 27–48 (2008)
7. Elghamry, K.: Using the Web in Building a Corpus-Based Hypernymy-Hyponymy Lexicon
with Hierarchical Structure for Arabic. Faculty of Computers and Information, 157–165
(2008)
8. Zhang, C.-x., Hao, T.-y.: The State of the Art and Difficulties in Automatic Chinese Word
Segmentation. Journal of System simulation 17(1), 138–143 (2005)
 Advanced Self-adaptation Learning and Inference
Techniques for Fuzzy Petri Net Expert System Units∗

Zipeng Zhang1, Shuqing Wang1, and Xiaohui Yuan2

1
Hubei University of Technology, Wuhan, 430068, China
2
Huazhong University of Science and Technology, Wuhan, 430074, China
zhzip1@163.com

Abstract. In a complicated expert reasoning system, it is inefficient for com-

monly fuzzy production rules to depict the vague and modified knowledge.
Fuzzy Petri nets are more accurate for dynamic knowledge proposition in de-
scribing expert knowledge. However, the bad learning ability of fuzzy Petri net
constrains its application in dynamic knowledge expert system. In this paper, an
advanced self-adaptation learning way based on error back-propagation is pro-
posed to train parameters of fuzzy production rules in fuzzy Petri net. In order
to enhance reasoning and learning efficiency, fuzzy Petri net is transformed into
hierarchy model and continuous functions are built to approximate transition
firing and fuzzy reasoning. Simulation results show that the designed advanced
learning way can make rule parameters arrive at optimization rapidly. These
techniques used in this paper are quite effective and can be applied to most
practical Petri net models and fuzzy expert systems.

Keywords: Fuzzy Petri net, dynamic fuzzy reasoning, neural network,

self-adaptation learning.

1 Introduction

In intelligent diagnoses system, since most knowledge is fuzzy and general logic rules
cannot describe it effectively. Fuzzy Petri net (FPN), which is based on fuzzy produc-
tion rules (FPRs), can provide us a useful way to properly represent fuzzy knowledge
[1], [2]. Many results prove that FPN is suitable to represent and reason misty logic
implication relations [3],[4].
However, Knowledge in expert systems is updated or modified frequently and ex-
pert systems may be regarded as dynamic systems. The models must have the ability
to adjust themselves according to the systems’ changes. But general FPN which lacks
learning ability cannot cope with potential changes of actual systems[5]. Artifical
neural network(ANN) has strong adaptation and learning ability[6]. It is effective to
combine the superiority of ANN into FPN to make FPN has self-adaptation learning
ability to adapt dynamic expert systems.
∗
This work is supported by NNSFC Grant 50539140 & 50779020 and NSF of Hubei Grant
2008CDB395.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 487–496, 2009.
© Springer-Verlag Berlin Heidelberg 2009
488 Z. Zhang, S. Wang, and X. Yuan

In order to overcome the disadvantage of learning speed slow and local optimiza-
tion in general back-propagation learning way, adaptive learning techniques are used
to learn and train parameters of FPRs in FPN. After a training process, an excellent
input-output map of the knowledge system can be gained.

2 Fuzzy Petri Nets

It is difficult to directly monitor the contamination degree of insulator. Here, an indi-
rect measurement method is used to diagnose insulator running state. In the diagnos-
ing, the easy measured leakage current, pulse frequency, the equivalent environment
humidity are used to analyse and diagnose insulator running state[8]. Those signals
are measured with on-line mode and then are used to analysis with off-line mode. The
correlative data may be analysed or processed and display through graphics in the
back computer.
FPRs may depict the fuzzy relationships between many propositions. Let R be a set
of FPRs, R = {R1 , R2 , L Rn } . The general formulation of the jth layer is given as fol-

lows: R j : IF d THEN d j +1 (CF = µ j )

j

and d j +1 are propositions, the truth of each proposition is a real value

j
Where, d
between zero and one. µ j is the value of the certainty factor(CF), µ j ∈ [0,1] [7].
If the antecedent portion or consequence portion of a FPR contains “and” or “or”
connectors, then it is called a composite FPRs[8],[9]. The structure of composite
FPRs is shown in Fig. 1.

j
p1 t11 λ11
d1j p j +1
µ 1j
p 2j µ 2j
d 2j ＋
t 2j λ 2j d j 1

p 3j
d 3j

Fig. 1. Structure a composite fuzzy production rule

A FPN is a bipartite directed graph containing two types of nodes: places and tran-
sitions, where circles represent places and bars represent transitions. Every FPR may
be expressed as a transition of FPN. The proposition of production rule is expressed
as place relatively. In a FPN, the relationships from places to transitions and from
transitions to places are represented by directed arcs.
Definition 1: A FPN has n layers noses and every layer has different places and transi-
tions. The FPN may be described using a 10-tuple[10],[11].
FPN = ( P, T , D, I , O, f , M ,W , Th, β ) (1)
 Advanced Self-adaptation Learning and Inference Techniques 489

Where, P = {p11,L, p1a1 , p12,L, pa22 ,L, p1n,L, pann } is a finite set of places.
T = {t11 ,L, tb11 , t12 ,L, tb22 ,L, t1n ,L, tbn
n
} is a finite set of transitions.
D = {d11,L, d a11 , d12 ,L, d a22 ,L, d1n ,L, d ann } is a finite set of propositions ,and
P IT I D = Φ , P = D .
：
I T → P ∞ is the input function, a mapping from transitions to bags of places,
which determines the input places to a transition, and P ∞ denotes bags of places.
：
O T → P ∞ is the output function, a mapping from transitions to bags of places,
which determines the output places form a transition.
：
f T → [0,1] is an association function, a mapping from transitions to real values
between 0 and 1, f (tbj ) = µbj .
M : P → [0,1] is an association function, every place pi ∈ P has a sign m( pi ) ,
which replaces the true degree.
W = {ω11 ,L, ωb11 , ω12 ,L, ωb22 ,L, ω1n ,L, ωbnn } is a set of input weights and output
weights which assign weights to all the arcs of a net.
（ ）{ }
Th t b = λ11 ,L, λ1b1 , λ12 ,L, λb22 ,L, λ1n ,L, λnbn is a set of threshold value from zero
j

：
to one to transition .
β P → D is an association function, a relationship mapping from places to
propositions.

3 Dynamic Reasoning Process of FPN

3.1 Fuzzy Reasoning Algorithm

Firstly, some basic definitions are given to explain the transition firing rule of FPN.

（ ）
Definition2: if ∀pij ∈ I (t j ) , m(pij ) > 0 i =1,2,L,a j = 1,2,L n then ∀t ∈ T , t is enabled.

Definition3: ∀t j ∈T , ∀p j ∈ I (t j ) ,if
i
∑m( pij ) ⋅ωij > Th(t j ) then transition fires, at the same
i
time, token transmission takes place.

Definition4: When transition is enabled, it produces a new certainty factor CF(t).

⎧∑m( pij ) ⋅ ωij , ∑m( pij ) ⋅ ωij > Th(t j )

⎪i
CF(t j ) = ⎨
i
(2)
⎪0 , ∑m( pij ) ⋅ ωij < Th(t j )
⎩ i

j
Sigmoid function F(x) is used to approximates the threshed of t .
490 Z. Zhang, S. Wang, and X. Yuan

F ( x ) = 1 /(1 + e − b ( x − Th (t
j
)) (3)
)

Where, x = ∑ m( pij ) ⋅ ωij , b is instant. When b is chosen properly, if x > Th(t j ) , then
i
− b ( x −Th (t j ))
e ≈ 0 and F ( x) = 1 ,which denotes transition t is enable. In the other hand,
if x < Th(t j ) , then e − b ( x −Th ( t )) ≈ 1 and F ( x) = 0 , which denotes transition t is not
j

enable and the sign of output place is 0.

3.2 Making Continuous Function of Fuzzy Reasoning

In order to implement learning and fuzzy reasoning effectively, continuous function is

used in FPN. Where, CF(t)(x) is utilized to approximate the new certainty factor of
output place.
CF (t )( x) = x ⋅ F ( x) (4)

Where, x = ∑ m( pij ) ⋅ ωi j , F ( x ) = 1 /(1 + e − b ( x − Th ( t

j
))
)
i

Definition 5: If transition fires, the token of input places does not vary and output
place produces new token.
1) If a place only has one input transition, a new token with certainty factor CF is
put into each output place, new token sign is given as:

）
m( p j +1 = f (t j ) × CF (t j )( x j ) (5)

）
Where pij ∈ I (t j , i = 1,2 L m, p j +1 ∈ O(t j ）
2) If a place has more than one input transitions t z ( z = 1,2, L , c) and more than
j

one route are active at the same time, then the new certainty factor is decided by the
maximal passed sign of the fired transitions:

）
m( p j +1 = max( f (t1j ) × CF (t1j )( x1j ), f (t 2j ) × CF(t 2j )( x2j ),
(6)
L, f (tcj ) × CF(tcj )(xcj ))

3.3 Process of Fuzzy Reasoning

Definition 6 (Source Places, Sink Places): A place p is called a source place if it has
no input transitions. It is called a sink place if it has no output transitions. A source
place corresponds to a precondition proposition in FPN, and a sink place corresponds
to a consequent.
Definition 7(input place, output place): The set of places P is divided into three
parts P = PUI U Pint U PO , where P is the set of places of FPN, PUI ={p∈P|.p = Φ} ,
p ∈ PUI is called a user input place; Pint = { p ∈ P|. p ≠ Φ and p. ≠ Φ} , p ∈ Pint is
 Advanced Self-adaptation Learning and Inference Techniques 491

called an interior place; PO ={p. =Φ}, p ∈ PO is called an output place. Where, Φ is an

empty set.
Definition 8(initially enabled transition, current enabled transition.): Let
Tinitital = {t ∈ T |.t I PUI ≠ Φ and .t I Pint = Φ}, t ∈ Tint itial is called an initially en-
abled transition. Let Tcurremnt={t ∈Tinitial| ∀pi∈. t,m( pi ) > 0, CF (t ) > Th(t )}, t ∈ Tcurrent is
called a current enabled transition [10],[11].
In order to describe the reasoning process of FPN, an FPN example is given here and
its structure is shown in Fig.2. The reasoning system contains three layers.
d11 , d 21 , d 12 , d 22 , d 32 , d 13 , d 23 , d 33 , d 43 , d 4 are related propositions of the expert system.
Between them there exist the following weighted FPRs.
R1 : IF d11 or d 21 THEN d 12 ( λ11 , µ11 , λ12 , µ12 )
R12 : IF d12 and d 22 THEN d 23 ( λ21 , µ 12 , ω 112 , ω122 )
R22 : IF d12 and d32 THEN d 43 ( λ22 , µ22 , ω 221 , ω22
2
)
R 3 : IF d13 and d 23 and d33 THEN
（，， ， ， ）
d 4 λ3 µ3 ω31 ω32 ω33

p32 3
ω 212 t 22 λ22 µ 2 p4
2
p11 t λ1
1 1 d 32
1
ω 222 d 43
µ 11 2
p1 p 3
d11 1
ω13
2 ω 11
p 1
t λ2 µ1
1 1 2
d13
2 2
2 d1 t λ
2 2
p23 t 3 λ3 p4
1 1 2
1µ ω 3
2
µ 3

d 21 p22 ω 122 d 3
2

d 2 p33 ω 33
2
d 33

Fig. 2. Structure of the conceived FPN

The conceived FPN according above FPRs is described as follow.

FPN = ( P, T , D, I , O, f , M ,W , Th, β )
In this FPN, p11 , p12 , p22 , p32 , p13 , p33 are initial places, p43 and p 4 are output places.
The first layer transition set contains t11 , t12 , the second layer transition set con-
tains t12 , t 22 and the third layer transition set contains t 3 .The process of fuzzy reasoning
is given as follow.
492 Z. Zhang, S. Wang, and X. Yuan

1) If m( p11 ) > λ11 then t11 fire, or if m( p12 ) > λ12 then t12 fire and then the true degree
of p12 may be calculated using follow formula.

）
m( p12 = max( f (t11 ) × x11 × F ( x11 ),
(7)
f (t12 ) × x12 × F ( x12 ))

Where, x11 = m( p11 ) , x12 = m( p12 ) , F(x) =1/(1+e−b(x−Th(t))) .

2) The second layer transitions fire, if t12 , t 22 fires singly, the true degree of p23 , p43
may be calculated singly through follow formula.

＝
m( p23 ) x22 × F ( x22 ) (8)

m ( p4 3 )＝x × F ( x )
2
4
2
4 (9)

Where x22 = m( p12 ) ⋅ ω11

2
+ ( p22 ) ⋅ ω12
2
, ω + ω ＝1 ,
2
11
2
12

F ( x22 ) = µ12 /(1 + e−b( x2 − λ1 ) ) , x42 = m( p32) ⋅ω21

2 2
2
+( p12) ⋅ω22
2
,
ω21
2
+ ω22
2
＝
1 , F ( x42 ) = µ 22 /(1 + e −b ( x 4 − λ 2 ) ) .
2 2

4
3) The last layer transitions fire, the true degree of p may be calculated through
follow formula.

＝
m( p 4 ) x 3 × F ( x 3 ) (10)

Where x 3 = m( p13 ) ⋅ ω13 + m( p23 ) ⋅ ω23 + m( p33 ) ⋅ ω33 ,

＝
ω13 + ω23 + ω33 1 , F ( x 3 ) = µ 3 /(1 + e −b ( x
3
− λ3 )
).

4 FPN Learning and Training

4.1 Translating into Neural Networks Structure

ANN has strong learning and adaptive ability. When weights can not be ascertained
via expert knowledge, FPN structure may be translated further into a neural networks-
like structure. The back-propagation way with multi-layered feed-forward networks
of ANN is used to training FPN model. The trained weight threshold value, and true
degree of FPRs may be used to analyze fuzzy reasoning system.
Supposing there are n layers networks, if t j is the jth layer transition of FPN, the
weights of input arc are supposed as ω1 j , ω 2 j L , ω m j , the threshold value of t j is
λ j and the true degree is µ j .the network output of nth layer node is expressed as:
 Advanced Self-adaptation Learning and Inference Techniques 493

G ( x j ) := F ( x j ) ⋅ x j (11)

m
Where, x j = W T × M = ∑ m( pi j )ωi j , j = 1,2 L , n ,
i =1
−b ( x j − λ j )
F ( x ) = µ /(1 + e
j j
) , W = [ω 1j , ω 2j , Lω mj ] ,
M = [m( p1j ), m( p 2j ), L m( p mj )] .
The last layer output may be computed via the last node of network.

O( R) = G ( x n ) = M ( p n ) (12)

4.2 Improved Learning Algorithm

The error function E is defined as:

E ＝ 12 ∑ ∑ (O(R) − O )
r b

l =1 o =1
* 2
(13)

Where, r is number of samples and b is output places. The learn ways are given as
follow.
dE
ωi j (k + 1) = ωi j (k ) − η , i = 1,2Lm − 1, j = 1,2Ln (14)
dωi j

m −1
ωm j ( k + 1) = 1 − ∑ ωi j (k + 1) (15)
i =1

dE
Where, may be computed through the follow way.
dωi
j

If the FPN model has n layers, the last layer weight varying dE may be com-
dω i
n

puted as follow.
dE dE dO ( R ) dE dG ( x n )
= × = ×
dω i n dO ( R ) dω i n dG ( x n ) dω i n
r b dG ( x n ) dx n (16)
= ∑ ∑ (O ( R ) − O * ) ×
l =1 o =1 dx n d ω i n
n n
Where, dG(x ) and dx may be calculated using follow formula:
dxn dω i n

µ n x nbe −b ( x − λ )
n n
dG ( x n ) µn
= + (17)
1 + e − b ( x − λ ) (1 + e − b ( x − λ ) ) 2
n n n n
dx n
494 Z. Zhang, S. Wang, and X. Yuan

dx n
= m ( pi n ) (18)
dωi n

The error term of the other layers may be computed through the same back-
propagation method.
In training, the choice of the learning rate η has an important effect on weights
convergence. If it is set too small, too many steps are needed to reach an acceptable
solution. On the contrary, a large learning rate will possibly lead to oscillation, pre-
venting the error to fall below a certain value. To overcome the inherent disadvan-
tages of pure gradient-descent, an adaptation way of the weight-updates according to
the behavior of the error function is applied to training weights.
Here, Individual update-value ∆ i (k ) is introduced for the weight-updating. The
adaptive update-value ∆ i (k ) evolves during the learning process based on its local
dE dE
(k − 1) . ∆ωi (k ) and ωi (k + 1) are
j j
sight on the error function (k ) and
dωi dωi j
j

obtained according to the following formula.

dE
∆ω i j ( k ) = − (k ) ∗ ∆ i j (k ) (19)
dω i j

ωi j (k + 1) = ωi j (k ) + ∆ωi j (k ) (20)

At the beginning, initial value ∆ i (0) should be set to all update-values ∆ i j (k ) . For
j

∆ i (0) directly determines the size of the first weight-step, it is preferably chosen in a
j

reasonably proportion to the size of the initial weights. Where, ∆ i j (0) = 0.19 is very
good.

5 Training Example
In this section, the above given fuzzy expert reasoning system (shown as Fig.2) is
used as example to illustrate the learning effect of above introduced way.
In the example, assume the ideal parameters are given as:

＝
λ11 = λ12 = λ12 = λ22 λ3 = 0.5 , µ11 = 0.8 , µ 12 = 0.9 , µ12 = 0.9 , µ 22 = 0.8 , µ 3 = 0.9 ,
ω 11
2
＝ 0.35 , ω12
2
= 0.65 , ω21
2
= 0.25 , ω22
2
= 0.75 , ω 13 = 0.15 , ω 32 = 0.55 , ω 33 = 0.3 .

If weights are unknown, neural networks technique are used to estimate the weights.
The learning part of the FPN may be formed as a standard sub single layer ANN and a
two layers sub neural network. Fig.3 and Fig.4 show the perfection of adaptation
learning techniques.
 Advanced Self-adaptation Learning and Inference Techniques 495

From the simulation example, we can see that the fuzzy reasoning algorithm and
the ANN training algorithm are very effectively when weights of FPN are not known.
Fig.3 and Fig.4 show that self-adaptation learning way is very quick to find optimiza-
tion weights to make error function E to least. Self-adaptation learning way over-
comes the disadvantage of learning speed slow and local optimization in common
learning way.

ω222

ω112

Fig. 3. Training perfection using adaptation way

ω122
ω 112 ω 32

ω 33
ω 13

Fig. 4. Training perfection using adaptation way

6 Conclusion
In expert system, the fuzzy reasoning process based on FPRs is difficult to obtain ef-
fective knowledge. In order to produce effective reasoning and learning ability, FPN
without loop is transformed into hierarchy model in this paper. Through building con-
tinuous functions, the approximate transition firing and fuzzy reasoning may be ac-
quire effectively, which provides powerful facility for FPN to reason and learn both
forward and backward. In training, self-adaptation learning way has been used in learn-
ing and training the parameters of FPN. Simulation results show that designed self-
adaptation learning way is very quick to find optimal weights. Therefore the designed
reasoning and learning way of FPN is effective for dynamic knowledge inference and
learning in expert system.
496 Z. Zhang, S. Wang, and X. Yuan

References
[1] Chen, S.M., Ke, J.S., Chang, J.F.: Knowledge representation using fuzzy Petri nets. IEEE
Trans. Knowledge Data Engineering 2, 311–319 (1990)
[2] Cao, T., Sanderson, A.C.: Representation and Analysis of Uncertainty Using Fuzzy Petri
Nets. Fuzzy System 3, 3–19 (1995)
[3] Lee, J., Liu, K.F.R., Chiang, W.: A Fuzzy Petri Net-Based Expert System and Its Appli-
cation to Damage Assessment of Bridges. IEEE Transactions on Systems, Man, and Cy-
bernetics -Part B: Cybernetics 29(3), 350–369 (1999)
[4] Scarpelli, H., Gomide, F., Yager, R.R.: A Reasoning Algorithm for High-level Fuzzy
Petri Nets. IEEE Trans. Fuzzy System 4(3), 282–293 (1996)
[5] Li, X., Lara-Rosano, F.: Adaptive Fuzzy Petri Nets for Dynamic Knowledge Representa-
tion and Inference. Expert Systems with Applications 19, 235–241 (2000)
[6] Zengren, Y.: Artificial Neural Network Application. Qinghua University Press (1999)
[7] Bugarn, A.J., Barro, S.: Fuzzy reasoning supported by Petri nets. IEEE Trans. Fuzzy Sys-
tem 2(2), 135–150 (1994)
[8] Pedrycz, W., Gomide, F.: A Generalized Fuzzy Petri Nets Model. IEEE Trans. Fuzzy
System 2, 295–301 (1994)
[9] Li, X., Yu, W.: Object Oriented Fuzzy Petri Net for Complex Knowledge System Model-
ing. In: IEEE international conference on control applications, September 2001, pp. 476–
481 (2001)
[10] Yeung, D.S., Tsang, E.C.C.: A Multilevel Weighted Fuzzy Reasoning Algorithm for Ex-
pert Systems. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and
Humans 28(2), 149–158 (1998)
[11] Scarepelli, H., Gomide, F.: A High Level Net Approach for Discovering Potential Incon-
sistencies in Fuzzy Knowledge Bases. Fuzzy Sets System 62(2), 175–193 (1994)
 MIMO Instantaneous Blind Identification Based on
Second-Order Temporal Structure and Newton’s Method

Xizhong Shen1,2,* and Guang Meng2

1
Mechanical & Electrical department, Shanghai Institute of Technology, Shanghai,
China 200235
2
State Key Laboratory for Mechanical System and Vibration, Shanghai Jiao Tong
University, Shanghai, China 200030
shenxizhong@online.sh.cn, gmeng@sjtu.edu.cn

Abstract. This paper presents a new MIMO instantaneous blind identification

algorithm based on second order temporal property and Newton’s Method.
Second order temporal structure is reformulated in a particular way such that
each column of the unknown mixing matrix satisfies a system of nonlinear mul-
tivariate homogeneous polynomial equations. The nonlinear system is solved
by Newton’s method for the equations. Our algorithm allows estimating the
mixing matrix for scenarios with 4 sources and 3 sensors, etc. Simulations and
comparisons show its effective with more accurate solutions than the algorithm
with homotopy method.

Keywords: instantaneous blind identification, homogeneous system, second

order temporal statistics, Newton’s method, nonlinear.

1 Introduction
Blind signal processing (BSP) is an important task for numerous applications such as
speech separation, dereverberation, communications, signal processing and control,
etc. Its research has obtained extensive interests all over the world and many exciting
results have been reported, [1] and references therein. Its task is to recover the source
signals only given the observations in the sense of some uncertain problems such as
scaling, permutation and/or delay.
Multiple-input multiple-output (MINO) instantaneous blind identification (MIBI)
is one of the attractive BSP problems, where a number of source signals are mixed by
an unknown MIMO instantaneous mixing system and only the mixed signals are
available, i.e., both the mixing system and the original source signals are unknown.
The goal of MIBI is to recover the instantaneous MIMO mixing system from the
observed mixtures of the source signals [1][2]. In this paper, we focus on developing
a new algorithm to solve the MIBI problem by using second-order statistics and New-
ton’s method.

*
Shen Xizhong, Ph.D, born in 1968, as a professor in Shanghai Institute of Technology, and
also as a visitor/post-doc in Shanghai Jiao Tong University. He is going on studying blind
signal processing, neural network, etc.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 497–506, 2009.
© Springer-Verlag Berlin Heidelberg 2009
498 X. Shen and G. Meng

Many researchers have investigated the use of second-order temporal structure

(SOTS) for MIBI, or BSP [1][3][4][5][6]. The greater majority of the available algo-
rithms are based on the generalized eigenvalue decomposition or joint approximate
diagonalization of two or more sensor correlation matrices for different lags and/or
times arranged in the conventional manner. An MIBI based on second order temporal
structure (SOTS) [2] has been proposed, which arrange the available sensor correla-
tion values in a particular fashion that allows a different and nature formulation of the
problem, as well as the estimation of the more columns than sensors. Our work is a
continuation and improvement of their work presented in [2].
In this paper, we further develop the algorithm proposed in [2] to obtain more ac-
curate and robust solution. At first, we exploit the SOTS by considering the sensor
correlation functions on a noise-free region of support (ROS). Then we project the
MIBI problem on the system of homogeneous polynomial equations of degree two.
At last Newton’s method is applied to estimate the columns of the mixing matrix,
which is different from the algorithm in [2] which applied homotopy method. The
MIBI method presented in this paper allows estimating the mixing matrix for several
underdetermined mixing scenarios with 4 sources and 3 sensors. Simulations and
comparisons show its effectivity with more accurate and robust solutions.

2 MIBI Model and Its Assumptions

Let us use the usual model [1][2] in MIBI problem as follows

x ( t ) = As ( t ) + ν ( t ) . (1)

where A = [ a1 , , am ] ∈ n×m
is an unknown mixing matrix with its

n -dimention array response vectors a j = ( a1 j anj ) , j = 1, 2,

T
,m ,
s ( t ) = ⎡⎣ s1 ( t ) , s2 ( t ) , , sm ( t ) ⎤⎦
T
is the vector of source signals,

ν ( t ) = ⎡⎣ν 1 ( t ) , ,ν n ( t ) ⎤⎦
T
is the vector of noises, and

x ( t ) = ⎡⎣ x1 ( t ) , x2 ( t ) , , xn ( t ) ⎤⎦ is the vector of observations.

T

Without knowing the source signals and the mixing matrix, the MIBI problem is to
identify the mixing matrix from the observations by estimating A as  , and if we
set the following linear transformation

y (t ) = A
ˆ + x (t ) . (2)

y ( t ) = ⎡⎣ y1 ( t ) , y2 ( t ) , , ym ( t ) ⎤⎦ is m -dimensional output vector at

T
where

time t , whose elements are the estimations of sources, and ˆ + is the Moore-Penrose
A
inverse of  , which is a demixing matrix named in BSP.
MIMO Instantaneous Blind Identification Based on Second-Order Temporal Structure 499

The mixing matrix is identifiable in the sense of two indeterminacies, which are
unknown permutation of indices of each column of the matrix and its unknown mag-
nitude. When signal si is multiplied by a scalar, this is equivalent to rescaling the
corresponding column of A by the scalar. Therefore, the scalar of each column
remains undetermined. The usual convention is to assume that each column satisfy
the normalization conditions, i.e., on the unit sphere,
n

∑a
i =1
ij
2
= 1, j = 1, 2, ,m . (3)

We set it as the following form for the usage in section IV,

S j ( a j ) = ∑ aij 2 − 1 = 0; j = 1, 2,
n
,m . (4)
i =1

It should be noted that m sources cannot be determined in their exact order. It is also
unknown which the first column of A is and which the second is, and thus its permu-
tation of indices of each column of the matrix is indeterminate.
To solve the MIBI problem, we first define the following concepts Def 1~2 for the
derivation of the algorithm, and then make the following assumptions AS 1~4 [2].

Definition 1. Autocorrelation function ru ( t ,τ ) of u ( t ) at time instant t and lag τ

is defined as

ru ( t ,τ ) E ⎡⎣u ( t ) u ( t − τ ) ⎤⎦ , ∀t ,τ ∈ . (5)

Definition 2. Cross-correlation function ru ,v ( t ,τ ) of u ( t ) , v ( t ) at time instant t

and lag τ is defined as

ru ,v ( t ,τ ) E ⎡⎣u ( t ) v ( t − τ ) ⎤⎦ , ∀t ,τ ∈ . (6)

AS 1. the source signals have zero cross-correlation on the noise-free ROS Ω:

rs , j1 j2 ( t ,τ ) = 0, ∀1 ≤ j1 ≠ j2 ≤ m . (7)

AS 2. the source autocorrelation functions are linearly independent on the noise-free

ROS Ω
m

∑ ξ r ( t ,τ ) = 0 ⇒ ξ
j =1
j s , jj j = 0, ∀j = 1, 2, ,m (8)

AS 3. the noise signals have zero auto- and cross- correlation functions on the noise-
free ROS Ω :

rn , j1 j2 ( t ,τ ) = 0, ∀1 ≤ j1 , j2 ≤ m . (9)
500 X. Shen and G. Meng

AS 4. the cross-correlation functions between the source and noise signals are zero on
the noise-free ROS Ω :

rν s ,ij ( t ,τ ) = rsν , ji ( t ,τ ) = 0, ∀1 ≤ i ≤ n,1 ≤ j ≤ m . (10)

Here, no assumptions are made on the mixing matrix, and even rank-deficient one can
be identified. This is a significant advantage with respect to other methods because of
the two reasons as follows: (1) most IBSS may not only require that the number of
sensors is larger than the number of sources, but also that the mixing matrix is full
rank; (2) no assumptions are made on the probability density functions of the noise
and source signals.
The procedure of our proposed algorithm includes two steps, that is, step 1 is that the
problem of MIBI is formulated as the problem of solving a system of homogeneous
polynomial equations; and step 2 is that Newton’s method is applied to solve the system
of polynomial equations. We detail these steps respectively in section III and IV.

3 Projection on Homogeneous Polynomial Equations

In this section, we will review the algebraic structure of MIBI problem derived under
the above assumptions, and some details can be referred to [2]. The correlation val-
ues of the observations are stacked as

R x ,◊ ⎡⎣rx ( t1 ,τ 1 ) rx ( t N ,τ N ) ⎤⎦ , (11)

where rx ( t N ,τ N ) = E ⎡⎣ x ( t N ) ⊗ x ( t N − τ N ) ⎤⎦ , and ⊗ denotes the Kronecker

product. Consider the eq. (1), we get

R x,◊ = A ◊ R s, . (12)

Here, A ◊ is the second-order Khatri-Rao product of A , which is defined as

A◊ [a1 ⊗ a1 a m ⊗ a m ] , and R s , ⎡⎣rs ( n1 , k1 ) rs ( nN , k N ) ⎤⎦ with
rs ( nq , kq ) E ⎡⎣s ( nq ) s ( nq − kq ) ⎤⎦ , q = 1, , N where denotes the elemen-
twise product. The homogeneous polynomial equations of degree two are expressed as

ΦA ◊ = 0 . (13)

Here, (
Φ = ϕq ,i1i2 ) q =1, ,Q ;i1 ,i2 =1, , n ;i1 ≤i2
is a matrix with

⎛1 ⎞
⎜ n ( n + 1) − rank ⎡⎣ R x ,◊ ⎤⎦ ⎟ × ( n ) dimensions, of which its rows form a basis for
2

⎝ 2 ⎠
MIMO Instantaneous Blind Identification Based on Second-Order Temporal Structure 501

the nonzero left null space Ν ( R x ,◊ ) . Therefore, there are Q equations about each
column of A in (13).
R x ,◊ is splitted into signal and noise subspace parts as

R x ,◊ = U s Σ s Vs T + Uν Σν Vν T . Φ can be obtained by SVD of R x ,◊ , that is, the

left null space of R x ,◊ is Φ = Uν T .
The zero contour level of a n -variate homogeneous function is a ( n − 1) -dimension
surface embedded in an n -dimension Euclidian space. In an n -dimension Euclidian
space, at least ( n − 1) surfaces of degree ( n − 1) are required to uniquely define one-
dimensional space intersections. The dimension of the solution set of a system of
( n − 1) homogeneous polynomials in n variables has one-dimension. By eq.(13), the
maximum number M max of columns that can be identified with n sensors equals

1
M max = n ( n + 1) − ( n − 1) . (14)
2
M max is the maximum number of the sources to be identified by SOTS.

4 Newton’s Method
In this section, we summarize the main ideas behind the so-called Newton’s method
that transforms a system of nonlinear equations into a convergent fixed-point problem
in a general way.
Newton’s method for nonlinear system is able to give quadratic convergence, pro-
vided that a sufficiently accurate starting value is known and the inversion of the
Jacobian matrix of the nonlinear equations exists [7]. Certainly, we could use the
quasi-Newton’s method to avoid the calculation of inversion; however we use New-
ton’s method in our algorithm for the number of equations and variables involved are
small and its inversion of the corresponding Jacobian matrix is simple.
We expand the expression in (13) as

fq (a j ) = ∑ ϕq;i i ai j ai j = 0;
12 1 2
i1 ≤i2 ;i1i2 =1, , n . (15)
q = 1, , Q; ∀j = 1, ,m

( )
F a j ( k ) = ⎡⎣ f1 ( a j ) f Q ( a j ) ⎤⎦ . By Newton’s method, for each
T
and set
column of the mixing matrix we have
502 X. Shen and G. Meng

( ) ( )
a j ( k +1) = a j ( k ) − J −1 a j ( k ) F a j ( k ) . (16)

where J ( a ) is the Jacobian matrix of F ( a ).

( ) k
j j

We employ the initial solutions as equal distributed vectors in the super sphere de-
fined on the super space of a j , for example, in our simulation of mixing matrix with
3 × 4 sizes,
⎡1 0 0 1⎤
A = ⎢⎢0 1 0 1⎥⎥ . (17)

⎣⎢0 0 1 1⎥⎦
Thus, the different solutions are easily obtained. Alternatively, to get all the solutions
of (16), we set the initial values given by 2 N +1 vectors with its entries random nor-
N +1
mally distributed values in [-1,1]. And then, after running the algorithm, we get 2
solutions and select the centers of all the solutions.
The procedure of Newton’s method for nonlinear equations can be described as
follows:
Step 1. Set initial approximation in (17) of an initial solutions
a j ( 0) , j = 1, 2, ,m;
Step 2. Set tolerance TOL and maximum number of iterations maxIteration;
Step 3. For each solution a j ( k ) , calculate a j ( k +1) by (16) until a j ( k +1) − a j ( k )
2
is less than a certain tolerance or the number of iterations reaches to the maximum
value maxIteration.

5 Examples with Speech and Three Sensors

To compare our proposed algorithm with the algorithm proposed in [2], we adopt
three mixtures of four speech signals the same example as in [2]. For convenience of
comparison, we name our improved algorithm in (16) as MIBI_NWT; and the algo-
rithm in [2] as MIBI_Homotopy. We also develop the algorithm by applying steepest
descent method [7] to solve the equations in (15), in which we construct objective
function as the sum of square of each function in (15). We name the algorithm with
steepest descent method as MIBI_SD, and it is not detailed here for we think it as an
intermediate result.
The speech source signals are sampled as 8kHz, consist of 10,000 samples with
1,250ms length, and are normalized to unit variance σ s = 1 . The signal sequences
are partitioned into five disjoint blocks consisting of 2000 samples, and for each
block, the one-dimensional sensor correlation functions are computed for lags 1, 2, 3,
4 and 5. Lag zeros are omitted because the corresponding correlation values are noise
contaminated. Hence, in total for each sensor correlation functions 25 values are
MIMO Instantaneous Blind Identification Based on Second-Order Temporal Structure 503

estimated and employed, i.e., the employed noise-free ROS in the domain of block-
lag pairs is given by

Ω = {(1,1) , , (1,5) , ( 2,1) , , ( 5,5)} ,

where the first index in each pair represents the block index and the second the lag
index. The sensor signals are obtained from (1) with 3 × 4 mixing matrix,

⎡0.6749 0.4082 0.8083 −0.1690 ⎤

A = ⎢⎢ 0.5808 −0.8165 0.1155 −0.5071⎥⎥ . (18)
⎢⎣0.4552 −0.4082 −0.5774 0.8452 ⎥⎦
The noise signals are mutually statistically independent white Gaussian noise se-
quences with variances σν 2 = 4 . We set the maximum iterative number is 20, and
stop the iteration step if the correction of the estimated is smaller than a certain toler-
ance 10-4.
The signal-to-noise ratio (SNR) is defined as the ratio of the variance of the noise-
free part of the sensor observation vector to the variance of the noise vector, i.e.,

∑σ i
si
2

SNR 10 log . (19)

∑ σνi
i
2

Hence, the SNR is -4.7881dB, which is quite bad.

Fig. 1 depicts the intersections of each zero contour level on the unit sphere. There
are two blue solid curves as the spherical ellipses (1 and 2 in the figure) correspond-
ing to the case q = 1 in eq.(15), and two magenta solid curves (3 and 4 in the figure)
to the caseq = 2 . There are also the column vectors of estimated mixing matrix Â
indicated by solid lines and dots, and the column vectors of A by dashed lines and
stars.
Let θ j , j = 1, 2,3, 4 be the included angle between the j -th column of A and
its estimate. The estimated mixing matrix is

⎡0.6347 0.3784 0.8379 −0.1198⎤

ˆ = ⎢ 0.5652 −0.8785 0.2034 −0.5111⎥ ,
A ⎢ ⎥
⎢⎣ 0.5270 −0.2915 −0.5066 0.8511 ⎥⎦
and the included angles are 4.8002, 7.7664, 6.6909 and 2.8494. We see that the
estimated columns approximately equal the ideal ones by comparison with the matrix
in (18).
504 X. Shen and G. Meng

1 1
2
3
0.5
4
A1
0 Estimated A 1
z

A2
Estimated A 2
-0.5
A3
Estimated A 3
-1 A4
-1 Estimated A 4

0.5 1
1 -0.5 0
x -1
y

Fig. 1. Zero contour spherical ellipses of q = 1, 2 , Q = 2

Solving a system consisting of trivariate homogeneous polynomial equations in

(13) of degree two is equivalent to finding the intersections between the correspond-
ing two-dimensional quadric cones embedded in a three-dimensional Euclidian space.
Because such cones can intersect in at most four directions, the maximum number of
columns that can be identified with three sensors equals four, and is equal to the
maximum number of different intersections between two spherical ellipses, depicted
in Fig. 1.
Fig.2 shows the Comparisons of MIBI_NWT with MIBI_Homotopy and
MIBI_SD. TL, TR, BL and BR in Fig.2 are respectively the estimated included an-
gles along different running times between the first, second, third and fourth columns
and their estimates. Red dot indicates the result of MIBI_NWT; magenta plus the
results of MIBI_SD and blue circle indicates the results of MIBI_homotopy. We run
them by 10 times with random generated noises each time, and use the same initial
N +1
values given by 2 vectors with its entries random normally distributed values in [-
1,1]. The parameters in each algorithm are adjusted for comparison to get the most
accurate solution. In MIBI_NWT, we set the maximum iterative number is 20, and
stop the iteration step if the correction of the estimated is smaller than a certain toler-
ance 10-4. In MIBI_homotopy, we set the start system as in [2], and λs = 0 ,
1
λe = 1 + j and ∆λ = (1 + j ) . From Fig.2, we see that the estimated included
10
angles by MIBI algorithm with Newton’s method are smaller than MIBI _homotopy
and MIBI_SD. Therefore, MIBI_NWT has better performance than MIBI_homotopy
and MIBI_SD.
MIMO Instantaneous Blind Identification Based on Second-Order Temporal Structure 505

Comparision of MIBI Newton method with MIBI Homotopy

80 (TL) 80 (TR)
Included Angle/° 60 60

40 40

20 20

0 0
2 4 6 8 10 2 4 6 8 10

80 (BL) 80 (BR)
Included Angle/°

60 60

40 40

20 20

0 0
2 4 6 8 10 2 4 6 8 10
Compared numbers Compared numbers

Fig. 2. Comparisons of MIBI_NWT with MIBI_SD and MIBI_Homotopy. TL, TR, BL and
BR are respectively the estimated included angles along different running times between the
first, second, third and fourth columns and their estimates. Red dot indicates the result of
MIBI_NWT; magenta plus the results of MIBI_SD and blue circle indicates the results of
MIBI_homotopy.

6 Conclusion
In this paper, we further develop the algorithm proposed in [2] to obtain more accu-
rate solution. The SOTS is considered only on a noise-free region of support (ROS).
We project the MIBI problem in (1) onto the nonlinear system of homogeneous poly-
nomial equations in (13) of degree two. The nonlinear system is solved by Newton’s
method for the equations, which is quite different from the algorithm in [2] which
applied homotopy method. Our algorithm allows estimating the mixing matrix for
scenarios with 4 sources and 3 sensors, etc. Simulations and comparisons show its
effective with more accurate solutions than the algorithm with homotopy method.
Acknowledgments. This paper is supported by the National Science Foundation of
China with the project number 10732060, and Shanghai Leading Academic Discipline
Project directed by Hu Dachao with the project number J51501, and also Shanghai
Education with No. ZX2006-01.

References
1. Cichocki, A., Amari, S.I.: Adaptive Blind Signal and Image Processing: Learning Algo-
rithms and Applications. Wiley, New York (2002)
2. van de Laar, J., Moonen, M., Sommen, P.C.W.: MIMO Instantaneous Blind Identification
Based on Second-Order Temporal Structure. IEEE Transactions on Signal Processing 56(9),
4354–4364 (2008)
506 X. Shen and G. Meng

3. Shen, X., Shi, X.: On-line Blind Equalization Algorithm of an FIR MIMO channel system
for Non-Stationary Signals. IEE Proceedings Vision, Image & Signal Processing 152(5),
575–581 (2005)
4. Shen, X., Haixiang, X., Cong, F., et al.: Blind Equalization Algorithm of FIR MIMO Sys-
tem in Frequency Domain. IEE Proceedings Vision, Image & Signal Processing 153(5),
703–710 (2006)
5. Hua, Y., Tugnait, J.K.: Blind identifiability of FIR-MIMO systems with colored input using
second order statistics. IEEE Signal Processing Letters 7(12), 348–350 (2000)
6. Lindgren, U., van der Veen, A.-J.: Source separation based on second order statistics—An
algebraic approach. In: Proc. IEEE SP Workshop Statistical Signal Array Processing, Corfu,
Greece, June 1996, pp. 324–327 (1996)
7. Burden, R.L., Faires, J.D.: Numerical Analysis, pp. 600–635. Thomson Learning, Inc.
(2001)
 Seed Point Detection of Multiple Cancers Based
on Empirical Domain Knowledge and K-means
in Ultrasound Breast Image

Lock-Jo Koo∗, Min-Suk Ko, Hee-Won Jo, Sang-Chul Park, and Gi-Nam Wang

Department of Industrial Engineering, Ajou University,

442-749, Suwon, Kyunggido, Korea
{lockjo9,sebaminsuk11,happy6654,scpark,gnwang}@ajou.ac.kr

Abstract. The objective of this paper is to remove noises of image based on the
heuristic noises filter and to automatically detect seed points of tumor region by
using K-MEANS in breast ultrasound. The proposed method is to use 4 differ-
ent kinds of process. First process is the pixel value which indicates the light
and shade of image is acquired as matrix type. Second process is an image pre-
processing phase that is aimed to maximize a contrast of image and to prevent a
leak of personal information. The next process is the heuristic noise filter which
is based on the opinion of medical specialist and it is applied to remove noises.
The last process is to detect a seed point automatically by applying K-MEANS
algorithm. As a result, the noise is effectively eliminated in all images and an
automated detection is possible by determining seed points on each tumor.

Keywords: Decision Support System, Breast ultrasound image, Seed points,

Computer Aided Diagnosis.

1 Introduction
Breast cancer has been becoming a growth rate of 2% in the world since 1990. Ac-
cording to the Ministry of Health & Welfare, South Korea and American Cancer
Society, breast cancer ranks second in the list of women’s cancers. Moreover, in
most countries of the world, it is on a higher position of women’s cancers [1]-[3].
Early detection of breast cancer is important for the identifiable treatment. How-
ever, the segmentation and interpretation of ultrasound image is based on the experi-
ence and knowledge of a medical specialist and speckles make a more difficult to
distinguish a ROI (Region Of Interests). As a result, more than 30% of masses
referred for surgical breast biopsy were actually malignant [4, 5].
CAD (Computer Aided Diagnosis) is defined as a diagnosis that is made by a
radiologist, who uses the output from a computerized analysis of medical images, as a
‘second opinion’ in detecting lesions and in making diagnostic decisions. Most of the
previous research related with CAD of breast cancer has been focused on segmentation
of lesion and eliminate noises on image. To remove noises, anisotropic diffusion filter,
∗
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 507–516, 2009.
© Springer-Verlag Berlin Heidelberg 2009
508 L.-J. Koo et al.

second order Butterworth filter, watershed transform etc. are used [4-7]. Each method
for removal is used by considering five features of ultrasound images which are con-
sisted area, circularity, protuberance, homogeneity, and acoustic shadow. For segmenta-
tion of lesion, algorithms based on region growing method are used because most
tumors are included in breast ultrasound images have a multiplicity of shapes & colors,
difference of pixel value between tumor and background image. According to the region
growing methods, in a case where an objective pixel and the adjacent pixels have an
identical feature using histogram, gradient based pixel value of image sequentially exe-
cuted processing integrates the pixels to form an area. Areas having the same feature are
gradually grown to achieve segmentation of the entire image. However, Most of the
previous researches have some limitation as follows.
1. The pixel value and tumor shape on ultrasound image are only considered for
noise removal, except for location of tumor on the image.
2. Only one ROI (Region of Interests) is segmented, or manual tasks are needed for
detection of plural tumors because one seed point is considered for region grow
method.

In this paper, the noise removal method based on the heuristic noise filtering related
to the lesion position on image and the detection method of one or more seed points
using K-MEANS algorithm is proposed. For accomplishment of this goal, the organi-
zation of paper is as follows. Section II is the introduction of the each step that is used
for noises removal. The detection of seed points in the breast ultrasound image is
present in Section III. Then the experimental results and conclusion are discussed in
Section IV.

2 Noise Removal

Our database consists of 40 consecutive ultrasound cases, being represented by 64

images. These images were obtained during diagnostic breast exams at the Park
Breast Clinic in Korea. Most of the cases were included either benignancy or malig-
nancy. In addition, the included information of these images is outpatient’s name,
distinction of sex, size of lesion and so on. Out of the images of 40 cases, 13 were
benign lesions; the other 27 were malignant lesions which have been chosen for the
purpose of testing. The size of the tumors were between 0.5~3cm.
In this section, methods for noises removal on this paper are briefly described by
step-by-step process. The proposed method is consisted of four steps. Each step is
explained as follows.
Before proceeding further, first, the mean (µ) variance (σ), minimum (Min) and
maximum (Max) pixel value are defined as follows. In Eq.(1), (2), (3), (4), I (i,j) de-
fines the i,jth pixel value the ni by nj matrix which is transformed from the ultrasound
image and this definition is applied to all equations of this paper.
ni nj

∑∑ I (i, j)
1
µ= (1)
ni n j i =1 j =1
 Seed Point Detection of Multiple Cancers Based on Empirical Domain Knowledge 509

ni nj

∑∑ I (i, j) − µ
1
σ2 =
2
(2)
ni n j i =1 j =1

Min( I ) = min
i∈ni , j∈n j
( I (i, j)) (3)

Max( I ) = max ( I (i, j ) ) (4)

i∈ni , j∈n j

2.1 Data Acquisition

The original breast ultrasound image that was used in this work is like a Fig. 1(a).
The level of pixel value information from each original image is transformed into
matrix for applying a CAD. The transformed matrix value that was shown in
Fig. 1(b) represents the light and shade value of each pixel on original image and
extra information such as outpatient’s name, distinction of sex, size of lesion. How-
ever, cutting of extra information is needed for protection of personal information.
So the extra information in images that were used in this work was cut as a result
we could only use breast ultrasound images without extra information.

(a) Original image

(b) The transformed matrix

Fig. 1. Original image and pixel information

510 L.-J. Koo et al.

2.2 Image Enhancement

Most pixel value of ultrasound image has all range of light and shade that distribute
from 0 to 255 values. But the distribution is a little too much to specific pixel value.
So, in this paper, ends-in search that makes more vivid the difference of the back-
ground and the Region Of Interest (ROI) is used for a image enhancement. A formula
of this method is as follow and the result is like Fig. 2.
Low = Min(Iorg ) + σ org
high = Max( Iorg ) − σ org
⎧ 0 Iorg (i, j ) ≤ Ιow (5)
⎪ Iorg (i, j ) − Low
I Endin (i, j) = ⎨255 × ( ) Low < Iorg (i, j ) > Ηigh
⎪255 High − Low
⎩ Iorg (i, j ) ≥ High

Where, Iorg is the ni by nj matrix which is transformed from the original image and σorg
is the standard deviation of Iorg acquired by Eq. (2).

Fig. 2. Applied Image of End-In search

The result of ends-in search was used for unnecessary image removal in domain
based heuristic noise filtering. The domain based heuristic noise filtering was basi-
cally designed by two domain known of ultrasound image as follows.
1. The shown lesion image on breast ultrasound is represented with a whole
feature.
2. The pixel value of noises related to unnecessary image and lesion has similar
range.
The proposed noise filter is based on distinct blocks operation of 13 by 13 matrixes as
shown in Fig. 2 and ostensive processes is operated by Eq. (6) as next step.
Step 1: The starting points of distinct blocks are four corners in Fig. 3.
Step 2: The mean of distinct block and the standard deviation of whole image
applied End-In search are computed and compared.
Step 3: If former is smaller than latter, all pixel values of distinct blocks are
transformed to 255. In against case, they are not changed by Eq. (6).
Step 4: If pixel values mentioned above are changed, the distinct blocks started by
1, 4 of Fig. 3 shift in the row direction. In against case, the progress direction
of blocks is opposite. Blocks have other starting points are operated in
reverse.
 Seed Point Detection of Multiple Cancers Based on Empirical Domain Knowledge 511

Fig. 3. The Principle of distinct blocks operation

∑I
1
µEndin = Endin (i, j )
NM i , j∈η
(6)
⎧ 255 µ < σ Endin
I Noi (i, j ) = ⎨ , where Endin
⎩ Endin
I (i , j ) µ Endin ≥ σ Endin

Where, η is the N by M district block of each pixel in the IEndin.

As a result, most shadowing artifacts are removed as like Fig. 4.

Fig. 4. Applied Image of noise filtering

2.3 Matrix Search Filtering

Generally, pixel value of lesion and surgical breast biopsy has low range but other
images that are not interfered have high range value. So we carry out Image negative
transformation for computing easily and clearly. A basic formula of Image negative
transformation is as follows.
I Neg (i, j ) = 255 − I Noi (i, j ) (7)
By executing the matrix search filtering, small noise in result image conducted image
negative transformation is removed. This filter used Sliding Neighborhood Operations
as 13 by 13 matrix of neighborhood blocks. Algorithms of this method are as follow
and the result is shown in Fig. 5.
512 L.-J. Koo et al.

∑
1
µNei = I Noi (i, j )
RC i, j∈ρ
Clow = µNeg − σ Neg
Chigh = µNeg + σ Neg
(8)

⎧⎪ 0 µNei < Clow or µNei < Chigh

I Mat (i, j ) = ⎨ , where
⎪⎩ I Neg (i, j ) Clow ≤ µNei ≤ Chigh

Where, ρ is the R by C local neighborhood of each pixel in the INoi and σNeg is the
standard deviation of INeg acquired by Eq. (2).

Fig. 5. Applied Image of matrix search filtering

2.4 Differential Image and 2D Adaptive Noise Removal Filtering

The differential image and final filtering are also used for noise removal. The differen-
tial image is the difference of two images that were conducted by matrix search filter-
ing and ends-in search. The result is eliminated noises that existed near the lesion
because the pixel values of the lesion image of matrix search filtering were high near to
255 but those of the lesion image of ends-in search were low near to 0. In this paper,
image of ends-in search is multiplied by 1.5 for elimination of noise around lesion.
The pixel value of ROI and background are clear certainly through the differential
image. Consequently, two-dimensional (2D) adaptive noise removal filter and local
mean filter is conducted for noise removal in this study. 2D adaptive noise removal
filter is methodology to remove noises using local mean and standard deviation [8]. In
this paper, pixel-wise Wiener filter is conducted including 100 by 100 neighborhood
blocks. The reason applying 100 by 100 matrix is that noises scatter but ROI concen-
trate. Pixel-wise Wiener filter is applied by Eq. (9) and the result is as Fig. 6.

Fig. 6. Applied image of pixel-wise Wiener filter

 Seed Point Detection of Multiple Cancers Based on Empirical Domain Knowledge 513

∑
1
µDiff = I Diff (i, j )
RC i , j∈ρ

∑
1 2
α 2 Diff = I Diff (i, j ) − µDiff 2 (9)
RC i , j∈ρ
α 2 Diff −ν 2
I 2 D (i, j ) = µDiff + I Diff (i, j ) − µDiff
α 2 Diff

Where, ν2 is the average of all the local estimated variances and ρ is the R by C local
neighborhood of each pixel in the IDiff.
All noises are not entirely removed by pixel-wise Wiener filter. So local-mean fil-
ter is conducted once again. The local-mean filter is operated by computing the local
mean and standard deviation based pixels that value was higher than 1 in 2D adaptive
noise removal filtered image. The process of this method is that the sum (CS) and
difference (CD) of the mean and standard deviation is compared with each pixel value
of image. Then, the pixel value determined based on compared result. The algorithms
of determining pixel value were as Eq. (10) and the result is like a Fig. 7.
⎧ 0 I (i, j ) < CD or I 2 D > CS
I Lcl (i, j ) = ⎨ , where 2 D (10)
⎩ 255 CS ≥ I 2 D ≥ CD

Fig. 7. Applied image of pixel-wise Wiener filter

3 Detection of Seed Points

3.1 Recovery of Tumor Area with Morphology

Original image of breast ultrasound is degraded through Mentioned steps above. So

we approached morphological combination of dilation and erosion technique for the
recovery of image, because dilation and erosion are often used in combination to
implement image processing operations with the same structuring element for re-
moval of some small size of noise. Dilation can be said to add pixels to an object, or
to make it bigger, and erosion will make an image smaller. In the simplest case, bi-
nary erosion will remove the outer layer of pixels from an object. An image Dilation
is using a structuring element. In this paper, we apply morphology combination as the
order of erosion and dilation after dilation and erosion.
514 L.-J. Koo et al.

In Dilation case, the origin of the structuring element is placed over the first white
pixel in the image, and the pixels in the structuring element are copied into their cor-
responding positions in the result image. Then the structuring element is placed over
the next white pixel in the image and the process is repeated. This is done for every
white pixel in the image.
In Erosion case, the structuring element is translated to the position of a white pixel
in the image. In this case, all members of the structuring element correspond to white
image pixels so the result is a white pixel. Now the structuring element is translated to
the next white pixel in the image, and there is one pixel that does not match. The
result is a black pixel. The remaining image pixels are black and could not match the
origin of the structuring element.
The result of morphology is shown Fig. 8.

Fig. 8. Reconstructed image using morphology

3.2 The Auto Detection of Seed Point Using K-MANS

In this section, seed points are detected for segmentation of lesion automatically.
Previous studies related to the detection of seed points had limitations that the number
of legions was one and the place of legion was center in breast ultrasound image. So,
K-MEANS algorithm was applied for solving these limitations above mentioned,
because this clustering method can best be described as a partitioning algorithm and is
suitable for auto detection of seed points if the number of tumors is set. The applica-
tion of this method is conducted under assumptions are as follows [9].
1. The number of tumors in breast ultrasound image was less than 10.
2. The pixel value of tumors after noise removal is close to 0.

K-means is conducted depending on the coordinates of image pixel, steps are as

follows.
Step 1: Coordinates of pixels more than 1 value are detected.
Step 2: The number of clustering group is defined as 1 and K-means conducted
based coordinates of step 1.
Step 3: If pixel value of the centroid coordinate is more than 1, K-means is stopped.
If pixel value indicated to result of step 2 is 0, K-means is iterated from step
1 by adding the number of clustering group automatically.
 Seed Point Detection of Multiple Cancers Based on Empirical Domain Knowledge 515

Fig. 9. Seed point detection using K-Means

The result is like as Fig. 9. In this paper, seed points mentioned results above are
used to improve boundary of ROI and background image because the segmentation of
ROI from seed points without additional processes take a long computing time.

4 Result and Conclusion

In this paper, the number of images applied to the proposed methods is 40, obtained
from two hospitals as shown Fig.10. As a result, the most of seed points are detected
by proposed methods.
Ultrasound creates a lot of noises and it was quite difficult to get an understandable
ROI. However, the image is extremely important and segmentation of ROI is neces-
sary for a surgical operation of breast cancer. So, many previous works have been
studied for solving problems mentioned above but it is really difficult to find out ROI
due to much noise generated.
CAD system has been developed for early diagnosis of cancers and improvement
of diagnostic exactitude. CAD based system is also used as an assistant of diagnosis
which is co-developed with medical specialists.

a) case1 b) case2

c) case3 d) case4

Fig. 10. Examples applying to other original images

516 L.-J. Koo et al.

The objective of this paper is to design noises removal methods and detect auto-
matically seed points for applying the breast ultrasound image to CAD. So, the whole
processes are conducted under the discussion with medical specialist and we pre-
sented auto detection of seed points preceded for segmentation with K-means algo-
rithm to find out the estimated central area of the tumor. Sometimes it is difficult to
differentiate between background color and tumor color, so we applied domain based
heuristic noise filter. As a result of applying proposed algorithm, the noise is effec-
tively eliminated in all images and an automated detection is possible by determining
seed points on each tumor.

References
1. Minister for Health: Welfare and Family Affairs (MIHWAF). Statics of Cancer, Republic of
Korea (2007)
2. Paulo, S.R., Gilson, A.G., Marcia, P., Marcelo, D.F., Chang, R.F., Suri, J.S.: A New Meth-
odology Based on q-Entropy for Breast Lesion Classification in 3-D Ultrasound Images. In:
Conf. Proc. IEEE Eng. Med. Biol. Soc., New York, vol. 1, pp. 1048–1051 (2006)
3. Moore, K.S.: Better Breast Cancer Detection. IEEE Spectrum 38, 51–54 (2001)
4. Madabhushi, A., Metaxas, D.N.: Automatic boundary extraction of ultrasonic breast lesions.
In: 2002 IEEE Int. Symp. Biomedical Imaging, pp. 601–604 (2002)
5. Huang, Y.L., Chen, D.R.: Automatic Contouring for Breast Tumors in 2-D Sonography. In:
Conf. Proc. IEEE Eng. Med. Biol. Soc., vol. 3, pp. 3225–3228 (2005)
6. Madabhushi, A., Metaxas, D.N.: Combining Low-, High-Level and Empirical Domain
knowledge for Automated Segmentation of Ultrasonic Breast Lesions. IEEE Trans. Medical
Imaging 22(2), 155–169 (2003)
7. Pietro, P., Jitendra, M.: Scale-Space and Edge Detection Using Anisotropic Diffusion. IEEE
Trans. Pattern Anal. Machine Intell. 12(7), 629–639 (1990)
8. Lim, J.S.: Two-Dimensional Signal and Image Processing, pp. 536–540. Prentice-Hall,
Englewood Cliffs (1990)
9. Guralnik, V., Karypis, G.: A scalable algorithm for clustering protein sequences. In: Proc.
Workshop Data Mining in Bioinformatics (BIOKDD), pp. 73–80 (2001)
 A Controlled Scheduling Algorithm Decreasing
the Incidence of Starvation in Grid Environments

Minna Liu, Kairi Ou, Yuelong Zhao, and Tian Sun

School of Computer Science and Engineering, South China University of Technology,

Guangzhou, 510006, China
liuminna@sina.com

Abstract. A fair scheduling algorithm accounting for the weight and execution
time of tasks is critical in the Grid environment. The MTWCT (Minimize Total
Weighted Completion Time) has been proved to minimize the total weighted
completion time of a set of independent tasks in a processor, but it results in an-
other problem that the response time of some tasks is far longer. To decrease
the incidence of the starvation phenomena, an improved algorithm named CSA
(Controlled Scheduling algorithm) based on MTWCT is proposed, which com-
putes the ρ factors of tasks by the execution time and weight of step chains, and
selects the unexecuted step chain in terms of ρ factor and the executed time of
task. Experimental results exhibit that CSA compared with the MTWCT, de-
creases the completion time of short tasks and the average turnaround time by
sacrificing a little total weighted completion time.

Keywords: grid environment, scheduling algorithm, starvation phenomena,

step chains.

1 Introduction
Efficient task scheduling can contribute significantly to the overall performance of the
Grid and Cluster computing systems [1] [2] since an inappropriate scheduling of tasks
can fail to exploit the true potential of the system [3]. Scheduling a set of independent
tasks in grid environment is NP-hard [4]. Among the algorithms that have been pro-
posed in the past, the Sufferage algorithm [5] has better performance. However, it
deals with one objective which minimizes the makespan in the set of independent tasks
[6]. In complex grid environment, the weight and execution time of takes are different.
It is unreasonable that assuming the importance and urgencies of all tasks are the same.
A scheduling algorithm called MTWCT (Minimize Total Weighted Completion Time)
[7], in the accordance to the fact that a task is made up of step chains of precedence
constraints, divides steps of each task into step chains by the ratios of weight and exe-
cution time of step chains and the step chain is to be executed as a whole. The ratios of
weight and execution time of step chains accounts for the importance of the task and
the length of execution time, which overcomes the limitation of Sufferage one-sided
pursuit of minimal makespan regardless of the importance of the task, but it results in
another problem that short tasks with a bit smaller ratios have to wait for long term.
That is because MTWCT deals with only one objective which is minimizing the total

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 517–525, 2009.
© Springer-Verlag Berlin Heidelberg 2009
518 M. Liu et al.

weighted completion time, neglecting the average turnaround time of tasks. When the
amount of long tasks is very great, it will exacerbate the extent of starvation [8]. Obvi-
ously it is unfair to the short tasks with a bit smaller ratios of weight and execution
time. In addition, it isn’t efficient for the system that all short tasks with a bit smaller
ratios haven’t chances to be executed until all long tasks have been completed; espe-
cially the execution time of long tasks is much longer than that of short tasks.
According to the above analysis, we propose a meaningful modified algorithm based
on MTWCT named CSA (Controlled Scheduling Algorithm) which is able to leverage
the strengths of MTWCT and avoid its weaknesses. CSA supports the following features.
(1) Avoiding short tasks with a bit smaller ratios waiting for long term.
(2) Decreasing the average turnaround time of all tasks. Namely, it is beneficial to
the majority of users and the system.
(3) Being able to make the total weighted completion time approximately optimal.
The rest of the paper is organized as follows: In section 2, we describe the strategies of
setting thresholds. The Controlled Scheduling Algorithm is introduced and we present
the analysis of the Controlled Scheduling Algorithm. Section 3 provides numerical
results to compare the task completion time, the average turnaround time and the total
weighted completion time under CSA and under MTWCT. Conclusions are drawn in
Section 4.

2 Controlled Scheduling Algorithm

2.1 Model Description

In this paper, we employ the triple used firstly by Graham [9] to depict the scheduling
1
∑
problem. 1|chains| ( w j p j )π ∪ ti : where 1 denotes all tasks are supposed to be
n
executed in a single processor and chains denotes a set of independent tasks.
1
∑
( w j p j )π ∪ ti is the objective of CSA that the total weighted completion time
n
and the average turnaround time are taking into account. Let (∑ w j p j )π denotes
1
the total weighted completion time and ti denotes the average turnaround time.
n
We summarize the notation used throughout the paper in Table 1.

Table 1. Notation

parameters descriptions
n Number of tasks in the task queue
k Number of steps every task
Tij the jth step of the ith task
pij execution time required by step Tij
wij weight of step Tij
ρ the ratio of step chain’s weight and execution time
t the limit executed time of a task
∆ρ the range of ρ factor
 A Controlled Scheduling Algorithm Decreasing the Incidence of Starvation 519

2.2 Definitions and Theorems Related with CSA

In this section, two definitions and theorems related with CSA are given.
：
Definition 1 For a task is composed by k steps at most, the rule of its ρ factors is
defined as follows:

∑ ⎧ ∑ l wi j ⎫
l*
j =1
wi j ⎪ j =1 ⎪
= max ⎨ l ⎬ (1)
∑
l*
j =1
pi j
1≤ l ≤ k
⎪
⎩ ∑ j =1
pi j ⎪
⎭

namely, ρi1 (1,2,…,l*), where l* is the deadline step of ρi1 . In a similar manner,
ρi 2 ...ρik can be obtained.
：
Definition 2 Assuming that there are two tasks i and j, task i contains k steps de-
noted by Ti1 → Ti 2 → ... → Tik , and task j is made up of k steps denoted
by Ti1 → Ti 2 → ... → Tik . Let their steps’ sequence denoted by π (Ti1 , Ti 2 ,..., Tik , T j1 ,...T jk ) , the
total weighted completion time is defined to be:.
k k k k
(∑ pij wij )π = wi1 pi1 + ... + wik ∑ pim + wj1 ( p j1 + ∑ pim ) + ... + wjk (∑ pim + ∑ p jf ) (2)
m =1 m =1 m =1 f =1

∑ ∑
m r
j =1
wij w fa
Theorem 1: If > a =1
, and the step chain made up of Ti1 , Ti 2 ,..., Tim z is
∑ ∑
m r
j =1
p ij a =1
p fa
executed prior to the step chain made up of T f 1 ,T f 2, ...T fr , which makes the total
weighted completion time of steps denoted by Ti1 ,Ti 2 ,..., Tim , T f 1 ,T f 2, ...T fr minimal.

Theorem 2: If Tl is the deadline step of ρ factor about the step chain denoted by
*

Ti1 , Ti 2 ,..., Tik , a step chain denoted by Ti1 , Ti 2 ,..., Til must be uninterrupted.
*

The proof of theorem 1 and theorem 2 is given in [7].

2.3 Strategy of Setting the ∆ρ and t Thresholds

Appropriate t and ∆ρ are critical to CSA. The objective of setting t is to make a step
chain represented by a smaller ∆ρ factor have a chance to be executed earlier than
that of a bigger ρ factor. Setting ∆ρ is aim to choose other ρ factors in
[max_ ρ − ∆ρ , max_ ρ ] , where max_ ρ denotes the maximum ρ factor, to ensure
the weighted total completion time as optimal as possible.

2.3.1 The Strategy of Setting t Thresholds

It should be noted that t thresholds come from the long tasks with bigger weights due
to the reason that setting t is aim to delay being executed of long tasks.
520 M. Liu et al.

We define the following arrays that are used in the description of setting t
thresholds.
t[][] : array of step chains execution time. where t[i][j] denotes the execution time
required by the jth step chain of the ith task.
j
sum_t[][] : array of total execution time of step chains, where sum _ t[i ][ j ] = ∑ t[i][ f ] .
f =1

tt[] : array of total thresholds.

The strategy of setting t thresholds is described as follows:
initialize t[][], sum_t[][], tt[];
for i=1 to n do
for j=2 to k do
sum_t[i][j] = sum_t[i][j-1]+t[i][j];
for i=1 to k do
for j=1 to n do
If(sum_t[i][j]>max) then max = sum_t[i][j];
If(sum_t[i][j]<min) then min = sum_t[i][j];
tt[++count]=max;
tt[++count]=min;

2.3.2 The Strategy of Setting ∆ρ Thresholds

Meaningful ∆ρ thresholds should make the short tasks with smaller weights execute
earlier which is helpful to decrease the average turnaround time while long tasks are
more than short tasks.
We define the following array and variants that are used in the description of set-
ting ∆ρ thresholds.
ρ[][] : array of ρ factors. Where ρ [i ][ j ] denotes the ρ factor of the task.
shorttasks _ min_ ρ : the minimum ρ factor in all the ρ factors from short tasks.
shorttasks _ max_ ρ : the maximum ρ factor in all the ρ factors from short tasks.
The description of strategy of setting ∆ρ thresholds as follows:

initialize ρ [][] , shorttasks _ min_ ρ , shorttasks _ max_ ρ ;

max = 0; min = ∞;
for i=1 to m do
for j=1 to k do
if( ρ[i][j]<0) break;
if( ρ[i][j]>max) max = ρ[i][j] ;
if( ρ[i][j]<min) min = ρ[i][j];
shorttasks _ min_ ρ = min;
shorttasks _ max_ ρ = max;
count = 0;
for i=1 to h do
for j=1 to k do
 A Controlled Scheduling Algorithm Decreasing the Incidence of Starvation 521

∆ρ[+ + count] = ρ[i][j] − shorttasks _ min_ ρ;

∆ρ[+ + count] = ρ[i][j] − shorttasks _ max_ ρ;

2.4 Description of CSA

Step 1: choose two critical values ∆ρ [i ] and ∆ρ [ j ] from the one-dimensional vector
∆ρ [] arbitrarily, and select a constant from the interval (∆ρ [i ], ∆ρ [ j ]) .
Step 2: Similar to step 1, set the initial limit executed time allocate to a task which
is called t .
Step 3: Sort all the ρ factors by non-ascend order, after that save them in the one-
dimensional vector ρρ [] .
Step 4: Set a flag to every ρ factor and initialize flag=0. Where flag=0 denotes the
step chain represented by ρ factor hasn’t been executed; otherwise, means it has
been executed.
Step 5: Set a temporary limit executed time temp_t, and initialize temp_t=t.
Step 6: Choose the maximum ρ factor whose value of flag is zero from one-
dimensional vector ρρ [] , which is called max_ ρ .
Step 7: Save all the ρ factors between max_ ρ − ∆ρ and max_ ρ whose values
of flags are zero in a one-dimensional vector candi _ ρ[] .
Step 8: If executed time of the task represented by max_ ρ less than temp_t, exe-
cute the step chain represented by max_ ρ , after that the value of max_ ρ ’s flag is
set to 1 and then go to step 9. Otherwise, if candi _ ρ[] is an empty vector, execute
the step chain represented by max_ ρ , else select a candi _ ρ[i ] with the minimum
task executed time from candi _ ρ[] and process the step chain represented by
candi _ ρ[i ] .
Step 9: Set temp _ t = t , go to step 11.
Step 10: Update temp _ t = temp _ t + min_ t , go to step 11.
Step 11: Repeat the whole procedures from step 5 through step 10 until all step
chains have been executed.

2.5 Analysis of CSA

The time complexity of CSA is mainly composed of three parts: setting t thresholds,
setting the ∆ρ thresholds and the choice of step chains. we can obtain that it takes
O(2nk) to set t thresholds and O(mk+hk) to set ∆ρ thresholds according to (3.3). the
choice of step chains costs O((mk+hk)*(mk+hk)). CSA costs O(2nk)+O(mk+hk) more
than MTWCT, which is ignored when m and h is very big.
According to the analysis of the CSA, we can obtain the following deductions:
Deduction 1: if t → +∞ , then CSA is equivalent to MTWCT.
522 M. Liu et al.

Proof: Accord to step 8 of the description of CSA, if t → +∞ , the task executed time
of max_ ρ always less than temp _ t , so the step chain represented by max_ ρ will
be executed at all times. ∆ρ and t has no impact on selecting unexecuted step chains ,
then CSA is equivalent to MTWCT.
Deduction 2: if ∆ρ → −∞ , then CSA is equivalent to MTWCT.

Proof: Accord to step 7 of the description of CSA, if ∆ρ → −∞ , candi _ ρ[] is an

empty vector. So the execution right is always given to the step chain represented by
max_ ρ according to step 8 of the description of CSA, then CSA is equivalent to
MTWCT.
Thanks to the task executed time limit t, CSA overcomes the starvation phenomena
under MTWCT, which chooses the unexecuted step chain represented by maximum
ρ factor every time resulting in other step chains denoted by a bit smaller ρ factors
non-execution long term.

3 Simulations and Analysis

The execution time and weights of steps can’t be identified in advance, which need
predicting. In this paper, experimental parameters are assumed to be obtained. We
focus on comparing MTWCT and CSA in terms of the task completion time, the av-
erage turnaround time and the total weighted completion time.
The execution time, weights and ρ factors of step chains of every task are given
in Table 2.
From Table 3, we can see that if t = 0.5 , ∆ρ = 0.32 , t = 0.5 , ∆ρ = 0.005 and
t = 0.06 , ∆ρ = 0.02 , CSA is equivalent to MTWCT. From deduction 1, it is obvious
that t is equivalent to +∞ while t >0.489 because of the maximum threshold t for
0.489 in this experiment. According to deduction 2, when ∆ρ <0.05, ∆ρ is equiva-
lent to −∞ due to the fact that the minimum threshold ∆ρ is 0.05 in this experiment.

Table 2. Experimental Parameters

category task pij (us ) wij ρij

0.163 0.25 1.53
0.116 0.16 1.38
task1
0.112 0.14 1.25
category1 0.098 0.12 1.22
0.078 0.11 1.41
task2 0.045 0.06 1.32
0.056 0.07 1.24
0.042 0.05 1.19
category2 task3
0.035 0.04 1.15
 A Controlled Scheduling Algorithm Decreasing the Incidence of Starvation 523

Table 3. The ρ factors sequence of MTWCT and CSA

t , ∆ρ Sequences of ρ factors
MTWCT ρ11 , ρ 21 , ρ12 , ρ 22 , ρ13 , ρ 23 , ρ14 , ρ31 , ρ32
t = 0.06, ∆ρ = 0.02 ρ11 , ρ 21 , ρ12 , ρ 22 , ρ13 , ρ 23 , ρ14 , ρ31 , ρ32
t = 0.06, ∆ρ = 0.3 ρ11 , ρ 21 , ρ31 , ρ 32 , ρ 22 , ρ12 , ρ13 , ρ 23 , ρ14
t = 0.11, ∆ρ = 0.15 ρ11 , ρ 21 , ρ 22 , ρ 23 , ρ12 , ρ31 , ρ32 , ρ13 , ρ14
CSA t = 0.11, ∆ρ = 0.1 ρ11 , ρ 21 , ρ 22 , ρ12 , ρ 31 , ρ 23 , ρ13 , ρ32 , ρ14
t = 0.162, ∆ρ = 0.25 ρ11 , ρ 21 , ρ31 , ρ12 , ρ 22 , ρ32 , ρ 23 , ρ13 , ρ14
t = 0.5, ∆ρ = 0.32 ρ11 , ρ 21 , ρ12 , ρ 22 , ρ13 , ρ 23 , ρ14 , ρ31 , ρ32
t = 0.5, ∆ρ = 0.005 ρ11 , ρ 21 , ρ12 , ρ 22 , ρ13 , ρ 23 , ρ14 , ρ31 , ρ32

The number of tasks is fixed to 100, the category1 tasks are assumed to long tasks
and the category2 task are supposed to short tasks. The percentage of the category1
tasks are in [10%,90%]. Since t = 0.5, ∆ρ = 0.32 , t = 0.5, ∆ρ = 0.005 and
t = 0.06, ∆ρ = 0.02 , CSA is equivalent to MTWCT, only t = 0.06, ∆ρ = 0.03 ,
t = 0.11, ∆ρ = 0.15 , t = 0.11, ∆ρ = 0.1 ,and t = 0.152, ∆ρ = 0.25 are adopted to test the
performance of CSA respectively.

3.1 Comparison of CSA and MTWCT in Terms of Task Completion Time

With the increasing of long tasks, the results of the short tasks completion time is
depicted in Fig.1. According to the simulation results, the performance of CSA varies
with different t and ∆ρ .
From Fig.1, Note that the short tasks completion time under CSA is obviously
smaller than that using MTWCT.

Fig. 1. Comparison of short tasks completion time using two algorithms

524 M. Liu et al.

3.2 Comparison of CSA and MTWCT in Terms of the Average Turnaround

Time

From Fig.1 it can be obtained easily that when short tasks are executed earlier, long
tasks are inevitably delayed. As we know, the average turnaround time can reflect the
overall performance of a scheduling algorithm. The average turnaround time is short
in favor of the majority of users, also reflects high system throughput, and high sys-
tem utilization rates. The simulation results is showed in Fig.2.
From Fig.2, we can see that either CSA or MTWCT increases the average turn-
around time with the increasing of long tasks. The values of t , ∆ρ have great impact
on CSA in the performance of average turnaround time. For t = 0.11, ∆ρ = 0.15 , CSA
is extremely efficient and when the percentage of long tasks is over 13%, CSA is
superior to MTWCT.

Fig. 2. Comparison of the average turnaround time using two algorithms

Fig. 3. Comparison of the weighted total completion time using two algorithms
 A Controlled Scheduling Algorithm Decreasing the Incidence of Starvation 525

3.3 Comparison of CSA and MTWCT in Terms of the Total Weighted

Completion Time

We can see that the weighted total completion time using CSA is definitely not better
than that using MTWCT by the previous theoretical analysis. The performance of the
total weighted completion time comparison of the CSA and MTWCT is depicted in
Figs.3.
From Fig.3, we can see that either CSA or MTWCT increases the weighted total
completion time in parallel to the increase in the number of long tasks. The weighted
total completion time using CSA is a bit greater than that using MTWCT.

4 Conclusion
In this paper, we presented a meaningful modified scheduling algorithm named Con-
trolled Scheduling Algorithm for a set of independent tasks in a processor of heteroge-
neous systems, which supports three features proposed in the section I. CSA takes into
account of fairness and effectiveness, decreasing the incidence of the starvation phe-
nomena produced by MTWCT. In addition, CSA has the flexibility to set t and ∆ρ to
meet various applications. A large number of experimental results indicate that in
general compared with MTWCT, the average turnaround time under CSA decreases
around 20% at the price of increasing 3% in the total weighted completion time.

References
1. Foster, I., Kesselman, C.: The Grid2, Blueprint for a New Computing Infrastructure. Mor-
gan Kaufmann, San Francisco (2004)
2. Abraham, A., Buyya, R., Nath, B.: Nature’s Heuristics for scheduling Jobs on Computa-
tional Grids. In: ADCOM 2000, Cochin, India, pp. 45–52 (2000)
3. Kwok, Y.-K., Ahmad, I.: Static Scheduling Algorithms for Allocating Directed Task Graphs
to Multiprocessors. In: ACM Computing Surveys, New York, pp. 406–471 (1999)
4. Maheswaran, M., Ali, S., Siegel, H.: Dynamic Matching and Scheduling of a Class of In-
dependent Tasks onto Heterogeneous Computing Systems. In: Eigth Heterogeneous Com-
puting Workshop, San Juan, Pueto Rico (April 1999)
5. Maheswaran, M., Ali, S., Siegel, H.J., Hensgen, D., Freund, R.F.: Dynamic Mapping of a
Class of Independent Tasks onto Heterogeneous Computing Systems. Journal of Parallel
and Distributed Computing 59(2), 107–131 (1999)
6. SaiRanga, P.C., Baskiyar, S.: A low complexity algorithm for dynamic scheduling of inde-
pendent tasks onto heterogeneous computing systems. In: Proceedings of the 43rd annual
Southeast regional conference, New York, pp. 63–68 (2005)
7. Pinedo, M.: Scheduling-theory, algorithms, and systems, ch. 4. Prentice Hall, Englewood
Cliffs (1995)
8. Cuesta, B., Robles, A., Duato, J.: An Effective Starvation Avoidance Mechanism to En-
hance the Token Coherence Protocol. In: IEICE Transactions on Fundamentals of Electron-
ics, Communications and Computer Sciences, Washington, pp. 47–54 (2007)
9. Graham, R.L., Lawler, E.L., Lenstra, J.K., et al.: Optimization and Approximation in De-
termi-nistic Sequencing and Scheduling. Annals of Discrete Mathematics 5, 287–326
(1979)
 A Space Allocation Algorithm for Minimal Makespan in
Space Scheduling Problems

Chyuan Perng1, Yi-Chiuan Lai 2, Zih-Ping Ho1, and Chin-Lun Ouyang1

1
Department of Industrial Engineering and Enterprise Information
Tunghai University, Taichung City 407, Taiwan
2
Department of Business Administration
Providence University, Shalu, Taichung County 433, Taiwan
ylai@pu.edu.tw

Abstract. The factory space is one of the critical resources for the machine as-
sembly industry. In machinery industry, space utilizations are critical to the ef-
ficiency of a schedule. The higher the utilization of a schedule is, the quicker
the jobs can be done. Therefore, the main purpose of this research is to derive a
method to allocate jobs into the shop floor to minimize the makespan for the
machinery industry. In this research, we develop an algorithm, Longest Contact
Edge Algorithm, to schedule jobs into the shop floor. We employed the algo-
rithm to allocate space for jobs and found that the Longest Contact Edge Algo-
rithm outperforms the Northwest Algorithm for obtaining better allocations.
However, the Longest Contact Edge Algorithm results in more time complexity
than that of the Northwest Algorithm.

Keywords: Space Allocation Algorithm, Resource Constrained Scheduling

Problem, Longest Contact Edge Algorithm, Dispatching Rules, Space Schedul-
ing Problem.

1 Introduction
Scheduling is an important tool for manufacturing and engineering, as it can have a
major impact on the productivity of a process. In manufacturing, the purpose of
scheduling is to minimize production time and costs by arranging a production facility
regarding what and when to make, with which staff, and on which equipment. Pro-
duction scheduling aims to maximize the efficiency of the operation and reduce costs
[13]. In general, the scheduling problem could be very complicated. Perng et al.
[2][3][4][5][6] proposed a space resource constrained job scheduling problem. In their
researches, utilization of a shop floor space was an important issue in a machinery
assembly factory. The assembly process of a machine required a certain amount of
space on a shop floor in the factory for a period of time. The sizes of the shop floor
and machines would determine the number of machines which can be assembled at
the same time on the shop floor. The sequences of jobs would affect the utilization of
the shop floor. The space on the shop floor could be divided into several chunks and
some of these jobs could be allocated for simultaneous assembling. When a new job

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 526–534, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 A Space Allocation Algorithm for Minimal Makespan 527

arrived, the factory has to offer a complete space to allocate the new job based on its
space requirement. If the factory was at its capacity, the new order must wait for the
space taken by jobs previously assigned to complete production. Due to space capac-
ity constraints, we have to schedule the sequence of jobs carefully in order to maxi-
mize space utilization.
The makespan is important when the number of jobs is finite. The makespan is de-
fined as the time the last job leaves the system, that is, the completion time of the last
job [8]. The makespan shows within how many days the factory can complete these
jobs. Nichols et al. [14] was the first paper to discuss scheduling with minimum
makespan as an objective. In their research, n independent, single operation jobs, all
available at time zero, on m identical processors; each job must be processed by ex-
actly one of these processors. Computational experience with the procedure indicated
that good solutions to large scale problems were easily obtained. Gallo and Scutella
[9] tried to minimize the makespan of scheduling problem with tree-type precedence
constraints. The assembly factories wanted to select a sequence from which all jobs
can be completed in time. Liao and Lin [1] considered the two-uniform-parallel-
machine problem with the objective of minimizing makespan. The problem can be
transformed into a special problem of two identical parallel machines from the view-
point of workload instead of completion time. An optimal algorithm was developed
for the transformed special problem. Although the proposed algorithm had an expo-
nential time complexity, the results showed that it could find the optimal solution for
large scale problems in a short time. Lian al. [15] presented a swarm optimization
algorithm to solve a job shop scheduling problem. The objective was to find a sched-
ule that minimizes the makespan. Perng et al. [5] proposed an algorithm based on a
container loading heuristic (CLH) approach [7] to a space scheduling problem with
minimal makespan. However, the original CLH could not find all the possible free
space, while this research could. Zobolas et al. [10] proposed a hybrid metaheuristic
for the minimization of makespan in permutation flow shop scheduling problems. A
genetic algorithm (GA) for solutions was selected. Computational experiments on
benchmark data sets demonstrated that the proposed hybrid metaheuristic reached
high-quality solutions in short computational times. Although numerous scheduling
problems discussed makespan, there is only a limited amount of literature available on
the space scheduling field.
One of the real world applications is the efficient space allocation algorithm. This
research proposes a new space allocation algorithm to minimize the makespan for a
space scheduling problem. In the other words, the main purpose of this research is to
derive a method to allocate the jobs into a shop floor more efficiently.

2 Problem Formulation
Let N denote a set of n jobs. Let s1, s2,…, sn denote the start time for each job. Let p1,
p2,…, pn denote the processing time for each job. Let dd1, dd2,…,ddn denote the due
date for each job. Let σ denote an arbitrary sequence. The rest of notations is shown
as follows:
528 C. Perng et al.

Cj: completion time of job j

Cj = sj + pj
Cmax: makespan
Cmax = Max{s j + p j }
σ : an arbitrary sequence
f (σ ) : makespan of σ
The objective is to find a sequence to minimize the makespan.

min f (σ ) = Cmax (1)

The function f (σ ) denotes the makespan of σ processing sequence. We assume

that all jobs are available at day one, all of the orders are rectangles, a job will not be
moved until it is done, there is no constraint on job’s height, and the buffer or storage
is available to fit in any number or any shape of jobs.

3 Longest Contact Edge Algorithm

The northwest algorithm (NWA) is a basic space allocation approach used in Perng
et al. [2][3][6]. It was used to allocate jobs into a factory. In this study, we intend to
propose a better space allocation algorithm, namely, longest contact edge algorithm
(LCEA), which differs from the northwest approach. The flow chat of LCEA is
shown in Figure 1.
The foundation of the Longest Contact Edge Algorithm is based on Perng et al.
[2][3][6]. The difference between LCEA and NWA is that LCEA evaluates every free
grid instead of finding only one northwest point and allocating the job at the north-
west points. The procedure of LCEA is as follows:

k: job number (k = 1, 2,…, n)

ak: the length of job k
bk: the width of job k
c: a candidate solution
(rpxc, rpyc): the coordinate of c
S: a set of c
⎧1, if the coordinate (i, j ) is occupied
grid ( i, j ) = ⎨
⎩0, otherwise

Step1: Search starting grids

At the beginning, the algorithm searches all free grids in the factory. The algorithm in
this step employs NWA in literature [2][3][6]. It scans (1,1), (1,2),… ,(m, n) for start-
ing points. The factory size is m x n grids.
Step 2: Search a job’s working space
The algorithm then searches ak x bk complete free grids for job k based on the points
obtained in step 1. If the ak x bk complete grids could be found, the starting point then
is defined as a reference point. It is also called a candidate solution c.
A Space Allocation Algorithm for Minimal Makespan 529

Start

Search starting grids

Search a job’s working space

Evaluate each candidate

solution

Choose the best c

Allocate the job

Finish

Fig. 1. Flowchart of LCEA

530 C. Perng et al.

Step 3: Evaluate each candidate solution

Based on Longest Contact Edge evaluation function (2), we tally the value of every c
found in step 2 by counting the contact length of the edge of the new job with the job
on the shop floor which is adjacent to it.
Step 4: Choose the best c
The largest value z (2) of candidate solution c in S will be chosen as a northwest point
to allocate the job. If there is a tie, the first candidate found in step 1 will be chosen.
Step 5: Allocate the job
Allocate the northwest point of the job on the point found in step 4.The reference
point of this job is (rpxc, rpyc).
The value z of the evaluation function is:
⎧⎪ rpxc + ak −1 rpxc + ak −1
z = Max ⎨
c∈S
⎪⎩
∑
i = rpxc
grid ( i, rpyc − 1) + ∑
i = rpxc
grid ( i, rpyc + bk )
(2)
rpyc + bk −1 rpyc + bk −1
⎫⎪
+ ∑
j = rpyc
grid ( rpxc − 1, j ) + ∑
j = rpyc
grid ( rpxc + ak , j ) ⎬
⎪⎭
.

In LCEA, we tally the number of grids that the perimeter of the new job we tend to
allocate if it is allocated from the reference point contacts with the jobs previously
allocated. We examine grid (i, j) around the new job’s working space. Let grid (i, j) =
1, if grid (i, j) is occupied by any job or obstacles in the shop floor. Let grid (i, j) = 0,
if grid (i, j) is an available space. The concept of longest contact edge is shown in
Figure 2. It shows that the perimeter of the new job’s working space is surrounded by
unavailable grids. Equation (2) tallies the number of contact grids if the job is allo-
cated at the reference point (northwest corner).

Fig. 2. Concept of LCEA

 A Space Allocation Algorithm for Minimal Makespan 531

4 Computation Results and Discussion

In order to evaluate the performance of the proposed algorithm under traditional dis-
patching rules, namely SPT, LPT, EDD, FCFS, SSR, and LSR, the algorithm was
developed by using PHP and Microsoft Visual Basic languages. Integrated software
was also developed for solving the space allocation problem. The current version of
the software includes both Northwest and Longest Contact Edge Algorithms. The
interface of the software is shown in Figure 3. We used a Pentium IV (Celeron CPU
2.40GHz) computer for the computations. We computed the makespans for different
number of jobs (25, 50 and 75). Thirty different data sets were obtained from the OR-
Library [1] [2] due to real data being too few for statistical analysis. The space re-
quirements of a job and order information were obtained from a company located in
central Taiwan. The results are listed in Table 1 to 3.
We employ t-test for matched samples of dispatching rules to compare the
makespans between the NWA and LCEA. The results indicate that LCEA outper-
forms NWA for every dispatching rule in makespan. However, LCEA consumes more
time complexity T [5]. All results can be obtained within 5 seconds. The LCEA could
find any possible free grids in the factory. Hence, it shortens the makespan in a space
scheduling problem.

Fig. 3. Interface of Program

532 C. Perng et al.

Table 1. The mean makespan and time complexity Ti for 25 Jobs

Sequencing
NWA LCEA T1 T2
rule
SPT 85.20 75.67 142455 1221082
LPT 67.43 63.90 140084.4 1232975
EDD 72.53 67.90 124962.9 1293922
FCFS 79.93 72.40 143633.4 1238529
SSR 80.37 75.47 141535.8 1679210
LSR 77.30 72.60 171410.2 921304.1

T1 NWA
T2 LCEA

Table 2. The mean makespan and time complexity Ti for 50 Jobs

Sequencing
NWA LCEA T1 T2
rule
SPT 306.33 232.73 324420.2 2012006
LPT 244.63 199.53 298566.8 2150939
EDD 262.53 198.60 296033.1 2056968
FCFS 282.63 211.43 308262 2081974
SSR 249.77 218.10 43540.9 2347393
LSR 314.60 259.60 304970.1 2039592

T1 NWA
T2 LCEA

Table 3. The mean makespan and time complexity Ti for 75 Jobs

Sequencing
NWA LCEA T1 T2
rule
SPT 151.5 130.83 345015.2 1873230
LPT 111.8 95.63 304846.5 1923400
EDD 119.37 103.77 315741.6 1919576
FCFS 134.57 115.57 327115.7 1888276
SSR 134.30 108.90 400728 2655211
LSR 130.80 108.90 369289.9 2655211

T1 NWA
T2 LCEA
 A Space Allocation Algorithm for Minimal Makespan 533

5 Concluding Remark
A space scheduling problem is an important issue in a machinery assembly factory. In
this study, we compared NWA with proposed LCEA. We employed the makespan as
the performance measure in the space scheduling problem. We employed the LCEA
to allocate space for jobs. We found that the LCEA can do the better assignments than
the NWA for all schedules obtained from traditional dispatching rules. However, the
computing complexity of the LCEA is a little worse than that of the NWA.
There are some assumptions in this study, for example, all jobs are available at day
one, shapes of the orders are all rectangles, a job will not be moved until it is done,
there is no constraint on job’s height, and the buffer or storage is available to fit in
any number or any shape of jobs. It may result in different conclusions if some of
assumptions are relaxed.

References
1. Liao, C.J., Lin, C.H.: Makespan minimization for two uniform parallel machines. Interna-
tional Journal of Production Economics 84(2), 205–213 (2003)
2. Perng, C., Lai, Y.C., Zhuang, Z.Y., Ho, Z.P.: Application of scheduling technique to job
space allocation problem in an assembly factory. In: The Third Conference on Operations
Research of Taiwan, vol. 59, pp. 1–7. Yuan-Zhe University, Tao-Yuan County (2006)
3. Perng, C., Lai, Y.C., Zhuang, Z.Y., Ho, Z.P.: Job scheduling in machinery industry with
space constrain. System Analysis Section. In: The Fourth Conference on Operations Re-
search of Taiwan, vol. 5, pp. 1–11. National Dong-Hwa University, Hwa-Lian City (2007)
4. Perng, C., Lai, Y.C., Ho, Z.P.: Jobs scheduling in an assembly factory with space obsta-
cles. In: The 18th International Conference on Flexible Automation and Intelligent Manu-
facturing, Skovde, Sweden, vol. 4B, pp. 1–9 (2008)
5. Perng, C., Lin, S.S., Ho, Z.P.: On space resource constrained job scheduling problems- A
container loading heuristic approach. In: The 4th International Conference on Natural
Computation, vol. 7, pp. 202–206. Shandong University, Jinan (2008),
doi:10.1109/ICNC.2008.419
6. Perng, C., Lai, Y.C., Ho, Z.P.: A space allocation algorithm for minimal early and tardy
costs in space scheduling. In: 3rd International Conference on New Trends in Information
and Service Science papers (NISS), Sec.1(6), Beijing Friendship Hotel, Beijing, China, pp.
1–4 (2009)
7. Pisinger, D.: Heuristics for the container loading problem. European Journal of Opera-
tional Research 141(2), 382–392 (2002)
8. Sule, D.R., Vijayasundaram, K.: A heuristic procedure for makespan minimization in job
shops with multiple identical processors. Computers and Industrial Engineering 35(3-4),
399–402 (1998)
9. Gallo, G., Scutella, M.G.: A note on minimum makespan assembly plans. European Jour-
nal of Operational Research 142(2), 309–320 (2002)
10. Zobolas, G.I., Tarantilis, C.D., Ioannou, G.: Minimizing makespan in permutation flow
shop scheduling problems using a hybrid metaheuristic algorithm. Computers & Opera-
tions Research 36(4), 1249–1267 (2009)
11. Beasley, J.E.: OR-Library: distributing test problems by electronic mail. Journal of Opera-
tional Research Society 41(11), 1069–1072 (1990)
534 C. Perng et al.

12. Beasley, J.E.: OR-Library (5/1/2009),

http://people.brunel.ac.uk/~mastjjb/jeb/info.htm
13. Pinedo, M.: Scheduling theory, algorithms and systems. Prentice Hall Press, New Jersey
(2002)
14. Nichols, R.A., Bulfin, R.L., Parker, R.G.: An interactive procedure for minimizing
makespan on parallel processors. International Journal of Production Research 16(1), 77–
81 (1978)
15. Lian, Z., Jiao, B., Gu, X.: A similar particle swarm optimization algorithm for job-shop
scheduling to minimize makespan. Applied Mathematics and Computation 183(2), 1008–
1017 (2006)

Biography of Authors
Chyuan Perng is an associate professor in the Department of Industrial Engineering and En-
terprise Information at Tunghai University, Taiwan. He received his Ph.D. degree in Industrial
Engineering from Texas Tech University, USA. He has also participated in numerous industrial
and governmental projects in Taiwan.

Yi-Chiuan Lai is an assistant professor in the Department of Business Administration at

Providence University, Taiwan. He received his Ph.D. degree in Industrial Engineering from
Iowa State University, USA. His primary research interests lie in simulation, scheduling, analy-
sis of stochastic system, and product realization processes.

Zih-Ping Ho is currently a Ph.D. candidate in the Department of Industrial Engineering and

Enterprise Information at Tunghai University, Taiwan. He also owns a system development
Studio in the Innovation Center of the same school. He has also participated in several indus-
trial and governmental projects in Taiwan.

Chin-Lun Ouyang graduated from the Department of Industrial Engineering and Enterprise
Information at Tunghai University, Taiwan. He received a Master and Bachelor degree. He
currently serves at the Conscription Agency Ministry of the Interior in Taiwan.
 A New Approach for Chest CT Image Retrieval*

Li-dong Wang1,2 and Zhou-xiang Shou1

1
College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang,
310027, China
2
Hangzhou Normal University, Hangzhou, ZheJiang, 310012, China
Violet_wld@163.com, shoupan@hz.cn

Abstract. A new approach for chest CT image retrieval is presented. The proposed
algorithm is based on a combination of low-level visual features and high-level
semantic information. According to the new algorithm, wavelet coefficients of the
image are computed first using a wavelet transform as texture feature vectors. The
zernike moment is then used as an effective descriptor of global shape of chest CT
images in database, and the semantic information is extracted to improve the ac-
curacy of retrieval. Finally, index vectors are constructed by the combination of
texture, shape and semantic information, and the technique of relevance feedback
is used in the algorithm to enhance the effectiveness of retrieval. The retrieval re-
sults obtained by application of our new method demonstrate an improvement in
effectiveness compared to other kinds of retrieval techniques.

Keywords: Medical image retrieval, Wavelet transform, Zernike moment,

Semantic information, Chest CT image.

1 Introduction
With DICOM (Digital Image Communication and Medician), information of patient
can be stored with the actual images. The digitally produced medical images are pro-
duced in everincreasing quantities and used for therapy and diagnostics. The medical
imaging field has generated additional interest in methods and tools for the manage-
ment, analysis, and communication of these medical images. Many diagnostic imaging
modalities are routinely used to support clinical decision making. It is important to
extend such applications by supporting the retrieval of medical images by content.
In recent years, much research has been done into specific medical image retrieval
system[1-4], and the research for medical image retrieval is based on the same anatomic
region. Chest CT is gray-scale image, so the retrieval method based on color feature can’t
attain effective performance. Meanwhile, extensive experiments in content-based image
retrieval(CBIR) systems have demonstrate that low-level features can not always de-
scribe high-level semantic concepts in the user’s mind[5]. Therefore, this paper presents a
new approach based on combination of low-level visual features and high-level semantic
information. Firstly, we extract stexture feature of images by wavelet transform. The
*
Sponsored by the Science and Technology Foundation of Hangzhou Normal University
(2009XJ065).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 535–543, 2009.
© Springer-Verlag Berlin Heidelberg 2009
536 L.-d. Wang and Z.-x. Shou

wavelet transform based texture analysis method has proven to be an efficient method for
texture analysis due to its property of both space and frequency localization in this
specified manner. Then a region-based shape[6] descriptor is presented, which utilizes a
set of the magnitudes of Zernike moments. We show that the Zernike moment descriptor
can be used effectively as a global shape descriptor of an image, especially for a large
medical image database, and then describe the process of extracting Zernike moment
descriptor. Finally, in order to improve the retrieval performance, texture and shape
features are combined with the high-level semantic feature to reduce the “semantic gap”
between visual features and the richness of human semantics, and the technique of rele-
vance feedback is used in the algorithm to enhance the effectiveness of retrieval.
Thus, based on the above technology, a simple prototype system is developed to
compare the retrieval accuracies. Experimental results show that the method discussed
in this paper is much more effective.
The paper is organized as follows. Section 2 gives a review of feature extraction,
including wavelet transform, Zernike moments and semantic feature. In section 3, we
proposed a method to combine texture, shape features with semantic information. In
section 4, experiments have been conducted and the performance of the proposed al-
gorithm is analyzed. The conclusions are presented in section 5.

2 Features

2.1 Wavelet Transform

It is found that various approaches to texture analysis are very diverse and in this re-
spect, four categories can be defined, which are namely statistical, geometrical, model
based and signal processing. The first three texture analysis methods are found suitable
for the regular and near regular methods, and the texture information of chest CT
images is focused on the middle region. Therefore, we utilize the signal processing
algorithm. The wavelet transform based technique performs the space-frequency de-
composition with low computational complexity, it has been proven to be an efficient
method for texture analysis[7]. In this paper, Daubechies’ wavelet is selected as bases
due to its orthonormal characteristics.
The experimental procedure for chest CT image texture feature extraction method is
as follows:
(1) Chest CT images are preprocessed and the sizes of the images are made to be same
to make the system computationally less complex.
(2) A two dimensional Daubechies’ wavelet transform is applied to decompose the
image into its sub-band images. Three-level pyramidal decomposition is used in
this work.
(3) All the sub-band images are stored for calculating the statistical features from the
sub-bands.

Fig.1 shows Daubechies’ wavelet based decomposed three sub-bands that are created from
512×512 chest CT image. In order to extract texture feature, all sub-band images from three
levels at different resolutions will be calculated for feature vector. During statistical feature
 A New Approach for Chest CT Image Retrieval 537

calculation, if the decomposed sub-band imagef(x,y) is with dimension(M,N), then the

various statistical features like mean, variance, energy and entropy feature vectors of
sub-band images are calculated using the notations[8] which are as follows:

Fig. 1. 3-scale 2D wavelet transform from chest CT image

M N
1
Mean =
MN
∑∑ f ( x, y ) (1)
x =1 y =1

M N
1
Variance = ∑∑ f ( x, y ) − Mean
2
2
(2)
( MN ) x =1 y =1

M N
1
Energy = ∑∑
2
2
f ( x, y ) (3)
( MN ) x =1 y =1

M N
Entropy = −∑∑ f ( x, y ) log f ( x, y ) (4)
x =1 y =1

These four values are considered to be beneficial to be used as texture features and then
calculated from all the sub-bands of chest CT images. Then the similarity between two
images can be measured by Euclidean distance.

2.2 Zernike Moment Descriptor

The Zernike moment descriptor has such desirable properties: rotation invariance,
robustness to noise, expression efficiency, fast computation and multi-level represen-
tation for describing the various shapes of patterns[9-10]. Therefore, we apply the Zernike
moments to chest CT image retrieval for shape feature extraction. We can calculate the
Zernike moment feature by (7) and (8).
538 L.-d. Wang and Z.-x. Shou

1 π
2n + 2
Cnl =
π ∫∫R nl
(r ) cos(lθ ) f (r , θ ) rdrdθ (7)
0 −π

1 π
2n + 2
S nl =
π ∫∫R nl
( r ) sin(lθ ) f ( r , θ ) rdrdθ (8)
0 −π

where l ≥ 0 and n > 0 .

In our experiment, 9 Zernike moments of order 0 to 4 in n and l are extracted from
the images, and the similarity between two images is measured by Euclidean distance.

2.3 Semantic Feature Extraction

Images in medical image database are in high similarity, it is difficult to retrieve the
images with the same pathologic feature. The low-level feature extraction can not attain
effective result. In order to retrieval accuracy of retrieval system, we try to reduce the
“semantic gap” between visual features and the richness of human semantics. In this
paper, we propose a method to extract semantic information.
Because the descriptions are subjective, different doctors have different diagnosis
words for each image, but they have the similar meaning. Considering the doctor’s
diagnostic description (Table 1) for the images, we can extract keywords from these
descriptions with doctor’s instruction. According to medical knowledge, we derive the
keyword sets(Table 2) as semantic information. Let Li be the semantic feature vectors,
Li =(Wi1,Wi2,…Wim), where m represents the number of keyword derived from de-
scription, Wij is the weight of keyword j from the semantic information of image i. In
order to reduce the computation complexity, the weight is equal to the number of times
that keyword appears in the semantic information, if it does not appear, Wij=0.

Table 1. Doctor’s diagnostic description of image

Image Doctor’s diagnostic description of Chest CT image

6003 superior lobe of right lung Ca with review of middle lobe of right lung, me-
tastatic carcinoma of mediastinal lymph node: chest symmetry, more textures
in pulmones, disorder, apicale segmentum of right lung’s superior lobe and
right lung’s middle lobe both have a nodular contour, no block in all levels of
bronchus, mediastina have multi-auxe lymphoid nodes, no adhaesio, effusion
in bilateral pleura.

Table 2. Keywords set

Keywords The number of images

0030 0031… 1001 1002
chest symmetry 1 1 1 1
more textures 0 0 1 0
shadow 1 0 0 0
nodular contour 0 0 1 1
thoracic deformity 0 0 0 0
 A New Approach for Chest CT Image Retrieval 539

Let p, q represent the texture features of the query image and the database image
respectively. The distance between them can be computed as:
m

∑W qk
× Wpk
D ( p, q ) = 1 − k =1 (9)
m m
(∑ Wqk 2 )(∑ Wpk 2 )
k =1 k =1

where m represents the number of keywords.

3 Feature Fusion

The retrieval method based on low-level features can not attain efficient performance.
To improve the effectiveness and accuracy, we propose a method to combine low-level
feature with high-level feature. The low-level features contain texture and shape. In
order to enhance the effectiveness of retrieval, the technique of relevance feedback is
used in the algorithm. The Rocchio algorithm[12], which uses user feedback on the
relevancy of retrieved images, has been shown to improve query results. Fig.2 illus-
trates the major steps of the integrated retrieval.

Chest CT image

Wavelet decomposition Size

in 3 scales normalizat
Construction of
Keyword sets
Square-circular
Construction of transformation
texture feature

Zernike moments
Similarity measure computation
Similarity measure

Similarity measure

Normalization

Retrieval Relevance Feedback

Fig. 2. Integrated retrieval algorithm

540 L.-d. Wang and Z.-x. Shou

Since different feature’s distance has different standard, before integrated retrieval
we should have a normalization.

4 Results and Discussion

The method described above has been extensively applied on 217 chest CT images,
which are from different patients in BMP format, represented by 512 × 512 size. In par-
ticular, a retrieved image is considered a match if it is in the same pathological modality
as the query. This assumption is reasonable, since the images in same anatomic region of
the same disease are chosen so that it can support clinical decision making.
Fig.3 illustrates a result based on wavelet transform. The retrieval result returns 30
similar images. The pathological modality of the query image is lung cancer. Accord-
ing to the medical knowledge, there are 24 returned images are in the same disease.
Meanwhile, Fig.4 demonstrates an example query based on feature fusion. It shows that
there are 27 similar images returned.
In system evaluation, the retrieval precision and average-r are calculated as statis-
tical parameters.
a
precision = (12)
a+b

Fig. 3. Example of a query based on wavelet transform

 A New Approach for Chest CT Image Retrieval 541

Fig. 4. Example of a query based on feature fusion

where a represents the number of similar images and b is the number of un-similar
images in the results. That is, the percentage of similar images retrieved with respect to
the number of retrieved images.
m
1
Average − r =
m
∑ρ r
(13)
r =1

where m is a given constant, and in this paper m = 8, ρ r is the rank of rth image in the
first m similar images. Therefore the ideal value of ρ r = 4.5, and the lower numbers
indicate better results. In our case, the retrieval results return 15, 30, 45 images for each
query. Tab.3-5 show the average precision, rank of different methods.
Tab.3 compares the results obtained by three texture extraction algorithms, which
are co-occurrence matrix, texture spectrum and wavelet transform. A total average
84.5% matched retrieved images is achievable using the wavelet transform algorithm.
The same experiment, performed using co-occurrence or texture spectrum, demon-
strated lower performance on average precision and average-r.
542 L.-d. Wang and Z.-x. Shou

Table 3. Results of texture indexing retrieval algorithms

parameters Co-occurrence matrix Texture spectrum Wavelet transform

15 30 45 15 30 45 15 30 45
precision 70.0% 67.8% 65.0% 80.5% 78.2% 75.0% 87.0% 84.5% 79.2%
Average-r 6.900 7.250 7.200 6.750 6.875 6.900 5.000 5.000 5.750

Table 4. Results of retrieval based on shape

parameters Hu moments Zernike moments

15 30 45 15 30 45
precision 59.0% 50.0% 45.8% 76.0% 70.8% 69.0%
Average-r 7.170 8.520 8.640 6.900 7.250 7.200

Table 5. Results of retrieval based on relevance feedback

parameters Feature fusion(first retrieval) Feature fusion(second retrieval)

15 30 45 15 30 45
precision 92.5% 88.2% 83.0% 93.3% 90.2% 85.0%
Average-r 4.500 4.500 5.500 4.500 4.500 5.000

Tab.4 shows the quantitative results obtained by the application of integrated re-
trieval and shape algorithms. As illustrated, the feature fusion algorithm has a fairly
high precision and small average-r, while demonstrating higher retrieval accuracy.
Meanwhile, Zernike moment performed better than Hu moment. This is expected and
mostly due to its property of multi-level representation for describing the various
shapes of patterns and its rotation invariance.
Tab.5 shows the precision of second retrieval is higher than original retrieval. And
the average rank reaches ideal value. Therefore the method of relevance feedback is
successful with high retrieval efficiency.

5 Conclusion
In this paper, a new approach for chest CT image is presented. Texture and shape features
are combined with semantic information for image retrieval. The experimental results
show that the proposed algorithm is effective. Applying this method to medical image
retrieval will support the clinical decision making and have an enormous potential.
However, further developments should be made in order to improve the retrieval
accuracy. The semantic feature vector should be modified to reduce its size and im-
prove its discriminative property.
 A New Approach for Chest CT Image Retrieval 543

Acknowledgement
The authors would like to thank Radiotherapic Department, Intervention Radiotherapic
Department and Endoscopic Department of Inner Mongolia Hospital.

References
[1] Wan, h.-l., et al.: Texture feature and its application in CBIR. Journal of Computer-aided
Design and Computer Graphics 15(2), 195–199 (2003)
[2] Shyu, C., Brodley, C., Kak, A., et al.: ASSERT, A physician-in-the-loop content-based
image retrieval system for HRCT image databases. Computer Vision and Image Under-
standing 75(1/2), 111–132 (1999)
[3] Aisen, A.M., Broderick, L.S., Winer-Muram, H., et al.: Automated storage and retrieval of
thin section CT images to assist diagnosis: System description and preliminary assessment.
Radiology 228(1), 265–270 (2003)
[4] Sun, J., Zhang, X., Cui, J., Zhou, L.: Image retrieval based on color distribution entropy.
Pattern Recognition Letters 27(10), 1122–1126 (2006)
[5] Liu, Y., Zhang, D., Lu, G., Ma, W.-Y.: A survey of content-based image retrieval with
high-level semantics. Pattern Recognition (2006)
[6] Wen, C.-Y., Yao, J.-Y.: Pistol image retrieval by shape representation. Forensic Science
International 155(1), 35–50 (2005)
[7] Paschos, G.: Fast Color: Texture Recognition Using Chromaticity Moments. Pattern
Recognition Letters 21, 837–841 (2000)
[8] Borah, S., Hines, E.L., Bhuyan, M.: Wavelet transform based image texture analysis for
size estimation applied to the sorting of tea granules. Journal of Food Engineering 79(2),
629–639 (2006)
[9] Mehrotra, R., Gary, J.E.: Similar-shape retrieval in shape data management. IEEE Com-
put. 28(9), 57–62 (1995)
[10] Kim, W.-Y., Kim, Y.-S.: A region-based shape descriptor using Zernike moments. Signal
Processing: Image Communication 16, 95–102 (2000)
[11] Ortega, M., et al.: Supporting similarity queries in MARS. In: ACM Conf. on Multimedia,
pp. 403–413 (1997)
[12] Baeza-Yates, R., Ribeiro-Neto, B.: ModernInformation Retrieval. Addison Wesley,
Reading (1999)
 Depth Extension for Multiple Focused Images by
Adaptive Point Spread Function*

Yuanqing Wang and Yijun Mei

Box 1209, Department of Electronic Science & Engineering, Nanjing University No. 22,
Hankou Road, Nanjing 210093, P.R. China
yqwang@nju.edu.cn

Abstract. A novel depth fusion algorithm is proposed for multi-focused images

based on point spread functions (PSF). In this paper, firstly, a discrete PSF
model is present and functions with different parameters are prepared for the
proposed algorithm. Based on the analysis of the effect of PSF, the detail proc-
ess to fuse the depth of multi-focused images is described. By employing PSF
to convolve each original images and comparing with its adjacent one in the
image sequence, the focused and defocused region in the original images may
be located. Combining the the focused region in the images according to
Choose Max rules, an all-in-focus image may be got. The complexity of the al-
gorithm for an image series with more than two original images is discussed at
the end of this paper. Experimental results show that the image is distinctly
segmented into multi-regions and the image edge is legible as well. The pro-
posed algorithm based on PSF convolvetion as a focus measure has been shown
to be experimentally valid. The fusion results are satisfactory with smooth tran-
sitions across region boundaries.

Keywords: Point spread functions (PSF), depth fusion, multi-focused images,

image fusion.

1 Introduction
Depth of field is the area in front of and behind the specimen that will be in accept-
able focus. In any imaging system(e.g. video camera, light optical microscope, etc.),
as an inherent feature, image blurring is inevitable for all of the lens. The system
works with finite depth of field which is relative to the numerical aperture (NA) and
can not image all objects at various distances with the same clarity. Everything im-
mediately in front of or in back of the focusing distance begins to lose sharpness.
Depth fusion is performed at three different processing levels according to the
stage at which the fusion takes place: pixel, feature and decision[1]. In general, the
techniques for multi-focused image fusion can be grouped into two classes: (1)
Color related techniques, and (2) Statistical/numerical methods. Selecting the ap-
propriate approach depends strongly on the actual applications. Some commonly

*
This work is Funded by the key program from the National Natural Science Foundation of
China(60832003).

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 544–554, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Depth Extension for Multiple Focused Images by Adaptive Point Spread Function 545

used techniques in pixel-level fusion are: weighted combination; ptimization ap-

proach and biologically-based approaches such as neural networks and multi-
resolution decompositions.[2]
Physically relevant features of an image often exist at its different scales or resolu-
tions and may be exploit by multi-scale and multi-resolution approaches. The basic
idea in multi-scale and multi-resolution techniques is to decompose the source images
at first by applying the Laplacian pyramid[3], wavelet[4] or complex wavelet[5]
transform, then the fusion operation on the transformed images is performed and
finally the fused image is reconstructed by inverse transform. These techniques may
obtain spatial and frequency domain localization of an image and can provide infor-
mation on the sharp contrast changes. Multi-scale methods involve processing and
storing of scaled data at various levels which are of the same size as that of the origi-
nal images. This results in a huge amount of memory and time requirement [2]. The
support vector machines (SVM) is a classification technique that has outperformed
many conventional approaches in various applications. Shutao Li use the discrete
wavelet frame transform(DWFT) for the multi-resolution decomposition, and then
replace the choose-max rule by support vector machines (SVM) for fusing the wave-
let coefficients.[6] Vivek Maik present a method to extend depth of field.[7] A single
fused image is obtained by decomposing the source images using filter subtract deci-
mate in Laplacian domain, and then extract the depth of focus in the source images
using sum-modified-Laplacian. K. Aizawa[8] present an algorithm by convolving the
different focused image with blurring functions (Gaussian functions). After the nth
iteration, the sharp image may be obtained. This method is formulated based on a
linear imaging model and does not need any object segmentation or complex scene
content-dependent processing. But the iterative generation algorithm will be complex
and time-consuming will be great for multi-focused images more than three images.
We have research a novel algorithm based on Point Spread Function (PSF) which has
strong immunity from noise. In our work, PSF is applied as a tool to purposely blur
some specified in-focus region in an image, just like a typical lens system to image the
objects in the scene out of focus. With proposed method, the in-focus region in an image
may be accurately located. In this paper, in Section 2, we gave a model of a PSF along
with analysis of its effect to a focused image, and the criterion is described to detect
the blurring region and estimate the blurring status. In Section 3, the proposed algorithm
to obtain a depth fusion image from a series of original images is described in detail.
Experimental results and conclusion are presented in Section 4.

2 Detecting the In-focused Image by PSF

2.1 Basic Idea
Most lenses are not perfect optical systems. The light from a single object point pass-
ing through a lens, which with aperture of finite size in diameter, is converged into a
cone with its tip at the CCD (Charge Couple Device) plane, if the point is perfectly in
focus; or slightly in front of or behind the CCD plane, if the object point is somewhat
out of focus. In the out-of-focus case the point is rendered as a circle where the CCD
plane cuts the converging cone or the diverging cone on the other side of the image
point. This circle of confusion will introduce a certain amount of blurring.
546 Y. Wang and Y. Mei

Blurring can be numerically characterized by a Point Spread Function (PSF). PSF

is the output of the imaging system for an input point source. Assuming that intensity
distribution of scene, I(xs, ys, zs), is recorded by a typical lens system. For the object
out of focus with the point spread function, Pz, the image function, f(xi, yi), may be
calculated as below:
f ( xi , yi ) = I ( xs , ys , zs ) ∗ Pz (1)
where, * is the convolving operator.
In the similar way, a blurry image may be reconstructed from a well focused image.
The discrete blurry image f(k, l) for digital image may be obtained from the original
sharp image g(n, m) by convolving it with discrete PSF:
+∞ +∞
f (k , l ) = ∑ ∑ g (k + i, l + j)h(−i, − j)
i =−∞ j =−∞
(2)

where the function, h, is the discrete PSF.

Comparing equation (2) with (1), they are different in approach but equivalent in
result. If the discrete PSF, h, is suitably choiced, blurry image, f(k, l), will be the same
blurry as image function, f(xi, yi).

2.2 Series of PSF

According to the isoplanatism condition, PSF has the same shape over the whole field
of view at the same focal plane. The complicated PSF may be simplified to a concise
expression as:
1 (3)
P(r ) =
4r 2
where, r is the radius to the center of the blur circle, and for discrete case, it may be
denoted as a symmetric matrix, h, which’s diagonal elements are confirmed as:
⎧ i2 , i ≤ ( n + 1) / 2
hii = ⎨ (4)
⎩( n + 1 − i ) , i > ( n + 1) / 2
2

where, i=1,2,3,……,n, and n is odd number.

Define a PSF vector, a, as below:

ai = hii
(5)
Then, the matrix of discrete PSF is:
h = kaT a (6)
where, k is normalized coefficient, and
n
k = 1 (∑ ai ) 2 (7)
i =1
 Depth Extension for Multiple Focused Images by Adaptive Point Spread Function 547

For example, define a PSF vector as below:

a = (1, 4, 6, 4, 1) (8)

Then, the PSF matrix, h, is a 5×5 mask:

⎡1 4 6 4 1⎤
⎢4 16 24 16 4 ⎥⎥
1 1 ⎢ (9)
h= n aT a = ⎢6 24 36 24 6⎥
256 ⎢ ⎥
(∑ ai ) 2
i =1
⎢4 16 24 16 4⎥
⎢⎣ 1 4 6 4 1 ⎥⎦

A series of PSF may be used in our work, which is listed in Table 1:

Table 1. PSF vectors employed to fuse multi-focused images (P0(r) = 1)

PSF P0 P1 P2 P3 P4 P5 P6 P7
a 1,3,11,3,1 1,3,10,3,1 1,3,9,3,1 1,3,8,3,1 1,4,8,4,1 1,3,6,3,1 1,4,6,4,1 1,2,3, 2,1
k 441 400 289 256 361 196 256 81
WH(1) 0.52 0.61 0.70 0.85 0.91 1.1 1.3 1.4
Sr(2) 0.4187 0.3460 0.2803 0.2500 0.2244 0.1837 0.1406 0.1111
(1)WHthe width of PSF, i. e. the Full Width at Half Maximum of the PSF, (2)Sr=P(r)/P0(r), i.e. the central inten-
sity of PSF of a defocused image compared to that of a focused image.

Fig.1 shows the profile of each

PSF in Table 1. When we
convolve a focused image g(n,m)
hi
P0
with a discrete PSF, it is
P1
equivalent to getting a new image
f(k, l) by low-pass filter. We can P2
set different components in the P3
PSF matrix in Equation (8), and P4
make the blurry image f(k, l) to be
blurred with different degree P5
relative to the original sharp P6
image g(n,m). Different PSF pro- P7
vides different results for the
same original image. i
Clearly, suitable PSF must be
selected to simulate the out-focus-
blurring by imaging lens, and Fig. 1. The profile of different PSF
theoretically, infinite amount of
discrete PSF is required. In most case, the eight preparative functions in Table 1
are enough for depth fusion. The required PSF can may be estimated by inspecting
the image.
548 Y. Wang and Y. Mei

3 Depth Fusion by PSF

3.1 Identification the In-focus Region

As show in Fig.2, original image g1 and g2 are two images respectively with different
focused region. For easier to describe our idea, the scenes is simplified that it only
consists of two objectives which are respectively imaged within circular and
triangular regions in the image, see Fig.2. We describe the focused region with black
color, defocused region with dark gray. In image g1, the triangular region is within the
focused range and the circular region is the defocused region. In image g2, the
triangular region is out of focus and the circular region is the in focus.
The blurry image fi(k, l) may be obtained as below:
fi = g i ∗ hi (10)

where i=1,2, hi is point spread function, and * represents the convolution.

Convolved by different PSF, the original images will be blurred at different degree.

g1 f1 D1

○
H ○
-

g2 f2 D2

○
H ○
-

Fig. 2. Forming candidates of focus region

At next step, we obtain the candidate graphics Di, as shown in Fig. 2, by the follow-
ing function:
D1 = (∑ g 2 j + ∑ g 2 j ) − (∑ g1k + ∑ g1k ) ∗ h1
j∈S j∈B k ∈S k∈B (11)
D2 = (∑ g1k + ∑ g1k ) − (∑ g 2 j + ∑ g 2 j ) ∗ h2
k∈S k ∈B j∈S j∈B

where symbols, S and B, respectively denote the sharp and blur regions in the images.
Equation (11) may be rewritten as:
D1 = (∑ g 2 j − ∑ g1k ∗ h1 ) + (∑ g 2 j − ∑ g1k ∗ h1 )
j∈S k∈B j∈B k∈S (12)
D2 = (∑ g1k − ∑ g 2 j ∗ h2 ) + (∑ g1k − ∑ g 2 j ∗ h2 )
k ∈S j∈B k ∈B j∈S
 Depth Extension for Multiple Focused Images by Adaptive Point Spread Function 549

In order to enhance the absolute dispersion and avoid the negative value, we use the
square number to replace the direct subtraction as below:

D1 = (∑ g 2 j − ∑ g1k ∗ h1 ) 2 + (∑ g 2 j − ∑ g1k ∗ h1 )
2

j∈S k∈B j∈B k∈S (13)

D2 = (∑ g1k − ∑ g 2 j ∗ h2 ) + (∑ g1k − ∑ g 2 j ∗ h2 ) 2
2

k∈S j∈B k∈B j∈S

The focused region in one image probably present as blurry region out of focus in
another image. We simply treat the area of the sharp region in one image as the same
as that of the blurry region in another image, which means that:
Area(∑ g 2 j ) ≈ Area(∑ g1k )
j∈S k ∈B (14)
Area(∑ g1k ) ≈ Area(∑ g 2 j )
k ∈S j∈B

It’s apparent that the more desirable the mask coefficients of PSF are, the lower
values of the second term in polynomial (13) will be. If set an identification thresh-
old, Vth, which is large than the values of the first term in Equation (13) and small
than that of the second term. All of the pixels in g1, g2, which behave with variant
larger than identification threshold in the candidate graphics, D2, D1, may be treated
as in-focus imaging region. The identification threshold is relative to the scene,
such as content, texture and illumination. For idiographic applied field, it may be
confirmed via statistic data from experiments or derived from experience. Then a
binary feature image, F1, F2, may be obtained in which the high gray level pixel is
sharp pixel for correspondence original image. The desirable PSF may be chosen
from one of the candidate vectors listed in Table 1, according to blurry degree of
the image.
In fact, if the discrete PSF is m×m matrix, the area of the high gray level pixel in
binary feature images, F1, F2, will be little large than that of the sharp region in origi-
nal images, g1, g2. Considering this condition, we use the fusion principle to obtain the
fusion image, I(x, y), as below:
⎧ g1 ( x, y ) ; F2 ( x, y ) = 1, F1 ( x, y ) = 0
⎪ (15)
I ( x, y ) = ⎨ g 2 ( x, y ) ; F2 ( x, y ) = 0, F1 ( x, y ) = 1
⎪[ g ( x, y) + g ( x, y )] / 2
⎩ 1 1 ; other

For the case of more than two original multi-focus images, an iterative algorithm may
be applied. Assume to there are n images, g1, g2, ……, gn, the steps of iterative algo-
rithm are taken as below:
STEP 0: g1= g1, g2= g2, i=2;
STEP 1: confirm the PSF mask according to the status of g1 and g2;
STEP 2: obtain binary feature images, F1, F2;
STEP 3: obtain the fusion image, I(x, y), as below:
550 Y. Wang and Y. Mei

⎧ g1 ( x, y ) ; F2 ( x, y ) = 1, F1 ( x, y ) = 0
⎪ ; F2 ( x, y ) = 0, F1 ( x, y ) = 1
⎪ g 2 ( x, y ) (16)
I ( x, y ) = ⎨
⎪ g 1 ( x , y ) ; F2 ( x, y ) = 0, F1 ( x, y ) = 0
⎪⎩[ g1 ( x, y ) + g1 ( x, y )]/ 2 ; F2 ( x, y ) = 1, F1 ( x, y ) = 1

STEP 4: i = i+1, if i >n go to STEP 6;

STEP 5: g1= I(x, y), g2= gi, go to STEP 1;
STEP 6: STOP;

3.2 Candidate of PSF

The selection of appropriate PSF has a close connection with blurring degree of the
region. So it’s important to estimate image blur by some appropriate method.
Suppose that Ii(x, y) and Ij(x, y) are two differently focused images in a series of multi-
focus images with different focus settings, a differential image is defined as below:
I i , j ( x, y ) =| I i ( x, y ) − I j ( x, y ) |
(17)
where Ik(x, y) is the pixel intensity value at position (x, y) in image k. Generally, set
j=i+1 correspond to the adjacent image of the ith one.
The value of Ii,j(x, y) may indicate the blur degree. But use of solely adjacent pixel-
level indicators can make decisions vulnerable to wide fluctuations dependent on spe-
cific parameters such as noise, brightness and local contrast. Hence, corroboration
from neighboring pixels of decision choices becomes necessary to maintain robustness
of the algorithm in the face of the above adverse effect. Adding this corroboration
while maintaining pixel-level decisions requires summing the Ii,j(x, y)s over an l×l
region surrounding each decision-point. This yields an indicator for focus measure:
l/2
Bi , j ( x, y ) =
τ
∑
=− l / 2
Ii , j ( x + τ , y + τ ) (18)

where, aggregation size, k, is constant number decided by image resolution. Range of

aggregation size is about 20~50 for different resolution.
The larger the relative magnitude of the indicator, the higher the probability that its
corresponding blur circle radius is larger, and the larger FW is of the appropriate PSF
in Table 1. The continuous value Bi,j(x, y) may be put into several discrete subsections
correspond to preset PSF.

4 Experiment Result
In this section, we experimentally demonstrate the effectiveness of the proposed algo-
rithm. Experiments were performed on 256-level original images of size 800×600 by
Matlab 6.0 in PC with Pentium 1.4GHz CPU and 628M RAM. Experiments were
carried out on two actual scenarios, which contain multiple objects at different dis-
tances from the camera one or more objects naturally become out of focus when the
image is taken. For example, the focus is on the trunk in Fig 3(a), while that in Fig
3(b) on the shrub and bench.
 Depth Extension for Multiple Focused Images by Adaptive Point Spread Function 551

(a) image A1 of scene A (b) image A2 of scene A

Fig. 3. Original image pair

Convolved by PSF, the original images will be blurred as shown in Fig.4. We use
two PSF to respectively blur the original image, by PSF with vector a = (1, 4, 6, 4,
1) to get blurry image B1 and a = (1, 2, 3, 2, 1) to get blurry image B2 as illustrated
in Fig.4.

(a) blurry image B1 from image A2 (b) blurry image B2 from image A1

Fig. 4. Blurry image pair

The fused image, with chosen identifica-

tion threshold equaling to 152, is shown in
Fig. 5. The fused image merges the respec-
tive portions that are in focus from the
original images and all objects in the im-
age are sharp. It can be seen in the final
fused image that any undesired disconti-
nuities at the region boundaries are pre-
vented. The elapsed time is 1.362s to
achieve the composite result.
Fig.6 shows some images with another
typical content to depict the similar process. Fig.5. Result image of fusion
552 Y. Wang and Y. Mei

The result of image fusion applied to two images having different depth of focus, with
the focus in on the scrip in Fig 6(a), while that in Fig.6(b) on the book. Please notice
that, in Fig.6(b), the scrip is blurred by Gaussion blur and little moving blur. The fused
image in Fig.6(c) is gotten by PSF with vector a = (1, 4, 6, 4, 1) to successively con-
volve the original images with two times, 178 for identification threshold. It takes
1.322s to finish the conduct.

(a) Original image 1 (b) Original image 2 (c) Fused image

Fig. 6. Second example for the proposed algorithm

Fig. 7. Image series of different focused range

 Depth Extension for Multiple Focused Images by Adaptive Point Spread Function 553

Fig.7 shows some images with another

typical content to depict the similar process.
The image series includes eight different
focused images captured by a microscope.
The depth distribution of the scene is shown
at the right-bottom, in which the more dark
is the gray level, the more close the object is
to the lens of imaging system. Result image
with depth extension is shown in the bottom
in Fig. 8. It is gotten by PSF with vector
a = (1, 4, 6, 4, 1), 178 for identification
threshold. Fig. 8. Depth extension for image series

5 Conclusion
The defocus portions in one image represent dissimilar blurry degree, thus the vector
of PSF is different for original image. The “ideal” way to choose the appropriate vec-
tor is accurately evaluate the blurry degree of each image’s region. The blurry degree
may be confirmed by different way such as gray gradient difference, wavelet trans-
form and so on. In mathematics, the local regularity of the image function is often
measures with Lipschitz exponent. Our method for blur estimation is based on calcu-
lating the Lipschitz exponent in all points where a change in intensity is found either
in the horizontal or vertical direction. The strategy for detecting and characterizing
singularities of the original images is important for this fusion theorem. The detail
method may be described in the other paper. In fact, assisting in correlative technol-
ogy, the proposed algorithm may provide acceptable results in most cases.

References
1. Zhang, Z., Blum, R.S.: A categorization of multiscale-decomposition-based image fusion
schemes with a performance study for a digital camera application. Proc. IEEE 87(8), 1315–
1326 (1999)
2. Piella, G.: A general framework for multiresolution image fusion: from pixels to regions.
Information Fusion 4, 259–280 (2003)
3. Maik, V., Shin, J., Paik, J.: Pattern Selective Image Fusion for Multi-focus Image Recon-
struction. In: Gagalowicz, A., Philips, W. (eds.) CAIP 2005. LNCS, vol. 3691, pp. 677–684.
Springer, Heidelberg (2005)
4. Pajares, G., de la Cruz, J.M.: A wavelet-based image fusion tutorial. Pattern Recognition
(2004)
5. Forster, B., Van de Vill, D., et al.: Complex Wavelets for Extended Depth-of-Field: A
NewMethod for the Fusion of Multichannel Microscopy Images. Microscopy Research and
Technique 65, 33–42 (2004)
6. Li, S., Kwok, J.T., Tsang, I.W., et al.: Fusing Images with Different Focuses. Using Support
Vector Machines 15(6), 1555–1561 (2004)
554 Y. Wang and Y. Mei

7. Maik, V., Shin, J., Paik, J.: Pattern Selective Image Fusion for Multi-focus Image Recon-
struction. In: Gagalowicz, A., Philips, W. (eds.) CAIP 2005. LNCS, vol. 3691, pp. 677–684.
Springer, Heidelberg (2005)
8. Aizawa, K., Kodama, K., Kubota, A.: Producing object-based special effects by fusing mul-
tiple differently focused images. IEEE Transactions on Circuits and Systems for Video
Technology 10(2), 323–330 (2000)

Wang Yuanqing graduated from Zhejiang University, Hangzhou,

Zhejiang, P.R.C. and received his Master degree from the same Uni-
versity, all in the optical & electrical engineering, received his Doc-
torial degree from Nanjing University,. Now, he acts as a professor in
the department of electronic science & engineering in Nanjing Univer-
sity. He has been involved in research work on auto-stereoscopic
display, non-intrusive human-computer interface, . His present re-
search interest includes the opti-electronics and information process,
especially in sensor, 3D vision and display. He has compered many
research projects such as National Hi-tech Research and Develop Plan
(863 Plan) of China, National Natural Science Foundation of China,
Jiangsu Hi-tech Plan, the 11th Five Years Research Plan and so on.
yqwang@nju.edu.cn
 Image Annotation Using Sub-block Energy of Color
Correlograms

Jingsheng Lei

School of Computer and Information Engineering, Shanghai University of Electric Power,

Shanghai 200090, China
jshlei@126.com

Abstract. This paper proposes an algorithm using local energy of color corre-
lograms without any explicit using sub-block energy of color correlograms. The
sub-block energy is defined as sub-windows from color correlograms informa-
tion based on color distribution of original image. The model for image annota-
tion involves computing histogram using color correlograms and analysis its
sub-block characteristics. Similarly, sub-block energy is applied to annotate im-
age’s class and got a satisfied result. The model is fast and invariant to image’s
size and rotation. The comparison with SVM is done by experiments demon-
strate the model is quite successful in annotation of image.

Keywords: Image Annotation, Color Correlogram, Sub-block Engery.

1 Introduction
With the development of network and multimedia technology, the storage of image
information is expanding quickly and image retrieval has become a hotspot of image
research field [1-2]. Content-based image retrieval (CBIR) which uses image content
such as color and texture to compute the similarity of images have succeeded in fin-
gerprint and logo recognition, etc. However, there is a huge gulf which is called “se-
mantic gap”, the lack of coincidence between the low-level features extracted from
the visual features and high-level semantics concept, result that CBIR cannot provide
meaningful results, and current state-of-art computer vision technology lags far be-
hind the human’s ability to assimilate information at a semantic level [3].
Using different models and machine learning methods to find the relation between
image visual features and keywords for them from labeled images, and propagating
keywords to unlabeled images, this is image annotation. There are three main meth-
ods for image annotation: annotation based on image segmentation, fixed size block
and annotation based on image classification. Annotation based on image segmenta-
tion [4] depends on image visual features and precise results of segmentation [5-7].
The main point is how to correspond features of the regions to keywords. In the ideal
condition, every segmented region corresponds to one object. However, the outcome
of image segmentation is not satisfied at present. So there exists a big diversity be-
tween the expression of image object level part and human vision system. And this
problem also exists in fixed size image division, for it may divide one object into

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 555–562, 2009.
© Springer-Verlag Berlin Heidelberg 2009
556 J. Lei

several blocks or put several objects into one block. Compared with the two methods,
image annotation based on image classification can avoid the low accuracy caused by
wrong image division. Cusano et al. [8] categorized images into various groups with
Support Vector Machines (SVMs) and counted the frequencies of co-occurrence be-
tween the keywords in each group. They then used them to annotate an input image.
Gauld et al. [9] treated features separately and trained the optimal classifier for each
feature in advance. But these methods using features extracted from whole images
that result in high computing cost which is not suitable for large data set. To avoid the
problem mentioned above, this paper realize an annotation procedure by sub-block
energy based on color correlograms [10] which indicates the color and spatial distri-
butions of the image directly.
To address the problems mentioned above, in this work, a fast algorithm for image
annotation based on color correlograms is proposed. Using sub-block energy of corre-
lograms, the algorithm can find the relative energy blocks in Euclid space. Differ
from the traditional annotation approach based on the whole image features, the pro-
posed method analyze the central area which associate with the image semantics and
only vision features of the area are extracted, then LIBSVM is used for image classi-
fication to get the relationship between image visual features and semantics, lastly
image annotation is executed.
The remainder of this paper is organized as follows: The sub-block energy of color
correlogram are introduced in section 2 and a fast image annotation approach based
on central area analysis and key issues are discussed in detail in section 3. In section
4, experimentation results are presented and the concluding remarks are given in the
last section.

2 Sub-block Energy of Color Correlograms

2.1 Color Correlograms

Traditionally the distance of points in an image can be measured by Euclid space.

Correlograms are graphs giving the information of auto-correlation changes with
distance, it captures the spatial correlation of colors in an image, rectifies the major
drawbacks of the classical histogram method
For an image distributed in a discrete domain I ⊂ Z 2 and its gray level is L,
ci , i = 1...C ,where ci is the i color index and C ≤ L is the gray level for statistic,.
hci is the histogram value of color correlogram for ci .
In Fig.1., a window wh × wv is used as a distance to compute histogram of gray
value of p1 with p2 and p3 . When moving the window in the image, we can get
a whole histogram table values which indicates the distance of ci → c j in C .
Therefore, If we want the color auto-correlograms, we can set i = j as the value of
histogram.
 Image Annotation Using Sub-block Energy of Color Correlograms 557

Fig. 1. Array image and its color correlograms window

2.2 Sub-block Energy

The correlograms shows the changes of the color histogram with the spatial distribu-
tion of pixels.
The color correlogram is selected to implement the sub-block energy model because
it is color histogram with spatial information. It signals the presence of local color dis-
tribution. In the sub-block model, a given image is computed with the feature of corre-
logram. Consequently, the correlogram is pieced into sub-blocks. The sub-blocks are

(a) dinasours

(b) horses

Fig. 2. Original images and its corresponding color correlograms

558 J. Lei

computed by absolute gradient of correlograms. Here we define the absolute gradient of

hci as below
Cov(k ) =| hci − hc | (1)
In order to use sub-blocks to annotate the image, we should erase the spatial changes
of sub-blocks. In Fig.2., we can note that the color correlograms is different and the
corresponding sub-block energy changes dynamically. The curve line of sub-block is
similar when they belong to a class while different classes are very different. Covari-
ance values of correlograms give variability of the statistic of gray color. The sub-
block energy is computed by summing the squared values of block.
( k )*blocksize
Ek = ∑
i = ( k −1)*blocksize
COV (i ) (2)

For the implementation of annotation of image, each type of images has their
uniquely identifiable features which can discriminate from others, surely other fea-
tures can be combined for annotation, but here we only put the sub-block vectors into
high-space. When a test sample is filtered into its sub-block energy, we can compute
its Euclid distance in high-dimensional space.

3 Sub-block Energy Algorithm

In this paper, the basic principle of sub-block energy algorithm is indicated as: firstly,
chose a proper gray level to compute color correlograms in order to produce the spa-
tial color distribution, and then compute the absolute gradient of color correlograms,
finally, select a block size in order to get the sub block energy and order them by the
energy size. The steps of sub-block energy are given as follows.

3.1 The Sub-block Energy Procedure

Input: The set of training images is Straining and the set of testing images is
Stesting . wh × wv is selected as window size. Levelwidth is the step size of L for
histogram and sub-block size is Sub − Block .
L = 256 , C = 256 / levelwidth .Image size is W × H and often W = H .
Output: Sub-Block Energy of Images

Begin
I ∈[S
training ,S
testing ]
For each
I ← I / levelwidth , C ← 256 / levelwidth ;
TempC ← [1: C ;1: C ] ← (0, 0);
i ∈ [1: H − wv ]
For each
j ∈ [1: W − wh ]
For each
 Image Annotation Using Sub-block Energy of Color Correlograms 559

m ← I (i, j ) + 1;
n ← I (i + wv , j + wh ) + 1;
TempC ( m, n) ← TempC (m, n) + 1;
TempC ( n, m) ← TempC (n, m) + 1;
End
End
i ∈ [1: H − wv ]
For each
j ∈ [ wh + 1: W ]
For each
m ← I (i, j ) + 1;
n ← I (i + wv , j − wh ) + 1;
TempC ( m, n) ← TempC (m, n) + 1;
TempC ( n, m) ← TempC (n, m) + 1;
End
End
For each i ∈ [1: C ]
C
hc (i) ← [TempC (i, i) / ∑ TempC (i, i )];
i =1
End
For each i ∈ [1: C ]
Cov(i ) ← hc(i ) − hc ;
End
For each k ∈ [1: C ]
( k )*blocksize
Ek ← ∑
i = ( k −1)*blocksize
COV (i)
;
End
E ← Sort ( E ) ;
End
End

After getting the results of sub-block energy, we can use it for annotating a testing set.
When an unknown image is input into the model, firstly, its sub-energy is calculated
base on the algorithm. For the reason of the sort of energy, there are is no need to
align at location. When we describe energy sub-block as the point of dimensional
space C / blocksize . We can use the Euclid distance to measure the class.

3.2 The Annotation Procedure

Input: The sub-block of training images with annotation and the sub-block of testing
images without annotation.
Output: The set of testing with annotation.
560 J. Lei

Begin
etesting ∈ Etesting
For each
D ← [];
etraining ∈ Etraining
For each
D ← [ D, Dis tan ceEculid (etesting − etraining )];
End
D ← Sort ( D );
L(etesting ) ← L({etraining | etraining = min( D)etraining })
End
End

After the annotation, images are labeled. In order to improve the performance, the
training set energy are represented by vectors. In Additional, the algorithm can paral-
lel executed to speed up the efficiency of computing.

4 Experimental Results
In order to verify the correctness and effectiveness of our proposed solution, an anno-
tation system is developed. The scheme is tested on the real images with different
semantic. The color correlogram parameter levelwidth is set as 8, oth-
ers( wh , wv , blocksize ) are variable. For the LIBSVM [11], its parameters are s=c-
svc, c=sigmoid, gama=1/k, degree=3. Firstly, the testing image is transformed into

Fig. 3. Annotation Result (a)image to be annotated; (b)first match image; (c) match image after
first; (d) match image after second
 Image Annotation Using Sub-block Energy of Color Correlograms 561

gray where range value of pixel is 0 to 255. Using the proposed annotation, we can
label the testing image properly and also we can get a list of compared images based
on distance from testing image to training one. And also, the list gives the information
of similar images. From the Fig.3., we can notice that the testing image(dinosaurs) is
labeled correctly and the found list have some similar images(buses are the second
and third match image). Therefore, the sub-block energy algorithm is suitable for
annotation of images. On the other hand, the annotation results between sub-block
algorithm and LIBSVM are compared. The test is performed using 80 items for train-
ing and 20 for each category with cross validation. Table.1 shows the result. From the
table, we can find the sub-block algorithm is better than LIBSVM. The result indi-
cates that the sub-block energy algorithm uses the characteristics of features, and can
achieve some better correct annotation for images with complex background. And the
small sub-block energy can gives more high speed for large dataset.

Table 1. Annotation Compare of Sub-Block Energy and SVM (E :Sum-Block Energy; S:SVM)

Train=240 BlockSize

Test = 40 16 8 6 4 2

E 67.5 67.5 65.0 70.0 67.5

2
S 58.5 50.0 50.0 60.0 50.0

E 70.0 62.5 62.5 65.0 65.0

4
wv
S 55.0 55.0 50.0 45.0 50.0
=
E 67.5 60.0 67.5 62.5 67.5
wh 8
S 57.5 50.0 50.0 40.0 50.0

E 70.0 82.5 62.5 70.0 60.0

16
S 60.0 50.0 50.0 42.5 47.5

The sub-block energy also has the ability of find similar images from the measure
energy list. Therefore, the result of annotation for images is satisfied and the comput-
ing cost is low based on the proposed algorithm.

5 Conclusions
In this paper, we present an algorithm for annotation of images. The contributions of
this paper can be summarized as follows:
562 J. Lei

(1) The approach uses an image’s histogram and its spatial information with
small feature vector, it can reduce computing cost dynamically. The proposed
algorithm can applied to large image dataset.
(2) This paper proposes a model using sub-block energy based on color correlo-
grams. The representation of the sub-block is a small vector. The vector can
be fast computed and can be regarded as a criterion for image’s annotation
using spatial information. The algorithm can annotate images with different
complex background. The algorithm is robust.
Experimental results indicate our method is superior in terms of accuracy, robustness,
and stability when compared to SVM method. The novelty of this solution is to find
the local relationship of image for annotation. In addition, the proposed method can
reduce computing cost. More studies in the aspect of image’s color correlograms and
its application should be invested in future.

References
1. Smeulders, A., et al.: Content-based image retrieval at the end of the early years. IEEE
Trans. PAMI 22(12), 1349–1380 (2000)
2. Zhang, R., Zhang, Z.: Effective image retrieval based on hidden concept discovery in im-
age database. IEEE Trans. on Image Processing 16(2), 562–572 (2007)
3. Naphade, M., Huang, T.: Extracting semantics from audiovisual content: The final frontier
in multimedia retrieval. IEEE Trans. on Neural Networks 13(4), 793–809 (2002)
4. Xuelong, H., Yuhui, Z., Li, Y.: A New Method for Semi-Automatic Image Annotation. In:
The Eighth International Conference on Electronic Measurement and Instruments, vol. 2,
pp. 866–869 (2007)
5. Mori, Y., Takahashi, H., Oka, R.: Image-to-word transformation based on dividing and
vector quantizing images with words. In: Proceedings of the International Workshop on
Multimedia Intelligent Storage and Retrieval Management (1999)
6. Duygulu, P., Barnard, K., Freitas, J., Forsyth, D.: Object recognition as machine transla-
tion: Learning a lexicon for a fixed image vocabulary. In: Heyden, A., Sparr, G., Nielsen,
M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2353, pp. 97–112. Springer, Heidelberg
(2002)
7. Jeon, J., Lavrenko, V., Manmatha, R.: Automatic image annotation and retrieval using
cross media relevance models. In: Proceedings of the ACM SIGIR Conference on Re-
search and Development in Information Retrieval (2003)
8. Cusano, C., Ciocca, G., Scettini, R.: Image annotation using SVM. In: Proceedings of
Internet Imaging IV (2004)
9. Gauld, M., Thies, C., Fischer, B., Lehmann, T.: Combining global features for content-
based retrieval of medical images. Cross Language Evaluation Forum (2005)
10. Huang, J., Kumar, S.R., Mitra, M., Zhu, W.J., Zabih, R.: Image indexing using color corre-
lograms. In: Proc. 16th IEEE Conf. on Computer Vision and Pattern Recognition, pp. 762–
768 (1997)
11. Chang, C., Lin, C.-J.: LIBSVM: - A Library for Support Vector Machines [EB/OL],
http://www.csie.ntu.edu.tw/~cjlin/libsvm
 Optimal Evolutionary-Based Deployment of Mobile
Sensor Networks

Niusha Shafiabady1 and Mohammad Ali Nekoui2

1
Azad University, Science and Research Branch
2
K.N.T. University of Technology
nshafiabady@yahoo.com, manekoui@eetd.kntu.ac.ir

Abstract. Sensor deployment is an important issue in designing sensor net-

works. Here both energy consumption and coverage factors of mobile sensor
networks are optimized in two phases using fully informed particle swarm op-
timization (FIPSO) that has proved to have a good achievement and given good
results. The method has been applied on three sets of data to test its efficiency.
All results have been perfectly good.

Keywords: Fully Informed Particle Swarm Optimization, Mobile sensor

networks.

1 Introduction

Mobile sensor networks consist of the sensor nodes that are deployed in a large area
collecting important information from the sensor field. The communication between
the nodes is wireless. Since energy is an important resource [4], the nodes’ energy
consumption must be kept at a minimum rate. Optimum placement of the sensors also
includes their coverage that means the sensors should be places in the best position
assigned by a coverage function to have the best functionality. Some random deploy-
ments [1] do not apply a uniform distribution over a surface that can be important in
some cases [3].
This paper uses FIPSO to find the optimal placement and energy usage of the mo-
bile sensor networks. In section 2 FIPSO is discussed. Section 3 describes applying
FIPSO for deployment of mobile sensor networks and section 4 describes the results
and section 5 denotes the conclusion.

2 Fully Informed Particle Swarm Optimization

The canonical particle swarm algorithm is a new approach to optimization, drawing

inspiration from group behavior and establishment of social norms [6]. It is gaining
popularity because of its being easy to use and its speed of convergence. This method
keep the individuals fully informed that means unlike simple PSO, every single mem-
ber of the population’s experience is taken into account.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 563–567, 2009.
© Springer-Verlag Berlin Heidelberg 2009
564 N. Shafiabady and M.A. Nekoui

r r rr r r rr r r
vi (t ) = φ1vi (t − 1) + r1c1 ⊗ ( x personalbest − xi (t )) + r2 c2 ⊗ ( x globalbest − xi (t ))
r r r
xi (t ) = vi (t ) + xi (t − 1) (1)
1− t
φ1 = 1 − 0.5 ×
1 − t max
As it is clear in Eq. (1), the experience of every member of the population is taken
into consideration and this method has a better performance than the simple PSO.

3 Applying FIPSO for Deployment of Mobile Sensor Networks

3.1 Optimization of Coverage

Coverage is considered to be the first optimization objective. The coverage of each

sensor can be defined either by a binary sensor model or a probabilistic sensor model
as described in Fig. 1.

Fig. 1. Sensor coverage models a) binary b) probabilistic

In the binary sensor models, the detection probability of the event of interest is 1
within the sensing range, otherwise, the probability is zero. Although the binary sen-
sor model is simpler, it is not realistic as it assumes that sensor readings have no asso-
ciated uncertainty. In reality sensor detections are imprecise, hence the coverage has
to be described in probabilistic terms. The probabilistic sensor model used in this
work is given in Eq. (2).

⎧ 0 , if r + re ≤ dij (x, y)
⎪ −λaβ
cij (x, y) = ⎨e , if r − re < dij (x, y) < r + re (2)
⎪ 1 , if r + re ≥ dij (x, y)
⎩
The sensor field is represented by a n × m grid where an individual sensor is placed
in point S at grid point (x,y). Each sensor has a detection range r. For any grid point P
at (i,j), the Euclidean distance between P and the grid point at (x,y) is denoted as
d ij ( x, y ) = ( x − i ) 2 + ( y − j ) 2 , the equation expresses the coverage
cij ( x, y ) of a grid point at (i,j) by a sensor S at (x,y). In this equation λ and β are the
 Optimal Evolutionary-Based Deployment of Mobile Sensor Networks 565

parameters that measure the detection probability when a target is at distance greater
than re but within a distance from the sensor and a = d ij ( x, y ) − ( r − re ) .

3.2 Optimization of Energy Consumption

After optimization of coverage, all the deployed sensor nodes move to their own posi-
tion. Now our goal is to minimize the energy usage in a cluster based sensor network
topology by finding the optimal cluster head positions. For this purpose a power con-
sumption model [5] for the radio hardware dissipation where the transmitter dissipates
energy to run the radio electronics and the power amplifier and the receiver dissipates
energy to run the radio electronics, is used to be optimized. The fitness function to be
minimized is given in Eq. (3).

m n ⎛ 1.3 × 10 −6 Dis 4j ⎞
f = ∑∑ 0.01disij +
⎜ 2 ⎟ (3)
⎜
j =1 i =1 ⎝ n ⎟
j ⎠
For this approach both the free space (distance2 power loss) and the multi-path fading
(distance4 power loss) channel models were used. Assume that the sensor nodes inside
a cluster have short distance dis to the cluster head but each cluster head has long dis-
tance Dis to the base station. The base station is situated at the position (25,80).

4 The Simulation Results

The program is simulated in MATLAB. The particles are randomly produced at the
beginning. Three different results are given in Fig. 2-5.

Fig. 2. The result showing the heads and the nodes

The red spots denote the chosen heads for each selected cluster among the four
chosen clusters.
566 N. Shafiabady and M.A. Nekoui

Fig. 3. The result showing the heads and the nodes

Fig. 4. The result showing the heads and the nodes

Fig. 5. The error of the first phase

 Optimal Evolutionary-Based Deployment of Mobile Sensor Networks 567

The error of the first phase is given in Fig. 5. It shows that it has efficiently placed
the nodes in the grid’s positions.
The results are achieved in two phases. At first the particles’ coverage cost func-
tion is minimized using FIPSO. Then FIPSO is applied to the achieved set of particles
to find the head and the best position of the node that will be chosen as the head.
FIPSO has been able to give better results in compare with the previous works [1-5].

5 Conclusion
The results show that the proposed method (FIPSO) has successfully placed the mobile
sensors and then chosen the head as an appropriate member of the clusters according to
the cost functions. This method is more efficient than the simple PSO as it uses the in-
formation from all the members rather than only some of the particles. The results are
achieved fast and the coverage and energy functions are optimized simultaneously.

References
1. Chakrabarty, K., Iyengar, S., Qi, H., Cho, E.: Grid coverage for surveillance and target loca-
tion in distributed sensor networks. IEEE Transactions on Computers 51, 1448–1453 (2002)
2. Jourdan, B., Weck, O.: Layout optimization for a wireless sensor network using multi-
objective genetic algorithm. In: IEEE VTC 2004 Conference, vol. 5, pp. 2466–2470 (2004)
3. Howard, A., Mataric, M.J., Sukhatme, G.S.: Mobile sensor network deployment using po-
tential field: a distributed, scalable solution to the area coverage problem. In: Proc. Int.
Conf. on Distributed Autonomous Robotics Systems, pp. 299–308 (2002)
4. Heo, N., Varshney, P.K.: Energy-efficient deployment of intelligent mobile sensor net-
works. IEEE Transaction on Systems, Man and Cybernetics 35(1), 78–92 (2005)
5. Heinzelman, W.B., Chandrakasan, A.P., Balakrishnan, H.: An application specific protocol
architecture for wireless microsensor networks. IEEE Transactions on Wireless Communi-
cations 1(4), 660–670 (2002)
6. Mendes, R., Kennedy, J., Neves, J.: The fully informed particle swarm. IEEE Transactions
on Evolutionary Computation 1(1) (January 2005)
 Research of Current Control Strategy of Hybrid Active
Power Filter Based on Fuzzy Recursive Integral PI
Control Algorithm

Zhong Tang and Daifa Liao

School of Computer and information Engineering, Shanghai University of Electric Power,

Shanghai 200090, China
tangzhong64@163.com, liaodaifa0724@163.com

Abstract. According to the current control characters of hybrid active power filter
(HAPF), the current control model of HAPF is designed. The fuzzy recursive in-
tegral PI control algorithm is presented when it is compared to conventional PI
control method. The control algorithm is applied to auto-regulate the proportional
and integral parameters of PI. Thus, the robustness and response speed is en-
hanced; the dynamic performance of the HAPF device is improved. Under Mat-
lab/Simulink background, a fuzzy recursive integral PI controller is designed and
it is applied in a HAPF model in PSCAD/EMTDC. The results prove the feasibil-
ity and effectiveness of this fuzzy recursive integral PI control algorithm.

Keywords: Hybrid active power filter, Recursive integral PI, Fuzzy control,
PSCAD/EMTDC.

1 Introduction
Combined with the advantages of passive power filter (PPF) and active power filter
(APF), hybrid active power filter (HAPF) is a powerful apparatus for reactive
power compensation, harmonic suppression; it is also an effective device to resolve
power quality problems [1-3] . It is a key link to keep DC voltage stable, ensure com-
pensate current signal timely, track instruction current correctly in active power filter-
ing technology [4].
Related work. Recently, there are many current control methods, mainly including
linear-current control, digital-deadbeat control, triangular-wave control, hysteresis
control and one-cycle control [5-6]. But because of the influences of detecting preci-
sion, phase shift of output filter and time delay of control method, the tracking ability
of the system is become non-ideal. Because of its advantages of simple algorithm,
good robustness and high reliability, PI control method has been widely used in indus-
try control system [7]. The model parameters are required invariable and proportional
integral parameters are hard to determine in conventional PI control. Because the peri-
odic characteristics of error signal in HAPF system, the application of conventional PI
control is limited seriously. Fuzzy logical control has ideal dynamic performance and
is insensitive of producer parameter; it has high robustness and can overcome the in-
fluence of non-linear factors. So, it has been widely used in control system [8-10].

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 568–578, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Research of Current Control Strategy of Hybrid Active Power Filter 569

In this paper, fuzzy logical control is applied to adjust the parameters of recursive
integral PI control method, the response speed is enhanced and dynamic performance
is improved. Based on fuzzy recursive integral PI control strategy, a HAPF current
controller is designed and it is applied to auto-regulate the proportional and integral
parameters. The simulation results prove the feasibility and effectiveness of this con-
trol algorithm.

2 Configuration of HAPF and Control Model

The HAPF circuit configuration suitable for medium voltage and low voltage system
is showed in Fig.1, which consists of PPF and APF, and is directly shunted to the
10KV bus. Fundamental resonance circuit (FRC) consists of L1 , C1 which is tuned at
fundamental frequency; an injection capacitance Cg is directly connected with FRC
in series which is used to suppress 6th harmonic current when it is combined with
FRC circuit. Because of the FRC suppressing fundamental current, the capacitance
and initial investment of APF are decreased sharply. In this circuit PPF are mainly
used to filter the vast majority of the5th, 7th, 11th, 13th harmonic currents and com-
pensate reactive power of the system.

Fig. 1. Shunt hybrid active power filter diagram

The single-phase equivalent circuit of HAPF is showed in Fig.2, in which the APF
is controlled as an ideal harmonic voltage resource U F and the load is controlled as a
harmonic current resource iL . Where I S and I L represent source current and load
current; I PF and I APF represent the current of PPF circuit and injected current of
APF. Z S and Z PF represent the equivalent impedance of source and PPF
branch; Z C and Z R represent the equivalent impedance of injected capacitance and
FRC branch; Z C 0 , LC 0 are capacitive reactance and inductance of output filter
branch. The ratio of coupling transformer is n: 1.
570 Z. Tang and D. Liao

Fig. 2. Single phase equivalent circuit of HAPF

In Fig.2 if we use Z1 represents the shunt impedance of n 2 Z C 0 and Z R , according

to KVL and KCL we can obtain:

⎧U S = U L + I S Z S
⎪U = I Z
⎪ L PF PF

⎪⎪ I APF Z C + I1Z1 = U L
⎨ (1)
⎪ I L 0 n Z L 0 = I1Z1 + nU F
2

⎪I + I = I
⎪ 1 L0 APF

⎪⎩ S
I = I APF + I PF + I L

When HAPF solve and compensate harmonic currents, harmonic voltage distortion is
very small and can be neglected, so from equation (1) we can obtain

( Z C Z PF + K 2 Z PF ) I L + nK1Z PFU F
IS = (2)
Z C Z PF + Z S Z C + Z S Z PF + K 2 ( Z PF + Z S )

where

⎪⎧ K1 = Z1 ( Z1 + n Z L 0 )
2

⎨
⎪⎩ K 2 = n Z1Z L 0 ( Z1 + n Z L 0 )
2 2

If we assume

nK1Z PF
G1 ( S ) = (3)
Z C Z PF + Z S Z C + Z S Z PF + K 2 ( Z PF + Z S )
 Research of Current Control Strategy of Hybrid Active Power Filter 571

Z C Z PF + K 2 Z PF
G2 ( S ) = (4)
Z C Z PF + Z S Z C + Z S Z PF + K 2 ( Z PF + Z S )

So we can obtain closed-loop control system of source current I sh showed in Fig.3.

Where I C∗ ( S ) represents reference signal of source current, Gcon ( S ) and
Ginv ( S ) represent controller function and transfer function of active inverter. If
I C∗ ( S ) = 0 that the closed-loop transfer function of I sh will be
I Sh ( S ) Gcon ( S )Ginv ( S )G1 ( S )
GS ( S ) = ∗
= (5)
I C ( S ) 1 + Gcon ( S )Ginv ( S )G1 ( S )
The current I Lh is a disturbance to this closed-loop control system of I sh , and the
transfer function of I sh to disturbance is

I Sh ( S ) G2 ( S )
GL ( S ) = = (6)
I Lh ( S ) 1 + Gcon ( S )Ginv ( S )G1 ( S )

I Lh ( S )
G2 ( S )
I C∗ ( S )
⊗ ⊗I
E (S ) Sh (S )
G con ( S ) Ginv ( S ) G1 ( S )

Fig. 3. Closed-loop control of the system current

3 The Design of Fuzzy Recursive Integral PI Controller

Conventional PI control method can reach deadbeat control when it control DC sig-
nals and changing slowly signals, but there are steady-state errors when it control
sinusoid signals directly. In this paper the system reference signals are combined with
all kinds of harmonic currents which are multiple of fundamental current whose pe-
riod is 20ms. Recursive integral PI control method based on conventional PI control
method is used to implement deadbeat control. Every corresponding sampling point
of error e(t ) is integrated in each cycle when use this kind of control algorithm, and
it is equivalent to N parallel PI controller to implement PI control of the system.
Equation (7) is recursive integral PI control algorithm, where sampling value of con-
troller output is U R ( K ) and sampling value of error is e( K ) at the moment of K; N
572 Z. Tang and D. Liao

is sampling point of a cycle, K P , K I respectively represents proportion and integral

coefficient. This kind of algorithm is equivalent to integrate error of each cycle.

⎧ C

⎪⎪ U R ( K ) = K P e ( K ) + ∑
i =0
K I e( K − iN )
⎨ (7)
⎪C = ent K
⎪⎩ N

In order to simplify the calculation, we use incremental form of U R ( K ) to calculate,

and we can obtain control low of incremental form

U R ( K ) = U R ( K − N ) + K P e( K ) − K P e( K − N ) + K I e( K ) (8)

From the equation (8) we can obtain pulse transfer function of recursive integral PI
controller

U R (S ) KI
Gcon ( S ) = = KP + (9)
E (S ) 1 − e − SNT
Where N is sampling points in each cycle of controlled object, T is sampling time. In
Fig.3 we use recursive integral PI controller and can obtain

K P Ginv ( S )G1 ( S )(1 − e− SNT ) + K I Ginv ( S )G1 ( S )

GS ( S ) = (10)
(1 + K PGinv ( S )G1 ( S )) × (1 − e− SNT ) + K I Ginv ( S )G1 ( S )

(11)
G2 ( S )(1 − e − SNT )
GL ( S ) =
(1 + K P Ginv ( S )G1 ( S )) × (1 − e − SNT ) + K I Ginv ( S )G1 ( S )
So the frequency characteristic equations of the two functions are

GS ( jω ) = [ K P Ginv ( jω )G1 ( jω )(1 − cos ω NT + j sin ω NT )

+ K I Ginv ( jω )G1 ( jω )] [(1 + K P Ginv ( jω )G1 ( jω )) (12)

×((1 − cos ω NT + j sin ω NT )) + K I Ginv ( jω )G1 ( jω )]

GL ( jω ) = [G2 ( jω )(1 − cos ω NT + j sin ω NT )] [(1 + K P Ginv ( jω )G1 ( jω )) (13)

×((1 − cos ω NT + j sin ω NT )) + K I Ginv ( jω )G1 ( jω )]

When ω = 2nπ f , from equation (12) (13) we can obtain

 Research of Current Control Strategy of Hybrid Active Power Filter 573

GS ( jn ⋅ 2π f ) = GL ( jn ⋅ 2π f ) = 1 (14)

Equation (14) illustrates that the amplitude of closed-loop system transfer function
and HAPF system to disturbance transfer function is 1, and the phase is 0; output of
the system can track the reference signal bitterly; the influence of load current become
0 with time increasing, when it’s frequency is power frequency or Integral Multiple
power frequency.
The systemic steady-state error can be eliminated using recursive integral PI con-
trol algorithm, but its robustness and dynamic performance are not ideal. So in this
paper fuzzy logical algorithm is applied to enhance the systemic response speed by
auto-regulating K P and K I parameters of recursive integral PI control method [9].
Compound control of fuzzy recursive integral PI is showed in figure 4. Where error e
and error rate ec is fuzzed into corresponding fuzzy variable E and EC . Firstly, it
seeks the fuzzy relation of K P , K I , E and EC , continuously detects E and EC within
operating time. Then, according to fuzzy logical control algorithm, K P and K I are
auto-regulated online to meet various requirement of recursive integral PI controller
in various E and EC . Thus makes the controlled object better robustness and dynamic
performance.

I Lh ( S )
G2 ( S )

IC (S ) E (S ) I Sh ( S )
Gcon ( S )Ginv ( S )G1 ( S )

Fig. 4. Configuration of fuzzy recursive integral PI control

Firstly, inputs are fuzzed by fuzzy controller; the linguistic values of input variable
and output variable are equally divided into 7 parts: {NB, NM, NS, 0, PS, PM, PB}.
According to engineering experience, using triangle membership functions, the uni-
verse of e and ec is [-6 6]. The proportion link is to proportionally reflect the systemic
error signal e. The integral link is to improve systemic indiscrimination degree. When
E is bigger, in order to have better tracking performance K P should be bigger. At
the same time, in order to avoid bigger overshoot of systemic response, the effect of
574 Z. Tang and D. Liao

integral should be restricted, usually we set K I = 0 . When E is medium, in order

to have smaller overshoot of systemic response, K P should be smaller. When E is
smaller, in order to have better steady-state performance, K P and K I both should be
bigger. So, according to the above analysis, we can obtain the fuzzy control rules of
K P and K I in Table 1 and Table 2.

Table 1. Fuzzy control rules of ∆K P

E ec
NB NM NS 0 PS PM PB
NB PB PB NB PM PS PM 0
NM PB PB NM PM PS 0 0
NS PM PM NS PS 0 NS NM
0 PM PS 0 0 NS NM NM
PS PS PS 0 NS NS NM NM
PM 0 0 NS NM NM NM NB
PB 0 NS NS NM NM NB NB

Table 2. Fuzzy control rules of ∆K I

E ec
NB NM NS 0 PS PM PB
NB 0 0 NB NM NM 0 0
NM 0 0 NM NM NS 0 0
NS 0 0 NS NS 0 0 0
0 0 0 NS NM PS 0 0
PS 0 0 0 PS PS 0 0
PM 0 0 PS PM PM 0 0
PB 0 0 NS PM PB 0 0

Fuzzy querying table reflects the final results of fuzzy control algorithm. It is
calculated off-line through afore-hand and stored in computer memory. In real-time
control system, the auto-regulate modification process of recursive integral PI
control parameters is translated into minute calculating querying process of control
rule table.
 Research of Current Control Strategy of Hybrid Active Power Filter 575

4 Simulation Results and Analysis

In this paper a HAPF model controlled by current tracking strategy is set up, and its
simulation is carried out, in which current detection use fuzzy recursive integral PI
control algorithm, compared with the results using conventional PI control algorithm.
Based on Matlab/Simulink toolbox and corresponding function, a fuzzy recursive
integral PI controller is established and it is applied in the HAPF model which is set
up based on PSCAD/EMTDC V4.2 of Manitoba HVDC research center.
In this simulation model, line voltage of the source is 10KV; capacitance of the
source is 100MVA; capacitance of coupling transformer is 4MVA, and its ratio is
10000:380; where the parameters of FRC are C1 = 960uF and L1 = 15.47 mH ,
injecting capacitance CG = 19.65uF . A single-tuned passive power filter to sup-
press 6th harmonic current is consisted of the injecting capacitance and FRC
branch. A six pulse rectifier bridge is used as a harmonic source and its power fre-
quency is 50HZ. The parameters of PPF are showed in table 3 and the high-pass
filter quality factor Q = 1.25 . The simulation circuit and configuration of HAPF
system is showed in Fig.5.
Load reactive power Ndc
Q B
Power Com Bus
active power

0.35 [ohm]
P A AM

0.47 [uF]
1 3 5
A A A Ia A A A A GM
RL 90[MVA]
Va 1.0 [ohm ]
PI / 180
Ib

0.0535 [H]
B B B B B B B Pi by 180
RL #1 #2
Vb 1.0 [ohm ]
C C C Ic C 10[kV] C C C AO *
RL 3[kV]
Vc 1.0 [ohm ]
4 6 2
KB
19.65[uF]

19.65[uF]

19.65[uF]

117.03 [uF] 0.00346 [H]

117.03 [uF] 0.00346 [H]

117.03 [uF] 0.00346 [H]

94.50 [uF]

94.50 [uF]

94.50 [uF]

Idc
70.46[uF]

70.46[uF]

70.46[uF]

34.30[uF]

34.30[uF]

34.30[uF]

20.00 [uF]

20.00 [uF]

20.00 [uF]
690 [uF]

690 [uF]

690 [uF]

0.00293 [H] 0.15 [ohm]

0.00293 [H] 0.15 [ohm]

0.00293 [H] 0.15 [ohm]

0.00088 [H]

0.00088 [H]

0.00088 [H]

0.00170 [H] 0.13 [ohm]

0.00170 [H] 0.13 [ohm]

0.00170 [H] 0.13 [ohm]

12.06 [ohm]

12.06 [ohm]

12.06 [ohm]
0.0031 [H]

0.0031 [H]

0.0031 [H]
0.01547 [H]

0.01547 [H]

0.01547 [H]

0.11[ohm]

0.11[ohm]

0.11[ohm]

0.15[ohm]

0.15[ohm]

0.15[ohm]

A P
RealPower
Power
B Q VFdc
APF reactive power

0.002 [ohm] 0.0008 [H] 1 3 5

0.001 [ohm]

A A A A 2 2 2
4[MVA] IFa G1 G3 G5
0.002 [ohm] 0.0008 [H]
B B B B
#1 #2 IFb
0.002 [ohm] 0.0008 [H]
C 10[kV] C C C
0.38[kV] IFc
4 6 2
2 2 2
G4 G6 G2
12.5 [uF]

12.5 [uF]

12.5 [uF]

Fig. 5. Hybrid active power filter model under PSCAD/ EMTDC

Table 3. Parameters of the LC passive power filter

Harmonics C L R
5th branch C5 = 117.03uF L5 = 3.46mH R5 = 0.11Ω
7th branch C7 = 70.46uF L7 = 2.93mH R7 = 0.15Ω

11th branch C11 = 94.50uF L11 = 0.89mH R11 = 0.10Ω

13th branch C13 = 34.30uF L13 = 1.70mH R13 = 0.13Ω

HPF CH = 20.00uF LH = 3.1mH RH = 12.06Ω

576 Z. Tang and D. Liao

phase A
ISa Ia
2.50
2.00
1.50
1.00
0.50

kA
0.00
-0.50
-1.00
-1.50
-2.00
IaP
0.200
0.150
0.100
0.050
kA 0.000
-0.050
-0.100
-0.150
-0.200
1.700 1.720 1.740 1.760 1.780 1.800

(a) source current and active power filter compensated current waveform using conventional PI
control

(b) the frequency spectrum of load current using conventional PI control

(c) the frequency spectrum of source current using conventional PI control

Fig. 6. The harmonic suppression and compensation results of HAPF using conventional PI
control

phase A
ISa Ia
2.50
2.00
1.50
1.00
0.50
kA

0.00
-0.50
-1.00
-1.50
-2.00
IaP
0.200
0.150
0.100
0.050
0.000
kA

-0.050
-0.100
-0.150
-0.200
1.700 1.720 1.740 1.760 1.780 1.800

(a) source current and active power filter compensated current waveform using fuzzy recursive
integral PI control

Fig. 7. The harmonic suppression and compensation results of HAPF using fuzzy recursive
integral PI control
 Research of Current Control Strategy of Hybrid Active Power Filter 577

(b) the frequency spectrum of load current using fuzzy recursive integral PI control

(c) the frequency spectrum of source current using fuzzy recursive integral PI control

Fig. 7. (Continued)

Table 4. Harmonic current analysis of two kinds of control methods

Current times 5th 7th 11th 13th

Pre-filtering(A) 116.69 77.79 39.61 28.29

Conventional PI (A) 50.13 30.21 15.55 9.12

Fuzzy recursive integral PI (A) 16.27 12.02 6.36 7.07

Fig.6 shows the harmonic suppression and compensation results of HAPF system
using conventional PI control method. Fig.7 shows the harmonic suppression and
compensation results when use fuzzy recursive integral PI control algorithm. ISa, Ia
and IaP respectively represents the source current of phase A, the load current of
phase A and APF compensated current of phase A. Table 4 shows the currents value
of 5th, 7th, 11th and 13th harmonics when use conventional PI control method and
fuzzy recursive integral PI control algorithm. Analyzed from Fig.6, Fig.7 and Table 4,
it is proved that fuzzy recursive integral PI control algorithm has better and more
ideal effect in HAPF harmonic suppression and compensation.

5 Conclusions
In this paper aiming at the harmonic detecting and control problems of HAPF system, a
fuzzy recursive integral PI control algorithm is been proposed based on the conventional
578 Z. Tang and D. Liao

PI control method, and effectively enhance the filtering performance, robustness and
dynamic response performance of HAPF system. A fuzzy recursive integral PI control-
ler is established based on Matlab/Simulink toolbox and corresponding function, and it
is applied in a HAPF model simulation under the PSCAD/EMTDC V4.2 of Manitoba
HVDC research center, the results illustrate the correctness and effectiveness of this
control algorithm.

Acknowledgement
The authors would like to thank Natural Capacity Discipline Project of Shanghai
Local High Institutions (No. 071605125) and Postgraduate Innovation Fund of
Shanghai University of Electric Power (No.D08116).

References
1. Luo, A.: Harmonic Suppression and Reactive Power Compensation Equipment and Tech-
nology. China Electric Power System Press, Beijing (2006)
2. Darwin, R., Luis, M., Juan, W.: Improving passive filter compensation performance with
active techniques. IEEE Trans on Industrial Electronics 50(1), 161–170 (2003)
3. He, N., Huang, L.: Multi-objective optimal design for passive part of hybrid active power
filter based on particle swarm optimization. Proceeding of the CSEE 28(27), 63–69 (2008)
4. Luo, A., Fu, Q., Wang, L.: High-capacity hybrid power filter for harmonic suppression and
reactive power compensation in the power substation. Proceeding of the CSEE 24(9), 115–
223 (2004)
5. Buso, s., Malesani, L.: Design and Fully Digital Control of Parallel Active Power Filters
for Thyristor Rectifiers to Comply with IEC 1000-3-2. IEEE Trans. on Industry Applica-
tion 34(2), 508–517 (1998)
6. Tang, X., Luo, A., Tu, C.: Recursive Integral PI for Current Control of Hybrid Active
Power Filter. Proceedings of the CSEE 23(10), 38–41 (2003)
7. Sun, M., Huang, B.: Iterative Learning Control. National Defense Industry Press, Beijing
(1999)
8. Zhou, K., Luo, A., Tang, J.: PI iterative learning for current-tracking control of active
power filter. Power Electronics 40(4), 53–55 (2006)
9. Xu, W.-f., Luo, A., Wang, L.: Development of hybrid active power filter using intelligent
controller. Automation of Electric Power Systems 27(10), 49–52 (2003)
10. Fukuda, S., Sugawa, S.: Adaptive signal processing based control of active power filters.
In: Proceeding of IEEE IAS Annual Meeting (1996)
 A Fuzzy Query Mechanism for Human Resource
Websites

Lien-Fu Lai1, Chao-Chin Wu1, Liang-Tsung Huang2, and Jung-Chih Kuo1

1
Department of Computer Science and Information Engineering,
National Changhua University of Education, Taiwan
{lflai,ccwu}@cc.ncue.edu.tw, verygoodtony@hotmail.com
2
Department of Information Communication, MingDao University, Taiwan
larry@mdu.edu.tw

Abstract. Users’ preferences often contain imprecision and uncertainty that are
difficult for traditional human resource websites to deal with. In this paper, we
apply the fuzzy logic theory to develop a fuzzy query mechanism for human re-
source websites. First, a storing mechanism is proposed to store fuzzy data into
conventional database management systems without modifying DBMS models.
Second, a fuzzy query language is proposed for users to make fuzzy queries on
fuzzy databases. User’s fuzzy requirement can be expressed by a fuzzy query
which consists of a set of fuzzy conditions. Third, each fuzzy condition associ-
ates with a fuzzy importance to differentiate between fuzzy conditions according
to their degrees of importance. Fourth, the fuzzy weighted average is utilized to
aggregate all fuzzy conditions based on their degrees of importance and degrees
of matching. Through the mutual compensation of all fuzzy conditions, the or-
dering of query results can be obtained according to user’s preference.

Keywords: Fuzzy Query, Fuzzy Weighted Average, Human Resource Websites.

1 Introduction
In traditional human resource websites [1,2,3,4,5], users must state clear and definite
conditions to make database queries. Unfortunately, users’ preferences often contain
imprecision and uncertainty that are difficult for traditional SQL queries to deal with.
For example, when a user hopes to find a job which is near Taipei City and pays good
salary, he can only make a SQL query like “SELECT * FROM Job WHERE (Loca-
≧
tion=‘Taipei City’ or Location= ‘Taipei County’) and Salary 40000”. However,
both ‘near Taipei City’ and ‘good salary’ are fuzzy terms and cannot be expressed
appropriately by merely crisp values. A job which locates in ‘Taoyuan County’ with
salary of 50000 may be acceptable in user’s original intention, but it would be ex-
cluded by the traditional SQL query. SQL queries fail to deal with the compensation
between different conditions. Moreover, traditional database queries cannot effec-
tively differentiate between the retrieved jobs according to the degrees of satisfaction.
The results to a query are very often a large amount of data, and the problem of the
information overload makes it difficult for users to find really useful information.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 579–589, 2009.
© Springer-Verlag Berlin Heidelberg 2009
580 L.-F. Lai et al.

Hence, it is required to sort results based on the degrees of satisfaction to the retrieved
jobs. Computing the degree of satisfaction to a job needs to aggregate all matching
degrees on individual conditions (e.g. location, salary, industry type, experience,
education etc.). It is insufficient for merely using the ORDER BY clause in SQL to
sort results based on some attribute. In addition, traditional database queries do not
differentiate between conditions according to the degrees of importance. One condi-
tion may be more important than another condition for some user (e.g. salary is more
important than location in someone’s opinion). Both the degree of importance and the
degree of matching to every condition should be considered to compute the degree of
satisfaction to a job. We summarize the problems of traditional human resource web-
sites as follows.
• Users’ preferences are usually imprecise and uncertain. Traditional database
queries are based on total matching which is limited in its ability to come to
grips with the issues of fuzziness.
• In users’ opinions, different conditions may have different degrees of impor-
tance. Traditional database queries treat all conditions as the same importance
and can not differentiate the importance of one condition from that of another.
• The problem of information overload makes it difficult for users to find really
useful information from a large amount of query results. Traditional database
queries do not support the ordering of query results by aggregating the degrees
of matching to all conditions (i.e. no compensation between conditions).
To solve the mentioned problems, we apply the fuzzy logic theory [15] to develop a
fuzzy query mechanism for human resource websites. First, a storing mechanism is
proposed to store fuzzy data into conventional database management systems without
modifying DBMS models. Second, a fuzzy query language is proposed for users to
make fuzzy queries on fuzzy databases. User’s fuzzy requirement can be expressed by
a fuzzy query which consists of a set of fuzzy conditions. Third, each fuzzy condition
associates with a fuzzy importance to differentiate between fuzzy conditions accord-
ing to their degrees of importance. Fourth, the fuzzy weighted average is utilized to
aggregate all fuzzy conditions based on their degrees of importance and degrees of
matching. Through the mutual compensation of all fuzzy conditions, the ordering of
query results can be obtained according to user’s preference.

2 The Fuzzy Query Mechanism for Human Resource Websites

Applying the fuzzy logic theory to develop fuzzy queries on human resource web-
sites, there are two issues that should be addressed:
• A storing mechanism is required to represent and store fuzzy data. It should be
applied directly to existing databases without modifying DBMS models. More-
over, it should not only deal with the continuous and numerical fuzzy data but
the discrete and lexical fuzzy data.
• A fuzzy query language is required to make fuzzy queries on fuzzy databases.
The degrees of importance should be considered to differentiate between fuzzy
conditions. The query results should be sorted by mutual compensation of all
fuzzy conditions.
 A Fuzzy Query Mechanism for Human Resource Websites 581

2.1 Storing Fuzzy Data into Databases

Galindo et al. [8] classify fuzzy data into four types: (1) Type 1 contains attributes
with precise data. This type of attributes is represented in the same way as crisp data,
but can be transformed or manipulated using fuzzy conditions. (2) Type 2 contains
attributes that gather imprecise data over an ordered referential. These attributes admit
both crisp and fuzzy data, in the form of possibility distributions over an underlying
ordered domain (fuzzy sets). (3) Type 3 contains attributes over data of discrete non-
ordered dominion with analogy. In these attributes, some labels are defined with a
similarity relationship defined over them. The similarity relationship indicates to a
degree that each pair of labels resembles each other. (4) Type 4 contains attributes
that are defined in the same way as Type 3 attributes, without being necessary for a
similarity relationship to exist between the labels.
By analyzing several popular human resource websites [1,2,3,4,5], we sum up that
the major fuzzy data needed to be stored are location preference, salary preference,
industry preference, job category preference, experience preference, education prefer-
ence, department preference, job seeker’s profile, and hiring company’s profile. We
adopt the notions of [6] to classify these fuzzy data into three types: (1) Discrete fuzzy
data include location preference, industry preference, job category preference, educa-
tion preference, and department preference. (2) Continuous fuzzy data include salary
preference and experience preference. (3) Crisp data include job seeker’s profile and
hiring company’s profile.

2.1.1 Discrete Fuzzy Data

The discrete fuzzy data (i.e. type 3 and type 4 fuzzy data) is represented by a discrete
fuzzy set which consists of a set of discrete data items with their degrees of confor-
mity. How to grade numeric degrees of conformity may confuse users. Hence, we use
linguistic degrees of conformity (i.e. totally unsatisfactory, unsatisfactory, rather un-
satisfactory, moderately satisfactory, rather satisfactory, very satisfactory, and totally
satisfactory) to make it easier and clearer for users to grade degrees. ‘Totally unsatis-
factory’ and ‘totally satisfactory’ stand for 0 and 1 respectively, while the others
grade values between 0 and 1. For example, a fuzzy term ‘near Taipei City’ can be
expressed as a discrete fuzzy set like {(Taipei City, totally satisfactory), (Taipei
County, very satisfactory), (Taoyuan County, moderately satisfactory)}. We utilize
the membership functions in [13] to define the linguistic degree of conformity be-
tween a discrete data item and a fuzzy term (see Figure 1).

rather
µ A~ ( x) rather
unsatisfactory moderately satisfactory very
satisfactory satisfactory
unsatisfactory
1

0 x
0 0.25 0.5 0.75 1

Fig. 1. Membership functions for degrees of conformity

582 L.-F. Lai et al.

A fuzzy number A~ can be defined by a triplet (a, b, c) and the membership func-
tion µ A~ ( x) is defined as:
⎧0 ,x < a
⎪x − a
⎪ ,a ≤ x ≤ b
⎪
µ A~ ( x) = ⎨ b − a
⎪c − x ,b ≤ x ≤ c
⎪c − b
⎪0 ,x > c
⎩

Therefore, each linguistic degree of conformity can be mapped to a triangular fuzzy

number [11], e.g. ‘rather satisfactory’ is mapped to (0.5,0.75,1).
To store the discrete fuzzy data in conventional databases, a new table is needed to
store the set of discrete data items with their degrees of conformity. For example, job
seekers’ profiles are stored in a Resume table (see Figure 2). If the location preference
of a job seeker numbered 10001 is ‘near Taipei City’, the discrete fuzzy set inputted
by users can be stored in a new Location_Preference table shown in Figure 3. In the
Location_Preference table, the RID attribute serves as a foreign key to reference the
primary key of the Resume table. Hence, we can retrieve a job seeker’s location pref-
erence via joining the two tables on equivalent RIDs.

RID Name SSN Education Experience …

10001 C.M. Wang N123456789 Bachelor 2 …

Fig. 2. The Resume table

RID Location SatisfactoryDegree

10001 Taipei City Totally satisfactory
10001 Taipei County very satisfactory
10001 Taoyuan County moderately satisfactory

Fig. 3. The Location_Preference table

2.1.2 Continuous Fuzzy Data

The continuous fuzzy data (i.e. type 2 fuzzy data) is represented by a continuous
fuzzy set which consists of a set of continuous data items with their degrees of con-
formity. For example, a fuzzy term ‘good salary’ can be expressed as a continuous
fuzzy set like {(50000, totally satisfactory), (45000, very satisfactory), (30000, totally
unsatisfactory)}. The degree of conformity can be defuzzified by using mathematical
integral to compute the center of the area that is covered by the corresponding trian-
gular fuzzy number [18], e.g. ‘very satisfactory’ is defuzzified by computing its center
of gravity of the triangular fuzzy number (0.75, 1, 1) as follows.
1
∫0.75
x ( 4 x − 3) dx
= 0.92
1
∫ 0.75
( 4 x − 3) dx
 A Fuzzy Query Mechanism for Human Resource Websites 583

Therefore, the membership function corresponding to the given ‘good salary’ can be
constructed by {(50000,1), (45000,0.92), (30000,0)} (see Figure 4).

Degree of Satisfaction

1
0.92

0
30000 45000 50000 Monthly Salary

Fig. 4. The membership function of ‘good salary’

To store the continuous fuzzy data in conventional databases, a new table is needed
to store the set of continuous data items with their degrees of conformity. A Sal-
ary_Preference table is used to store the corresponding membership function in which
the degrees of conformity have been defuzzified (see Figure 5). In the Salary
_Preference table, the RID attribute serves as a foreign key to reference the primary
key of the Resume table.

RID Salary SatisfactoryDegree

10001 50000 1
10001 45000 0.92
10001 30000 0

Fig. 5. The Salary_Preference table

2.1.3 Crisp Data

Crisp data (i.e. type 1 fuzzy data) contain precise and certain information and can be
directly stored in conventional databases. For example, the Resume table in Figure 2
stores job seeker’s profile including Name, BirthDate, SSN, Education, Department,
Experience, Address, and Phone etc.

2.2 Fuzzy Queries on Web Databases

To deal with fuzzy queries on web databases, three major tasks are required to be
accomplished: (1) a fuzzy query language for users to make fuzzy queries on fuzzy
databases, (2) matching fuzzy conditions in a fuzzy query with fuzzy data in fuzzy
databases, and (3) aggregating all fuzzy conditions based on their degrees of impor-
tance and degrees of matching.

2.2.1 Making Fuzzy Queries on Web Databases

A fuzzy query consists of a set of fuzzy conditions. Each fuzzy condition associates
with a fuzzy importance to differentiate between fuzzy conditions according to the de-
grees of importance and uses a fuzzy set to state the degrees of conformity for different
584 L.-F. Lai et al.

attribute values. We use linguistic degrees of importance (i.e. don’t care, unimportant,
rather unimportant, moderately important, rather important, very important, and most
important) to make it easier for users to grade relative importance. Each linguistic de-
gree of importance can be mapped to a triangular fuzzy number as shown in Figure 7.

rather rather
µ A~ ( x) unimportant moderately important
very
unimportant important important
1

0 x
0 0.25 0.5 0.75 1

Fig. 7. Membership functions for degrees of importance

The fuzzy set that defines a fuzzy condition could be discrete, continuous, or crisp. In
the example shown in Figure 8, a fuzzy query for the resume search consists of 7 fuzzy
conditions with their degrees of importance. The hiring company may consider that job
seeker’s department and salary preference are the most important, experiences and

Fig. 8. A fuzzy query for resume search

 A Fuzzy Query Mechanism for Human Resource Websites 585

location preference are very important, the education level is rather important, and the
company doesn’t care about other fuzzy conditions. In Figure 8, experience preference
is defined by a continuous fuzzy set {(5,1), (2,0.92), (0,0)}, while education level pref-
erence is defined by a discrete fuzzy set {(Master, very satisfactory), (Bachelor, rather
satisfactory)}. Job seekers and hiring companies can make their own fuzzy queries to
search jobs and resumes via selecting options on web pages. In addition, users can set
the least matching degree and the most amount of data listing to reduce the search space
and to avoid the information overload.

2.2.2 Matching Fuzzy Conditions with Fuzzy Data

Once the user makes a fuzzy query, our system would execute a preliminary SQL
query to filter out the data that are totally unsatisfactory to some fuzzy condition in
the fuzzy query. Reducing the search space could effectively save the execution time
of fuzzy matching and fuzzy computation. For example, when a job seeker makes a
fuzzy query that the location is near Taipei City with {(Taipei City, totally satisfac-
tory), (Taipei County, very satisfactory), (Taoyuan County, moderately satisfactory)}
and the salary is good with {(50000, totally satisfactory), (45000, very satisfac-
tory),(30000, totally unsatisfactory)}, our system would automatically execute a SQL
query “CREATE VIEW Job_List AS SELECT * FROM Job WHERE JID in (SE-
LECT DISTINCT JID FROM Job_ Location WHERE Location=‘Taipei City’ or
Location= ‘Taipei County’ or Location=‘Taoyuan County’) and JID in (SELECT JID
FROM Job_Salary_Offer WHERE SalaryOffer>30000 and SatisfactoryDegree=0)” to
exclude those totally unsatisfactory data.
For matching fuzzy conditions with fuzzy data, we adopt the possibility measure in
the fuzzy logic theory [16] to compute the degree of matching between a fuzzy condi-
~ ~
tion A and the fuzzy data B :
~ ~
Poss{B is A} = sup[min(µ A~ ( x), µ B~ ( x))]
x∈U
~ ~
Where the possibility of B being A is obtained by: for each data item x∈U, we get a
minimum of two degrees of conformity µ ~ ( x ) and µ ~ ( x ) , and the possibility measure
A B

is the maximum of these minimums. As fuzzy conditions and fuzzy data may be con-
tinues, discrete or crisp, the matching of a fuzzy condition and a fuzzy data could be
classified into 9 types according to the possibility measure (see Table 1).

Table 1. Different types of matching

Fuzzy Condition Fuzzy Data Result of Matching

continuous continuous a crisp value [0,1]
continuous crisp a crisp value [0,1]
continuous discrete not exist
discrete continuous not exist
discrete crisp a triangular fuzzy number
discrete discrete a triangular fuzzy number
crisp continuous a crisp value [0,1]
crisp crisp a crisp value 1 or 0
crisp discrete a triangular fuzzy number
586 L.-F. Lai et al.

• In the case of the fuzzy condition and the fuzzy data being two continuous fuzzy
sets, the intersection point of two membership functions is the result of match-
ing. For example, when a job seeker’s salary preference is {(60000,1),
(50000,0.92), (30000,0)} and a job’s salary offer is {(20000,1), (30000,0.75),
(40000,0)}, we obtain the degree of matching is 0.285 (see Figure 9).
• In the case of one continuous fuzzy set and one crisp value, the point of the crisp
value mapping to the continuous fuzzy set is the result of matching. For exam-
ple, when a hiring company’s experience preference is {(5,1), (4,0.92), (1,0)}
and a job seeker’s experience is 3 years, we obtain the degree of matching is
0.613 (see Figure 10).

Fig. 9. Matching of two continuous fuzzy sets Fig. 10. Matching of a continuous fuzzy set
with a crisp value

• In the case of one continuous fuzzy set and one crisp value, the point of the crisp
value mapping to the continuous fuzzy set is the result of matching. For exam-
ple, when a hiring company’s experience preference is {(5,1), (4,0.92), (1,0)}
and a job seeker’s experience is 3 years, we obtain the degree of matching is
0.613 (see Figure 10).
• The case of one continuous fuzzy set and one discrete fuzzy set does not exist,
since an attribute cannot contain both continuous and discrete fuzzy data simul-
taneously.
• In the case of two discrete fuzzy sets, the possibility measure is a triangular
fuzzy number. For example, when a job seeker’s location preference is {(Taipei
City, totally satisfactory), (Taipei County, very satisfactory), (Taoyuan County,
moderately satisfactory)} and a job’s location offer is {(Taichung City, very sat-
isfactory), (Taipei city, rather satisfactory), (Taoyuan County, rather satisfac-
tory)}, we can obtain the minimum degree for each data item {(Taipei City,
rather satisfactory), (Taipei County, 0), (Taoyuan County, moderately satisfac-
tory), (Taichung City, 0)}. Hence, the result of matching is the maximum ‘rather
satisfactory’ which can be represented by a triangular fuzzy number
(0.5,0.75,1).
• In the case of one discrete fuzzy set and one crisp value, the result of matching
is the mapping of the crisp value to the discrete fuzzy set. For example, when a
hiring company’s education preference is {(Master, very satisfactory), (Bache-
lor, rather satisfactory)} and a job seeker’s education is ‘Master’, we obtain the
result of matching is ‘very satisfactory’ (0.75,1,1).
 A Fuzzy Query Mechanism for Human Resource Websites 587

• In the case of two crisp values, the result of matching is either 1 (the fuzzy
data satisfies the fuzzy condition) or 0 (the fuzzy data doesn’t satisfy the fuzzy
condition).

2.2.3 Aggregating All Fuzzy Conditions

Computing the overall degree of satisfaction between a fuzzy query and the fuzzy
data needs to aggregate all fuzzy conditions based on their degrees of importance and
degrees of matching. We apply the fuzzy weighted average (FWA) [6,9,10] to calcu-
late the overall degree of satisfaction using triangular fuzzy numbers. In our approach,
the matching degrees of all fuzzy conditions are indicators (xi) that rate the overall
degree of satisfaction between a fuzzy query and the fuzzy data. The degrees of im-
portance are weights (wi) that act upon indicators. Therefore, the fuzzy weighted
average y can be defined as:
∑ wi xi
n

y = f ( xi ,..., xn , w1 ,..., wn ) = i =n1

∑i=1 wi
Where there is n fuzzy conditions, the degree of matching xi, 1 ≤ i ≤ n is represented
by a crisp value or a triangular fuzzy number, and the degree of importance wi, 1 ≤ i ≤
n is represented by a triangular fuzzy number. Since wi are no longer crisp numbers,
∑ wi = 1 is not imposed as a requirement. We adopt the approximate expressions
n
i =1

on ⊕ and ⊗ operators for the computation of L-R fuzzy numbers, which is suggested
by Dubois and Prade [7]. Consider this example: (1) A job seeker makes a fuzzy
query containing a salary preference {(60000,1), (50000,0.92), (30000,0)} with
‘rather important’, a location preference {(Taipei City, totally satisfactory), (Taipei
County, very satisfactory), (Taoyuan County, moderately satisfactory)} with ‘most
important’, his education ‘Master’ with ‘moderately important’, and other fuzzy con-
ditions with the default importance ‘don’t care’. (2) A job stored in databases contains
the salary offer {(20000,1), (30000,0.75), (40000,0)}, the location offer {(Taichung
City, very satisfactory), (Taipei city, rather satisfactory), (Taoyuan County, rather
satisfactory)}, the education preference {(Master, very satisfactory), (Bachelor, rather
satisfactory)}, and a set of attributes with fuzzy data. We apply FWA to calculate the
overall degree of satisfaction between the fuzzy query and the fuzzy data as follows.
(0.5,0.75,1) ⊗ 0.285 ⊕ (1,1,1) ⊗ (0.5,0.75,1) ⊕ (0.25,0.5,0.75) ⊗ (0.75,1,1)
y=
(0.5, 0.75, 1) ⊕ (1, 1, 1) ⊕ (0.25, 0.5, 0.75)
(0.1425, 0.21375, 0.285) ⊕ (0.5, 0.75, 1) ⊕ (0.125, 0.5, 0.75)
=
(0.5, 0.75, 1) ⊕ (1, 1, 1) ⊕ (0.25, 0.5, 0.75)
(0.7675, 1.46375, 2.035) 1.431765
= ≈ = 0.63634
(1.75, 2.25, 2.75) 2.25
Applying the mathematical operations on fuzzy numbers [7,12,16], we get two fuzzy
numbers (0.7675, 1.46375, 2.035) and (1.75, 2.25, 2.75). The center of gravity is
adopted to defuzzify a fuzzy number [14], which is achieved by mathematical inte-
gral. Therefore, the overall degree of satisfaction between the fuzzy query and the
fuzzy data is 0.63634 (i.e. 63.634%).
588 L.-F. Lai et al.

By applying FWA to calculate fuzzy data’s overall degrees of satisfaction to a

fuzzy query, the ordering of all fuzzy data is obtained according to their overall de-
grees of satisfaction. Figure 11 shows the ordering result of a fuzzy query on the re-
sume search, and users can click the resume code for more detailed information.

Fig. 11. The result of a fuzzy query on resume search

3 Conclusion
In this paper, we apply the fuzzy logic theory to develop a fuzzy query mechanism for
human resource websites. The advantages of the proposed approach are as follows.
• Users’ preferences often contain imprecision and uncertainty. Our approach
provides a mechanism to express fuzzy data in human resource websites and to
store fuzzy data into conventional database management systems without modi-
fying DBMS models.
• Traditional SQL queries are based on total matching which is limited in its abil-
ity to come to grips with the issues of fuzziness. Our approach provides a
mechanism to state fuzzy queries by fuzzy conditions and to differentiate be-
tween fuzzy conditions according to their degrees of importance.
• Traditional SQL queries fail to deal with the compensation between different
conditions. Our approach provides a mechanism to aggregate all fuzzy conditions
based on their degrees of importance and degrees of matching. The ordering of
query results via the mutual compensation of all fuzzy conditions is helpful to al-
leviate the problem of the information overload.

References
1. http://hotjobs.yahoo.com/
2. http://www.104.com.tw/
3. http://www.1111.com.tw/
4. http://www.find-job.net/
5. http://www.monster.com/
6. Chang, P.T., Hung, K.C., Lin, K.P., Chang, C.H.: A Comparison of Discrete Algorithms
for Fuzzy Weighted Average. IEEE Transactions on Fuzzy Systems 14(5), 663–675 (2006)
7. Dubois, D., Prade, H.: Fuzzy Sets and Systems: Theory and Applications, New York,
London (1980)
 A Fuzzy Query Mechanism for Human Resource Websites 589

8. Galindo, J., Urrutia, A., Piattini, M.: Fuzzy Databases: Modeling, Design and Implementa-
tion. Idea Group Publishing, Hershey (2005)
9. Guu, S.M.: Fuzzy Weighted Averages Revisited. Fuzzy Sets and Systems 126, 411–414
(2002)
10. Kao, C., Liu, S.T.: Competitiveness of Manufacturing Firms: An Application of Fuzzy
Weighted Average. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Sys-
tems and Humans 29(6), 661–667 (1999)
11. Kaufmann, A., Gupta, M.M.: Introduction to Fuzzy Arithmetic: Theory and Applications.
Van Nostrand Reinhold, New York (1985)
12. Lai, Y.J., Hwang, C.L.: Fuzzy Mathematical Programming, Methods and Applications.
Springer, Heidelberg (1992)
13. Ngai, E.W.T., Wat, F.K.T.: Fuzzy Decision Support System for Risk Analysis in E-
Commerce Development. Decision Support Systems 40(2), 235–255 (2005)
14. Tseng, T.Y., Klein, C.M.: A New Algorithm for Fuzzy Multicriteria Decision Making. In-
ternational Journal of Approximate Reasoning 6, 45–66 (1992)
15. Zadeh, L.A.: Fuzzy Sets. Information and Control 8, 338–353 (1965)
16. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications, 2nd revised edn. Kluwer Aca-
demic Publishers, Dordrecht (1991)
 Selecting Cooperative Enterprise in Dynamic Enterprise
Alliance Based on Fuzzy Comprehensive Evaluation

Yan Li1, Guoxing Zhang2, and Jun Liu1

1
College of Mechanical & Electrical Engineering, Henan Agriculture University
2
Information Center, Henan University of Urban Construction
liyanliulizhi@sina.com, zgxyx@hncj.edu.cn, liujunhenau@163.com

Abstract. Dynamic enterprise alliance is a complex organization system. It has

the virtual enterprise establishment, virtual operation and virtual management
functions. The suitable cooperative partner is the necessary condition to ensure
the efficient functioning of the dynamic enterprises alliance. It is the key tech-
nology of selecting the suitable partner reasonably and forming the dynamic
enterprises alliance to realize the network manufacturing. In this paper, one rea-
sonable model of fuzzy comprehensive evaluation on the cooperative partner of
the dynamic enterprises alliance is established, using the fuzzy multi-grade
comprehensive evaluation method based on the theories of fuzzy logic. It can
offer a scientific method to select the optimum cooperative partner for the dy-
namic enterprises alliance, and improve the comparison level for alternatives.

Keywords: cooperative enterprise, dynamic enterprise alliance, fuzzy compre-

hensive evaluation.

1 Introduction
Dynamic enterprise alliance is a complex organization system. It is one organization of
cooperation and competition composed by two or more than two enterprises which have
the common strategic interests and share the enterprise resources each other in order to
achieve the business strategies and certain objectives, and restricted by the various
agreements and contracts in one certain period each other [1-3]. The development of
network manufacturing technology is restricted by the network manufacturing environ-
ment resource management, and suitable cooperative partner is the necessary condition
to ensure the efficient functioning of the dynamic enterprises alliance. It is the key tech-
nology to evaluate the cooperative partner synthetically and select the optimum coop-
erative partner scientifically to realize the network manufacturing.
Selecting the optimum cooperative partner is a typical combinatorial optimization
problem, and it also is one key problem in the process of establishing the dynamic
enterprise alliance. It is very difficult to select the optimum cooperative partner sim-
ply depending on the experience of the influence factors. The fuzzy comprehensive
evaluation is one kind of effective multitudinous factors decision method used to
evaluate the object synthetically affected by various factors, and the influence factors
of selecting the cooperative partner always are uncertain, so the fuzzy comprehensive

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 590–598, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Selecting Cooperative Enterprise in Dynamic Enterprise Alliance 591

evaluation is the optimum method to evaluate the cooperative partner synthetically,

and select the optimum cooperative partner [4-7].

2 Decision Making of Fuzzy Comprehensive Evaluation on the

Cooperative Partners of the Dynamic Enterprise Alliance
Dynamic enterprise alliance is a generalized business organization. It includes multi-
tudinous cooperative partners, such as design firms, manufacturing enterprises, com-
ponent manufacturers, assembly plant, wrapping enterprise, advertising company,
logistics enterprise, sales company and after-sale service company, and so on. The
influence factors on selecting the optimum cooperative partner can be divided into
internal factors and external factors. The internal factors mainly include design,
innovation, manufacturing, cost, time, quality, sale, serve, informatization, staff,
management, culture and credibility, and so on. The external factors mainly include
geography condition, cooperation ability, logistical support system and social envi-
ronment, and so on. Towards the different kind of cooperative enterprise, the thinking
emphasis of the influence factors is different. For example, toward the manufacturing
enterprise, it mainly thinks of the technological levels of workers, the applicability
and advanced degree of the manufacturing equipments, manufacturing technique, the
rationality of process and the credibility of the manufacturing enterprise in the entire
manufacturing industry. In this paper, the sales enterprise is taken as the example to
research the method of selecting cooperative enterprise in dynamic enterprise alliance
based on fuzzy comprehensive evaluation.

2.1 The Influence Factor Set Establishment

The influence factor set is one ordinary set comprised by various factors influencing
the fuzzy comprehensive evaluation on the cooperative partner. It mainly takes into
account six factors while the sales enterprise is evaluated, such as the economic
strength, marketing strategy, economic benefit, development prospect, staff quality
and sales channels.
The economic strength factor mainly includes registered capital, permanent assets,
bank loans and liquid assets. The marketing strategy factor mainly includes marketing
purposes, marketing planning, market research and market positioning. The economic
benefit factor mainly includes return on assets, sales net profit rate, profit-tax rate of
cost and velocity of liquid assets. The development prospect factor mainly includes
enterprise culture, customer relationship, sales achievement, service quality and strain
capacity. The staff quality factor mainly includes professional dedication, professional
skill, insight, psychology bearing capacity and decision-making ability. And the sales
channels factor mainly includes the economic benefits of sales channels, the ability of
enterprise to control the sales channels and the adaptability of the sales channels to
the market environment. Thus, the influence factor set U can be represented as
U = ( u1 u2 u3 u4 u5 u6 ) (1)
592 Y. Li, G. Zhang, and J. Liu

In (1), the subset u1 indicates economic strength, the subset u 2 indicates marketing
strategy, the subset u 3 indicates economic benefit, the subset u4 indicates develop-
ment prospect, the subset u5 indicates staff quality and the subset u6 indicates sales
channels.
The economic strength subset u1 can be represented as

u1 = ( u11 u12 u13 u14 ) (2)

In (2), u11 indicates registered capital, u12 indicates permanent assets, u13 indicates
bank loans and u14 indicates liquid assets.
The marketing strategy subset u2 can be represented as

u2 = ( u21 u22 u23 u24 ) (3)

In (3), u21 indicates marketing purposes, u22 indicates marketing planning, u23 indi-
cates market research and u24 indicates market positioning.
The economic benefit subset u3 can be represented as

u3 = ( u31 u32 u33 u34 ) (4)

In (4), u31 indicates return on assets, u32 indicates sales net profit rate, u33 indicates
profit-tax rate of cost and u34 indicates velocity of liquid assets.
The development prospect subset u4 can be represented as

u4 = ( u41 u42 u43 u44 u45 ) (5)

In (5), u41 indicates enterprise culture, u42 indicates customer relationship, u43 indicates
sales achievement, u44 indicates quality of service and u45 indicates strain capacity.
The quality of staff subset u5 can be represented as

u5 = ( u51 u52 u53 u54 u55 ) (6)

In (6), u51 indicates professional dedication, u52 indicates professional skill, u53 indi-
cates insight, u54 indicates psychology bearing capacity and u55 indicates decision-
making ability.
The sales channels subset u6 can be represented as

u6 = ( u61 u62 u63 ) (7)

In (7), u61 indicates economic benefits of sales channels, u62 indicates ability of en-
terprise to control the sales channels and u63 indicates adaptability of the sales chan-
nels to the market environment.
 Selecting Cooperative Enterprise in Dynamic Enterprise Alliance 593

2.2 The Evaluation Set Establishment

The evaluation set is comprised of various possible evaluate results made by the deci-
sion makers. The economic strength subset u1 , marketing strategy subset u2 , eco-
nomic benefit subset u3 , development prospect subset u4 , staff quality subset u5 and
sales channels subset u6 can be separately evaluated by the set of {good, relatively
good, general, bad, sucks}, {high, relatively high, general, low, very low} and {long,
relatively long, general, short, very short}, and so on. In this paper, the evaluation set
of economic strength subset u1 , marketing strategy subset u2 , economic benefit
subset u3 , development prospect subset u4 , staff quality subset u5 and sales channels
subset u6 are unified into one evaluation set V , it can be represented as

V = ( v1 v2 v3 v4 v5 ) (8)

In (8), v1 means excellent, v2 means good, v3 means middle, v4 means pass muster
and v5 means bad.

2.3 The Weight Set Establishment

The weight set is comprised of every influence factor’s weight number. It can reflect
every influence factor’s importance. Assuming ai is the weight number of the influ-
ence factor ui , thus, the weight set A can be represented as

A = ( a1 a2 L am ) (9)

Usually, every influence factor’s weight number should meet polarity and non-
negativity constraint. i.e.
⎧ m
⎪ ∑ ai = 1
⎨ i =1 (10)
⎪0 ≤ a ≤ 1
⎩ i

Different evaluator maybe has the different attitude toward the same thing, and the
weight numbers offered by them also are different. In this paper, the weighted statis-
tics method is adopted to determine the weight number of every influence factor [8].
Firstly, make a weight distribution questionnaire (shown as table 1), then ask some
experts or related people fill in the optimum weight number who believe, after taking
bake the weight distribution questionnaires, adopt the weighted statistics method to
calculate the weight number A .

Table 1. Weight Distribution Questionnaire

influence factor ui u1 u2 u3 ∑
weight number ai a1 a2 a3 1
594 Y. Li, G. Zhang, and J. Liu

The weight number set A can be calculated through the statistical investigation on
influence degree of the six subsets u1 , u 2 , u3 , u 4 , u 5 and u6 on the making-
decision of selecting the cooperative partner decision.
A = ( 0.2 0.23 0.12 0.08 0.15 0.22 ) (11)

In the same way, the weight number sets of the six subsets u1 , u 2 , u3 , u 4 , u 5 and
u6 also can be calculated as
A1 = ( 0.23 0.18 0.27 0.32 ) (12)

A2 = ( 0.22 0.28 0.19 0.31) (13)

A3 = ( 0.22 0.28 0.19 0.31) (14)

A4 = ( 0.09 0.22 0.27 0.23 0.19 ) (15)

A5 = ( 0.29 0.22 0.18 0.12 0.19 ) (16)

A6 = ( 0.39 0.33 0.28 ) (17)

3 The Membership Function Establishment

It is the key point to determine the membership degree for the fuzzy synthetic evalua-
tion, and it can directly influence the result of evaluation. The membership degree of
the j-th evaluating indicator of the influence factor ui can be calculated through the
membership function.
Using the method of fuzzy statistics, to calculate the membership degree of the j-th
evaluating indicator of the influence factor ui offered by every expert, and then the
weight of the evaluating indicator aij or the span of the value aij of the membership
degree µij can be calculated using statistical analysis method [9]. The value aij of
the membership degree µij is shown as table 2.

Table 2. The Value aij of the Membership Degree µij

factor set factor subset v1 v2 L vm

ui1 a10 − a11 a11 − a12 L a1m −1 − a1m
ui 2 a20 − a21 a21 − a22 L a2 m −1 − a2 m
ui
M M M M M
uin an 0 − an1 an1 − an 2 L anm −1 − anm
 Selecting Cooperative Enterprise in Dynamic Enterprise Alliance 595

Based on the table 2, the membership function can be structured and represented as

⎧ 1 (ai 0 ≤ x ≤ ai1 )
⎪
µi1 ( x ) = ⎨(ai 2 − x) (ai 2 − ai1 ) (ai1 ≤ x ≤ ai 2 ) (18)
⎪ (ai 2 ≤ x ≤ aim )
⎩ 0

In (18), i = 1, 2,L , n .

⎧0 (ai 0 ≤ x ≤ aij − 2 )
⎪( x − a ) ( a − a ) (aij − 2 ≤ x ≤ aij −1 )
⎪ ij − 2 ij −1 ij − 2
⎪
µi j ( x ) = ⎨1 (aij −1 ≤ x ≤ aij ) (19)
⎪( a − x ) ( a − a ) (aij +1 ≤ x ≤ aim )
⎪ ij +1 ij +1 ij

⎪⎩0 (aij ≤ x ≤ aij +1 )

In (19), i = 1, 2,L , n and j = 2,3,L , m − 1 .

⎧0 (ai 0 ≤ x ≤ aim − 2 )
⎪
µim ( x ) = ⎨( x − aim − 2 ) (aim −1 − aim − 2 ) (aim − 2 ≤ x ≤ aim −1 ) (20)
⎪1
⎩ (aim −1 ≤ x ≤ aim )

In (20), i = 1, 2,L , n .
The membership degree µij of the j-th evaluating indicator of the influence factor
ui on one certain technological design scheme can be calculated through the mem-
bership function.

4 Fuzzy Comprehensive Evaluation

4.1 The Primary Fuzzy Comprehensive Evaluation

Assuming the primary fuzzy comprehensive evaluation is carried out on the influence
factor ui in the influence set U , thus, the membership degree µij of the j-th evaluat-
ing indicator of the influence factor ui can be calculated through the membership
function, and the evaluation result of the single factor ui can be represented by the
fuzzy set as Rij

µi1 µi 2 µim
Rij = + +L + (21)
v1 v2 vn

In (21), Rij is the single factor evaluation set, it can be simply represented as

Rij = ( ri1 ri 2 L rim ) (22)

596 Y. Li, G. Zhang, and J. Liu

At the same way, the evaluation set corresponding to every influence factor ui can be
obtained. And the single factor evaluation matrix Ri can be represented as

Ri = [ Ri1 Ri 2 L Ri 4 ]
T
(23)

In (23), Ri is the single factor fuzzy evaluation matrix.

The primary evaluation set Bi can be obtained through the fuzzy comprehensive
evaluation on the influence factor ui .

Bi = Ai o Ri = ( bi1 bi 2 L bim ) (24)

4.2 The Secondary Fuzzy Comprehensive Evaluation

With the primary evaluation set B1 , B2 , L , and Bk , the single factor evaluation
matrix R of the influence factor set U can be represented as

⎡ B1 ⎤ ⎡ b11 b12 L b1m ⎤

⎢ B ⎥ ⎢b b L b2 m ⎥⎥
R = ⎢ 2 ⎥ = ⎢ 21 22 (25)
⎢ M ⎥ ⎢M M M M ⎥
⎢ ⎥ ⎢ ⎥
⎣ BK ⎦ ⎣bk 1 bk 2 L bkm ⎦
The fuzzy comprehensive evaluation set B can be obtained through the fuzzy com-
prehensive evaluation on the influence factor set U , and it can be represented as
B = A o R = ( b1 b2 L bm ) (26)

Through the normalization of the fuzzy comprehensive evaluation set B , the fuzzy
comprehensive evaluation on one certain technological design scheme can be carried
out with the maximum membership degree method, the weighted average method or
the fuzzy distribution method.

5 An Example of Fuzzy Comprehensive Evaluation

The fuzzy comprehensive evaluation on one certain sales enterprise is carried out.
Based on the statistical results, the evaluation matrix of the subset u1 can be obtained
and represented as
⎡ 0.35 0.42 0.18 0.04 0.01⎤
⎢0.20 0.53 0.23 0.02 0.02 ⎥⎥
R=⎢ (27)
⎢ 0.41 0.37 0.16 0.06 0 ⎥
⎢ ⎥
⎣ 0.38 0.49 0.10 0.02 0.01⎦
 Selecting Cooperative Enterprise in Dynamic Enterprise Alliance 597

The comprehensive evaluation result of the subset u1 can be calculated and repre-
sented as B1

B1 = A1 o R1 = ( 0.20 0.37 0.23 0.18 0.18 ) (28)

At the same way, the comprehensive evaluation results of the subset u2 , u3 , u4 , u5

and u6 can be calculated and represented as B2 , B3 , B4 , B5 , and B6 . Then, the
evaluation matrix R of the factor set U can be calculated and represented as

⎡ B1 ⎤ ⎡ 0.20 0.37 0.23 0.18 0.18 ⎤

⎢ B ⎥ ⎢0.27 0.32 0.30 0.17 0.17 ⎥⎥
⎢ 2⎥ ⎢
⎢ B ⎥ ⎢ 0.23 0.21 0.22 0.19 0.19 ⎥
R = ⎢ 3⎥ = ⎢ ⎥ (29)
⎢ B4 ⎥ ⎢ 0.22 0.29 0.25 0.09 0.09 ⎥
⎢ B5 ⎥ ⎢ 0.19 0.21 0.22 0.12 0.12 ⎥
⎢ ⎥ ⎢ ⎥
⎣⎢ B6 ⎦⎥ ⎣⎢ 0.28 0.34 0.32 0.28 0.28⎦⎥

The comprehensive evaluation result of the factor set U can be calculated and repre-
sented as B
B = A o R = ( 0.20 0.21 0.22 0.09 0.09 ) (30)

The normalized result of the comprehensive evaluation result B can be calculated

and represented as B '
B ' = ( b1 ' b2 ' b3 ' b4 ' b5 ') = ( 0.247 0.259 0.272 0.111 0.111) (31)

To assign the value to the evaluation set V , assuming the value of the evaluation set
V is
V = (1 0.85 0.75 0.6 0.5 ) (32)

Think of b j is the weight of the evaluation target v j , and the final evaluation result of
the certain sales enterprise can be calculated with the weighted average method.
6
V = ∑ bj ⋅ v j =
j =1 (33)
1× 0.247 + 0.85 × 0.259 + 0.75 × 0.272 + 0.6 × 0.111 + 0.5 × 0.111 = 0.793
It means the fuzzy comprehensive evaluation value of the certain sales enterprise is
0.793.
At the same way, the fuzzy comprehensive evaluation value of all the sales enter-
prises which have the cooperation intent can be obtained. And the sales enterprise it
has the maximal fuzzy comprehensive evaluation value can be selected as the opti-
mum cooperative partner.
598 Y. Li, G. Zhang, and J. Liu

Similarly, the fuzzy comprehensive evaluation on the other cooperative enterprise

in the other business area can be carried out, and the optimum cooperative partner
also can be selected.

6 Conclusion
It can play the role of experts adequately, reduce the harm caused by the personal
subjective assume, and provide the scientific basis for evaluating the enterprises
which have the cooperation intent based on comprehensive evaluation of the coopera-
tive partners in the dynamic enterprise alliance with the fuzzy comprehensive evalua-
tion method. It can improve the assessment level on the cooperative partners and
make the evaluation result more scientific with the fuzzy comprehensive evaluation
method.

References
1. Wanshan, W., Yadong, G., Peili, Y.: Networked manufacturing. Northeast University Press,
Shenyang (2003) (in Chinese)
2. Jicheng, L., Jianxun, Q.: Dynamic alliance synergistic decision-making model based on
business intelligence center. In: International Symposium on Information Processing, ISIP
2008 and International Pacific Workshop on Web Mining and Web-Based Application,
WMWA 2008, Moscow, Russia, May 23-25, pp. 219–223 (2008)
3. Congqian, Q., Yi, G.: Partner optimization of enterprises dynamic alliance. Journal of
Tongji University 35(12), 1674–1679 (2007) (in Chinese)
4. Lixin, W.: Fuzzy System & Fuzzy Control Tutorial. Tsinghua University Press, Beijing
(2003) (in Chinese)
5. Baoqing, H.: The basis of fuzzy theory. Wuhan University Press, Wuhan (2004)
6. Bing, Z., Zhang, R.: Research on fuzzy-grey comprehensive evaluation of software process
modeling methods. In: Proceedings - 2008 International Symposium on Knowledge Acqui-
sition and Modeling, KAM 2008, Wuhan, China, December 21-22 (2008)
7. Xulin, L., Baowei, S.: Three level fuzzy comprehensive evaluation based on Grey Rela-
tional Analysis and Entropy weights. In: 2008 International Symposium on Information
Science and Engineering, ISISE 2008, Shanghai, China, December 20-22, vol. 2, pp. 32–35
(2008)
8. Korotchenko, M.A., Mikhailov, G.A., Rogazinskii, S.V.: Value modifications of weighted
statistical modelling for solving nonlinear kinetic equations. Russian Journal of Numerical
Analysis and Mathematical Modelling 22(5), 471–486 (2007)
9. Shinguang, C., Yikuei, L.: On performance evaluation of ERP systems with fuzzy mathe-
matics. Expert Systems with Applications 36(3 Part 2), 6362–6367 (2009)
 An Evolutionary Solution for Cooperative
and Competitive Mobile Agents

Jiancong Fan1, Jiuhong Ruan2, and Yongquan Liang1

1
College of Information Science and Engineering, Shandong University of Science and
Technology, Qingdao 266510, China
howdoyoudo07@yahoo.com.cn
2
Scientific Research Department, Shandong Jiaotong University, Jinan 250023, China
jh_ruan@bit.edu.cn

Abstract. The cooperation and competition among mobile agents using evolu-
tionary strategy is an important domain in Agent theory and application. With
evolutionary strategy the cooperation process is achieved by training and iterat-
ing many times. From evolutionary solution of cooperative and competitive
mobile agents (CCMA), a group of mobile agents are partitioned into two popu-
lations, cooperative agents group and competitive agent group. Cooperative
agents are treated as several pursuers, while a competitive agent is viewed as
the pursuers’ competitor called evader. The cooperation actions take place
among the pursuers in order to capture the evader as rapidly as possible. An
agent individual (chromosome) is encoded based on a kind of two-dimensional
random moving. The next moving direction is encoded as chromosome. The
chromosome can be crossed over and mutated according to designed operators
and fitness function. An evolutionary algorithm for cooperation and competi-
tion of mobile agents is proposed. The experiments show that the algorithm for
this evolutionary solution is effective, and it has better time performances and
convergence.

Keywords: evolution, mobile agent, cooperative agents, competitive agents.

1 Introduction
Mobile Agent is a kind of software Agent. Mobile Agent has many characteristics,
such as autonomy, sociality, self-learning, and importantly mobility[1]. Mobile Agent
can move from one position to another. The moving involves in the states transition
and Agent’s states change. Under current context and state mobile Agent autono-
mously determines what time to move and what place it moves to[2,3].
Cooperative mobile agents have the ability of cooperation and adaptability[4,5].
Firstly, agents can cooperate by exchanging data and/or code when they meet on a
given condition. Secondly, the behavior of an agent can change based on its current
state and the information it has gathered while traveling.
In this paper the cooperation among mobile agents is studied using evolutionary
strategy. The determination of cooperation process and form is achieved by training
and iterations many times. The evolvement result is that a reasonable cooperation
strategy is obtained and stable.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 599–607, 2009.
© Springer-Verlag Berlin Heidelberg 2009
600 J. Fan, J. Ruan, and Y. Liang

2 Relative Work
The combination of evolutionary algorithm and cooperative mobile agents has been
studied in some aspects, most of which use agents to study or implement evolutionary
computation processes. The literature [6,7] presented an agent-based version of coop-
erative co-evolutionary algorithm. This type of systems has been already applied to
multi-objective optimization. Evolution learning has also been studied for multi-agent
with strategic coalition [8], which used the iterated prisoner’s dilemma game to model
the dynamic system in an evolutionary learning environment. In literature [9], an evo-
lution strategy is introduced on the basis of cooperative behaviors in each group of
agents. The evolution strategy helps each agent to be self-defendable and self-
maintainable, and agents in same group cooperate with each other. This method use
reinforcement learning, enhance neural network and artificial life. In literature [10] a
neural network is used for the behavior decision controller. The input of the neural
network is decided by the existence of other agents and the distance to the other agents.
The output determines the directions in which the agent moves. The connection weight
values of this neural network are encoded as genes, and the fitness of individuals is
determined using a genetic algorithm. There are also other studies from the perspec-
tives of game theory [11,12,13]. Other studies concentrating on neural computing can
be obtained in literatures [14,15,16].

3 Cooperative and Competitive Mobile Agents

In this paper, cooperative agents are treated as several pursuers, while a competitive
agent is viewed as the pursuers’ competitor called evader. The cooperation actions
take place among the pursuers in order to capture the evader as rapidly as possible.
However, the competitive agent evader doesn’t cooperate with the pursuers so as to
avoid the capturing.
The action area of pursuers P={p1, p2, …, pn} and evaders E={e1, e2, …, em} is pro-
jected to a two-dimension coordination plane. P and E are represented as particles in
the coordinate in our application experiments. The locations of P and E are repre-
sented as coordinate variables (x, y). For convenience but not influencing the prob-
lem-solving results, the coordinate variables (x, y) are discrete and expanded as square
regions. The edge length of each square is ε (ε >0). And pi(i =1,2,…,n) and
ej(j=1,2,…,m) move between two adjacent squares, which is called two-dimensional
random moving.
Definition 1. reflection barriers: in two-dimension coordinate plane XOY, (X,0) and
(0,Y) are called reflection barriers.
Definition 2. two-dimensional random moving: in two-dimension coordinate plane
XOY, Agents move autonomously and accord with the following terms:
(1) Agents can only move to adjacent coordinate cells for each moving;
At random circumstances,
(2) If one side of Agents current location is reflection barriers, Agents have k (k =3 or
5) alternative moving locations. They move to the adjacent areas with probability 1/k.
 An Evolutionary Solution for Cooperative and Competitive Mobile Agents 601

(3) If there don’t exist reflection barriers around the Agents, the Agents move to the
adjacent areas with probability 1/8.
If the Agents have preferences under particular conditions,
(4) The Agents move to the adjacent areas with probability λi P(di), where
λi is called preference coefficient. λi denotes a measure of Agenti’s customs or cogni-
tive styles of moving directions,
di denotes moving direction of Agenti at particular moment, and at this moment
there exist reflection barriers adjacent to Agenti, then
∑i=1 to 3 λi P(di) =1 . or ∑i=1 to5 λi P(di) =1 . (1)
else,
∑i=1 to 8 λi P(di) =1 . (2)
The moving is called two-dimensional random moving if it is accord with the above
items.
Example 1. There are three Agents showed in figure 1 (a) (b) (c) respectively. (a)
shows an Agent locates at a place where there is no reflection barriers around. There
are 8 alternative directions to move towards. It probably likes to move upward verti-
cally or others. (b) shows left side of an Agent’s location is reflection barrier. There
are 5 directions to choose. (c) shows there are only three directions to choose because
there are two reflection barriers around the Agent.

(a) no reflection barriers (b) reflection barriers at one side (c) reflection barriers at two sides

Fig. 1. Agents’ two-dimensional random moving

4 Evolutionary Solution for CCMA

There are two populations designed in this paper. One population has k individuals
cooperating with each other, that is, population size is k. The other has one individual
which is a competitor of the above k cooperative individuals.

4.1 Encoding Schemes for Individual

The chromosome is encoded based on two-dimensional random moving. The next

moving direction of individual is encoded as chromosome. One agent locates at one
of the two positions, no reflection barriers and reflection barriers. Figure 2 (a) and (b)
illustrate respectively. In Figure 2 (a) there are eight selected directions. An agent’s
next moving scheme can be represented as 8 bits binary code. The codes from left to
602 J. Fan, J. Ruan, and Y. Liang

right correspond to the locations 0 to 7. The bit 1 denotes the agent moves to the di-
rection, and 0 doesn’t. On no barriers circumstances the encoding resembles with the
above. The only difference is that the directions with barriers are coded with a charac-
ter neither 1 nor 0, such as an asterisk.

0 1 2 1 2
7 · 3 · 3
6 5 4 5 4

(a) no barriers (b)reflection barriers

Fig. 2. Numbering of moving directions

For example, the chromosome (0 1 0 0 0 0 0 0) represents an agent moves verti-

cally upward because the bit of location 1 is 1 while the other locations are 0. The
chromosome (* 0 0 1 0 0 * *) represents the adjacent left side is reflection barrier and
an agent moves right horizontally.

4.2 Evolutionary Operators

Selection. The chromosomes compete against each other and the chromosome with
the highest fitness value is selected. In this paper distance-based function is adopted
as the fitness function, that is, the Euclid distance between pursuer and evader is the
fitness value. The smaller fitness value is better for pursuing agent while the larger
fitness value is better for evading agent.
Crossover. The crossover operation exchanges parts of a pair of chromosomes, creat-
ing new chromosomes called children or offspring (If crossover does not succeed in
the probability test, then the children will be identical copies of their parents). The
crossover operation needs two chromosomes, so the population with over two agents
can be applied with crossover operator. An agent moves to only one direction. There
should be strict limitation that the offspring must have one bit 1 only. Figure 3 illus-
trates a crossover operation in the solution.

Parents A: 1 0 0 0 0 0 0 0 B: 0 0 0 1 0 0 0 0
Offspring C: 0 1 0 0 0 0 0 0 D: 0 0 1 0 0 0 0 0

Fig. 3. Crossover operation

Mutation. The mutation operator aims to increase the variability of the population,
allowing the evolution to simultaneously search different areas of the solution space.
This operator changes at random the value of a chromosome gene, also randomly
chosen with a given probability (named mutation rate). Also, an agent moves to only
one direction. If one bit 1 becomes 0, some other bit 0 should become 1. Figure 4
shows a mutation operation that changes the value of the forth gene from 0 to 1.
 An Evolutionary Solution for Cooperative and Competitive Mobile Agents 603

Before mutation 0 0 1 0 0 0 0 0
After mutation 0 0 0 0 1 0 0 0

Fig. 4. Mutation operation

4.3 Evolutionary Algorithm for CCMA

Let k cooperative agents (a cooperative agent abbr. cooAg) and a competitive agent
(abbr. comAg) consist of a niche in which k cooAgs pursue a comAg. In the process
of all agents’ two-dimensional random moving cooAgs pursue comAg and want to
capture comAg as rapidly as possible. The comAg tries its best not to be captured.
Once one of the cooAgs captures comAg, the algorithm halts. A penalty factor is
designed in the algorithm in order to evaluate the quality of cooAgs in pursuing co-
mAg. The penalty factor α is defined as formula (3).
Let
x’ = P(cooAgi).x - P(cooAgi).x
y’ = P(cooAgi).y - P(cooAgi).y
D = (x’2 + y’2 )1/2
Then
α, D ≤ θ0 .
α(cooAgi) = α+1, θ0 < D ≤ θ1 . (3)
α+2, θ1 < D ≤ θ2 .

(1) Initialize distance threshold value Ti, time interval T, capture distance d;
//initialize penalty factor D
(2) for (int i=0; i<=k; i++)
(3) D(cooAg[i]) = 0;
(4) t = 0; //start from time 0
(5) while (true){
t = t+T;
(6) for (int j=0; j<=k; j++)
(7) Compute D(cooAg[i], comAg);
(8) if(D<=d){
(9) Capture succeed and break; }
(10) else{
(11) Compute D(cooAg[i]) according to D and formula (1); }
(12) Select the chromosome with the smallest D to the next generation;
(13) Crossover the chromosomes with bigger D to produce offspring;
(14) Mutate the chromosome with the largest D;
(15) if (t exceeds the longest time threshold) break;
(16) } //The algorithm halt

Fig. 5. Algorithm description for CCMA

604 J. Fan, J. Ruan, and Y. Liang

In formula (3), P( · ) is agent’s position in coordinate system. θ0, θ1, θ2 are threshold
values of distance. The distances between cooAgs and comAg are computed at a fixed
interval T. From time 0, at time kT (k=1, 2, …) the distance D between each cooAg
and comAg is computed and the penalty factor α is computed according to D. The
chromosome with smaller α has better fitness. The chromosomes of cooAgs with
bigger α are crossed over or mutated. The chromosome with the smallest α is copied
to the next generation.
In our algorithm, it is assumed that if the distance D between some cooAg and co-
mAg is less than a threshold value d, the cooperative capture succeeds. The threshold
d is called capture distance. The algorithm description is given by figure 5.

5 Experimental Analysis
Assume that XOY shown in figure 1 is participants’ active area. The whole active
area is discrete, and the participants perform two-dimensional random moving in
XOY. In our experiments, each cell is represented by coordinate value, that is, a cell
as a coordinate point. Although the agents move randomly, they have their own pref-
erences, which are illustrated in figure 6 and 7. In figure 6 two persuading agents and

P1
P1

P2
P2

(a) no capturing (b) capturing occurring

Fig. 6. Area of two cooperative agents and one competitive agent

E E
P1

P3
P3
P1
P2
P2

(a) no capturing (b) capturing occurring

Fig. 7. Area of three cooperative agents and one competitive agent

 An Evolutionary Solution for Cooperative and Competitive Mobile Agents 605

an evading agent are interacting in some niche. In figure 6(a) the evader prefers to
move to the direction on the whole as the arrow shows, but two cooperative persuad-
ers P1 and P2 search within the scope of area they like. Under this circumstance, P1
and P2 can’t capture the evader however they endeavor. On the contrary, figure 6(b)
shows that the evader E has preferential moving direction meeting with the pursuer
P1’s. Figure 7 shows the similar situation but with three persuading agents.
In this experiments, the evading agent moves at random. The algorithm proposed in
this paper is mainly used to solve the problems shown in figure 6(a) and figure 7(a).
During evolution the point with the next moving direction towards the evading
agent’s preferential direction and the nearest distance with the evading agent is se-
lected to continue the persuader’s moving. Under other circumstances the agents’
chromosome is crossed over or mutated.
Table 1 shows the comparison of random capture times and evolutionary capture
times with three cooperative agents. The capture distance d takes values 50, 100 and
200. Maximized moving times are designed with six values. With each d and all de-
signed moving times 50 mobile experiments are implemented with random and evolu-
tionary modes respective. The experiment results show that the times of evolutionary
cooperative mobile agents capture successfully are more than random mobile agent
obviously.

Table 1. Comparison of random capture times and evolutionary capture times

moving times random mobile evolutionary mobile

d
succeed fail succeed fail
5000 1 49 5 45
10000 7 43 13 37
20000 17 33 29 21
50 50000 32 18 39 11
100000 37 13 45 5
200000 41 9 48 2
5000 6 44 18 32
10000 15 35 31 19
20000 28 22 40 10
100 50000 33 17 44 6
100000 41 9 50 0
200000 45 5 50 0
5000 25 25 37 13
10000 31 19 42 8
20000 33 17 47 3
200 50000 37 13 50 0
100000 42 18 50 0
200000 45 5 50 0
606 J. Fan, J. Ruan, and Y. Liang

The average iterative generations and average capture time according to different
capture distances 50, 100 and 200 listed in table 2. From the table we can see that at
each distance less moving times and more moving times can lead to more average
iterative generations and capture time than medium moving times. This is because
less moving times make the agents’ mobility ceasing much faster, and more moving
times make their mobile preferences more diverse.

Table 2. Average iterative generations and average capture time according to different capture
distances

Capture Moving times Average Iterative Average

distance (103) generations capture time (ms)
5 4301 243
10 3747 202
50 20 3194 145
50 7868 798
100 20075 1892
5 6862 386
10 5014 299
100 20 4341 214
50 9881 1319
100 17150 3008
5 4844 312
10 3892 260
200 20 2368 192
50 2663 194
100 6554 702

6 Conclusions

As it can be observed in section 5 analyzing the quality of obtained solutions as the

function of iteration and time, evolutionary multi-agent system with co-operation is
very attractive alternative since initially solutions proposed by this algorithm were
better than solutions proposed by “classical” (non evolution-based) algorithms. Evo-
lution process can assure higher cooperative efficiency and optimal strategy.
The future plans include further investigation of the proposed mechanism. It could
be interesting to modify the co-operation proposed in this paper in such a way that
agents from different sub-populations (species) would specialize in different criteria
and form aggregates (teams) composed of the represents of different species in order
to solve the problem, for example, more competitive agents and more cooperative
agents, even more populations.
 An Evolutionary Solution for Cooperative and Competitive Mobile Agents 607

Acknowledgments. This work is supported by Natural Science Foundation of Shan-

dong Province of China (NO G2007Y07) and Spring Bud Planning of SUST(NO
06540040512).

References
1. Rao, A.S., Georgeff, M.P.: BDI agents: From theory to practice. In: Proc. the First Interna-
tional Conference on Multi-Agent Systems, San Franciso, CA, USA, pp. 312–319 (1995)
2. Lange, D.B., Oshima, M.: Seven good reasons for mobile agents. Communications of the
ACM 42(3), 88–89 (1999)
3. Kotz, D., Gray, R.S.: Mobile Agents and the Future on the Internet. ACM Operating Sys-
tems Review 33(3), 7–13 (1999)
4. Bassett, J., De Jong, K.: Evolving behaviors for cooperating agents. In: Ras, Z. (ed.) Pro-
ceedings from the Twelfth International Symposium on Methodologies for Intelligent Sys-
tems, Charlotte, NC, pp. 157–165. Springer, Heidelberg (2000)
5. Berenji, H., Vengerov, D.: Learning, cooperation, and coordination in multi-agent systems,
Technical Report IIS-00-10, Intelligent Inference Systems Corp., 333 W. Maude Avennue,
Suite 107, Sunnyvale, CA 94085–4367 (2000)
6. Dreżewski, R., Siwik, L.: Agent-based co-operative co-evolutionary algorithm for multi-
objective optimization. In: Rutkowski, L., Tadeusiewicz, R., Zadeh, L.A., Zurada, J.M.
(eds.) ICAISC 2008. LNCS (LNAI), vol. 5097, pp. 388–397. Springer, Heidelberg (2008)
7. Dreżewski, R., Siwik, L.: Multi-objective optimization technique based on co-evolutionary
interactions in multi-agent system. In: Giacobini, M. (ed.) EvoWorkshops 2007. LNCS,
vol. 4448, pp. 179–188. Springer, Heidelberg (2007)
8. Yang, S.-R., Cho, S.-B.: Co-evolutionary learning with strategic coalition for multiagents.
Appl. Soft Comput. 5(2), 193–203 (2005)
9. Lee, M.: A study of evolution strategy based cooperative behavior in collective agents. Ar-
tificial Intelligence Review 25(3), 195–209 (2006)
10. Lee, M., Chang, O.-b., Yoo, C.-J., et al.: Behavior evolution of multiple mobile agents un-
der solving a continuous pursuit problem using artificial life concept. Journal of Intelligent
and Robotic Systems 39(4), 433–445 (2004)
11. Nitschke, G.: Designing emergent cooperation: a pursuit- evasion game case study. Artifi-
cial Life Robotics 9(4), 222–233 (2005)
12. Tanev, I., Shimohara, K.: Effects of learning to interact on the evolution of social behavior
of agents in continuous predators-prey pursuit problem. In: Banzhaf, W., Ziegler, J.,
Christaller, T., Dittrich, P., Kim, J.T. (eds.) ECAL 2003. LNCS (LNAI), vol. 2801, pp.
138–145. Springer, Heidelberg (2003)
13. Tanev, I., Brzozowski, M., Shimohara, K.: Evolution, generality and robustness of
emerged surrounding behavior in continuous predators-prey pursuit problem. Genetic Pro-
gramming and Evolvable Machines 6(3), 301–318 (2005)
14. Lee, M., kang, E.-K.: Learning enabled cooperative agent behavior in an evolutionary and
competitive environment. Neural compu. & applic. 15(2), 124–135 (2006)
15. Lee, M.: Evolution of behaviors in autonomous robot using artificial neural network and
genetic algorithm. Information Sciences 155(1-2), 43–60 (2003)
16. Caleanu, C.-D., Tiponut, V., et al.: Emergent behavior evolution in collective autonomous
mobile robots. In: 12th WSEAS international conference on system, Heraklion, Greece,
July 2008, pp. 428–433 (2008)
 Towards an Extended Evolutionary Game Theory with
Survival Analysis and Agreement Algorithms for
Modeling Uncertainty, Vulnerability, and Deception

Zhanshan (Sam) Ma

IBEST (Initiative for Bioinformatics and Evolutionary Studies) & Departments of

Computer Science and Biological Sciences, University of Idaho, Moscow, ID, USA
ma@vandals.uidaho.com

Abstract. Competition, cooperation and communication are the three funda-

mental relationships upon which natural selection acts in the evolution of life.
Evolutionary game theory (EGT) is a 'marriage' between game theory and Dar-
win's evolution theory; it gains additional modeling power and flexibility by
adopting population dynamics theory. In EGT, natural selection acts as optimi-
zation agents and produces inherent strategies, which eliminates some essential
assumptions in traditional game theory such as rationality and allows more real-
istic modeling of many problems. Prisoner's Dilemma (PD) and Sir Philip Sid-
ney (SPS) games are two well-known examples of EGT, which are formulated
to study cooperation and communication, respectively. Despite its huge
success, EGT exposes a certain degree of weakness in dealing with time-,
space- and covariate-dependent (i.e., dynamic) uncertainty, vulnerability and
deception. In this paper, I propose to extend EGT in two ways to overcome the
weakness. First, I introduce survival analysis modeling to describe the lifetime
or fitness of game players. This extension allows more flexible and powerful
modeling of the dynamic uncertainty and vulnerability (collectively equivalent
to the dynamic frailty in survival analysis). Secondly, I introduce agreement al-
gorithms, which can be the Agreement algorithms in distributed computing
(e.g., Byzantine Generals Problem [6][8], Dynamic Hybrid Fault Models [12])
or any algorithms that set and enforce the rules for players to determine their
consensus. The second extension is particularly useful for modeling dynamic
deception (e.g., asymmetric faults in fault tolerance and deception in animal
communication). From a computational perspective, the extended evolutionary
game theory (EEGT) modeling, when implemented in simulation, is equivalent
to an optimization methodology that is similar to evolutionary computing ap-
proaches such as Genetic algorithms with dynamic populations [15][17].

Keywords: Evolutionary Game Theory, Extended Evolutionary Game Theory,

Survival Analysis, Agreement Algorithms, Frailty, Deception, Dynamic Hybrid
Fault Models, UUUR (Uncertain, latent, Unobservable or Unobserved Risks),
Strategic Information Warfare, Reliability, Security, Survivability.

1 Background
The endeavor to develop evolutionary game theory (EGT) was initiated in the 1970s
by John Maynard-Smith and Peter Price with the application of game theory to the

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 608–618, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Towards an Extended EGT with Survival Analysis and Agreement Algorithms 609

modeling of animal conflict resolution [21]. Their pioneering research not only
provided a refreshing mathematical theory for modeling animal behavior, but also
revealed novel insights to evolution theory and greatly enriched the evolutionary
ecology. It also quickly became one of the most important modeling approaches for
emerging behavioral ecology in the 1980s. The well-known prisoner's dilemma (PD)
game (especially its latter versions), which plays a critical role in the study of coop-
eration or altruism, is essentially an evolutionary game [1][3]. Today, the applications
of EGT have been expanded to fields well beyond behavioral ecology.
Evolutionary game is a hybrid of continuous static and continuous dynamic games
(differential games) in the sense that its fitness (payoff) functions are continuous, but
the strategies are inheritable phenotypes, which usually are discrete. In the study of
animal communication or engineering communication (such as wireless sensor net-
works), the strategy can be the behavior of animals or communication nodes. Another
feature is that evolutionary game is played by a population of players repeatedly.
Although the latter is not a unique feature of evolutionary games, the integration with
population dynamics theory is somewhat unique and makes it particularly powerful
for modeling biological systems or any systems where population paradigm is impor-
tant [4][9][21][23]. One of the most frequently used models is the replicator dynam-
ics model [4][23].
In evolutionary game theory, replicator dynamics is described with differential
equations. For example, if a population consists of n types E1 , E2 ,..., En with frequen-
cies x1, x2 ,..., xn . . The fitness fi (x) of Ei will be a function of the population structure,
or the vector, x = ( x1, x2 ,..., xn ) . Following the basic tenet of Darwinism, one may
define the success as the difference between the fitness fi (x) of Ei and the average
fitness of the population, which is defined as:

f ( x) = ∑ x f ( x).
i i (1)

Then, the simplest replicator model can be defined as:

dxi / dt = xi [ f i ( x) − f ( x)] (2)

for i = 1,2,..., n. The population x(t ) ∈ Sn, where S n is a simplex, which is the space
for population composition, is similar to mixed strategies in traditional games [4][23].
Another formulation is the Fitness Generating Function (G-function) and is invented
by Vincent and Brown (2005) [23] to specify groups of individuals within a popula-
tion. Individuals are assigned the same G-function if they possess the same set of
evolutionarily feasible strategies and experience the same fitness consequences within
a given environment [23].
EGT still exposes certain limitations in dealing with dynamic (time-, space-, and/or
covariate-dependent) uncertainty, vulnerability, and deception of game players. For
example, the so-called UUUR (Unknown, latent, Unobserved, or Unobservable Risks)
events exist in many problems such as the study of reliability, security and survivabil-
ity of computer networks, and UUUR events are largely associated with uncertainty,
vulnerability, or their combinations (the frailty) [9][19]. Another challenge is the
modeling of deception in communication among the players: the players not only
610 Z. (Sam) Ma

communicate, but also need to reach consensus (agreement) or make decisions with
the existence of dynamic frailty and deception [9][10].
In this paper, I introduce two extensions to EGT to primarily deal with the above
mentioned limitations. In the first extension, I introduce survival analysis and its
'sister' subjects (competing risks analysis and multivariate survival analysis), which
have some uniquely powerful features in dealing with UUUR events, to deal with
dynamic uncertainty and vulnerability. The second extension is designed to address
deception of game players: the consequence of dynamic deception and frailty on the
game strategies, as well as the capability for game players to reach an agreement
under the influences of deception and frailty. There is a third extension—using hedg-
ing principle from mathematical finance for decision-making, which is necessary for
some applications such as survivability analysis [9], prognostic and health manage-
ment in aerospace engineering [19], and strategic information warfare [10] [18], but is
skipped here.
The significance of EGT is probably best revealed by a brief analysis of the fun-
damental elements that Darwin's evolutionary theory addresses. Competition, coop-
eration and communication were clearly in Darwin's primary concerns. According to
an historical analysis by Dugatkin [3] in his landmark volume "Origin of Species"
(Darwin 1859), Darwin focused on the competition, or the struggle for life, but Dar-
win was also clearly concerned with the cooperation or altruism in nature. In his
"Origin of Species," Darwin's solution to altruism was that selection may be applied
to the family, and individual may get the desired benefit ultimately. It took near a
century for scientists to formalize Darwin's idea mathematically with a simple equa-
tion, known as Hamilton's Rule, first formulated by William Hamilton (1964). About
two decades later, Hamilton's collaborative work with political scientist Robert Ax-
elrod ([1]), which implemented Robert Trivers' (1971) suggestion of using PD game
to study altruism, led to the eruption of the study of cooperation in the last three dec-
ades [3]. Today, there is hardly a major scientific field to which PD game has not
been applied. The PD game, especially its latter versions, is essentially the evolution-
ary game. Besides competition and cooperation, Darwin also published a volume on
communication, titled "The Expression of the Emotions in Man and Animals" in 1872. A
century later, "The Handicap principle: a missing piece of Darwin's Puzzle" by Zahavi
(1997) opened a new chapter in the study of animal communication. It was also the
EGT modeling that led to the wide acceptance of the Handicap principle. The sim-
plest EGT model for this principle is the Sir Philip Sydney (SPS) game, which has
been used to elucidate the handicap principle—that animal communication is honest
(reliable) as long as the signaling is costly in a proper way. Therefore, the phenome-
non that EGT seems to offer the most lucid modeling elucidation for both cooperation
and communication shows its power. Of course, the fact that EGT is a 'marriage' be-
tween game theory and Darwin's evolution theory is the best evidence of its impor-
tance in the study of competition (the struggle for life).
The far reaching cross-disciplinary influences of Darwin's evolution theory and PD
game also suggest that the study of communication, especially with EGT (e.g., SPS
game) may spawn important interdisciplinary research topics and generate significant
cross-disciplinary ramifications that are comparable to the study of cooperation and
PD games [9][16]. It is hoped that the extensions to EGT in this paper will also be
helpful to the cross-disciplinary expansion because these extensions address some
 Towards an Extended EGT with Survival Analysis and Agreement Algorithms 611

critical issues that exist across several fields from reliability and survivability of dis-
tributed networks, prognostics and health management in aerospace engineering,
machine learning, to animal communication networks [9][10][16]–[20].
In the following, we assume that the players of an evolutionary game form a popu-
lation, similar to a biological population. Population dynamics over space and time is
influenced by environmental covariates (factors). The covariates often affect the life-
time or fitness of game players by introducing uncertainty and/or frailty. Players may
have stochastically variable vulnerabilities. In addition, the individual players may
form various spatial structures or topologies, which can be captured with a graph
model or evolutionary games on graphs. Players can form a complex communication
network; in the networks, some act as eavesdroppers, bystanders or audience. The
communication signals may be honest or deceitful depending on circumstance, or the
so-called dynamic deception. Finally, the 'failure' (termination of lifetime or a proc-
ess) of players in engineering applications is often more complex than biological
death in the sense that a failing computer node may send conflicting signals to its
neighbor nodes, or the so-called asymmetric faults. Indeed, the challenge in dealing
with somewhat arbitrary Byzantine fault as well as dynamic deception is the major
motivation why we introduce agreement algorithm extension to EGT.
The remainder of this paper is organized as follows: Section 2 and 3 summarize the
first extension with survival analysis, and the second extension with agreement algo-
rithms, to EGT, respectively. Section 4 discusses implementation issues related to the
EEGT, with the modeling of animal communication networks as an example. Due to
the complexity involved, the EEGT has to be implemented as a simulation environ-
ment in the form of software. The page space limitation makes it only feasible to
sketch a framework of the EGGT. In the following, I often use the network settings
for survivability analysis, strategic information warfare, or animal communication as
examples, and use the terms node, individual, and player interchangeably.

2 The Extension to EGT with Survival Analysis

Survival analysis has become a de facto standard in much of the biomedical research
and also found applications in many other disciplines. There are more than a dozen
monographs of survival analysis since the 1980s: e.g., [7] for (univariate) survival
analysis; [5] for multivariate survival analysis; [2] for competing risks analysis. A
series of reviews on survival analysis in reliability, network survivability and com-
puter science can be found in [11][13][14].
Survival analysis can be considered as the study of time-to-event random variables
(also known as failure or survival time) with censoring. Time-to-event random vari-
ables exist widely, e.g., lifetime of animals or wireless sensor network nodes, duration
of eavesdropping in animal communication, duration of aggression displays in con-
flict resolution. The survivor function S(t) is defined as the probability that survival
time (or failure time) T exceeds t; i.e.,
S (t ) = P (T ≥ t ), 0<t<∞ (3)

This definition is exactly the same as the traditional reliability function, but survival
analysis has much rich models and statistical procedures such as Cox's proportional
612 Z. (Sam) Ma

hazard modeling [equ. (4) & (5)] and accelerated failure time modeling. Dedicated
volumes have been written to extend the Cox model (see citations in [11]).
Cox proportional hazard model represents conditional survivor function:

S (t | z ) = [S0 (t )]exp(zβ ) (4)

⎡ t ⎤
where S 0 ( t ) = exp ⎢ −
⎣
∫0
λ 0 ( u ) du ⎥ .
⎦
(5)

and z is the vector of covariates, which can be any factors that influence the survival
or lifetime of game players.
Competing risks analysis is concerned with the scenario that multiple risks exists but
only one of the risks leads to the failure ([2][13]). The single risk that leads to the failure
becomes the failure cause aftermath, and the other risks that competed for the 'cause' are
just latent risks. Competing risks analysis is neither univariate nor truly multivariate
given its univariate cause and multivariate latent risks. Multivariate survival analysis
deal with truly multivariate systems where multiple failure-causes and multiple failures
(modes or failure times) may exist/occur simultaneously [5][14]. Furthermore, the mul-
tiple failure causes (risks) may be dependent with each other, and so do the multiple
failures (modes or times). Indeed, multivariate survival analysis often offers the most
effective statistical approaches to study dependent failures [5][9][14].
Observation (or information) censoring refers to the incomplete observation of ei-
ther failure events or failure causes or both. Censoring is often unavoidable in the
studies of time-to-event random variables such as reliability analysis. However, tradi-
tional reliability analysis does not have a rigorous procedure to handle censored ob-
servations; either including or excluding the censored individuals may cause bias in
statistical inference.
In network reliability and survivability analysis, Ma (2008) proposed to utilize sur-
vival analysis to assess the consequences of the UUUR (Unpredictable, latent, Unob-
served or Unobservable Risks) events [9]. Mathematically, although the probabilities
of UUUR events are unknown, survival analysis does provide procedures to assess
their consequences, e.g., in the form of the survivor function variations at various
censoring level, the influences of latent risks on failure time and/or modes, or the
effects of shared frailty (the unobserved or unobservable risks) on failure times,
modes, and/or dependencies. From the perspective of application problems such as
network reliability and survivability, UUUR covers a class of risks that are particu-
larly difficult to characterize, such as malicious intrusions, virus or worm infections,
software vulnerabilities that are poorly understood. The event probabilities associated
with those kinds of risks are often impossible to obtain in practice. Furthermore, these
risks are time, space and covariate dependent or dynamic, and this further complicates
the problem. On the other hand, survival analysis models such as Cox proportional
hazard models are designed to deal with covariate dependent dynamic hazards (risks).
Therefore, survival analysis provides a set of ideal tools to quantitatively assess the
consequences of UUUR events, which is recognized as one of the biggest challenges
in network security and survivability research.
In the context of evolutionary games, the concept of UUUR events can be used to
describe 'risk' factors that affect the lifetime or fitness of game players. Of course,
 Towards an Extended EGT with Survival Analysis and Agreement Algorithms 613

those factors may affect the fitness positively, negatively, or variable (nonlinearly) at
different times. Therefore, the term 'risk' may bear a neutral consequence, rather than
consistently negative. In general, UUUR captures uncertainty or frailty. The latter can
be considered as the combinations or mixture of uncertainty and vulnerability. For
example, information can dispense uncertainty and consequently eliminate some
perceived vulnerability [9]–[11]. In addition, deception in some applications such as
the study of animal communication networks or strategic information warfare can also
be described with UUUR risks from the perspective of the players [10][16][18].
From above discussion it can be seen that survival analysis models can be har-
nessed to describe the lifetime and/or survivor probability of individual players,
which can often be transformed into the individual fitness. An alternative modeling
strategy is to utilize survival analysis models for population dynamics of the game
players, or the dynamics of meta-populations. For example, in the Dove-Hawk game,
there are two populations: one is the hawk population and another is the dove popula-
tion. Both populations can be described with survival analysis models. The population
modeling with survival analysis should be similar to the modeling of biological popu-
lation dynamics [9].
In summary, the advantages from introducing survival analysis include: (i) flexible
time-, space- and covariate-dependent fitness function; (ii) unique approaches to as-
sess the consequences of UUUR events; (iii) deception modeling, which can be recast
as a reliability problem; (iv) effective modeling of the dependency between 'failure'
events or between the 'risks' that influence the 'failure' events.

3 The Extension to EGT with Agreement Algorithms

The second extension to EGT with agreement algorithms is actually built upon the
first extension with survival analysis. It was initially conducted with the Agreement
algorithms in distributed computing ([6][8]) and the resulting models were termed
Dynamic Hybrid Fault (DHF) Models ([9][12]). One of the earliest Agreement
algorithm problems was formulated as the Byzantine Generals problem (BGP) by
Lamport (1982)[6], in which the components of a computer system are abstracted as
generals of an army. Loyal generals (nodes or players) need to find a way (algorithm)
to reach a consensus (e.g., to attack or retreat) while traitors (or bad nodes) would try
to confuse others by sending conflicting messages. Because the focus of Agreement
algorithms is to reach a consensus, in the traditional hybrid fault models (which are
simply the constraints, often in the form of inequalities, used to classify faults based
on the Agreement algorithms), the failure rate (λ) is often ignored or is implicitly
assumed to be constant. In other words, the hybrid fault models only specify whether
or not an agreement can be reached, given a certain number of traitors, but they do not
keep track of when the generals committed treason. This assumption is appropriate in
the study of Agreement algorithms because, even if the voting is dynamic, they are
abstracted as multiple rounds of voting to study the possibility to reach a consensus.
However, when the fault models are applied to analyze a real-world system consisting
of multiple components (generals), the history of the generals must be considered.
Some generals may be loyal for their entire lifetimes; some may quickly become
'corrupted;' still others may be loyal for a long time but ultimately become 'corrupted.'
614 Z. (Sam) Ma

Each of the generals may have different (inhomogeneous) time-variant (not constant)
hazard functions, λi(t), i =1, 2, ..., g, where g is the number of generals.
To overcome this limitation of lacking real time notion in tradition hybrid fault
models, Ma and Krings (2008) extended the traditional hybrid fault models with the
so-called Dynamic Hybrid Fault models (DHF) [9][12]. Essentially, there are two
limitations with the traditional hybrid fault models, when they are applied to reliabil-
ity analysis. The first is the lack of the notion of real time, as explained in the previ-
ous paragraph, and the second is the lack of approaches to incorporate hybrid fault
models into reliability analysis after the issues associated with the first limitation is
resolved. The latter depends on the integration of the dynamic hybrid fault models
with the evolutionary game theory.
The solution to the first aspect, the missing notion of real time, or the Agreement-
algorithm aspect of the problem, is to introduce survival analysis. In particularly, time
and covariate dependent survivor functions or hazard functions can be used to de-
scribe the survival of the Byzantine generals. For example, the constraint of the BGP
under oral message assumption is,
N ≥ 3m + 1 (6)
is replaced with the following model in the dynamic hybrid fault models:
N (t ) ≥ 3m(t ) + 1 . (7)
Further assuming that the survivor function of generals is S(t|z), a simplified concep-
tual model can be:
N (t ) = N (t − 1) * S (t | z ) (8)
m(t ) = m(t − 1) * S m (t | z ) (9)

where N(t) and m(t) are the number of total generals and treacherous generals (trai-
tors) at time t, respectively. S(t|z) and Sm(t|z) are the corresponding conditional survi-
vor functions for the total number of generals and traitors, respectively. They can use
any major survival analysis models, such as Cox proportional hazard model, ex-
pressed in equation (4). One immediate benefit of the dynamic hybrid fault models is
that it is now possible to predict the real-time fault tolerance level in a system.
The above extension with survival analysis models is necessary, but not sufficient
for applying the dynamic hybrid fault models to reliability analysis, except for ex-
tremely simple cases. The difficulty arises when there are multiple types of failure
behaviors. This is typical in real world dynamic hybrid fault models. For example, the
failure modes could be symmetric vs. asymmetric, transmissive vs. omissive, benign
vs. malicious, etc. Besides different failure modes, node behaviors can also include:
cooperative vs. non-cooperative, mobile nodes vs. access points (which might be
sessile), etc. To model the different behaviors, we will need multiple groups, or sys-
tem of equations (8) and (9). The challenge is that we lack an approach to synthesize
the models to study reliability and survivability. This is the second limitation (aspect)
of traditional hybrid fault models. The solution to the second limitation (aspect), is a
new notion termed 'Byzantine Generals Playing Evolutionary Games,' first outlined in
Ma (2007, unpublished dissertation proposal).
 Towards an Extended EGT with Survival Analysis and Agreement Algorithms 615

Evolutionary game theory models such as represented with replicator dynamics

[(1) and (2)] can be used to synthesize the fitness functions for various failure behav-
iors in dynamic hybrid fault models [(4)–(9)]. In the case of BGP, there are only two
strategies (behaviors): loyal (reliable) or treacherous (deceitful), i.e., types E1 and E2
in (1) and (2). The fitness function fi(x) can adopt survivor functions such as repre-
sented with Cox models [(4)–(5), (8)–(9)]. Since strategies in EGT are discrete,
Agreement algorithms are introduced to perform dynamic 'voting'. In the case of BGP,
the constraint [inequality (7)] to reach an agreement is dynamically checked, which is
a reason for the notion of 'Byzantine generals playing evolutionary games' [9][12].
In principle, the similar procedure with dynamic hybrid fault models, briefly sum-
marized above, can be generalized to any evolutionary games. Of course, the con-
straint for reaching an agreement does not have to be BGP constraints or any of the
Agreement algorithms in distributed computing [6][8] at all. The constraints can be
any algorithms that set the rules for players to assess their global status or reach an
agreement if requested. Therefore, the term agreement algorithm in the context of
EEGT can refer to either the existing Agreement algorithms (from distributed com-
puting [6][8]) or any user-defined agreement algorithms 'customized' for a specific
problem. Nevertheless, I do suggest that the agreement algorithms take similar forms
with dynamic hybrid fault models, i.e., taking advantages of survival analysis based
fitness function. In other words, the extension with agreement algorithms is built upon
the first extension with survival analysis.
In the following, I use an even simpler legislature scenario to illustrate the general
idea for extending evolutionary game theory with agreement algorithms. Assume a
legislature bill is presented to the Congress. Two opposing parties may try to defeat
each other. Lobbyists may try to influence the delegates. Also assume the bill has to
be voted before the end of the summer recess. It is a rule that 2/3 majority is needed
to pass the bill. However, there are time-dependent covariates that influence the dele-
gates. The minds of the delegates may change from day to day. Each delegate has a
different fitness function, which determines how long it takes for a delegate to make
the decision. Since the time-to-decision is a time-to-event (survival or failure time)
random variable, survival analysis is ideally applicable. The fitness function can be,
λi(t) =f (effort of lobbyist, media influence, constituent, health, scandal, etc). (10)
The voting result or the agreement should depend on time. At the time of voting,
determining the result is simple; for example, a simple 2/3 majority rule. However,
assessing probability to reach an agreement each day is not trivial at all, e.g.,
λ (t ) = g[λi (t )] (11)

or λ (t ) = g[λR (t ), λD (t ), λ I (t )], z ] (12)

z could be lobbyists or other factors that influence the fitness functions.

Up to this point, we only completed the extension of evolutionary game theory with
survival analysis. The real complexity lies in the secondary extension with agreement
algorithms. The extension starts with the function g. What is the form of function g?
The delegates may be 'playing' the game, and we assume that they 'play' evolutionary
game. Simple addition or even complex survival function is not sufficient because the
behavior (strategy) change is not a “continuous” function and furthermore, the players
616 Z. (Sam) Ma

may form complex coalitions. Beyond those complexities, there are conditional com-
mitments, which evolve daily. To deal with all the complexity, the 'Byzantine generals
playing evolutionary games' paradigm, which turns the problem into an evolutionary
game, is necessary. Different problems may have different function g, but the principle
should be similar.
In principle, there are four major types of function g: (i) borrow from traditional
games; e.g., Hawk-Dove game; (ii) utilize the replicator dynamics modeling [4]; (iii)
Fitness Generating Function (G-function) [23]; (iv) 'Byzantine Generals Playing
EG'—the approaches outlined above.
Actually, there is a third extension to the EGT with the principle of hedging, which
is necessary when UUUR events dominantly influence decision-making, e.g., dealing
with malicious actions in survivability analysis. The three-layer survivability analysis
is essentially the application of the EEGT to survivable network systems at the tacti-
cal and strategic levels, and further extended the EGT with the hedging principle at
the operational level, which is responsible for the decision-making in managing sur-
vivable network systems or planning information warfare [9][10][18][19].

4 Modeling Animal Communication Networks with the EEGT

One of the most important results from EGT is evolutionary stable strategies (ESS).
Obtaining ESS can be a challenging computational problem for complex evolutionary
game models. This is yet another reason that simulation and heuristic algorithms are
usually necessary for the EEGT modeling, especially given the added complexity
from the extensions. Evolutionary computing is particularly suitable for implementing
the EEGT. When implemented in simulation, EEGT itself is equivalent to an optimi-
zation methodology [15][17]. The paradigm of EEGT has been applied to the studies
of reliability and survivability of wireless sensor networks [9], strategic information
warfare research [10][19], and aerospace PHM engineering [18].
Animal communication has been studied with the sender-receiver dyadic paradigm
traditionally, and so does the theory of honesty (reliability) in animal communication.
In the last decade, the network paradigm for animal communication has emerged, in
which the third-party receivers, such as eavesdroppers, bystanders and audience, are
part of an animal communication network and their roles are considered [22]. How-
ever, much of the current research in animal communication networks is still qualita-
tive. Particularly, the modeling of the reliability (honesty) of animal communication
networks, similar to the SPS game for dyadic paradigm, is still largely an uncharted
area. As argued previously, both the studies of animal communication networks and
their reliability may spawn exciting cross-disciplinary research, such as offering a
paradigm for resilient and pervasive computing, serving as a trust model in computer
science, which are further expanded as seven open problems in [16].
Here, I suggest that the EEGT, which has been applied to modeling of reliability and
survivability of distributed computer networks [9], should provide an effective model-
ing architecture for studying animal communication networks and their reliability,
resilience, and evolution. The following diagram (Fig.1) shows a simulation modeling
architecture for such a study. In Fig. 1, the top rectangle box represents a population of
animals (or game players) consisting of senders, receivers, eavesdroppers, bystanders,
 Towards an Extended EGT with Survival Analysis and Agreement Algorithms 617

audience, etc. The second row of five boxes lists the major issues of animal communi-
cation. The third row is the major models (algorithms) involved, such as those from the
EEGT, models for behavior adaptation and evolution (e.g., the Handicap principle),
and other constraints (competition and cooperation). The last row shows the major
results obtainable from the EEGT modeling.
In perspective, it is my opinion that the study of animal communication networks
and their reliability is on the verge to spawn cross-disciplinary research similar to that
generated by the study of cooperation and PD games. Honesty or deception is an issue
potentially interested in by researchers from many disciplines, such as psychology,
sociology, criminal justice study, political and military sciences, economics, and
computer science. Deception can create uncertainty, vulnerability and frailty, but what
makes it particularly hard to model is its asymmetric or Byzantine nature. Further-
more, Byzantine deception and frailty are likely to be the major factors that often
contribute to the unexpected but catastrophic 'crashes.' This may explain the phe-
nomenon that the fraud and corruption in social and economic systems, as well as
deception in information warfare are often among the most illusive and difficult ob-
jects to describe in any quantitative modeling research. It is hoped that the extensions
to EGT introduced in this paper will relieve some of the difficulties.

Fig. 1. The major modules of modeling animal communication networks with the EEGT

References
1. Axelrod, R., Hamilton, W.D.: The evolution of cooperation. Science 211, 1390–1396
(1981)
2. Crowder, M.J.: Classical Competing Risks Analysis, p. 200. Chapman & Hall, Boca Raton
(2001)
3. Dugatkin, L.A.: The Altruism Equation. Princeton University Press, Princeton (2006)
4. Hofbauer, J., Sigmund, K.: Evolutionary Games and Population Dynamics, p. 323. Cam-
bridge University Press, Cambridge (1998)
5. Hougaard, P.: Analysis of Multivariate Survival Data, p. 560. Springer, Heidelberg (2000)
6. Lamport, L., Shostak, R., Pease, M.: The Byzantine Generals Problem. ACM Transactions
on Programming Languages and Systems 4(3), 382–401 (1982)
618 Z. (Sam) Ma

7. Lawless, J.F.: Statistical models and methods for lifetime data, 2nd edn. Wiley, Chichester
(2003)
8. Lynch, N.: Distributed Algorithms. Morgan Kaufmann Press, San Francisco (1997)
9. Ma, Z.S.: New Approaches to Reliability and Survivability with Survival Analysis, Dy-
namic Hybrid Fault Models, and Evolutionary Game Theory. PhD Dissertation, University
of Idaho (2008)
10. Ma, Z.S.: Extended Evolutionary Game Theory Approach to Strategic Information War-
fare Research. Journal of Information Warfare 8(2), 25–43 (2009)
11. Ma, Z.S., Krings, A.W.: Survival Analysis Approach to Reliability Analysis and Prognos-
tics and Health Management (PHM). In: Proc. 29th IEEE–AIAA AeroSpace Conference,
p. 20 (2008a)
12. Ma, Z.S., Krings, A.W.: Dynamic Hybrid Fault Models and their Applications to Wireless
Sensor Networks (WSNs). In: The 11-th ACM/IEEE International Symposium on Model-
ing, Analysis and Simulation of Wireless and Mobile Systems, ACM MSWiM 2008, Van-
couver, Canada, p. 9 (2008)
13. Ma, Z.S., Krings, A.W.: Competing Risks Analysis of Reliability, Survivability, and Prog-
nostics and Health Management (PHM). In: Proc. 29th IEEE–AIAA AeroSpace Confer-
ence, p. 21 (2008)
14. Ma, Z.S., Krings, A.W.: Multivariate Survival Analysis (I): Shared Frailty Approaches to
Reliability and Dependence Modeling. In: Proc. 29th IEEE–AIAA AeroSpace Conference,
p. 21 (2008)
15. Ma, Z.S., Krings, A.W.: Dynamic Populations in Genetic Algorithms, SIGAPP. In: 23rd
Annual ACM Symposium on Applied Computing (ACM SAC 2008), Brazil, March 16-20,
p. 5 (2008)
16. Ma, Z.S.: The Handicap Principle for Trust in Computer Security, the Semantic Web and
Social Networking. In: Liu, W., Luo, X., Wang, F.L., Lei, J. (eds.) WISM 2009. LNCS,
vol. 5854, pp. 458–468. Springer, Heidelberg (2009)
17. Ma, Z.S.: Towards a Population Dynamics Theory for Evolutionary Computing: Learning
from Biological Population Dynamics in Nature. In: Deng, H., Wang, L., Wang, F.L., Lei,
J. (eds.) AICI 2009. LNCS (LNAI), vol. 5855, pp. 195–205. Springer, Heidelberg (2009)
18. Ma, Z.S., Krings, A.W., Sheldon, F.T.: An outline of the three-layer survivability analysis
architecture for modeling strategic information warfare. In: Fifth ACM CSIIRW, Oak
Ridge National Lab (2009)
19. Ma, Z.S.: A New Life System Approach to the Prognostic and Health Management (PHM)
with the Three-Layer Survivability Analysis. In: The 30th IEEE-AIAA AeroSpace Confer-
ence, p. 20 (2009)
20. Ma, Z.S., Krings, A.W.: Insect Sensory Systems Inspired Computing and Communica-
tions. Ad Hoc Networks 7(4), 742–755 (2009)
21. Maynard Smith, J., Price, G.R.: The logic of animal conflict. Nature 246, 15–18 (1973)
22. McGregor, P.K. (ed.): Animal Communication Networks. Cambridge University Press,
Cambridge (2005)
23. Vincent, T.L., Brown, J.L.: Evolutionary Game Theory, Natural Selection and Darwinian
Dynamics, p. 382. Cambridge University Press, Cambridge (2005)
24. Zahavi, A., Zahavi, A.: The Handicap Principle: A Missing Piece of Darwin’s Puzzle. Ox-
ford University Press, Oxford (1997)
 Uncovering Overlap Community Structure in
Complex Networks Using Particle Competition

Fabricio Breve, Liang Zhao, and Marcos Quiles

Institute of Mathematics and Computer Science, University of São Paulo,

São Carlos SP 13560-970, Brazil
{fabricio,zhao,quiles}@icmc.usp.br

Abstract. Identification and classification of overlap nodes in communi-

ties is an important topic in data mining. In this paper, a new clustering
method to uncover overlap nodes in complex networks is proposed. It is
based on particles walking and competing with each other, using random-
deterministic movement. The new community detection algorithm can
output not only hard labels, but also continuous-valued output (soft la-
bels), which corresponds to the levels of membership from the nodes
to each of the communities. Computer simulations were performed with
synthetic and real data and good results were achieved.

Keywords: complex networks, community detection, overlap commu-

nity structure, particle competition.

1 Introduction
In the last years, the advances and the convergence of computing and commu-
nication has rapidly increased our capacities of generating and collecting data.
However, most of this data is in its raw form, and it is not useful until it is
discovered and articulated. Data Mining is the process of extracting the implicit
potentially useful information from the data. It is a multidisciplinary field, draw-
ing works from areas including statistics, machine learning, artificial intelligence,
data management and databases, pattern recognition, information retrieval, neu-
ral networks, data visualization, and others [1,2,3,4,5].
Community Detection is one of the data mining problems that arose with
the advances in computing and the increasingly interest in complex networks,
which studies large scale networks with non-trivial topological structures, such
as social networks, computer networks, telecommunication networks, transporta-
tion networks, and biological networks [6,7,8]. Many of these networks are found
to be divided naturally into communities or modules, therefore discovering of
these communities structure became an important topic of study [9,10,11,12,13].
Recently, a particle competition approach was successfully applied to detect
communities modeled in networks [14].
The notion of communities in networks is straightforward, they are defined
as a subgraph whose nodes are densely connected within itself but sparsely
connected with the rest of the network. However, in practice there are common

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 619–628, 2009.


c Springer-Verlag Berlin Heidelberg 2009
620 F. Breve, L. Zhao, and M. Quiles

cases where some nodes in a network can belong to more than one community.
For example: in a social network of friendship, individuals often belong to several
communities: their families, their colleagues, their classmates, etc. These nodes
are often called overlap nodes, and most known community detection algorithms
cannot detect them. Therefore, uncovering the overlapping community structure
of complex networks becomes an important topic in data mining [15,16,17].
In this paper we present a new clustering technique, based on particle walking
and competition. We have extended the model proposed in [14] to output not
only hard labels, but also a fuzzy output (soft labels) for each node in the
network. The continuous-valued output can be seen as the levels of membership
from each node to each community. Therefore, the new model is able to uncover
the overlap community structure in complex networks.
The rest of this paper is organized as follows: Section 2 describes the model in
details. Section 3 shows some experimental results from computer simulations,
and in Section 4 we draw some conclusions.

2 Model Description

The model we propose in this paper is an extension of the particle competition

approach proposed by [14], which is used to detect communities in networks.
Some particles walk in a network, competing with each other for the possession
of nodes, while rejecting intruder particles. After a number of iterations, each
particle will be confined within a community of the network, so the communities
can be divided by examining the nodes ownership. The new model is not only
suitable to detect community structure, but it can also uncover overlap commu-
nity structure. In order to achieve that, we have changed the nodes and particles
dynamics, and introduced a few new variables, among other details that will
follow.
Let the network structure be represented as a graph G = (V, E), with
V = {v1 , v2 , . . . , vn }, where each node vi is an element from the network. An
adjacency matrix W defines which network nodes are interconnected:

Wij = 1, if there is an edge between nodes i and j, (1)

Wij = 0, otherwise, (2)

and Wii = 0.
Then, we create a set of particles P = (ρ1 , ρ2 , . . . , ρc ), in which each particle
corresponds to a different community. Each particle ρj has a variable ρω j (t) ∈
[ωmin ωmax ] is the particle potential characterizing how much the particle can
affect a node at time t, in this paper we set the constants ωmin = 0 and ωmax = 1.
Each node vi has two variables viω (t), and viλ (t). The first variable is a vector
vi (t) = {viω1 (t), viω2 (t), . . . , viωc (t)} of the same size of P, where each element
ω
ω
vi j (t) ∈ [ωmin ωmax ] corresponds to the instantaneous level of ownership by
particle ρj over node vi . The sum of the levels of ownership of each node is
always a constant, because a particle increases its own ownership level and, at
 Uncovering Overlap Community Structure in Complex Networks 621

the same time, decreases the other particles ownership levels. Thus, the following
equations always holds:
c
ω
vi j = ωmax + ωmin (c − 1). (3)
j=1

The second variable is also a vector viλ (t) = {viλ1 (t), viλ2 (t), . . . , viλc (t)} of the
same size of P and it also represents ownership levels, but unlike viω (t) which
λ
denotes the instantaneous ownership levels, vi j (t) ∈ [0 ∞] rather denotes long
term ownership levels, accumulated through the whole process. The particle with
higher ownership level in a given non-overlap node after the last iteration of the
algorithm is usually the particle which have visited that node more times, but
that does not always apply to overlap nodes, in which sometimes the dominant
particle could easily change in the last iterations, and thus it would not corre-
spond to the particle which have dominated that node for more instants of time.
Therefore, the new variable viλ (t) was introduced in order to define the owner-
ship of nodes considering the whole process. Using a simple analogy, we can say
that now the champion is not the one who have won the last games, but rather
the one who have won more games in the whole championship. Notice that the
long term ownership levels only increases and their sum is not constant, they
are normalized only at the end of the iterations.
We begin the algorithm by setting the initial level of instantaneous ownership
vector viω by each particle ρj as follows:

ω ωmax − ωmin
vi j (0) = ωmin + ( ), (4)
c
which means that all nodes starts with all particles instantaneous ownership
levels equally set. Meanwhile, the long term ownership levels viλ (t) are all set to
zero:
λ
vi j (0) = 0. (5)
The initial position of each particle ρvj (0) is set randomly, to any node in V, and
the initial potential of each particle is set to its minimum value, as follows:

ρω
j (0) = ωmin . (6)

Each particle will choose a neighbor to visit based in a random-deterministic rule.

At each iteration, each particle will chose between random walk or deterministic
walk, where random walk means the particle will try to move to any neighbor
randomly chosen, i.e., the particle ρj will try to move to any node vi chosen with
the probabilities defined by:
Wki
p(vi |ρj ) = n , (7)
q=1 Wqi

where k is the index of the node node being visited by particle ρj , so Wki = 1
if there is an edge between the current node and vi , and Wki = 0 otherwise.
622 F. Breve, L. Zhao, and M. Quiles

The deterministic walk means that the particle will try to move to a neighbor
with probabilities according to the nodes instantaneous ownership levels, i.e., the
particle ρj will try to move to any neighbor vi chosen with probabilities defined
by:
ω
Wki vi j
p(vi |ρj ) = n ωj , (8)
q=1 Wqi vi

again, k is the index of the node stored being visited by particle ρj .

At each iteration, each particle has probability pdet of taking deterministic
movement and probability 1 − pdet of taking random movement, with 0 ≤ pdet ≤
1. Once the random movement or deterministic movement is chosen, a target
neighbor ρτj (t) will be randomly chosen with probabilities defined by Eq. 7 or
Eq. 8 respectively.
Regarding the node dynamics, at time t, each instantaneous ownership level
viωk (t) of each node vi , which was chosen by a particle ρj as its target ρτj (t), is
defined as follows:


∆v ρω
j (t)
max{ωmin, viωk (t) − c−1 } if k = j
ωk
vi (t + 1) =  ωq ωq , (9)
vi (t) + q=k vi (t) − vi (t + 1) if k = j
ωk

where 0 < ∆v ≤ 1 is a parameter to control the changing rate of the instanta-

neous ownership levels. If ∆v takes a low value, the node ownership levels change
slowly, while if it takes a high value, the node ownership levels change quickly.
Each particle ρj will increase their corresponding instantaneous ownership level
ω
vi j of the node vi they are targeting, while decreasing the instantaneous own-
ership levels (of this same node) that corresponds to the other particles, always
respecting the conservation law defined by Eq. 3.
Regarding the particle dynamics, at time t, each particle potential ρω
j (t) is set
as:
ωj
j (t + 1) = ρj (t) + ∆ρ (vi (t + 1) − ρj (t))
ρω ω ω
(10)
where vi (t + 1) is the node rhoj is targeting, 0 < ∆ρ ≤ 1 is a parameter to
control the particle potential changing rate. Therefore, every particle ρj have
their potential ρωj set to approximate the value of instantaneous ownership level
ω
vi j from the node it is currently targeting. In this sense, a particle gets stronger
when it is visiting a node with higher ownership level of its own, but it will be
weakened if it tries to invade a node dominated by other particle.
The long term ownership levels are adjusted only when the particle selects the
random movement. This rule is important because although the deterministic
movement is useful to prevent particles from abandoning their neighborhood,
which would let it susceptible to other particles attack, it is also a mechanism
that makes a node gets more visits from the particle that currently dominates it.
We consider only when the random movement was chosen because, in this case,
particles will choose a target node based only in their current neighborhood, and
not in their instantaneous ownership levels that are important for community
detection, but are too volatile in overlap nodes. Therefore, for each particle
 Uncovering Overlap Community Structure in Complex Networks 623

selected in random movement by a particle ρj , the long term ownership levels

λ
vi j are update as follows:
λ λ
vi j (t + 1) = vi j (t) + ρω
j (t), (11)

where vi is the node ρj is targeting. The increase will always be proportional to

the current particle potential, which is a desirable feature because the particle
will probably have a higher potential when it is arriving from its own neighbor-
hood, while it will have a lower potential when it is arriving from a node from
other particles neighborhoods.
It should be noted that a particle really visits a target node only if its own-
ership level in that node is higher than the others; otherwise, a shock happens
and the particle stays at the current node until next iteration.
At the end of the iterations, the degrees of membership fij ∈ [0 1] for each
node vi are calculated using the long term ownership levels, as follows:
λ
v j (∞)
fij = c i λq (12)
q=1 vi (∞)

where fij represents the final membership level of the node vi to community j.
In summary, our algorithm works as follows:
1. Build the adjacency matrix W by using Eq. 1,
2. Set nodes ownership levels by using Eq. 4 and Eq. 5,
3. Set particles initial positions randomly and their potentials by using Eq. 6,
4. Repeat steps 5 to 8 until convergence or for a pre-defined number of steps,
5. Select the target node for each particle by using Eq. 8 or Eq. 7 for determin-
istic movement or random movement respectively,
6. Update nodes ownership levels by using Eq. 9,
7. If the random movement was chosen, update the long term ownership levels
by using Eq. 11,
8. Update particles potentials by using Eq. 10,
9. Calculate the membership levels (fuzzy classification) by using Eq. 12.

3 Computer Simulations
In order to test the overlap detection capabilities of our algorithm, we generate
a set of networks with community structure using the method proposed by [13].
Here, all the generated networks have n = 128 nodes, split into four communities
containing 32 nodes each. Pairs of nodes which belongs to the same community
are linked with probability pin , whereas pairs belonging to different communities
are joined with probability pout . The total average node degree k is constant
and set to 16. The value of pout is taken so the average number of links a node
has to nodes of any other community, zout , can be controlled. Meanwhile, the
value of pin is chosen to keep the average node degree k constant. Therefore,
zout /k defines the mixture of the communities, and as zout /k increases from zero,
624 F. Breve, L. Zhao, and M. Quiles

the communities become more diffuse and harder to identify. In each of these
generated networks we have added a 129th node and created 16 links between
the new node and nodes from the communities, so we could easily determine an
expected “fuzzy” classification for this new node based on the count of its links
with each community.
The networks were generated with zout /k = 0.125, 0.250, and 0.375 and the
results are shown in Tables 1, 2, and 3 respectively. The first column of these

Table 1. Fuzzy classification of a node connected to network with 4 communities

generated with zout /k = 0.125

Connections Fuzzy Classification

A-B-C-D A B C D
16-0-0-0 0.9928 0.0017 0.0010 0.0046
15-1-0-0 0.9210 0.0646 0.0079 0.0065
14-2-0-0 0.8520 0.1150 0.0081 0.0248
13-3-0-0 0.8031 0.1778 0.0107 0.0084
12-4-0-0 0.7498 0.2456 0.0032 0.0014
11-5-0-0 0.6875 0.3101 0.0016 0.0008
10-6-0-0 0.6211 0.3577 0.0111 0.0101
9-7-0-0 0.5584 0.4302 0.0011 0.0103
8-8-0-0 0.4949 0.4944 0.0090 0.0017
8-4-4-0 0.5025 0.2493 0.2461 0.0021
7-4-4-1 0.4397 0.2439 0.2491 0.0672
6-4-4-2 0.3694 0.2501 0.2549 0.1256
5-4-4-3 0.3144 0.2491 0.2537 0.1828
4-4-4-4 0.2512 0.2506 0.2504 0.2478

Table 2. Fuzzy classification of a node connected to network with 4 communities

generated with zout /k = 0.250

Connections Fuzzy Classification

A-B-C-D A B C D
16-0-0-0 0.9912 0.0027 0.0024 0.0037
15-1-0-0 0.9318 0.0634 0.0026 0.0023
14-2-0-0 0.8715 0.1219 0.0023 0.0044
13-3-0-0 0.8107 0.1827 0.0036 0.0030
12-4-0-0 0.7497 0.2437 0.0044 0.0022
11-5-0-0 0.6901 0.3036 0.0034 0.0029
10-6-0-0 0.6298 0.3654 0.0020 0.0028
9-7-0-0 0.5584 0.4360 0.0026 0.0030
8-8-0-0 0.4952 0.4985 0.0027 0.0036
8-4-4-0 0.5060 0.2485 0.2427 0.0028
7-4-4-1 0.4442 0.2477 0.2429 0.0652
6-4-4-2 0.3762 0.2465 0.2514 0.1259
5-4-4-3 0.3178 0.2500 0.2473 0.1849
4-4-4-4 0.2470 0.2518 0.2489 0.2523
 Uncovering Overlap Community Structure in Complex Networks 625

Table 3. Fuzzy classification of a node connected to network with 4 communities

generated with zout /k = 0.375

Connections Fuzzy Classification

A-B-C-D A B C D
16-0-0-0 0.9709 0.0092 0.0108 0.0091
15-1-0-0 0.9160 0.0647 0.0093 0.0101
14-2-0-0 0.8571 0.1228 0.0104 0.0097
13-3-0-0 0.8008 0.1802 0.0100 0.0090
12-4-0-0 0.7422 0.2385 0.0095 0.0098
11-5-0-0 0.6825 0.2958 0.0123 0.0093
10-6-0-0 0.6200 0.3566 0.0111 0.0123
9-7-0-0 0.5582 0.4181 0.0128 0.0109
8-8-0-0 0.4891 0.4846 0.0130 0.0133
8-4-4-0 0.5045 0.2437 0.2406 0.0113
7-4-4-1 0.4397 0.2461 0.2436 0.0705
6-4-4-2 0.3797 0.2471 0.2445 0.1287
5-4-4-3 0.3175 0.2439 0.2473 0.1913
4-4-4-4 0.2462 0.2494 0.2549 0.2495

tables shows the number of links the 129th node has to communities A, B, C, and
D, respectively. Notice that in each configuration the 129th node has different
overlap levels, varying from the case where it fully belongs to a single community
up to the case where it belongs to the four communities almost equally. From
2nd to 5th column we have the fuzzy degree of membership of the 129th node
relative to communities A, B, C, and D respectively, obtained by our algorithm.
The presented values are the average of 100 realizations with different networks.
For these simulations, the parameters were set as follows: pdet = 0.5, ∆v = 0.4
and ∆ρ = 0.9.
The results shown that the method was able to accurately identify the fuzzy
communities of the overlap nodes. The accuracy gets lower as zout /k increases,
this was expected, since a higher zout /k means that the communities are more
diffuse and the observed node can be connected to nodes that are overlap nodes
themselves.
Based on this data, we have created an overlap measure in order to easily
illustrate the application of the algorithm in more complex networks with lots of
overlap nodes. Therefore, the overlap index oi for a node vi is defined as follow:

fij∗∗
oi = (13)
fij∗

where j∗ = arg maxj fij and j ∗ ∗ = arg maxj,j=j∗ fij , and oi ∈ [0 1], where
oi = 0 means completely confidence that the node belongs to a single commu-
nity, while oi = 1 means the node is completely undefined among two or more
communities.
626 F. Breve, L. Zhao, and M. Quiles

0.9

4
0.8

0.7
2

0.6

0 0.5

0.4

−2
0.3

0.2
−4

0.1

−6 0
−5 −4 −3 −2 −1 0 1 2 3 4 5

Fig. 1. Problem with 1000 elements split into four communities, colors represent the
overlap index from each node, detected by the proposed method

Fig. 2. The karate club network, colors represent the overlap index from each node,
detected by the proposed method

Then, we have applied the algorithm to a problem with 1000 elements, split
into four communities with 250 elements each. There are four gaussian kernels in
a two dimensional plane and the elements are distributed around them. To build
the network, each element is transformed into a network node. Two elements
i and j are connected if their Euclidean distance d(i, j) < 1. The algorithm
parameters were set as follows: pdet = 0.6, ∆v = 0.4 and ∆ρ = 0.9. In Figure 1
the overlap index of each node is indicated by their colors. It is easy to realize
that the closer to the communities frontier the nodes are, the higher are their
respective overlap indexes.
Finally, the algorithm was applied to the famous Zachary’s Karate Club Net-
work [18] and the results are shown in Figure 2. The algorithm parameters were
set as follows: pdet = 0.6, ∆v = 0.4 and ∆ρ = 0.9. Again, the overlap index of
each node is indicated by their colors. In Table 4 the fuzzy classification of all
the nodes on this network are shown.
 Uncovering Overlap Community Structure in Complex Networks 627

Table 4. Fuzzy classification of the Zachary’s Karate Club Network achieved by the
proposed method

Node Community A Community B Node Community A Community B

1 0,8934 0,1066 18 0,9861 0,0139
2 0,8744 0,1256 19 0,0126 0,9874
3 0,5727 0,4273 20 0,6452 0,3548
4 0,9470 0,0530 21 0,0140 0,9860
5 0,9960 0,0040 22 0,9877 0,0123
6 0,9979 0,0021 23 0,0168 0,9832
7 0,9979 0,0021 24 0,0106 0,9894
8 0,9282 0,0718 25 0,0275 0,9725
9 0,3703 0,6297 26 0,0262 0,9738
10 0,2750 0,7250 27 0,0050 0,9950
11 0,9968 0,0032 28 0,1319 0,8681
12 0,9957 0,0043 29 0,2298 0,7702
13 0,9791 0,0209 30 0,0123 0,9877
14 0,7510 0,2490 31 0,3293 0,6707
15 0,0238 0,9762 32 0,2188 0,7812
16 0,0100 0,9900 33 0,0878 0,9122
17 1,0000 0,0000 34 0,1342 0,8658

4 Conclusions
This paper presents a new clustering technique using combined random-
deterministic walking and competition among particles, where each particle cor-
responds to a class of the problem. The algorithm outputs not only hard labels,
but also soft labels (fuzzy values) for each node in the network, which corre-
sponds to the levels of membership from that node to each community. Com-
puter simulations were performed in both synthetic and real data, and the results
shows that our model is a promising mechanism to uncover overlap community
structure in complex networks.

Acknowledgements
This work is supported by the State of São Paulo Research Foundation (FAPESP)
and the Brazilian National Council of Technological and Scientific Development
(CNPq).

References
1. Han, J., Kamber, M.: Data Mining: Concepts and Techniques, 2nd edn. Morgan
Kaufmann, San Francisco (2006)
2. Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Tech-
niques, 2nd edn. Morgan Kauffman, San Francisco (2005)
628 F. Breve, L. Zhao, and M. Quiles

3. Hand, D.J., Mannila, H., Smyth, P.: Principles of Data Mining. MIT Press, Cam-
bridge (2001)
4. Weiss, S.M., Indurkhya, N.: Predictive Data Mining: A Practical Guide. Morgan
Kaufmann, San Francisco (1998)
5. Tan, P.N., Steinbach, M., Kumar, V.: Introduction to Data Mining. Pear-
son/Addison Wesley (2005)
6. Newman, M.E.J.: The structure and function of complex networks. SIAM Re-
view 45, 167–256 (2003)
7. Dorogovtsev, S., Mendes, F.: Evolution of Networks: From Biological Nets to the
Internet and WWW. Oxford University Press, Oxford (2003)
8. Bornholdt, S., Schuster, H.: Handbook of Graphs and Networks: From the Genome
to the Internet. Wiley-VCH (2006)
9. Newman, M.E.J., Girvan, M.: Finding and evaluating community structure in net-
works. Physical Review E 69, 026113 (1–15) (2004)
10. Newman, M.: Modularity and community structure in networks. Proceedings of
the National Academy of Science of the United States of America 103, 8577–8582
(2006)
11. Duch, J., Arenas, A.: Community detection in complex networks using extremal
optimization. Physical Review E 72, 027104 (1–4) (2006)
12. Reichardt, J., Bornholdt, S.: Detecting fuzzy community structures in complex
networks with a potts model. Physical Review Letters 93, 218701 (1–4) (2004)
13. Danon, L., Dı́az-Guilera, A., Duch, J., Arenas, A.: Comparing community structure
identification. Journal of Statistical Mechanics: Theory and Experiment 9, P09008
(1–10) (2005)
14. Quiles, M.G., Zhao, L., Alonso, R.L., Romero, R.A.F.: Particle competition for
complex network community detection. Chaos 18, 033107 (1–10) (2008)
15. Zhang, S., Wang, R.S., Zhang, X.S.: Identification of overlapping community struc-
ture in complex networks using fuzzy c-means clustering. Physica A Statistical
Mechanics and its Applications 374, 483–490 (2007)
16. Palla, G., Derényi, I., Farkas, I., Vicsek, T.: Uncovering the overlapping community
structure of complex networks in nature and society. Nature 435, 814–818 (2005)
17. Zhang, S., Wang, R.S., Zhang, X.S.: Uncovering fuzzy community structure in
complex networks. Physical Review E 76, 046103 (1–7) (2007)
18. Zachary, W.W.: An information flow model for conflict and fission in small groups.
Journal of Anthropological Research 33, 452–473 (1977)
 Semi-supervised Classification Based on Clustering
Ensembles

Si Chen, Gongde Guo, and Lifei Chen

School of Mathematics and Computer Science,

Fujian Normal University, Fuzhou, China
chensi072@163.com, {ggd,clfei}@fjnu.edu.cn

Abstract. In many real-world applications, there only exist very few labeled
samples, while a large number of unlabeled samples are available. Therefore, it
is difficult for some traditional semi-supervised algorithms to generate the useful
classifiers to evaluate the labeling confidence of unlabeled samples. In this pa-
per, a new semi-supervised classification based on clustering ensembles named
SSCCE is proposed. It takes advantages of clustering ensembles to generate mul-
tiple partitions for a given dataset, and then uses the clustering consistency index
to determine the labeling confidence of unlabeled samples. The algorithm can
overcome some defects about the traditional semi-supervised classification algo-
rithms, and enhance the performance of the hypothesis trained on very few la-
beled samples by exploiting a large number of unlabeled samples. Experiments
carried out on ten public data sets from UCI machine learning repository show
that this method is effective and feasible.

Keywords: semi-supervised learning, classification, clustering ensembles, clus-

tering consistency.

1 Introduction
In traditional supervised learning, a large number of labeled samples are required to
learn a well-performed hypothesis, which is used to predict the class labels of unla-
beled samples. In many real-word applications such as toxicity prediction of chemical
products, computer-aided medical diagnosis, labeled samples are very difficult to
obtain because labeling samples needs to spend much time and effort [1]. However, a
large number of unlabeled samples are easily available. If only using very few labeled
samples, it is difficult to generate a good classification model, and a lot of available
unlabeled samples are wasted. Thus, when the labeled samples are very limited, how
to enhance the performance of the hypothesis trained on very few labeled samples, by
making use of a large number of unlabeled samples, has become a very hot topic in
machine learning and data mining [1].
Semi-supervised learning is one of the main approaches exploiting unlabeled sam-
ples [1]. This paper mainly focuses on semi-supervised classification algorithms
which aim to improve the performance of the classifier by making fully use of the
labeled and unlabeled data. A kind of semi-supervised classification learning ap-
proaches are called pre-labelling approaches [2]. The method firstly trains an initial

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 629–638, 2009.
© Springer-Verlag Berlin Heidelberg 2009
630 S. Chen, G. Guo, and L. Chen

classifier on a few labeled samples and then the unlabeled samples are utilized to
improve the performance of the classifier. For example, self-training [5] is a well-
known pre-labelling method. In this method, only a classifier is trained on the labeled
training set. In some pre-labelling methods such as co-training [7], tri-training [9], and
co-forest [6], more classifiers are used in their training processes. However, when
there only exist very few labeled samples, some pre-labelling approaches [5, 6, 7, 9]
are difficult to generate the initial useful classifier to estimate the labeling confidence
of unlabeled samples, so that a certain number of unlabeled samples may be misclas-
sified. Additionally, some algorithms place constraints to the base learner [6, 8], or
require the sufficient and redundant attribute sets [7].
Another kind of semi-supervised classification learning methods are called post-
labelling approaches [2, 17]. The method firstly generates a data model on all the
available data by applying a data density estimation method or a clustering algorithm,
and then labels the samples in every cluster or estimates the class conditional densi-
ties. However, using only a single clustering algorithm may generate the bad-
performed data model, and thus affect the performance of the final classifier.
In this paper, a new post-labelling algorithm named SSCCE, i.e., Semi-Supervised
Classification based on Clustering Ensembles, is proposed. Based on clustering en-
sembles, SSCCE firstly generates multiple partitions on all the samples, and matches
clusters in different partitions. Then by analyzing the multiple clustering results, the
unlabeled samples with high labeling confidence are selected to label and added into
the labeled training set. Finally, a learner is trained on the enlarged labeled training
set. Experiments carried out on ten UCI data sets show that this method is effective
and feasible. SSCCE is easier to understand, which uses the clustering consistency
index to estimate the labeling confidence of unlabeled samples, and requires neither
special classifiers nor the sufficient and redundant attribute sets.
The rest of the paper is organized as follows: Section 2 gives a brief review on
semi-supervised learning and clustering ensembles. Section 3 presents the SSCCE
algorithm. Section 4 reports the experimental results carried out on ten UCI data sets.
Section 5 concludes this paper and raises several issues for future work.

2 Related Work
2.1 Semi-supervised Learning

Semi-supervised learning methods include pre-labelling approaches and post-labelling

approaches [2], etc. For example, self-training [5] is a pre-labelling method, in which,
however, the unlabeled samples may be misclassified when the initial labeled samples
are insufficient for learning a well-performed hypothesis [14]. Some pre-labelling
methods [6, 7, 8, 9], such as co-training, more classifiers are used in the training proc-
ess. However, co-training requires the sufficient and redundant attribute sets. In [8],
Goldman et al. proposed a new algorithm extending co-training which requires the
two special classifiers, each of which can divide the data space into several equiva-
lence classes. Zhou and Li developed the tri-training algorithm [9], which does not
require the sufficient and redundant attribute sets and special classifiers. The algo-
rithm implicitly determines the labeling confidence by prediction consistency for
unlabeled samples. But this implicit determination may not accuracy enough, and the
 Semi-supervised Classification Based on Clustering Ensembles 631

initial labeled samples are not large enough to train a high accuracy classifier, which
is likely to lead to misclassify a certain number of unlabeled samples. In [6], Li et al.
proposed the co-forest algorithm which incorporates Random Forest [18] algorithm to
extend the co-training paradigm. But this algorithm tends to under-estimate the error
rates of concomitant ensembles, and it places constraints to the base learner and en-
semble size.
The post-labelling approach [2, 17] firstly generates a data model on all the avail-
able data by applying a data density estimation method or a clustering algorithm, and
then labels the samples in every cluster or estimates the class conditional densities.
An unlabeled sample is labeled mainly according to its relative position with the la-
beled data. The kind of algorithm is formally described in [2]. In [17], Parzen win-
dows are used for estimating the class conditional distribution and a genetic algorithm
is applied to maximize the posteriori classification of the labeled samples.

2.2 Clustering Ensembles

Recently, a number of clustering algorithms [10] have been proposed. However, it is

difficult to find a single clustering algorithm which can handle all types of cluster
shapes and sizes, and it is hard to determine what kind of clustering algorithm should
be used for a particular dataset [15]. Therefore, many scholars begin to study cluster-
ing ensembles methods. In [11], it shows that clustering ensembles can go beyond a
single clustering algorithm in robustness, novelty, stability, confidence estimation,
parallelization and scalability.
Let X = { x1 , x2 , ..., xn } be a data set with n samples in multidimensional space.
Consider H as ensemble size (i.e., the number of partitions), and π = {π 1 , π 2 ,..., π H }
as the set of H partitions. π i = {C1 , C2 ,..., Cn } ( i ∈ {1, 2, … , H } ) denotes a data parti-
i i i

tion, where C j is the j cluster in the partition π i , and n i is the number of the cluster
i
i th

labels in π i . Let π i ( x ) be the cluster label of a sample x in partition π i . Clustering

ensembles methods aim to combine H partitions π 1 , π 2 ,..., π H into a better final parti-
tion, where the samples in the same cluster are more similar while the samples be-
longing to different clusters are less similar. Recently, the researches for clustering
ensembles mainly focus on how to generate effective multiple partitions [16], and
how to combine multiple partitions [12].
The k-means algorithm [10] with different initial cluster centers is often used to
generate multiple partitions for a given dataset. The k-means algorithm has low compu-
tational complexity, and does not require many user-defined parameters. Because the
k-means algorithm is sensitive to the choice of the initial cluster centers, SSCCE gener-
ates H different partitions by using the k-means algorithm which randomly selects k
different initial cluster centers each time, and then combines the H different partitions.

3 SSCCE Algorithm

3.1 Outline of SSCCE

In this paper, a novel semi-supervised classification based on clustering ensembles

named SSCCE is proposed. SSCCE firstly generates H different partitions on all the
632 S. Chen, G. Guo, and L. Chen

samples using the k-means algorithm [10], and then matches clusters in different par-
titions. Secondly, the samples with high labeling confidence are selected and the clus-
ter labels of these selected unlabeled samples are matched with the real class labels of
the labeled samples. Then the selected unlabeled samples are labeled and added into
the initial labeled training set. Finally, a hypothesis is trained on the enlarged training
set. The SSCCE algorithm is summarized as follows.

Algorithm: SSCCE
Input: labeled training set L={(x1, y1),(x2, y2),...,(x|L|, y|L| )};
unlabeled training set U = {x1′ , x2′ ,..., x|U |′} ;
X =L ∪ U ;
H: number of partitions (i.e., ensemble size);
π a : reference partition;
α : threshold value of clustering consistency index;
k: number of cluster centers of k-means algorithm;
Output: a hypothesis h.
Step1: for i=1 to H
1.1. Randomly select k initial cluster centers from the data set X ;
1.2. Use the k-means algorithm to generate partition π i on the data set X ;
end for
Step2: Match clusters in different partitions according to the matching clusters
method described in Section 3.2.
Step3: Compute clustering consistency index (CI) and clustering consistency label
(CL) of each sample in the data set X .
Step4: Select the samples with CI > α from the data set X , and these samples are
divided into clusters C1 , C2 ,..., Cm according to their corresponding cluster-
ing consistency labels CL1 , CL2 ,..., CLm (m is the number of labels).
Step5: Using the methods described in Section 3.4, match the corresponding clus-
tering consistency labels CL1 , CL2 ,..., CLm and the real class labels of the la-
beled training set L.
Step6: Assign the re-matched clustering consistency labels to all the selected unla-
beled samples in clusters C1 , C2 ,..., Cm , and add them into the labeled train-
ing set L.
Step7: A hypothesis h is trained on the enlarged labeled training set L.
According to the above outline of SSCCE, the following three main issues are re-
quired to address: 1) how to match the clusters in different partitions? , 2) how to
define the clustering consistency index (CI) and clustering consistency label (CL)? ,
and 3) how to match the cluster labels of unlabeled samples and the real class labels?
We will discuss these issues in the following sections.

3.2 Matching Clusters

In the multiple different partitions, the similar clusters in different partitions may
be assigned with different labels. For example, the cluster labels of five samples are
(1, 2, 2, 1, 2) and (2, 1, 1, 2, 1) in two different partitions, respectively. In fact, these
two partitions are exactly the same, and are only assigned with different cluster labels.
 Semi-supervised Classification Based on Clustering Ensembles 633

How to match clusters in different partitions is one of the most important problems
needed to be resolved.
Based on the methods proposed in [12, 19], SSCCE uses Jaccard index to deter-
mine the best matching pair of clusters in different partitions after converting clusters
C ij (the j cluster in partition π i ) into a binary valued vector X j . The value of l
th i th
th
entry of X ij , say 0 or 1, denotes absence or presence of l sample in the cluster C ij .
Note that any partition can be arbitrarily selected as the reference partition from H
partitions, so that the clusters of the other partitions are matched with those of the
selected reference partition.

3.3 Clustering Consistency Index and Clustering Consistency Label

After matching the clusters in different partitions, it is easy to find that some samples
are steadily assigned to the same cluster, whereas the clusters are frequently changed
for some examples in H partitions. In this paper, the samples with the stable cluster
assignment are considered as the ones with high labeling confidence, and then these
unlabeled samples are selected to label. Therefore, SSCCE introduces clustering con-
sistency index (CI) to measure the labeling confidence. Clustering consistency index
(CI) [16] is defined as the radio of the maximal number of times which the sample is
assigned in a certain cluster to the total number of partitions H. The clustering consis-
tency index CI( x ) of a sample x is defined by Eq. (1).

{ } ,with δ ( a , b ) = ⎧
1, if a = b
H
1
CI ( x ) = max ∑ δ (π ( x ), C ) ⎨ . (1)
⎩ 0, otherwise
i
H i =1 C ∈ cluster labels

By analyzing these partitions π 1 , π 2 ,..., π H , the samples with the clustering consis-
tency index greater than the given threshold value α are considered to be steadily
assigned to a certain cluster. The label of this cluster is defined as clustering consis-
tency label (CL). For a sample x , if CI( x ) > α , then its corresponding clustering
consistency label CL( x ) is defined by Eq. (2).

{∑ }
H

CL ( x ) = arg max δ (π i ( x ), C ) . (2)

C i =1 C ∈ cluster labels

The greater the clustering consistency index of a sample is, the higher the labeling
confidence of the sample is. SSCCE needs to label the samples with high clustering
consistency index. However, their corresponding clustering consistency labels do not
match with the real class labels of the labeled samples. Therefore, we should match
clustering consistency labels and real class labels.

3.4 Matching Clustering Consistency Labels and Real Class Labels

Suppose that the clustering consistency labels of the selected samples with high clus-
tering consistency index are respectively CL1 , CL2 ,..., CLm (m is the total number of
CL). Then the clusters C 1 , C 2 , ..., C m only including the selected samples, correspond-
ing to the clustering consistency labels C L1 , C L 2 , ..., C L m , may be the following three
cases: (1) both labeled samples and unlabeled samples are contained; (2) only the
634 S. Chen, G. Guo, and L. Chen

unlabeled samples are contained; (3) only the labeled samples are contained. The third
case will not be taken into account because there are not any unlabeled samples in
clusters. The first two cases will be analyzed as follows.
In the first case, according to the principle of majority vote, if cluster Ci
( i ∈ {1,...., m} ) contains both labeled samples and unlabeled samples, then its corre-
sponding label CLi is re-matched with the majority label of the labeled samples in
cluster Ci . In the second case, according to k-nearest-neighbor [13], if cluster Ci
contains only the unlabeled samples, then its corresponding label CLi is re-matched
with the majority label of the labeled nearest neighbors of all the unlabeled samples in
cluster Ci . Finally, all the unlabeled samples in cluster Ci are labeled with the re-
matched clustering consistency labels, and then added into the labeled training set.

4 Experiments
In order to test the performance of SSCCE, we use ten public data sets from the UCI
machine learning repository [4], named solar-flare, breast-w, Australian, credit-a, tic-
tac-toe, Liver, heart-statlog, heart-c, colic, and auto-mpg, respectively. For each data
set, 10-fold cross validation is employed for evaluation. In each fold, training data are
randomly partitioned into labeled training set L and unlabeled training set U under the
label rate only 1%, i.e., just 1% training data (of the 90% data) are used as labeled
samples while the remaining 99% training data are used as unlabeled samples. The
entire 10-fold cross validation process is repeated ten times. In each time, randomize
the ordering of samples, and then run a tested algorithm to generate an experimental
result using 10-fold cross validation. Finally, the ten results of 10-fold cross valida-
tion are averaged for each test.
For SSCCE, SimpleKMeans algorithm from the Weka Software package [3] is
used to generate multiple partitions, and the number of clusters is set to the number of
classes of the labeled training data. When using the matching clusters algorithm, ref-
erence partition is set to partition π 1 . Use 1-nearest-neighbor to deal with the second
case that a cluster only contains unlabeled samples in Section 3.4.
For comparison, we evaluate the performances of SSCCE, the supervised learning
algorithm trained on only the labeled training set L, co-forest [6] and self-training [5].
For SSCCE, the ensemble size H of clustering ensembles is set to 100. In co-forest,
the number of classifiers is set to 10. For fair comparison, Random Forest [18] in
WEKA [3] is used as the classifier in all the above algorithms. In either co-forest or
self-training, the threshold value of labeling confidence is set to 0.75. For each algo-
rithm, the average classification accuracy, i.e., the ratio of the number of test samples
correctly classified to the total number of test samples, is obtained with 10 times of
10-fold cross validation on each data set. Table 1 illustrates the average classification
accuracies of the supervised learning algorithm using Random Forest (denoted by
RF), SSCCE using Random Forest denoted by SSCCE (RF), co-forest, and self-
training using Random Forest denoted by self-training (RF). The last row Average
shows the average results over all the experimental data sets using Random Forest.
The highest average accuracy is shown in bold style on each dataset.
 Semi-supervised Classification Based on Clustering Ensembles 635

Table 1. Average classification accuracies of the compared algorithms using RF (%)

Data set RF SSCCE (RF) co-forest self-training (RF)

solar-flare 77.92 80.04 77.00 77.11
breast-w 83.23 84.72 82.58 79.07
Australian 61.43 65.19 59.58 56.28
credit-a 59.62 61.48 59.52 58.06
tic-tac-toe 61.44 62.91 61.47 60.62
Liver 50.19 50.63 51.71 49.96
heart-statlog 48.96 53.44 47.00 47.04
heart-c 52.97 59.03 52.18 52.18
colic 58.42 58.16 58.64 55.69
auto-mpg 52.36 55.86 57.00 52.36
Average 60.66 63.15 60.67 58.84

Table 1 shows that SSCCE benefits much from the unlabeled samples since the
performances of SSCCE (RF) are better on 7 data sets than RF, co-forest, and self-
training (RF). Furthermore, Table 1 also shows that SSCCE outperforms self-training
on all the data sets, and the average result of SSCCE (RF) over all the experimental
data sets is highest. Therefore, Table 1 supports that SSCCE is effective and feasible,
and can improve the performance of the classifier trained on very few labeled samples
by utilizing a large number of unlabeled data under the label rate 1%.
For further analysis on SSCCE, Table 2 tabulates the average threshold value α
of clustering consistency index. In the experiments, set the threshold value α to the
arithmetical mean of the clustering consistency indices of all the training data on each
data set. The column add_ratio shows the ratio of the number of unlabeled samples
added into the labeled training set L to the total number of unlabeled samples. In

Table 2. Further analysis of SSCCE using RF

Data set α add_ratio (%) one-k-means (%) SSCCE (RF) (%)

solar-flare 0.21 42.79 77.92 80.04
breast-w 0.54 58.02 83.23 84.72
Australian 0.54 49.49 61.43 65.19
credit-a 0.54 47.46 59.62 61.48
tic-tac-toe 0.58 48.32 61.44 62.91
Liver 0.54 42.48 50.19 50.63
heart-statlog 0.53 49.32 48.96 53.44
heart-c 0.25 45.72 52.97 59.03
colic 0.54 46.72 58.42 58.16
auto-mpg 0.3 46.82 52.36 55.86
Average 0.46 47.71 60.66 63.15
636 S. Chen, G. Guo, and L. Chen

addition, the column one-k-means shows the average classification accuracy of the
extreme case of SSCCE using Random Forest when H = 1, i.e., the k-means algorithm
is run only one time to generate a partition, and then all the unlabeled samples are
given the matched class labels and added into the training set L. In Table 2, SSCCE
(RF) is the same as those of Table 1. All the parameters of the algorithms still keep
unchanged.
From Table 2, it can be observed that the average 47.71% of unlabeled samples are
added into the labeled training set L. And when H=100, the average threshold value
α is about 0.46. It can be concluded that SSCCE can effectively select a certain
number of unlabeled samples with high labeling confidence. Moreover, the average
accuracies of SSCCE when H=100 are better than those of one-k-means on 9 data
sets. Therefore, it validates that clustering ensembles can outperform the individual
clustering, and thus enhance the performance of SSCCE.
In the experiments, we also use J48 [32] from the Weka Software package [3] to
observe the impact of different classifiers on the average classification accuracies of
these compared algorithms. The results show that the performances of SSCCE (J48)
are better than the supervised algorithm J48 and self-training using J48 on 6 data sets
among 10 data sets, and the average result of SSCCE using J48 over all the experi-
mental data sets is highest.
Therefore, SSCCE is effective when using either Random Forest or J48, and co-
forest and self-training fail to perform well on the most data sets. One possible expla-
nation is when the number of labeled samples is very few, i.e., the label ratio only 1%,
it is difficult to generate the initial useful classifier to estimate the labeling confidence
of unlabeled samples, and a certain number of noises can be added into the initial
labeled training set for co-forest and self-training. However, SSCCE can benefit much
from a large number of unlabeled samples based on clustering ensembles, which can
select a certain number of unlabeled samples with high labeling confidence to label
and enhance the performance of the initial classifier.

Table 3. Average classification accuracies of SSCCE (RF) over different ensemble sizes (%)

Data set H=50 H=100 H=150

solar-flare 80.20 80.04 81.16
breast-w 85.46 84.72 86.69
Australian 64.93 65.19 68.45
credit-a 61.07 61.48 60.25
tic-tac-toe 60.59 62.91 60.36
Liver 51.58 50.63 50.16
heart-statlog 55.74 53.44 56.59
heart-c 57.43 59.03 56.06
colic 58.64 58.16 57.43
auto-mpg 57.30 55.86 52.60
Average 63.29 63.15 62.98
 Semi-supervised Classification Based on Clustering Ensembles 637

Note that in previous experiments, the ensemble size H of SSCCE is fixed to 100.
But the different ensemble size might affect the performance of SSCCE. Table 3
tabulates the performances of SSCCE using Random Forest on H=50, H=100, and
H=150. From Table 3, it can be seen that the last row average accuracy on H=50 is
better than that on H=100 and H=150. Ensemble sizes of the highest average accura-
cies of SSCCE (RF) are the same on most data sets except heart-statlog.

5 Conclusion
In this paper, a new semi-supervised classification based on clustering ensembles
named SSCCE is proposed. It firstly generates multiple different partitions on all the
samples using the k-means algorithm which uses the different cluster centers each
time, and then matches clusters in different partitions. Moreover, the samples with
high clustering consistency index are selected, and their corresponding clustering
consistency labels are matched with the real class labels. Then the selected unlabeled
samples are labeled with the re-matched clustering consistency labels. Finally, these
unlabeled samples are added into the initial labeled training set, and then a hypothesis
is trained on the enlarged labeled training set.
Using an easily understandable method of estimating the labeling confidence of
unlabeled samples, SSCCE can make full use of the labeled and unlabeled samples,
and overcome some defects in some traditional semi-supervised learning algorithms.
Experiments carried out on ten UCI data sets show that SSCCE outperforms the su-
pervised learning algorithm only trained on labeled training data. Compared with
some traditional semi-supervised learning algorithms such as co-forest and self-
training, SSCCE can also perform better on most data sets, when the number of the
labeled training data is very few.
The ensemble size H, i.e., the number of partitions, might impact the performance
of SSCCE. Therefore, how to overcome the effect of ensemble size and further im-
prove the performance of SSCCE will be studied in the further work.

Acknowledgments. This work was supported by the Natural Science Foundation of

Fujian Province of China under Grants No. 2007J0016 and No. 2009J01273, and the
Scientific Research Foundation for the Returned Overseas Chinese Scholars, Ministry
of Education of China under Grant No. [2008] 890.

References
1. Zhou, Z.H., Li, M.: Semi-Supervised Regression with Co-Training Style Algorithms. IEEE
Transactions on Knowledge and Data Engineering 19, 1479–1493 (2007)
2. Gabrys, B., Petrakieva, L.: Combining Labelled and Unlabelled Data in the Design of Pat-
tern Classification Systems. International Journal of Approximate Reasoning, 251–273
(2004)
3. Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques,
2nd edn. Morgan Kaufmann, San Francisco (2005)
4. Asuncion, A., Newman, D.J.: UCI Machine Learning Repository. School of Information
and Computer Science. University of California, Irvine, CA,
http://www.ics.uci.edu/~mlearn/MLRepository.html
638 S. Chen, G. Guo, and L. Chen

5. Nigam, K., Ghani, R.: Analyzing the Effectiveness and Applicability of Co-Training. In:
Proceedings of the 9th International Conference on Information and Knowledge Manage-
ment, pp. 86–93 (2000)
6. Li, M., Zhou, Z.H.: Improve Computer-Aided Diagnosis with Machine Learning Tech-
niques Using Undiagnosed Samples. IEEE Transactions on Systems, Man and Cybernetics
– Part A: Systems and Humans 37, 1088–1098 (2007)
7. Blum, A., Mitchell, T.: Combining Labeled and Unlabeled Data with Co-Training. In: Pro-
ceedings of the 11th Annual Conference on Computational Learning Theory (COLT
1998), MI, Wisconsin, pp. 92–100 (1998)
8. Goldman, S., Zhou, Y.: Enhancing Supervised Learning with Unlabeled Data. In: Proceed-
ings of the 17th International Conference on Machine Learning (ICML 2000), CA, San
Francisco, pp. 327–334 (2000)
9. Zhou, Z.H., Li, M.: Tri-training: Exploiting Unlabeled Data Using Three Classifiers. IEEE
Transactions on Knowledge and Data Engineering 17, 1529–1541 (2005)
10. Han, J., Kamber, M.: Data Mining: Concepts and Techniques, 2nd edn. Morgan Kauf-
mann, San Francisco (2006)
11. Topchy, A., Jain, A.K., Punch, W.: A Mixture Model for Clustering Ensembles. In: Pro-
ceeding of the 4th SIAM International Conference on Data Mining, pp. 379–390 (2004)
12. Fred, A.: Finding Consistent Clusters in Data Partitions. In: Kittler, J., Roli, F. (eds.) MCS
2001. LNCS, vol. 2096, pp. 309–318. Springer, Heidelberg (2001)
13. Roussopoulos, N., Kelly, S., Vincent, F.: Nearest Neighbor Queries. In: Proceedings of the
1995 ACM SIGMOD International Conference on Management of Data, pp. 71–79 (1995)
14. Li, M., Zhou, Z.H.: SETRED: Self-Training with Editing. In: Ho, T.-B., Cheung, D., Liu,
H. (eds.) PAKDD 2005. LNCS (LNAI), vol. 3518, pp. 611–621. Springer, Heidelberg
(2005)
15. Dubes, R., Jain, A.K.: Clustering Techniques: the User’s Dilemma. Pattern Recogni-
tion 41, 578–588 (1998)
16. Topchy, A., Minaei-Bidgoli, B., Jain, A.K., Punch, W.F.: Adaptive Clustering Ensembles.
In: Proceedings of the 17th International Conference on Pattern Recognition (ICPR 2004),
vol. 1, pp. 272–275 (2004)
17. Kothari, R., Jain, V.: Learning from Labeled and Unlabeled Data. In: IEEE World Con-
gress on Computational Intelligence, IEEE International Joint Conference on Neural Net-
works, HI, Honolulu, USA, pp. 1468–1474 (2002)
18. Breiman, L.: Random Forests. Machine Learning 45, 5–32 (2001)
19. Zhou, Z.H., Tang, W.: Clusterer Ensemble. Knowledge-Based Systems 9, 77–83 (2006)
 Research on a Novel Data Mining Method Based on the
Rough Sets and Neural Network

Weijin Jiang1, Yusheng Xu2, Jing He1, Dejia Shi1, and Yuhui Xu1
1
School of computer, Hunan University of Commerce, Changsha 410205, China
jwjnudt@163.com
2
China north optical-electrical technology, Beijing 100000, China
yshxu520@163.com

Abstract. As both rough sets theory and neural network in data mining have
special advantages and exiting problems, this paper presented a combined algo-
rithm based rough sets theory and BP neural network. This algorithm deducts
data from data warehouse by using rough sets’ deduct function, and then moves
the deducted data to the BP neural network as training data. By data deduct, the
expression of training will become clearer, and the scale of neural network can
be simplified. At the same time, neural network can easy up rough set’s sensi-
tivity for noise data. This paper presents a cost function to express the relation-
ship between the amount of training data and the precision of neural network,
and to supply a standard for the change from rough set deduct to neural network
training.

1 Introduction
As a product of the combination of many kinds of knowledge and technology, Data
mining (acronym: DM) is a new technology rose in 90’s of twentieth century. On
August, 1989, Detroit of American, the concept of KDD (vivid form: Knowledge
Discovery in Database) is presented firstly on the eleventh international artificial
intelligence union meeting[1,2]. 1995, American computer meeting (ACM) proposed
Data Mining, and it vividly described large database as a valuable mineral resources,
from which useful information can be found by efficient knowledge-finding technol-
ogy[3,4]. Because Data Mining is the key step of KDD, people usually will not distin-
guish Knowledge Discovery in Database (KDD) from Data Mining (DM).
Data Mining is a technology which is used to analyze observed data set, usually
very large, the intention of DM is to find unknown relation and summarize data under
an efficient way, which can be understood by data owner. The common algorithm and
theory of Data Mining are Rough Sets[5], Artificial Neural Networks, Decision trees,
Genetic algorithm and etc. this paper mainly discussed Rough Sets theory and BP
Neural Networks in Artificial Neural Networks[6].
Both Rough sets and BP Neural Networks have classification function in Data
Mining. The advantage of Rough Sets is that it is good at parallel execution, descrip-
tion of uncertain information and the strategy dealing of redundant data, and the prob-
lem is it is sensitive with object noise[7,8]. BP Neural Networks is the most popular
Neural Networks, whose main merits is high precision and non-sensitive with noise,

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 639–648, 2009.
© Springer-Verlag Berlin Heidelberg 2009
640 W. Jiang et al.

but its problem is that redundant data can easily cause over-training of neural net-
works, besides, networks scale and the amount of training sample’s influence on the
speed of network training and training time are headache problem[9].
As to the merits and the demerits of Rough Set theory and BP Neural Networks,
this paper proposed a new Data Mining algorithm that combines Rough Sets theory
and BP Neural Network[10]. The algorithm overcame rough sets’ sensitive on noise
data; meanwhile, it reduced the training time of BP Neural Network, provided
constringency, which improved efficiency much.

2 Rough Sets Theory

Rough Sets first proposed by a Poland scholar Z.Pawlak in 1982[3,4]. One of the char-
acteristic is its powerful qualitative analysis ability; it has no need to give Rough Sets
quantity description for some character or property in advance, such as probability
classification in statistics, subjection degree or subjection function in Blur Sets theory,
rough sets can find problem’s inner rule directly from the given problem’s description
set under the way that certain the given problem’s approximate area by non-
distinguished relation and non-distinguished class.
According to Pawlak’s Rough Sets model, the following definition and description
are produced.
， ，
Definition 1. Decision System regard S=(U A{Va} a) as knowledge expression
system, and:
S is a non-empty limit set, and it is called topic area;
A is a non-empty limit set, and it is called property set;
∈
Va is property a A ‘s value area;
：
a U→a U→Va： 为
is one-to-many reflection, which makes every element of
topic area U only has a reflection in Va.
if A is composed by condition property set C and conclusion property set D, and C,
D satisfy C∩D=φ, then S can be called strategy system.
In a strategy system, every element of U represents a rule; the precondition of rule
is decided by C and its value, and the post-condition is decided by D and its value.
：
Under the situation that Va and a U→Va will not rise confuse, people usually use
(, ∪)
U C D to represent strategy system
Definition 2. The non-discriminability relation of DS as for strategy system
(, ∪）
S= U C D , B0C is a subset of condition property set, duality relation IND
( , ) ( )∈
B D ={ x,y ， ∈
U×U f(y,a) ∀ a B} is S’s non-discriminability relation.
Non-discriminability relation is an equal relation, from which classification of strat-
egy system can be obtained, and here, we called the divided equal class non-
discriminability class. Usually, [x]IND（B denotes x ‘s non-discriminability class, under
（）
the unconfused situation, IND B is used to present non-discriminability relation
IND(B, D).
Definition 3. Upper-approximate and lower-approximate as for knowledge expres-
（，） ，
sion system S U A , we set B ⊆ A X ⊆ U, and:
 Research on a Novel Data Mining Method Based on the Rough Sets 641

BX={x ∈U[x] ∈X}IND(B)

B X={ x∈U[x] ∩X≠φ}

IND(B)

They are X’s B-Lower approximation and B-upper approximation. Lower-

approximate BX is union set of all the atom set of X subset, it is the biggest sets of X;
upper-approximate B X is union set of atom sets, which has no empty intersect set
with X, and it is the smallest set of X.
Definition 4. Positive area, negative area and boundary the upper-approximate and
lower-approximate of set X ⊆ U divided topic area into three uncross area; positive
（） （）
area POS X , negative area NEG X and boundary area BND X : （）
POS（X）=B（X）
NEG（X）=U- B （X）
BMD（X）= B （X）- B（X）
Any element x which belongs to POS（X）or NEG（X）certainly not belongs to X,
but belongs to X’s supplementary set; when an element belongs to BND（X）, it is
hard to decide whether it belongs to X or X’s supplementary set, as a results, upper
boundary area is a kinds of uncertain area to some extent, upper-approximate of one
set is the union set of positive area and boundary area, it is B X=POS （X）UBND
() （）
X . If BND X =φ, then X is precise set, otherwise, X is rough set.
Definition 5. Tthe dependency degree of property the dependency degree of two
，
sets B R ⊆ U can be valued by property dependency function ( rough membership
function), the definition of which is as follows:

card ( POS R ( B ))
yR ( B ) =
card ( U )
POS R ( B ) = UR X
X ∈U / IND ( B )

（）
And card · denotes the base of function, POSR （B）is property set R’s positive
（）
area in U/IND B .
Definition 6. The importance of property different property has different influence
on dependency relation between condition and decision property. The meaning of
Adding property a to R for classification U/IND B is: （）
SGF （ a，R，B）= y (B)-y R R-|a| (B)
The importance of property a is relative, and it is dependent on property set B and R.
so, the importance of property may different in different situation. If set D as decision
（，， ）
property, then the meaning of SGF a R D is: After adding property a in prop-
erty set R, dependency degree between R and D change, and this reflects the impor-
tance of property a.
642 W. Jiang et al.

Definition 7. Redundant property as for property sets D and R, property a R, if ∈

（） （）
POSR D = POSR-{a} D ,then a is redundant in property set R, otherwise, a is
necessary for Set D.

3 The Design of BP Neural Network

Although BP neural network is the most popular neural network, its design is still
dependent on experience.
Input level design: the input level is mainly used to input data, so the design of in-
put level is dependent on input information, for example, object ‘person’ has two
properties: height and weight. As for object ‘person’, two inputs neural can be de-
fined, inputting height and weight as training information, and one input neural is
define will be ok, property ‘height’ will be used as training information. As for data
table in database, input information can be the properties which not belongs to deci-
sion property. As results, in data mining, as for data table in database, the number of
neural is less than the amount of non-decision property.
Hidden level design: the main task of hidden level design is to confirm the amount
of hidden level and neural. This problem is being studied by many people, so the
design is mainly dependent on experience. There are two views which are important
for the design of hidden level. The first one is: double hidden neural network can
solve super-dimensional classification problem, which can overcome almost every
classification problem; the second is: the simple hidden level neural network, which
contains enough neural, can solve almost every reality classification problem.
The amount of hidden level neural can be made sure from three experience for-
mula:

k < ∑ ( in )
n
1
(1)
i=0

k is the amount of training sample, n is the number of input neural, ni is the number of
hidden level neural, when i>ni, ( ) =0.
n1
i

n1 = n − m + a (2)
m is the number of output neural, n is the number of input neural, a’s value is between
1 and 10, n1 is the number of hidden level neural.

n1 = log 2 n (3)

n is the number of input neural, n1 is the number of hidden level neural.

When it come to design the number of hidden level neural, these three formulas are
always taken in consider, and the most suitable one will be chose.
The design of output level: a neural network with m output neural can provide two
kinds of classification. For example, as we mentioned, ‘weight’ and ‘height’, ‘sex’ is
needed to output. Here, only one neural can realize the discrimination of man or
woman. In order to improve precision, more neural is adopted; such as two outputs
 Research on a Novel Data Mining Method Based on the Rough Sets 643

neural are adopt to discriminate sex. As a result, the number of output must equal with
log2X, or bigger than log2X, in the formula, X represents the needed classification
model. In common situation, the number of output neural always equal to the standard
of classification model.

4 The Combined Algorithm of Rough Set and BP Neural Network

A lot of problems will be arisen to access the tremendous data set in data-house: the
most conspicuous is problem caused by data gathering, such as, producing redundant
data will make the data increased rapidly; secondly, the distortion of big data set, such
as noise data. Almost every process of data mining can not avoid these two problems.
As for rough set theory, dealing redundant data is its advantage, but it is not sensitive
with noise data. As for BP neural network, training speed and training sample is its
troublesome problem.
Usually, the clearer the method of data expression, the less the redundant data, then
the neural network will be easier to studied, but the number of neural and power value
will increase, its training time will be longer, and the extension ability of neural net-
work will be weaker.
One idea is: using rough set to reduce data, then using the reduced data set as the
design evidence and training data of BP neural network. And this can make training
data clear, and absorbs both advantages of these two methods. While reducing the
training time of BP neural network, the BP neural network can reduce the influence of
noise. Based on this, the paper proposed a combined algorithm of rough set and neu-
ral network.
Data mining based on rough set theory always are property reduction, the steps of
property reduction are: first, finding the kernel of property reduction set by discrimi-
nation matrix, then using reduction algorithm to calculate reduction set, and deciding
the best reduction set by some judge standard. The data mining process based on
rough set theory is illustrated in fig. 1.

Concept property
Level valve

Initial Extension Calculate Discrimina- Calculate Property Property

Da Extension Relation tion Kernel Reduction
Matrix Kernel Reduction

Simplify Simplified Combine Visualize

extension Meta Macro Rule
extension
Relation Relation Production
Relation Rule store

Fig. 1.
644 W. Jiang et al.

The difficulty of the algorithm is the end condition of rough set reduction, in an-
other words, it is hard to decide the selection condition of the amount of BP neural
network training data. According to the statement of third part, selecting the amount
of training data has big influence for the training time of BP neural network. At pre-
sent, there is no certain way to decide the amount of training data, just a cursory
evaluation method: the amount of training data is twice as many as connect power.
For example, if a BP neural network has n input knots, n1 hidden level knots and m
（ ）
output knots, then it need 2× n×n1+ n1×m training data[11].
The selection of the amount of training data is concern with the precision of neural
network. Usually, error is used to reflect the capability of study. The definition of
error is:

m n

∑∑ ( d
i =1 j =1
ij
− yij )
e
m⋅n
m denotes the sample number of training set, n is the amount of the output unit of
neural network.
When the amount is increased, the error will be smaller; as a result, adding more
training data will help to avoid error. But at the same time, when the training data is
increased, the training time also will be longer. Based on this, a cost function is pro-
posed.
Cost function is to describe the relation between the amount of training data and er-
ror, the error function can be modified as:

m n

∑∑ ( d
i =1 j =1
ij
− yij )
e
Xm ⋅ n
A variable X is added into this formula, when X=1, the function will be turned back.
The form of cost function is:

1
y = x /(1 − )
x
x is coefficient, the value area is >1, y is cost guideline. Table 1 illustrates the relation
between cost function, its differential coefficient and the value of x.
From the table, when x‘s value is 2.25, cost function’s one differential coefficient
is 0, which is the minimum value of cost function. When cost function’s differential
coefficient is smaller than -1, or bigger than 1, we can deem that the cost change too
fast, and the coefficient is wrong. As a result, the coefficient should be in the area
between 1.93 and 4, the optimal selection is 2.25.
 Research on a Novel Data Mining Method Based on the Rough Sets 645

Table 1.

x 1.25 1.5 1.75 2 2.25 2.5 2.75 3

y 11.84 8.17 7.17 6.83 6.75 6.2 6.80 7.10
- -
y’ -30 -6.7 0 0.38 0.61 0.75
2.24 0.71
x 3.25 3.5 3.75 4 4.25 4.5 4.75 5
y 7.30 7.52 7.75 8 8.25 8.51 8.78 9.05
y’ 0.85 0.91 0.96 1 1.02 1.05 1.06 1.08

Cost function can be regarded as the selection guideline or the end rule of rough set
reduction. As for the sample with many properties, the optimal cost coefficient 2.25
will be chose to be the selection guideline, as for sample with few properties, cost
higher than 2.25 will be chose. Data mining mainly deal with tremendous data, cost
2.25 will be absolutely the best answer, in order to mind special situation, here, the
situation with few data has been taken in to consider. The algorithm is written as
follows:
Step 1: sampling data, pointing mining condition, and deciding the goal of mining.
Step 2: deleting redundant property by following rough set theory.
Step3: doing property reduction under rough set theory.

Begin The needed information

Design of neural
network
Problem definition, deciding
goal Input training data Obtain the training data

Data gaining Net training

Data reduction Training results

〈〈
Training data amount
Reduced data amount
Results
Cost calculating, comparing
training data and reduced
Minimum property
set, training data has
a big ratio in reduced

Fig. 2.
646 W. Jiang et al.

Step 4: if the minimum property set has been get, choosing training data set by cost
2.25.otherwise, using the highest cost 4 and reduction property to calculate training
sample data, if the results of calculating is smaller than the reduced amount, then turn
to step 3, otherwise, choosing training data by the definition and cost function.
Step 5: designing neural network by training data, and training these training data
sample.
Step 6: outputting the final results.
The flow chart is illustrated by fig.2[12].

5 Applications
Here, a car data table in reference [2] is used to illustrate the algorithm in table 2.
The decision property is Make-model, others are condition properties. Using rough
set to reduce the redundant propertied, table 3 is got.
After reducing the redundant properties, two properties are deleted. Doing data re-
duction for table 3, taking the user’s request that the property reduction set must con-
tain displace and weight, table 4 is obtained.
Then building the neural network, and selecting training sample. The neural net has
4 inputs neural, 3 outputs neural, the hidden level has 4 neural, the structure of the
network is illustrated by fig.3.
Following network structure and cost coefficient, the number of training sample is
（ ）
4×2× 4×4+4×3 =224. Then training these samples, the final results will be output
as table 5.

Table 2.

Plate# Make- Color Cy1 Door Displace

model
BCT89C Ford Silver 6 2 Medium
escoort
RST45W Dodeg Green 4 2 Big
IUTY56 Benz Green 4 2 Big
…… …… …… …… …… ……
PLMJH9 Honda Brown 4 2 Medium
DSA321 Toyota Black 4 2 Small
paso

Compress Power Trans Weight Mileage

High High Auto 1020 Medium
High Medium Manual 1200 Medium
Medium Medium Manual 1230 High
…… …… …… …… ……
Medium Medium Manual 900 Medium
Medium Low Manual 850 High
 Research on a Novel Data Mining Method Based on the Rough Sets 647

Disdlac

Trans
Make-
Weight

Mile-

Fig. 3.

Table 3.

Make-
Obj cy1 Door Compress Power Trans Weight Mileage
model
1 USA 6 2 Medium High Auto Auto Medium
2 Germany 4 2 Big Medium Manual Heavy High
3 Japan 4 2 Small Medium Low Light High
… … … … … … … … …

Table 4.

Make-model Displace Trans Weight Mileage

USA Medium Auto Medium Medium
USA Big Manual Heavy Medium
Germany Big Manual Heavy High
Japan Small Manual Light High
…… …… …… …… ……

Table 5.

Make-model Displace Trans Weight Mileage

USA Medium Auto Medium
USA Medium Medium
USA Big Heary Medium
Germany Big Heary High
Germany Big Auto Medium
Japan Small High
Japan Manual Light

6 Conclusions
In the mining process of data house which has tremendous data and many properties,
this algorithm possesses the advantages of both rough set theory and BP neural network.
It can overcome the noise’s influence for data sensation, at the same time; it can delete
redundant data, provide clearer training data, lessen the scale of network, improves
efficiency of mining. The proposal of cost function not only resolved the relation
648 W. Jiang et al.

between training data and mining precision, but also provided guideline for the trans-
formation from rough set to neural network. Unfortunately, data mining is aiming to big
data warehouse, so the algorithm is not suitable for data mining with small scale.

Acknowledgement
This paper is supported by the Society Science Foundation of Hunan Province of
China No. 07YBB239.

References
1. Bazan, J., Son, N.H., Skowron, A., Szczuka, M.S.: A View on Rough Set Concept Ap-
proximations. In: RSFDGrC, Chongqing, China (2003)
2. Bazan, J.: Dynamic reducts and statistical inference. In: Sixth International Conference on
IPMU, pp. 1147–1152 (1996)
3. Kai, Z., Jue, W.: A Reduction Algorithm Meeting Users’ Requirements. Journal Computer
Science and Technology 17(15), 578P–P593 (2002)
4. Hu, K., Lu, Y., Shi, C.: Feature Ranking in Rough Sets (2002)
5. Nguyen, H.S., Ślęzak, D.: Approximate Reducts and Association Rules. In: Zhong, N.,
Skowron, A., Ohsuga, S. (eds.) RSFDGrC 1999. LNCS (LNAI), vol. 1711, pp. 137–145.
Springer, Heidelberg (1999)
6. Nguyen, T.T., Skowron, A.: Rough set Approach to Domain Knowledge Approxima-tion.
In: RSFDGrC 2003, Chongqing, China (2003)
7. Pawlak, Z.: Rough Sets. Int. J. Comput. In-form. Sci. 11(5), P341–P356 (1982)
8. Pawlak, Z.: Rough sets-Theoretical Aspects of reasoning about data. Kluwer Academic
Publishers, Dordrecht (1991)
9. Polkowski, L.: A Rough Set Para-digm for Unifying Rough Set Theory and Fuzzy Set
Theory. In: Wang, G., Liu, Q., Yao, Y., Skowron, A. (eds.) RSFDGrC 2003. LNCS
(LNAI), vol. 2639, pp. 70–77. Springer, Heidelberg (2003)
10. Polkowski, L., Tsumoto, S., Lin, T.Y.: Rough Set Methods and Applications: New Devel-
opments. In: Knowledge Discovery in Information Systems. Springer, Heidelberg (2000)
11. Shaffer, C.A.: A Practical Introduc-tion to Data Structures and algorithm analysis. Prentice
Hall, Englewood Cliffs (1998)
12. Weijin, J.: Research on Extracting Medical Diagnosis Rules Based on Rough Sets Theory.
Computer Science 31(11), 97–101 (2004)
 The Techno-Economic Analysis of Reducing NOx of
Industrial Boiler by Recycled Flue Gas

Caiqing Zhang1, Wenpei Hu1, Yu Sun2, and Bo Yang2

1
School of Business Adnimistration, North China Electric Power University,
Baoding 071003, Hebei Province, China
2
School of Environmental Science and Engineering,
North China Electric Power University,
Baoding 071003, Hebei Province, China
sunyu1031@126.com

Abstract. The quantity of industrial boiler is huge in our country, consuming

the coal is about 350Mt/a, so a large number of NOx is emitted every year. It
is a complex technology to reduce the emission of NOx. There are various
methods according to different control mechanism, from the angel of tech-
nology and economy, the feasibility of industrial boilers reducing NOx by re-
cycled flue gas is ananlyzed and get the decrease of NOx is between 30% and
40%. It is a mature, simple technology, lower investment and low cost of
runs, what’s more offers the reference of the NOx emission control technical
reformation for industrial boilers, at last make some suggestions for the emis-
sion of NOx.

Keywords: industrial boilers, recycled flue gas, NOx, technicality, economy.

1 Introduction

Coal is the main fuel of industrial boilers in our country, and the primarily boilers are
the low-power grate furnace. It is estimated that the number of all kinds of boilers

％ ～
which are being used is more than 50 ten thousand, the quantity is large, and mean-
while, more than 60 is chain boilers that is 0.5 75T/h. The consumption of coal is
350Mt/a [1.2], and 1000 kilogram coal could produce 7.4 kilogram NOx, thus it leads
to make 2600 thousand NOx annual. Although our country does not have the standard
about the emission of NOx that the industry boilers make, however, the quantity is
quite considerable, and its number is still increasing year by year, the capacity is also
expanding, and the emissions of nitrogen oxide compound (NOx)is increasing. How
to reduce NOx which the industrial boilers emit economically is extremely important
to reduce the air pollution.
Reducing the NOx that the industry boilers emit is a complex technology, lots of
methods according to different control mechanism. The technology of “selectivity
catalytic combustion SCR” is best to reduce NOx, but the investment and operating

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 649–657, 2009.
© Springer-Verlag Berlin Heidelberg 2009
650 C. Zhang et al.

Table 1. Comprehensive comparison table of Low N0x control technology

Appraisal index
Degree
Craft plan The level Reduction
of Initial Operation
of craft Site area amount
technical investment cost
difficulty of NOx
mature
Grading
Fuel A little plant
Mature Complex ignored 30ˉ40
graduation large equipment
(big)
Inleting the Air pipe
secondary Mature Simple none expend ignored 20
wind (small)
Recycled
Recycled Relativel
Mature Simple pipe expend ignored 30ˉ40
flue gas y small
(small)

cost are extremely high, a lot of small and medium-sized enterprise can not afford it,
therefore it is not realistic. From economic and technological angel, the technology of
recycled flue gas was selected to reduce the NOx emissions as much as possible. It
will provide the reality basis for government to establish the policy formulation on the
NOx emission.
From table 1, which shows that the comparison of low N0x control technology:
The technology of recycled flue gas is mature, the craft is simple, at the premise of
the NOx which drops largely, and it only adds the recycled pipeline, the cost of the
invest at the beginning could be ignored, so it is a feasible method for reducing NOx.
The feasibility of the technology was analyzed that uses recycled flue gas to reduce
NOx, which is emitted by industrial boiler from the economical and technological
points.

2 The Technological Analysis of Recycled Flue Gas

2.1 The Introduction of the Computation Mode

The simulation object is a 35t/h chain boiler in this article. This is a boiler with a sin-
gle barrel and a natural circulation water-tube boiler, using the lamination coaling
installment, the burning way of scale type chain link fire grate level, with the suitable
coal of bituminous coal, the chamber size 7500mm×459mm×13500mm .The first
inlet sets six wind areas to enter the wind separately, the amount of wind is different,
and around the arches are two inlet areas. The functions of the secondary wind is
mainly to disturb the above the smoke current, supplying the oxygen promptly, burn-
ing the aerosol combustible substance, to enhance the combustion efficiency, thus the
design request of the position cannot cause the first wind and the wind pressure fluc-
tuate to influence the combustion. Part flue gas after dust removal is sent to the first
 The Techno-Economic Analysis of Reducing NOx of Industrial Boiler 651

wind pipe through the recycling fan from the induced draft fan, then it is sent to the
furnace after mixed with the first wind, and the quality of the recycled wind can be
adjusted by the both sides of the door of the recycled flue gas.
Because the reaction that occurs in the entrance and combustion mainly concen-
trates above the chain, the grids of this region are relatively crowded to others in
order to increase computation precision; the grid division can not cause the false
proliferation in numerical simulation. In order to reduce the false proliferation, the
grids that are in the calculate area is supposed to adapt the development of mobile
as far as possible, causing the streamline enter the grid in vertical direction to con-
trol false proliferation occur. In this article, dividing the grids of the burning area
follows this thought. Liking figure 1, firstly, decomposing the entire chamber into
5 parts, because the reaction in above individual is very few, the computation re-
quest is not so accurate and the size of the grids is 0.3 meter. The downward has
two gusty areas, therefore the request is high, the size of grids is 0.1 meter, the
coal entrance, the first wind area as well as the chain are main reaction regions,
which are needed to concentrate the encryption, the dense of the grids is
dense is 0.05, the diameter of the second wind areas itself only has 0.05 meter,
because of its particularity, the grids are 0.02 meter to guarantee continuously and
uniformity.

Fig. 1. Grids of the boiler

2.2 The Introduction of Computation Method

The simulation of the combustion process mainly has the mutual couplings of the
gas and solid, the combustion reaction also with the turbulence, momentum, energy
and mass transfer computation to draw the distribution of the temperature of boiler,
652 C. Zhang et al.

the flowing and the component. The solid phase pellet uses the stochastic model,
adds to the computation as the separation, the pellet has the independence equations
to carry on the computation, considering the mutual functions between the two
phase; the coal pellet in the combustion process is thought that the particle size is
invariable, along with burning the quality of them reduces. The turbulence model
uses the K-ε model, the burning computation uses the simply PDF to compute com-
ponent score that is in the reaction process, the fuel and the oxidant are thought to
react in a rapid rate in this computation, the percent of the fuel and the oxidant is
mixed score. The double mixed score is used in this paper, and the experiential
form is used in PDF calculation, thinking that the fuel and the oxidant fully contact
and react. 90% of thermal transmission comes from radiation; the radiation uses the
separate coordinate mode, considering the radiation exchanging. The temperature of
surface of the wall is 600K, according to the temperature of saturated water add to
revise temperature that takes entire, the revise temperature is considered the general
burning situation that the slagging causes heating. According to the function of
standard wall, we process the boundary condition of the surface. The entrance of air
and the coal pellet entrance are established separately, the inlet area is thought to be
the speed entrance. The outlet is established as the free, and meanwhile we suppose
that the quality that enters and flows out the chamber is constant. The parameter of
the entrance is illustrated in Table 2, the kind of the coal: bituminous coal. The pa-
rameters are listed in table 3.

Table 2. Parameters of the simulation of the combustion (kg/s)

The The The The The The The

quality quality quality quality quality quality quality
of the of the of the of the of the of the of the
The
first first first first first first secondary
operation of
wind in wind in wind in wind in wind in wind in wind
boiler
inlet inlet inlet inlet inlet inlet in
1 2 3 4 5 6 inlet

The
combustion
of the
0.731 1.219 1.920 2.194 1.950 0.731 0.455
normal
operation
of boiler
The
recycled
0.731 1.700 2.720 3.094 2.650 0.731 0.455
flue gas
rate is 30%

Indication the temperature of the first and secondary wind is 423K.

 The Techno-Economic Analysis of Reducing NOx of Industrial Boiler 653

Table 3. Analytical table of the characteristic of coal

Name Mark Unit Value

amount of carbon for receiving Cy % 55.19
amount of hydrogen for receiving Hy % 2.38
amount of oxygen for receiving Oy % 1.51
amount of nitrogen for receiving Ny % 0.74
amount of sulfur for receiving Sy % 2.51
amount of ash for receiving Ay % 28.67
amount of water for receiving Wy % 9.0
the volatility of combustable part V % 26.7
the low calorific capacity for
Qnet,v,ar kJ/kg 20900
receiving

The Fluent software is applied in this computation. Firstly solve the flow field of
the uniform temperature when starting calculating, after the momentum equation
restraining, then couple pellet field, burning and the radiation heat transfer, sec-
ondly carry on the iteration; When the residual error of the equation of continuity
and the energy all no longer reduces, each parameter is constant along with the it-
erative, the computation namely is restraining. At last the numerical simulation of
NOx emission is the post-processing of the combustion, after restraining calculate
the amount of the NOx, at that time the flow model, the rapids model, the energy
equation, the radiation model and the PDF mode of the component do not partici-
pate in the computation, the calculate region, the parameter value of other each field
are frozen except the component of NOx , that is the NOx reaction and the process
of the combustion are not coupled.

2.3 The Numerical Simulation and Discussion of the Result

2.3.1 The Comparison of the Distribution of the Temperature Field with and
without Recycled Flue Gas
Figure 2 (a) is the distribution of the temperature field without recycled flue gas,
Figure 2 (b) is map that is with recycled flue gas. By contrast we could observe: the
average temperature of the boiler with recycled flue gas decreases to some extent, it
～ ℃ ℃
drops 30 80 , the central region decreases 90 .This is because the quantity of
the flue which enters to the chamber increases, with it the flue that needs to be
heated also increased with recycled flue gas. Moreover, the concentration of the O2
that is in the recycled flue gas drops obviously, but the concentration of CO2 rises,
this has some influence to burning. Therefore the average temperature of the boiler
decreases on the whole, it is an advantage to reducing of NOx, and the temperature
that reduces does not affect the normal combustion as well as the movement.
654 C. Zhang et al.

（a） The distribution of the temperature field without recycled flue gas

（b） The distribution of the temperature field with recycled flue gas

Fig. 2. The distribution of the temperature field

2.3.2 The Comparison of the Concentration of NOx with and without Recycled
Flue Gas
The change of NOx is illustrated in chart 2-3, the total concentration of NOx is about
394mg/m3 (192ppm) without recycled flue gas while the concentration reduces to
 The Techno-Economic Analysis of Reducing NOx of Industrial Boiler 655

281mg/m3 (137ppm) with recycled flue gas, in certain region, the emissions of NOx
drops from 610mg/m3 to 450mg/m3. Speaking of the total quantity of NOx, the emis-
sion of NOx improves greatly.

（c） The quality score of NOx at the outlet along the height of the boiler without recycled flue gas

（d） The quality score of NOx at the outlet along the height of the boiler with recycled flue gas

Fig. 3. Distribution of the quality score of NOx at the outlet along the height of the boiler
656 C. Zhang et al.

3 The Analysis of Economical Property

The pressure of the air blower is obtained from:

L ω2
hL = λ • (1)
D 2g
From the Table 4 conclude that: the dropped pressure which the recycled pipe causes
is probably ranging 0.71 mm water column to 8.09 mm water column, its respective
dropped pressure does not affect the use of air blower comparing with the selected
blower, therefore the investment of this technical t only increases the expense of the
pipeline and labor, the cost of the operation can be ignored.
In order to study the optimizing control countermeasure, the cost of information on
the control technology related nitrogen oxide compound is essential. Under ordinary
circumstances, the total expense includes the expense of investment and annual operat-
ing, but the cost still does not have stipulation. Normally, the expense of the investment
contains react equipment, pipeline and so on; but the annual cost includes depreciation,
service, and electric power and so on. table 3 is illustrated Several kinds of low NOx
control technology of industrial chain boiler are illustrated in table 3, the beginning of
investment, cost of movement, reducing of NOx and so on are compared in it.

Table 4. Parameters of recycled pipe

The pressure
The hL(mm of the air
quantity ¬ Q(m3/s) V(m/s) D(m) L(m) water blower˄mm
˄T˅ column ) water
column˅
10 0.018 7.29 12-15 0.49-0.61 8-16 3.25-8.09 107
35 0.018 19.7 12-15 1.3-1.45 10-20 1.10-3.81 170
60 0.018 33.33 12-15 2.22-2.78 15-25 0.71-2.32 214

4 Conclusions
1. Comparing with the boiler without recycled flue gas, the boiler with recycled flue
gas, the average temperature in internal boiler drops to some extent, the emission
concentration of NOx decreases about 30% 40%. －
2. Regarding the middle and small scale industrial boilers, the use of recycled flue gas
does not affect the use of the original air blower.
3. The technology of recycled flue gas is mature, the craft is simple, the investment is
small and the cost of operation is little.

References
1. Commandr, J.-M., Stanmore, B.R., Salvadore, S.: The high temperature reaction of carbon
with nitric oxide. Combust Flame, 128 (2005)
2. Glarborg, P., Jensen, A.D., Johnsson, J.E.: Fuel nitrogen conversion in solid fuel fired sys-
tems. Progress in Energy and Combustion Science, 29 (2003)
 The Techno-Economic Analysis of Reducing NOx of Industrial Boiler 657

3. Jones, J.M., Patterson, P.M., Pourkashanian, M., Williams, A., Arenillas, F.R., Pis, J.J.:
Modeling NOxformation in coal particle combustion at high temperature an investigation of
the devolatilisation kinetic factors. Fuel, 88 (1999)
4. Xu, X.C., Chen, C.H., Qi, H.Y., et al.: Development of coal combustion pollution control
for SO2 and NOx in China. Fuel Processing Technology 62, 153–160 (2000)
5. Yin, C., Cailat, S., Harion, J.-l.: Bernard Baudo Everest Perez Investigation of the flow util-
ity tangentialy fired pulverized-coal boiler. Fuel, 81 (2002)
 Urban Traffic Flow Forecasting Based on Adaptive
Hinging Hyperplanes

Yang Lu1,2, Jianming Hu1,2, Jun Xu1,2, and Shuning Wang1,2

1
Department of Automation, Tsinghua University, Beijing, China
2
Tsinghua National Laboratory for Information Science and Technology
{lu-y08,yun-xu05}@mails.tsinghua.edu.cn,
{hujm,swang}@mail.tsinghua.edu.cn

Abstract. In this paper, after a review of traffic forecasting methods and the
development of piecewise linear functions, a new traffic flow forecasting model
based on adaptive hinging hyperplanes was proposed. Adaptive hinging hyper-
planes (AHH) is a kind of piecewise linear models which can decide its division
of the domain and the parameters adaptively. Acceptable results (forecasting er-
ror is smaller than 15%) were obtained in the test of the real traffic data in Bei-
jing. After comparison with the results of prediction model base on MARS, the
following conclusions can be drawn. First, the two methods have almost the
same performance in prediction precision. Second, AHH will be a little more
stable and cost less computing time. Thus, AHH model may be more applicable
in practical engineering.

Keywords: adaptive hinging hyperplanes, MARS, traffic forecasting.

1 Introduction
Traffic flow forecasting is a very important aspect in Advanced Traffic Information
System (ATIS). The quality of the system’s service depends on the accuracy of the
traffic prediction to a great extent. Over the past 30 years, traffic flow forecasting has
been a very active issue but without perfect solutions. There are a variety of forecast-
ing methods[1,2,3,4,5], such as: historical average, time series models (ARIMA),
neural network and nonparametric models (k-NN method), which all have their own
shortage. Historical average method cannot respond to dynamic changes, ARIMA
model is very sensitive to disturbance which cannot be avoided in traffic environment,
neural network is possible to overtrain the network and k-NN model has complicated
way in looking for neighbors. As we know, the main application of the above meth-
ods is in highway traffic flow prediction, but the urban transportation system is more
complex. To overcome the drawbacks above, Ye applied MARS in urban traffic flow
forecasting[6], and obtained good results.
In another aspect, during the past 30 years, the research about piecewise linear
function has been developing gradually and there have been some kinds of represent-
ing models, such as canonical piecewise linear representation [7,8,9,10], lattice PWL
function model[11] and hinging hyperplanes model[12]. Among them, the HH model
is well-known and comparatively practical model which was proposed by Breiman in

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 658–667, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Urban Traffic Flow Forecasting Based on Adaptive Hinging Hyperplanes 659

1993 with a good capacity in nonlinear function approximation. Due to its special
structure, the least square method can be used to make identification. Therefore, the
HH model is a very useful tool in black-box modeling. However, in fact, HH model is
equivalent with the canonical PWL model. To improve the model’s representation
ability, Wang introduced generalized hinging hyperplanes (GHH) in 2005, which is
actually derived from the lattice PWL function, so the model’s general representation
ability is proved directly[13].
Adaptive Hinging Hyperplanes model[14], which was proposed by Xu in 2008,
based on MARS and GHH and then proved to be a special case of GHH, shares the
advantages of the two approaches. The model is adaptive, flexible and its basic func-
tion is linear in each subregion, so the acceptable results in traffic flow prediction
using AHH can be expected because the model based on MARS has already obtained
good results. In this paper, we proposed an urban short-term traffic flow forecasting
model based on adaptive hinging hyperplanes, and the real traffic data in Beijing was
used to test its performance. Furthermore, a comparison between AHH and MARS is
drawn to demonstrate AHH’s effectiveness in forecasting.
The paper is organized as follows. Section 2 introduced AHH model and its algo-
rithm after a review of MARS and GHH. How to use AHH model to carry out the
traffic flow prediction was described in detail in Section 3. We then compared the
results of AHH model and MARS model in Section 4 and made a brief conclusion in
Section 5.

2 Adaptive Hinging Hyperplanes

2.1 Review of MARS and GHH

2.1.1 MARS
MARS (Multivariate Adaptive Regression Splines) was first introduced by Friedman
in 1991[15]. It is developed from recursive partitioning regression as generally con-
sidered. The basic idea of the latter method is to find the most appropriate division of
the domain through successive partition. In the beginning of the procedure, there is
only one father node and it can be divided into two child nodes, each of them denotes
one subregion. One subregion does not intersect with each other and the union of all
subregions constructs the whole domain. The above step is repeated until getting the
terminal node which cannot be split anymore, while the region it denotes corresponds
with one basic function. The partitioning is based on a greedy algorithm where in
each step it selects the split which can give the best fit to the data.
Recursive partitioning regression can be written as the following forms of a set of
basic functions:
M
fˆ ( x) = ∑ am Bm ( x) (1)
m =1

where the basic function takes the form

Bm ( x) = I ( x ∈ Rm ) (2)
660 Y. Lu et al.

in which I is an indicator function having the value one when x ∈ Rm and zero
otherwise.
From the splitting procedure, the disadvantage of recursive partitioning regression
is obvious to be known. The subregion is discontinuous and the method cannot repre-
sent the functions which have no interactions between their variables. To overcome
the problems above, firstly, MARS uses truncated power function ⎡⎣ ± ( x − t ) ⎤⎦ + to
q

replace the step function, where q = 1 and the form [ ]q+ denotes the positive part of
the expression. The basic function is constructed in the terms of tensor product of the
truncated power function. Second, MARS allows the parent basic function which has
already been split to be involved in the further splitting, so it has the ability to ap-
proximate the additive functions whose variables do not interact with each other.
MARS model can be written as follows:
M Km
f ˆ( x) = a1 + ∑ am ∏ ⎡⎣ skm ⋅ ( xv (k ,m) − tkm )⎤⎦ + (3)
m= 2 k =1

where K m is the number of split of the basic function Bm , skm = ±1 and xv ( k ,m ) is

the split variable, tkm is the split location in the k-th split of the m-th basic function.
a1 , a2 ,..., am is the coefficient of each basic function.

2.1.2 GHH
Generalized hinging hyperplanes was first introduced by Wang in 2005 as an extension
of the hinging hyperplanes. Although HH model can approximate many nonlinear func-

：
tions well, it has been proved that the model is lack of the ability to represent CPWL
functions in high dimensions. HH model can be written as follows

± max {l ( x, a ), l ( x, b)} (4)

Wang made some modification to the above form and rewrote it as follows ：
∑σ i { }
max l ( x, θ1 (i )), l ( x, θ 2 (i )),..., l ( x, θ ki +1 (i )) (5)
i

where l ( x, θ (i )) is linear function, ki ≤ n , σ i = ±1 . This model’s representation

ability was proved in [13].

2.2 AHH Model

The truncated power splines ⎡⎣ skm ⋅ ( xv ( k ,m ) − tkm ) ⎤⎦ in MARS basic function can be
+
written as max{0, skm ⋅ ( xv ( k , m) − tkm )} , and can be seen as a special HH model. To
improve the stability when the dimension of the predictive variable is high, after
 Urban Traffic Flow Forecasting Based on Adaptive Hinging Hyperplanes 661

making a modification to the MARS basic function, replacing the operator “∏” by
“min”, the AHH model is obtained[14]. AHH model can be written as follows:
M
f ˆ( x) = a1 + min {max{0, skm ⋅ ( xv ( k , m) − tkm )}}
∑ am k∈{1,2..., (6)
Km}
m= 2

where K m is the number of splits of the basic function Bm , skm = ±1 and xv ( k ,m ) is

the split variable, tkm is split location in the k-th split of the m-th basic function.
a1 , a2 ,..., am is the coefficient of each basic function.
AHH’s algorithm was developed analog to the MARS algorithm and the general-
ized cross validation (GCV) [16] which is shown as following formula was still used
as estimating standard of errors.

2
1 N
∑ [ yi − fˆM ( xi ) ] [1−C% ( M ) N]
2 (7)
LOF ( fˆ ) = GCV ( M ) =
M N i =1

Forward Procedure

Step 1: Initialize B1 ( x ) as a positive integer which ensures

min( B1 ( x ), Bm ( x )) = Bm ( x ) , set M = 2
Step 2: Loop from B1 ( x ) to BM −1 ( x ) , in each step of the loop changes xv and t ,let

M −1

g = ∑ ai Bi ( x) + aM min {Bm ( x) , max{( xv − t ), 0}} + a M +1

min {B ( x) , max{−( x − t ), 0}}
m v
(8)
i =1

use least square method to decide a , a ,..., a 1 2 M +1

to get the least LOF ( g ) of the step.
After finishing the loop, set m = m; v = v; t = t , which can lead to the least
* * *

LOF ( g ) of all possibilities appeared in Step 2.

Step 3: Set

BM ( x ) = min{Bm ( x ), max{( xv − t ), 0}}

*
* * (9)

BM +1 ( x ) = min{Bm ( x ), max{−( xv − t ), 0}}

* (10)
* *

Setp 4: M = M + 2 . If M < M max

, go to step 2; else finish the algorithm.

The AHH algorithm also needs a backward step just as MARS algorithm. In the
backward procedure, the basic functions which do not make contributions to the fit-
ting precision will be removed.
662 Y. Lu et al.

3 AHH Based Traffic Flow Forecasting

3.1 Data Description

In this paper, the real traffic flow and occupancy data collected from 50 detectors are
provided by Beijing Traffic Management Bureau, which last for 16 days from Oct.22,
2006 to Nov.6, 2006. The sampling interval is 10mins, which means 144 sample
points per day for each detector.
A small part of the network, which is shown in Figure 1, will be used to test the
prediction effect of the model. The flow and occupancy data in link1 and its
neighbors, such as link2, 3, 4, 5 and 6, will be used to forecast the traffic flow in
link1.

Fig. 1. The topological structure of the test network

3.2 Traffic Flow Forecasting Model and Evaluating Criterions

3.2.1 Forecasting Model

The data lasting for 14 days from Oct.22th, 2006 to Nov.4th, 2006 is used to train the
AHH model, and the data in Nov.6th is applied to examine the model’s prediction
accuracy and obtain the forecasting error at the same time.
The traffic forecasting model which is to predict the flow of link i in t+1 time
interval can be written as follows:
∧
flowit +1 = f ( flowit , flowit −1 ,..., flowit − N , flowit+1, flowit+ 2, ..., flowit+ L , , occit σ i ) (11)

where flowit , flowit −1 ,..., flowit − N denotes the link’s own historic flow data before the
prediction interval, in which t , t − 1,..., t − N denote different time intervals;
flowit+1, flowit+ 2, ..., flowit+ L denote the neighbors’ flow data in t interval, in which L de-
cides how many neighbors should be considered; and occit σ i is the link’s own historic
occupancy data in t interval, σ i can be chosen as 1 or 0.
Some explanation is made here to illustrate why the formula above should contain
these factors as predictive variable. First, because the traffic flow can be recognized
as a time series which changed gradually and not very fast, the data which is even N
 Urban Traffic Flow Forecasting Based on Adaptive Hinging Hyperplanes 663

time intervals before the prediction time can be considered to be related, not only one
interval before. Second, it is obvious that the link to be predicted has correlation with
its neighbors, so the neighbor’s data should also be taken into account. Third, the
situation in the network obtained through prediction with both the flow and occu-
pancy data may be more accurate than with only one of them. However, what value N
and L should be is an important issue worth discussing later. If the value is small, the
trained network may be incomplete while too large value will cause the overfitting
problem in the process of training.

3.2.2 Evaluating Criterions

Two criterions will be used in this paper to evaluate the AHH model’s predictive
performance. One is mean absolute percentage error (MAPE) which can provide
useful information in comparison when the traffic flow changes in a large scale. It can
be calculated as following formula:

1 N
predict (i ) − obversed (i )
MAPE (%) =
N
∑ obversed (i )
× 100% (12)
i =1

Another is called PT criterion which denotes the Percentage of the sample points
whose predictive error is smaller than the given Threshold. The criterion can be writ-
ten as following formula:
N
predict (i ) − obversed (i )
∑ predict (i) |
1
PT (%) = < α × 100% (13)
N i =1 obversed (i )

where the expression = 1 if the condition in is satisfied, otherwise = 0 and α is

the given threshold.

3.3 Prediction Results and Analysis

3.3.1 The Length of Time Intervals

Table 1 shows the prediction results while only the link’s own traffic flow informa-
tion is considered as predictive variables in the forecasting procedure. N, which repre-
sents the length of time intervals being considered before the prediction time, changes
from 0 to 9. The threshold in PT criterion is set to 0.2.

Table 1. Results using own flow information in AHH

N 0 1 2 3 4 5 6 7 8 9
MAPE(%) 16.5 14.4 13.9 13.9 14.1 13.3 12.9 13.4 13.5 13.6
PT(%) 69.9 78.9 79.4 80.0 77.7 76.0 77.4 77.9 77.8 77.6

From Table 1, a tendency of MAPE that first becomes smaller and then larger can
be observed with the increasing of N. When N is small, the historic information which
is needed to make prediction is deficient. But when N is too large, the prediction ef-
fect is also worse than the results when N=6. For example, if N=9, the traffic informa-
tion which is nearly 2 hours before the prediction time will be taken into account.
664 Y. Lu et al.

Actually, the data 2 hours ago does not have much relationship with current situation.
However, if these data is seen as predictive variable, the network trained will get
accustomed to the situations with no sense and reduce the prediction accuracy.

3.3.2 The Effect of Occupancy

Table 2 can give an answer about whether the occupancy of the link should be a pre-
dictive variable. In the procedure of prediction, N also changes from 0 to 9.
After comparing Table 1 and 2, the following conclusions can be drawn. First, when
N is small, which means the deficiency of information, the addition of occupancy is
useful to improve the precision. However, when N is large enough, which means the
information getting from the link itself is sufficient, the results are almost the same. But
in practice, considering the computation speed of the computer, the time interval won’t
be too long. So, in sum, occupancy data should be a predictive variable.

Table 2. Results using own flow and occupancy information in AHH

N 0 1 2 3 4 5 6 7 8 9
MAPE(%) 15.6 14.3 13.8 13.9 13.7 13.3 12.9 13.4 13.5 13.6
PT(%) 73.4 78.8 78.7 80.0 78.4 76.1 77.4 77.9 77.8 77.6

3.3.3 The Number of Neighbor Detectors

Table 3 shows different prediction results with different neighbors’ combinations
being considered to forecast the flow of link 1 which are link 2, link2+3, link 2+3+4,
link 2+3+4+5 and link 2+3+4+5+6, respectively.

Table 3. Mean absolute percentage error results using own flow, occupancy and neighbors’
flow information in AHH

Link N 0 1 2 3 4 5 6 7 8 9
2 16.0 14.3 13.7 13.9 13.0 12.6 13.0 13.4 13.5 13.6
2,3 14.8 13.9 13.5 13.2 13.3 12.8 13.0 13.4 13.8 13.9
2,3,4 13.8 13.4 13.0 13.2 12.7 12.5 13.5 13.5 13.2 13.2
2,3,4,5 14.3 14.2 13.5 13.2 12.7 12.5 13.5 13.5 13.2 13.2
2,3,4,5,6 14.2 14.1 13.8 13.5 13.5 12.8 13.4 13.5 13.2 13.2

In Table 3, there are no obvious advantages or disadvantages among the prediction

results of different combinations, but in general the result of link 2+3+4 is relatively
best. The reason to explain this phenomenon is just like the reason to decide time
intervals. There is no doubt that with the neighbors’ information the accuracy will be
better. However, the specific relationship between the neighbors’ data set is not clear
and worth of further research.

3.3.4 Typical Results

In sum, the most precise result is obtained when N=5 and with the contribution of link
1’s own occupancy and link 2, 3 and 4’s flow information. The typical result is shown
in Figure 2.
 Urban Traffic Flow Forecasting Based on Adaptive Hinging Hyperplanes 665

Forecasted flow VS Real flow Using AHH

1200
Forecasted
Real
1000

800
flow

600

400

200

0
0 20 40 60 80 100 120 140
Samples

Fig. 2. The typical prediction result

4 Compared with Results Based on MARS

Based on MARS algorithm and adopts the same traffic prediction model form of
AHH, the prediction results based on MARS has been obtained and will be compared
with the results of AHH in this section.

4.1 General Performance

As seen in Figure 3, in general, the predictive performance of AHH and MARS model
using their own flow information are almost the same. In specific terms, when N is
small, the precision of AHH model is better with smaller error. But when N becomes
large, the MARS model will have some advantages gradually.

Error using AHH and MARS Time cost between AHH and MARS
0.17 200

0.165 180 AHH

AHH
MARS
MARS 160
0.16
140
0.155
120
Time cost/s

0.15
error

100
0.145
80
0.14
60
0.135
40
0.13 20

0.125 0
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
N N

Fig. 3. General Comparison between AHH and Fig. 4. Time Cost Comparison between AHH
MARS and MARS
666 Y. Lu et al.

From Figure 3, we can also observe that there exists more fluctuation in the per-
formance of MARS. For instance, when N=7, 8 and 9, a slightly increase in error is
intended to be seen because of the overfitting problem, however, a sharp increase
followed with dramatic decrease in error have been found instead.

4.2 Time Cost

As seen in Figure 4, in most circumstances, especially when N is large, the time

which MARS model needs to accomplish the computation is longer than AHH model.
However, MARS can only get better results with larger N which denotes much more
computing time.

5 Conclusions
There have been a variety of traffic forecasting methods during these years, but every
approach has its own limitation, so the research about this issue is still very active.
We combine the MARS algorithm and the thinking of piecewise linear functions,
introducing a new traffic prediction method.
In this paper, after a review of MARS algorithm and the development of piecewise
linear functions, a traffic forecasting model based on adaptive hinging hyperplanes
was proposed. Acceptable results (error is smaller than 15%) were obtained in the test
of the real traffic data in Beijing. After comparison with the results of MARS model,
the following conclusions can be drawn. First, the two methods have almost the same
performance in precision. Second, AHH will be more stable and cost less computing
time so it is more applicable in practice.
However, there are still some problems unsolved. The prediction results based on
AHH which is nonparametric model cannot always follow the real data when the
traffic flow is becoming very fast. In further research, some other methods may be
adopted together with AHH as a combination model.
Acknowledgement. This work described in the paper is partially supported by Na-
tional Natural Science Foundation of China (NSFC) 50708054, 60674025 and
60534060, Hi-Tech Research and Development Program of China (863Project)
2007AA11Z222 and 2007AA04Z193, National Basic Research Program of China
(973Project) 2006CB705506, National Key Technology Research and Development
Program 2006BAJ18B02 and The Research Fund for the Doctoral Program of Higher
Education 200800030029.

References
1. Smith, B.L., Williams, B.M., Oswald, R.K.: Comparison of parametric and nonparametric
models for traffic flow forecasting. Transportation Research Part C 10, 303–321 (2002)
2. Okutani, I., Stephanedes, Y.J.: Dynamic prediction of traffic volume through Kalman fil-
tering theory. Transportation Research Part B 18B, 1–11 (1984)
 Urban Traffic Flow Forecasting Based on Adaptive Hinging Hyperplanes 667

3. Williams, B.M., Hoel, L.A.: Modeling and Forecasting Vehicular Traffic Flow as a Sea-
sonal ARIMA Process: Theoretical Basis and Empirical Results. ASCE Journal of Trans-
portation Engineering 129(6), 664–672 (2003)
4. Davis, G.A., Nihan, N.L.: Nonparametric Regression and Short-term Freeway Traffic
Forecasting. ASCE Journal of Transportation Engineering 117(2), 178–188 (1991)
5. Smith, B.L., Demetsky, M.J.: Traffic Flow Forecasting: Comparison of Modeling Ap-
proaches. ASCE Journal of Transportation Engineering 123(4), 261–266 (1997)
6. Ye, S.Q., et al.: Short-Term Traffic Flow Forecasting Based on MARS. In: The 4th Inter-
national Conference on Natural Computation, pp. 669–675. IEEE Press, Jinan (2008)
7. Chua, L.O., Kang, S.M.: Section-wise piecewise-linear functions: canonical representation,
properties, and applications. IEEE Transactions on Circuits and Systems 30(3), 125–140
(1977)
8. Kang, S.M., Chua, L.O.: Global representation of multidimensional piecewise-linear func-
tions with linear partitions. IEEE Transactions on Circuits and Systems 25(11), 938–940
(1978)
9. Kahlert, C., Chua, L.O.: A generalized canonical piecewise linear representation. IEEE
Transactions on Circuits and Systems 37(3), 373–382 (1990)
10. Lin, J.N., Unbehauen, R.: Canonical piecewise-linear networks. IEEE Transactions on
Neural Networks 6(1), 43–50 (1995)
11. Tarela, J.M., Alonso, E., Martinez, M.V.: A representation method for PWL functions ori-
ented to parallel processing. Mathematical & Computer Modelling 13(10), 75–83 (1990)
12. Breiman, L.: Hinging hyperplanes for regression, classification and function approxima-
tion. IEEE Transactions on Information Theory 39(3), 999–1013 (1993)
13. Wang, S.N., Sun, X.S.: Generalization of Hinging Hyperplanes. IEEE Transactions on In-
formation Theory 12(51), 4425–4431 (2005)
14. Xu, J., Huang, X.L., Wang, S.N.: Adaptive Hinging Hyperplanes. In: The 17th World
Congress of International Federation of Automatic Control, World Congress, Seoul, Korea,
pp. 4036–4041 (2008)
15. Friedman, J.H.: Multivariate adptive regression splines. The Annuals of Statistics 19(1), 1–
61 (1991)
16. Friedman, J.H., Silverman, B.W.: Flexible parsimonious smoothing and additive modeling.
Technometrics 31, 3–39 (1989)
 Turbine Fault Diagnosis Based on
Fuzzy Theory and SVM

Fei Xia1, Hao Zhang1, Daogang Peng1, Hui Li1, and Yikang Su2
1
College of Electric Power and Automation Engineering,
Shanghai University of Electric Power,
200090 Shanghai, China
2
Nanchang Power Supply Corporation,
330006 Nanchang, JiangXi, China
{Fei.Xia,Hao.Zhang,Daogang.Peng,Hui.Li,Yikang.Su,
xiafei}@shiep.edu.cn

Abstract. A method based on fuzzy and support vector machine (SVM) is pro-
posed to focus on the lack of samples in fault diagnosis of turbine. Typical fault
symptoms firstly are normalized by the membership functions perceptively.
Then some samples are used to train SVM of fault diagnosis. With the trained
SVM, the correct fault type can be recognized. In the application of condenser
fault diagnosis, the approach enhances successfully the accuracy of fault diag-
nosis with small samples. Compared with the general method of BP neural net-
work, the method combining advantages of fuzzy theory and SVM makes the
diagnosis results have higher credibility.

Keywords: turbine fault diagnosis, fuzzy theory, SVM.

1 Introduction
The turbine on electric power plant is the important equipment [1]. Its structure is
complex and the operation environment is particular. It works under the particular
operation environment which is under high temperature, high-pressured, and high
speed revolving. Obliviously, the steam turbine has high fault rate and hazardous
breakdown. These breakdowns will make the heavy economic loss and the social
consequence [2]. Therefore, it needs to use the advanced intelligent technology to
monitor and analysis device status parameters in order to judge the corresponding
equipment whether it is in a health state.
Condenser is the primarily auxiliary equipment of steam turbine whose quality of
work situation affects the safety and economic operation of generator. Therefore, the
monitoring and diagnosis of condenser running state are generally concerned by the
operation department of the power plant. The research of condenser system and its
fault diagnosis has great significance to reduce the downtime of units and improve
availability of units.
In the application of modern fault diagnosis, the method of neural network has be-
come an important tool with the characteristics of self-learning capability, fault-tolerant
capability and parallel calculation capability. Various methods of neural network have

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 668–676, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Turbine Fault Diagnosis Based on Fuzzy Theory and SVM 669

been applied to fault diagnosis of the condenser [3-6]. As the result of nonlinear rela-
tionship between the failures and the symptoms, complexity, ambiguity and random-
ness, it is hard to construct the precise mathematical model for the condenser system.
Therefore, the values of symptoms can be transformed by the corresponding member-
ship function with fuzzy theory. After that, more methods of fuzzy neural network are
used in the fault diagnosis of condenser [7-8]. Although these methods take into account
the merits of the two principles, the neural network is still vulnerable to fall into
local minimum point. Even though, the network structure should be determined by the
experience.
For solving these problems, the approach combined with fuzzy theory and Support
Vector Machine (SVM) is proposed to achieve the fault diagnosis results in the con-
denser system. SVM is a new machine learning method developed on basis of the
statistical learning theory [9]. Based on the principle of structural risk minimization, it
can effectively address the learning problem and have better classification accuracy
[10]. SVM has been used for face detection, speech recognition and medical diagnosis
etc. Through the study of this paper, it can be proved successfully that SVM does very
well in the tasks of fault diagnosis of condenser.

2 Fault Analysis of Condenser

Condenser plays an important role in the power plants. However, the condenser often
has some faults in the operation process due to the design, the installation, the main-
tenance ant other reasons, such as low vacuum operation, cooling pipe leakage, super
cool condensate water, higher oxygen content in condensate etc.. One of the most
common is running low vacuum condenser. When condenser vacuum is too low, the
thermal efficiency and output of unit is reduced. Even some security issues are caused
by this state. The details are as following.
Due to the increased exhaust pressure, the turbine low pressure rotor blades level
chatter and the stress increased which reduce the security of unit. Because of increased
exhaust pressure and exhaust steam temperature, the speed ratio increases for the end
of lower-level. In the reason of the thermal expansion of the mental increasing, the
rotor center line is changed and the unit has the vibration. The increased exhaust tem-
perature results in increase of exhaust hood temperature and condenser temperature.
And the relative expansion between the cylinder and the condenser is up to 3 times
than the normal value. The relative difference between brass condenser tube and plate
rises to bulging population relaxation and destruction of water-tight side which result
in that condensation water is contaminated.
Thus the condenser vacuum monitoring is very important in actual operation. In
the standard operation of 300MW unit, the value of low vacuum condenser alarm is
87kpa and the value of down trip is 67kpa. According to the survey, the unit availabil-
ity in thermal power plants over 600 MW declined 3.8% due to the equipment of
condenser having the poor reliability.
Fault diagnosis of condenser is mainly discussed in this paper according to previous
fault diagnosis methods. In order to carry out fault diagnosis for condenser, fault uni-
verse of condenser are divided into three typical fault fuzzy modes through selection
combining experience of engineer experts and technicians in this paper, they are
670 F. Xia et al.

Y1:serious fault of cycle pump; Y2:full water of condenser; Y3:not working of con-
densate pump. Eight symptoms relating with condenser fault are extracted as follows:
X1: vacuum declined significantly and sharply; X2: vacuum declined slowly and
slightly; X3: output pressure of condensate pump increased; X4: output pressure of
condensate pump decreased; X5: temperature rise of cycle water decreased; X6: termi-
nal difference of condenser increased; X7: coldness of condensation water increased;
X8: pressure difference between pumping port and input of extractor decreased.

3 Definition of Membership Function

3.1 Fuzzy Theory

The In the daily life, many concepts are vague. It is far insufficient to describe with
belonged absolutely or not belong absolutely. Thus, it has the necessity to break the
relations of absolutely subordinates. University of California Professor Zadeh has
introduced the fuzzy set theory in 1965. The fuzzy set theory develops the ordinary
set theory concept, value scope of characteristic function is expanded from gathers
{0,1} to [0,1]. Some object of the universe is no longer regarded as belongs to this set
or not belong to this set, but said that the degree of subordinates in this set is how
much [4].
Assuming that a mapping is assigned on universe U:
A : U → [0,1] (1)
u → A(u )
(2)
So, A is fuzzy set on the U, all fuzzy set in the U are recorded as (U), that is,
(U ) = { A | A : U → [0,1]}
(3)
A fuzzy set definite in the universe, its membership function have many different
form. Determines the membership function correctly is the foundation for using fuzzy
set to describe fuzzy concept appropriately. One basic step of the application of fuzzy
mathematics to solve the practical problem is to find one or a several membership
function. If this question has been solved, other questions are easily solved.

3.2 Membership Function

In order to disperse data from sensors, proper membership function should be con-
structed. Membership function is the foundation of fuzzy sets applied to practical
problems, correctly constructing membership function is the key of properly using
fuzzy sets. While there is no mature and effective method in the existing membership
functions, they are all determined by experience, and then corrected by experiments
or feedback information by computer simulating. In the turbine fault diagnosis, if the
corresponding membership function is unable to determined, the following three
methods can be adopted according to specific conditions:
 Turbine Fault Diagnosis Based on Fuzzy Theory and SVM 671

1 ）Triangular fuzzy function

⎧ ( x − a ) / (b − a ) a≤ x≤b
⎪
µ A ( x ) = ⎨ ( c − x ) / (c − b ) b≤ x≤c
⎪0 < x>c
⎩ x a or
(4)
）
2 Trapezoidal fuzzy function

⎧ 0 x<a
⎪ ( x − a ) / (b − a ) a≤ x≤b
⎪⎪
µ A ( x) = ⎨ 1 b≤ x≤c
⎪( d − x ) / ( d − c ) c<x≤d
⎪
⎪⎩ 0 x>d
(5)
3 ）Normal-shaped fuzzy function
x −a 2
−( )
µ A ( x) = e b
(6)

3.3 Definition

Judging whether fault symptom exist or not directly by values of each thermal pa-
rameters of condenser is not accurate at all, because one fault symptom may corre-
spond to different fault, and although the changing trend of same fault symptom of
different faults are the same, the degree of changing are different. Therefore, thermal
parameters should be fuzzy treated with the concept of fuzzy mathematics. Mean-
while, data from sensors can be dispersed. The changing trends are shown natural and
veritable by blurring each input parameters using membership function.
According [11], the following three types of membership functions are applicable
to condenser fault symptom:
1) Smaller sized

⎧(1 + a( x − c)b )−1 ( x > c)

µ A ( x) = ⎨
⎩ 1 ( x ≤ c)
(7)
In the formula, c ∈ U is an arbitrary point,(a>0,b>0).
2) Larger sized
⎧(1 + a( x − c)b )−1 ( x > c)
µ A ( x) = ⎨
⎩ 0 ( x ≤ c)
(8)
In the formula, c ∈ U is an arbitrary point,(a>0,b>0).
3) Medium sized
µ A ( x) = e − k ( x − c )
2

(9)
In the formula, c∈U is an arbitrary point; k is a parameter larger than zero (k>0).
672 F. Xia et al.

According to the upper formulas and combining practical situation, the member-
ship function of operating parameters of 300MW unit condenser corresponding symp-
tom can be obtained and they are shown in TABLE I.

Table 1. Membership function of each fault symptom

Fault symptom Membership function

（ ）
vacuum declined significantly and sharply X1 exp[−0.3( x − 87) 2 ]
vacuum declined slowly and slightly（X2） exp[−0.3( x − 97)2 ]
output pressure of condensate pump increased（X3） ( x − 0.5) / 3.7
output pressure of condensate pump decreased（X4） exp[−4( x − 3.7) 2 ]
temperature rise of cycle water decreased（X5） exp[−0.2( x − 8) 2 ]
terminal difference of condenser increased（X6） ( x − 3) / 3.56
coldness of condensation water increased（X7） ( x − 2) / 0.5
pressure difference between pumping port and input of
extractor decreased（X8）
exp[−0.05( x − 8.2) 2 ]

4 Support Vector Machine

4.1 Fundamentals

SVM is a class of supervised learning algorithms introduced firstly by Vapnik. SVM

has excellent performance in some pattern recognition application fields. Also, it has
obtained much success in a variety of domains based on the principle of structural risk
minimization. The basic idea of SVM in the application of classification is introduced
by the simple example with 2-D space. Considered as shown in Figure 1 corresponds
to 2-dimensional space on a simple classification to discuss the function for the linear
function f (x) = (w, x) + b. When the classification problem is to be found a suitable
straight-line which divided the entire 2-D plane, that is, to determine the direction of
w and the intercept method b.

Fig. 1. Schematic Diagram of Classification Plane

 Turbine Fault Diagnosis Based on Fuzzy Theory and SVM 673

Two types of points can be separated from a right line which exists numerously.
The classification function is approximated by the following function:
f ( x, a ) = w ⋅ x + b (10)

Coefficients w and b are estimated by minimizing the regularized risk function:

1
min φ ( x) =
2
w (11)
2
s.t. yi [( w ⋅ xi ) − b] ≥ 1, i = 1, 2,L , n
According to Lagrange principle, the above problem can be transformed to solve its
antithesis problem. Finally we can get the decision function as follows:

⎧ n ⎫
f ( x) = sgn ⎨∑ α i* yi ( xi ⋅ x) + b* ⎬ (12)
⎩ i =1 ⎭

4.2 Method of SVM

With the deep study on support vector machine, many researchers increased and
changed the function, the variable coefficient method or formula for the deformation,
resulting in advantages that there is a certain application or algorithm with a number
of support vector machines deformation algorithm. The main algorithms are the fol-
lowing: C-SVM Series, ν-SVM Series, One-class SVM, RSVM (reduced SVM),
WSVM (weighted SVM) and LS-SVM (least-square SVM).
C-SVM algorithm for the only adjustable parameter C is not an intuitive explana-
tion, it is very difficult in practice to choose the appropriate value of the defection,
Schblkoph proposed ν-SVM algorithm. ν-SVM with the new parameter ν to replace the
C, so that we can control the number of support vector or error, but also easy to choose
the value than that of C parameter. It has very clear physical meaning for lν which
represents the maximum amount of support vector borders and the lower limit of the
number of support vector. l is the size of samples. Although the standard ν-SVM algo-
rithm is more complex than C-SVM algorithm, it has the effective way to solve the
problem of classification with small samples. Chang et al proposed a deformed ν-SVM

Fig. 2. Processing of ν-SVM Algorithm

674 F. Xia et al.

algorithm which has original objective function with the increment of variable a. In
order to distinguish the two ν-SVM methods, the later is called Bν-SVM algorithm. For
simplification, the ν-SVM algorithm is adopted in the fault diagnosis analysis task. A
simple process of ν-SVM algorithm is displayed in Figure2.

5 Simulation Experiments

5.1 SVM Training

The three classical fault types of condenser mentioned are used in the stimulation of
fault diagnosis. The fault set is expressed by Y={Y1, Y2, Y3}. In the set, Y1 repre-
sents serious fault of cycle pump, Y2 represents full water of condenser and Y3 repre-
sents not working of condensate pump.
For different fault types, 40 samples are collected respectively. Then 20 samples
are used in SVM train and the other 20 samples are to be measured the SVM through
the previous train. Some examples in rain set and the corresponding fault type are
shown in Table 2.

Table 2. Examples of SVM Train Set

1 2 3 4
X1 0.05 0.06 0.03 0.01
X2 0.65 0.60 0.53 0.75
X3 0.15 0.11 0.24 0.12
X4 0.18 0.17 0.42 0.15
X5 0.56 0.50 0.67 0.86
X6 0.43 0.40 0.58 0.69
X7 0.11 0.09 0.13 0.22
X8 0.37 0.28 0.41 0.73
Fault Type Y1 Y1 Y2 Y3

There is not a standard to choose the various parameters of SVM in theory. So it is

hard to get the correct value in the practical problem. In the problem of condenser
fault diagnosis, Gaussian function is used as the kernel function:
⎧ 2⎫
⎪ − xi − x j ⎪
k ( xi , x j ) = exp ⎨ ⎬, i = 1, 2, L, n (13)
⎪ σ2 ⎪
⎩ ⎭

σ, which is the perception variable, determines the width of Gaussian function. The
value of σ can control the number of support vector machine. When the number of
support vector machine is too much, it can be appropriate by reducing the value of σ.
Generally the value is based on data analysis and experiences. To simplify, the value
of σ in the work is determined by the follow formula:
 Turbine Fault Diagnosis Based on Fuzzy Theory and SVM 675

⎛ 2⎞
σ 2 = E ⎜ xi − x j ⎟ (14)
⎝ ⎠

In the above formulation, E represents mathematical expectation. And the value of ν

is also obtained by the test. In the work of condenser fault diagnosis, ν equals 10.

5.2 Stimulation Test

According to the trained SVM, 60 samples of condenser fault are adopted for the
verification. And some of the test results are shown in Table 3.

Table 3. Examples of SVM Test Set

1 2 3 4
X1 0.05 0.03 0.02 0.01
X2 0.62 0.55 0.70 0.72
X3 0.10 0.26 0.11 0.13
X4 0.16 0.41 0.13 0.15
X5 0.52 0.66 0.79 0.84
X6 0.43 0.58 0.71 0.69
X7 0.09 0.14 0.18 0.21
X8 0.30 0.43 0.78 0.71
Fault Type Y1 Y2 Y3 Y3
Actual Situation Y1 Y2 Y3 Y3

The results of stimulation test shows that SVM has the ability to classify correctly
the fault samples especially for the small samples due to the strong capability of data
processing.
In order to compare with the method of SVM proposed in this paper, the same
eight fault characteristics and three fault types of condenser are also adopted in the
method of BP neural network. For compared the result, the same membership func-
tion and three-level BP neural network [12-13] are adopted in tests. Eight fault char-
acteristics are inputs of the neural network and three fault types are the outputs.
Furthermore, the above two methods were tested through 30 simulation data. The
results are shown in Table 4.

Table 4. Results of Comparing Experiments

Diagnosis method True(%) Fault(%)

SVM 91.3 8.7
BP neural network 83.3 16.7

Diagnostic results show that SVM has a higher diagnostic accuracy than that of BP
neural network in the application of small and complex fault samples. Generally the
correct classification rate is improved 8%.
676 F. Xia et al.

6 Conclusion
SVM is a learning algorithm based on the principle of structural risk minimization,
which has stronger theoretical basis and better generalization ability than neural network
algorithm based on the experience minimization. Steam condenser fault diagnosis is
taken as an example to verify the approach of SVM with fuzzy theory in intelligent
steam turbine fault diagnosis. Experiment results show that SVM had higher diagnostic
accuracy and stronger ability to classify than BP neural network in the complex fault
diagnosis with small samples. In the future research, the approach will be applied in the
more complicated fault diagnosis to verify the applicability. Furthermore, other intelli-
gent diagnosis methods can be combined with SVM to enhance the accuracy of fault
diagnosis.
Acknowledgments. This work is supported by Program of Shanghai Subject Chief
Scientist with (09XD1401900) and Natural Science Foundation of Shanghai (No.
09ZR1413300).

References
1. Diao, Y., Passinb, K.M.: Fault diagnosis for a turbine engine. Control Engineering Prac-
tice, 1151–1165 (December 2004)
2. Zeng, X., Li, K.K., Chan, W.L., Yin, X., Chen, D., Lin, G.: Discussion on application of
information fusion techniques in electric power system fault detection. Electric Power 36,
8–12 (2003)
3. Zhao, H., Li, W., Sheng, D., et al.: Study on Fault Diagnosis of Condenser Based on BP
Neural Network. Power System Engineering 20, 32–34 (2004)
4. Chen, Z., Xu, J.: Condenser fault diagnosis based on Elman networks. East China Electric
Power 35, 871–874 (2007)
5. Wang, J., Li, L., Tang, G.: Study on Fault Diagnosis of Condenser Based on RBF Neural
Network. Electric Power Science and Engineering 23, 27–31 (2007)
6. Ma, Y., Yin, Z., Ma, L.: Study on fault diagnosis of condenser based on SOM neural net-
work. Journal of North China Electric Power University 33, 5–8 (2006)
7. Wang, X., Wang, Q.: A Study on the Diagnosis of the Condenser Faults Based on the
Fuzzy Neural Network. Modern Electric Power 18, 12–17 (2001)
8. Wu, Z., Lin, Z.: FNN-based fault diagnosis system for condenser. Gas Turbine Technol-
ogy 21, 42–45 (2008)
9. Burges, C.J.: A Tutorial on Support Vector Machine for Pattern Recogntion. Data Mining
and Knowledge Discovery 2(2) (1998)
10. Vapnik, V.N.: The Nature of Statistical Learning. Tsinghua University Press, Beijing
(2004)
11. Jia, X.: Fault Diagnosis in Turbine of Condenser. Harbin Engineering University, Haerbin
(2004)
12. FeiSi Center of Technology and Research, Neural Network Theory and MATLAB 7 Ap-
plication, pp. 44–51. Publishing House of Electric Industry, Beijing (2005)
13. Wu, Z., Lin, Z.: FNN-based fault diagnosis system for condenser. Gas Turbine Technol-
ogy 21, 42–45 (2008)
 Architecture of Multiple Algorithm Integration
for Real-Time Image Understanding Application

Bobo Duan, Chunyang Yang, Wei Liu, and Huai Yuan

Software Center, Northeastern University,

110179 Shenyang, China
{duanbb,yangcy,lwei,yuanh}@neusoft.com

Abstract. Robustness and real-time usually are the main challenges when de-
signing image understanding approach for practical application. To achieve ro-
bustness, integrating multiple algorithms to make a special hybrid approach are
becoming a popular way and there have been a lot of successful hybrid ap-
proaches. But this aggravates the difficulties in achieving real-time because of
heavy computational workload of multiple algorithms. To design a hybrid ap-
proach with real-time constraint more easily, some theoretical researches about
multiple algorithm integration are very necessary. This paper presents a common
multiple algorithm integration model and architecture for typical image under-
standing applications. To achieve robustness and real-time in a hybrid approach,
the strategies for increasing robustness and speed up are analyzed. Finally a ro-
bust hybrid approach for rear vehicle and motorcycle detection and tracking is
introduced as a sample.

Keywords: Image Understanding, Robustness, Real-Time, Hybrid, Fusion.

1 Introduction
When developing an image understanding system for practical applications, follow
challenges usually are faced: robustness and real-time. For example, an obstacle de-
tection approach for driver assistance system need to provide robust perception in wide
variety of outdoor environment, and this procedure need to be finished as quickly as
possible to save time for driver reaction [1]. A successful image understanding system
should meet the two requirements at the same time, but it is very difficult to design such
an approach because the two requirements are usually with strong conflicts.
In image understanding procedure, it is often the case that many algorithms exist to
acquire same special information, and each possessing different performance and
computing workload characteristics. Since no single algorithm will be robust enough to
deal with a wide variety of environment conditions, integrating multiple algorithms to
make a special hybrid approach is becoming a popular way to increase robustness. But
image understanding algorithms usually have high computational workload, this inte-
gration aggravates the difficulty in achieving real-time. So it become very important to
do some common research about how to design a hybrid image understanding approach
which can achieve robustness and real-time at the same.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 677–684, 2009.
© Springer-Verlag Berlin Heidelberg 2009
678 B. Duan et al.

Up to now, there have been a lot of researches about multiple algorithms integration
in image understanding. Most of them propose their hybrid approaches for special
problems, which usually are with a form of multiple kinds of technologies hybrid or
multiple visual cues integration. For example, P. Zhang et al. [2] propose a hybrid
classifier for handwritten numeral recognition, which uses neural network as coarse
classifier and used decision tree as fine one. K. Toyama et al. [3] propose a hybrid visual
tracking approach, which includes multiple trackers such as edge of polygon tracker,
rectangle tracker and etc. B. Leibe et al. [4] propose a segmentation approach for object
recognition that integrates multiple visual cues and combines multiple interest region
detectors. Multiple algorithms integration has been proved to be an effective way for
increasing robustness, and some theoretical researches have been performed. For ex-
ample, M. Spengler et al. [5] propose a common framework for visual tracking that can
support the integration of multiple cues. These researches can help to design robust
approach for some special purposes such as tracking, but deeper researches are still
needed to be performed about how to design a robust image understanding hybrid
approach for common purpose with real-time constraint.
This paper is organized as follow, firstly a common multiple algorithm integration
model is proposed, which can support common purpose such as segmentation, classi-
fying, tracking and etc. Then some strategies are introduced about how to design a
robust and real-time hybrid approach base on the model, and design strategies and
executing strategies are also analyzed about increasing robustness and speeding up.
Then architecture for typical applications is introduced, which can help design a total
solution for image understanding task. At last a hybrid approach for rear vehicle
detection and tracking is introduced as a sample.

2 Hybrid Model and Architecture

2.1 Hybrid Model

The kernel of multiple algorithm integration is to dynamically select and execute

appropriate algorithm sequence with expected performance and executing time. So it
is important to analyze those factors that can affect algorithm sequence. Performance
and executing time of algorithm usually are related with input data, so input data is
the first important factor. Besides, computing power is another important factor.
When there is abundance computing power, all algorithms could be executed to get
expected information as robust as possible in real-time. But when computing power is
limited it is difficult to archive robustness and real-time at the same time, a tradeoff
has to been made to determine the weightiness for robustness and real-time. So the
third factor is executing strategies which implement the tradeoff. According to above
analysis, a multiple algorithm integration model can be given and we call it
“HYBRID”. As in Fig.1, the model includes follow parts: “Algorithm Pool”,
“Scheduler”, “Evaluator” and “Fuser”, and real arrows indicate data flow and dash
arrows indicate control flow.
 Architecture of Multiple Algorithm Integration 679

Fig. 1. Multiple Algorithm Integration Model - “HYBRID”

Algorithm Pool includes a set of different algorithms which can provide same output
information and each algorithm has a description about its performance and computing
workload characteristics. In most cases, quality of output information could be as a
usual description of performance characteristic of algorithm, and computing cost could
be a usual description of computing workload characteristic. These descriptions may be
express in different forms, for example, the average detection rate of segmentation
algorithm in some illumination condition may be as the expression of quality of its
output information, and average executing time may be as the expression of its com-
puting cost. And this information can be acquired from statistic.
Evaluator is in charge of providing the evaluation values of performance and
computing workload characteristics in current input data. For example, to above seg-
mentation algorithms, the evaluation value could be given according to current illu-
mination condition which could be calculated base on current input data.
Fuser is in charge of fusing output information of multiple algorithms to give final
output. And most technologies about multi-sensor data fusion and multi-cues integra-
tion could be applied very well for this procedure.
Scheduler is in charge of scheduling algorithms according to current available
computing power, executing strategy and evaluation values of performance and com-
puting workload characteristics of each algorithm. Usually computing power indicates
available executing time of hybrid approach and executing strategy indicates the pri-
ority of performance and computing workload requirement, and special executing
strategy will be implemented with special scheduling method.
The main advantage of HYBRID model is that it can support common image un-
derstanding task. On the one hand, the algorithms of “Algorithm Pool” may be detector,
classifier and tracker, so the model can also support different purposes such as detec-
tion, classification, tracking and etc. On the other hand, when taking visual cue ob-
servation procedures as algorithms of “Algorithm Pool”, HYBRID procedure is
equivalent to multiple visual cues integration. And when taking information acquiring
procedures from sensors as algorithms of “Algorithm Pool”, HYBRID procedure is
680 B. Duan et al.

equivalent to multiple sensor data fusion. It shows the model can support multi-cues
integration even multi-sensor data fusion. The analyzing indicates that HYBRID model
can provide support for current main image understanding purposes.

2.2 Strategies for Hybrid Approach

One important aim of this research is to find some strategies that can help to design a
robust and real-time approach base on above HYBRID model. For such an approach,
strategies about increasing robustness can be analyzed base on three parts. The first part
is “Algorithm Pool”. For single algorithm, an adaptive algorithm such as adaptive
SOBEL edge detector is helpful for robustness in complexity environment. The second
part is “Scheduler”. The selection of appropriate algorithms sequence is helpful for
increasing robustness. The third part is “Fuser”. The fusion of multiple visual cues and
multiple sensor data can provide huge help for robustness.
Since the procedure of information fusion usually needs less computing power than
image understanding algorithm, the strategies about speeding up hybrid approach will
be analyzed base on only “Algorithm Pool” and “Scheduler”. For single algorithm in
Algorithm Pool, the speeding up can be implemented in two levels. The first level is
software implementation of single algorithm. And high efficient implementation in
special hardware is an important way to achieve real-time, such as high performance
program base on data parallel technology. The second level is base on the reducing of
computing complexity of single algorithm, multi-resolution technology is popular way
to reduce complexity. For multiple algorithm integration in Scheduler, the determining
of appropriate algorithms sequence is helpful for archive real-time.
Above strategies can be implemented in two phases: designing phase and executing
phase. Strategy implementations in design phase are mainly focusing on choosing
efficient algorithms for “Algorithm Pool”. And strategy implementations in executing
phase are mainly focusing on scheduling algorithms and fusing their output informa-
tion. And this paper will mainly provide some analysis about algorithms scheduling
method. In actual applications the scheduling strategies usually are various, but they
still can be summarized into two patterns in the rough as follow.

2.2.1 Scheduling Pattern – Select-Do

This pattern usually has another name – Adaptive Pattern. This pattern means the
scheduler needs to select and execute an algorithm sequence which can provide ex-
pected information with enough quality and with least computing cost. To reduce the
difficulty of scheduling, this paper will consider only selecting one algorithm from
“Algorithm Pool” as a sequence. And an available scheduling procedure is shown as
Fig.2. In this procedure, selecting step will choose one algorithm that can provide
output information with enough quality for current input and is with lowest computing
cost. If there is not an available algorithm for current input, the robustness will not been
archived. And when the executing time of selected algorithm is more than available
executing time, real-time will not been achieved. A typical example of this pattern has
been shown in article [6].
 Architecture of Multiple Algorithm Integration 681

Fig. 2. Scheduling Pattern: Select-Do

Fig. 3. Scheduling: Do-Till

2.2.2 Scheduling Pattern – Do-Till

This pattern usually means scheduler needs to select and execute a sequence of algo-
rithms which can provide expected information with best quality in a time constraint.
And an available scheduling procedure is shown as Fig.3.In this procedure, selecting
step will select one algorithm whose expected executing time is less than current
available executing time and with highest robustness. And the selecting step will be
repeated till there is not available executing time. If there is not any available algorithm,
the procedure will also be terminated. After the selected algorithm has been executed,
its executing time should been subtract from available executing time. Then check step
will judge if the special task has been finished. If the task has been finished, the pro-
cedure will be terminated. Two typical samples of this pattern have been shown in
article [7] [8].
682 B. Duan et al.

2.3 Hierarchical Hybrid Framework for Typical Application

The HYBRID model and above strategies can help design one approach for some
purpose, but they are not enough for designing a total solution of special image un-
derstanding application. So a framework is necessary and there are a lot of researches
that provide their frameworks [9] [10]. Base on above researches, this paper customizes
a hierarchical framework for common image understanding task and give an analysis
about how to design a hybrid approach base on the framework. As Fig.4, the framework
includes six kinds of information (sensor data, data, feature, object hypotheses, object
observation, object tracking), and it includes five typical processing procedures (pre-
process, feature extraction, hypothesis generation (HG), hypothesis verification (HV),
tracking), which can transform one or multiple kinds of information to another. One
ellipse is a kind of information, and one a dash arrow indicates there is procedure that
can implement the transform along the direction of arrow. One circle is a special case of
some kind of information, and a real arrow indicates there is an algorithm which can
implement the transform along the direction of arrow. If there is a circle towards which
there are multiple real arrows, it means we can design a hybrid approach base on
HYBRID model for acquiring the information. Then once the expected information is
specified, if a data chain from sensor data to expected information could be constructed,
a solution of this image understanding task could be determined. And a total solution
could be enhanced robustness and speed up using above strategies.

Fig. 4. Framework for Multiple Algorithm Integration

 Architecture of Multiple Algorithm Integration 683

3 Hybrid Approach for Rear Vehicle Detection and Tracking

To validate above model and framework, an approach about rear vehicle and motor-
cycle detection and tracking base on image understanding is developed, which need to
provide robust result information in real-time (33frame/s) and in different situations
such as day or night, sunshine or rain, etc. A part of our system has been shown in Fig 5.
Shadow and wheel based hypothesis generation method is applied for daytime and
headlight-based method is applied for night, and it is a sample of scheduling strategy
pattern 1. And knowledge-based classifier is applied as coarse one and SVM classifier
is applied as fine one, and it is a sample of scheduling strategy pattern 2. A real-time
prototype of our approach has been finished which can process a frame (640*480) at an
average processing time less than 33 ms (with NEC SIMD Computer IMAP-CAR) and
the DR (Detection Rate) is more than 95% and FAR (False Detection Rate) is less than
5%. Some detail information about this approach can be fund in [11]. The test result
shows the efficiency of multiple algorithms integration and validate our framework and
HYBRID Model.

Fig. 5. A Part of the Our Vehicle and Motorcycle Recognition Approach

4 Conclusions
To increase robustness of image understanding approach, integrating multiple algo-
rithms to make a special hybrid approach are becoming popular. But this aggravates the
difficulties in achieving real-time because of heavy computational workload of multi-
ple image understanding algorithms. Some theoretical researches about how to design a
robust and real-time hybrid approach are necessary.
684 B. Duan et al.

This paper presents a common multiple algorithm integration model and analyzes
the strategies and scheduling method for increasing robustness and speeding up. Then
architecture for typical applications is introduced and it can help to design a total image
understanding approach for special application. At last the prototype of rear vehicle and
motorcycle detection and tracking validate efficiency of the framework and HYBRID
model for achieving robustness and real-time.
Research about multiple algorithms integration can help people to apply image
understanding in real applications more easily and it is necessary to do deeper re-
searches. Our next step is to validate them using a multi-sensor system.

References
1. Sun, Z., Bebis, G., Miller, R.: On-road vehicle detection using optical sensors a review. In:
IEEE International Conference on Intelligent Transportation Systems, pp. 585–590. IEEE
Press, New York (2004)
2. Heidemann, G., Kummert, F., Ritter, H., Sagerer, G.: A Hybrid Object Recognition Archi-
tecture. In: Vorbrüggen, J.C., von Seelen, W., Sendhoff, B. (eds.) ICANN 1996. LNCS,
vol. 1112, pp. 305–310. Springer, Heidelberg (1996)
3. Tay, Y.H., Khalid, M., Yusof, R., Viard-Gaudin, C.: Offline Cursive Handwriting Recog-
nition System base on Hybrid Markov Pattern and Neural Networks. In: IEEE International
Symposium on Computational Intelligence in Robotics and Automation, vol. 3, pp.
1190–1195. IEEE Press, New York (2003)
4. Leibe, B., Mikolajczyk, K., Schiele, B.: Segmentation based multi-cue integration for object
detection. In: 17th British Machine Vision Conference, vol. 3, p. 1169. BMVA Press (2006)
5. Spengler, M., Schiele, B.: Towards Robust Multi-Cue Integration fot Visual Tracking. In-
ternational Journal of Machine Vision and Applications 14, 50–58 (2003)
6. Cao, X., Balakrishnan, R.: Evaluation of an On-line Adaptive Gesture Interface with
Command Prediction. In: Graphics Interface Conference, pp. 187–194. Canadian Hu-
man-Computer Communications Society (2005)
7. Teng, L., Jin, L.W.: Hybrid Recognition for One Stroke Style Cursive Handwriting Char-
acters. In: 8th International Conference on Document Analysis and Recognition, pp.
232–236. IEEE Computer Press, Los Alamitos (2005)
8. Koerich, A.L., Leydier, Y., Sabourin, R., Suen, C.Y.: A Hybrid Large Vocabulary Hand-
written Word Recognition System using Neural Networks with Hidden Markov Patterns. In:
8th International Workshop on Frontiers in Handwriting Recognition, p. 99. IEEE Press,
New York (2002)
9. Hall, D.L.: Mathematical Techniques in Multisensor Data Fusion. Artech House, Boston
(2002)
10. Scheunert, U., Lindner, P., Richter, E., Tatschke, T., Schestauber, D., Fuchs, E.: Early and
Multi Level Fusion for Reliable Automotive Safety Systems. In: IEEE Intelligent Vehicles
Symposium, pp. 196–201. IEEE Press, New York (2007)
11. Duan, B., Liu, W., Fu, P., Yang, C., Wen, X., Yuan, H.: Real-Time On-Road Vehicle and
Motorcycle Detection Using a Single Camera. In: IEEE International Conference on In-
dustrial Technology, pp. 1–6. IEEE Press, New York (2009)
 Formalizing the Modeling Process of Physical Systems
in MBD

Nan Wang1,2, Dantong OuYang1,3,*, Shanwu Sun2, and Chengli Zhao1,2

1
Department of Computer Sciences and Technology, Jilin University, Jilin, Changchun 130012
2
Department of Information, Changchun Taxation College, Jilin, Changchun, 130117
3
Key Laboratory of Symbol Computation and Knowledge Engineering of the Ministry of
Education, Jilin, Changchun 130012

Abstract. Many researchers have proposed several theories to capture the es-
sence of abstraction. The G-KRA model(Genera KRA model), based on the
GRA model which offers a framework R to represent the world W where a set of
generic abstraction operators allows abstraction to be automated, can represent
the world from different abstraction granularity. This paper shows how to model
a physical system in model-based diagnosis within the G-KRA model framework
using various kinds of knowledge. It investigates, with the generic theory of ab-
straction, how to automatically generate different knowledge models of the same
system. The present work formalizes the process of constructing an abstract
model of the considered system (e.g., using functional knowledge) based on the
fundamental model and abstract objects database and expects that formalizing
the modeling process of physical systems in MBD within the G-KRA framework
will open the way to explore richer and better founded kinds of abstraction to
apply to the MBD task.

Keywords: G-KRA Model, Knowledge Modeling, Model-Based Diagnosis.

1 Introduction
Abstraction is an essential activity in human perception and reasoning. In the Artificial
Intelligence community, abstraction has been investigated mostly in problem solving
[1], planning [2], diagnosis [3], and problem reformulation [4]. In general, the proposed
approaches differ either in the kind of abstractions they use or in how they formally
represent them. They all fail to characterize the practical aspects of the abstraction
process. Saitta and Zucker [5] have proposed a model of representation change that
includes both syntactic reformulation and abstraction. The model, called KRA
(Knowledge Reformulation and Abstraction), is designed to help both the conceptu-
alization phase of a problem and the automatic application of abstraction operators.
Compared to the KRA model, the G-KRA model(Genera KRA model) [6] is more
general and flexible to represent the world. It can represent the world from different
abstraction granularity.

*
Corresponding author: ouyangdantong@163.com

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 685–695, 2009.
© Springer-Verlag Berlin Heidelberg 2009
686 N. Wang et al.

The major efforts in model-based reasoning have concentrated on paradigms for

qualitative modeling and qualitative reasoning about physical systems [7]. Some ap-
proaches have exploited structural and behavioral knowledge in order to support a
variety of tasks, such as prediction, diagnosis [8,9], and so on. Some researchers fo-
cused on theories of qualitative reasoning based on the use of knowledge about func-
tions and goals [10]. At the same time, the issue of cooperation of multiple models of
the same system has been causing increasing attention in order to improve the effec-
tiveness and efficiency of reasoning processes. These approaches failed to investigate
the relations between proposed methods and general theories of abstraction.
In this paper, we simply introduce G-KRA model which is based on KRA model and
extends it. We show how to model a physical system in model-based diagnosis within
the G-KRA model framework using various kinds of knowledge. We investigate with
the generic theory of abstraction how to automatically generate different knowledge
models of the same system. We expect that formalizing the modeling process of
physical systems in MBD within the G-KRA framework will open the way to explore
richer and better founded kinds of abstraction to apply to the MBD task.

2 The G-KRA Model Applied to Knowledge Modeling in MBD

The modeling process within the G-KRA model can be divided into two correlative
phases: fundamental modeling and abstract modeling. The former one uses the funda-
mental knowledge, i.e. structural and behavioral knowledge, to derive the fundamental
model of the system based on some kind of ontology and representational assumption
which is relatively simple and intuitive. While the abstract modeling process automati-
cally constructs a(some) more abstract model(models) for the same system utilizing other
knowledge such as functional knowledge, teleological knowledge.

2.1 The General KRA Model of Abstraction

In this section we simply introduce the definition of G-KRA abstraction model, see [6]
for details.
Definition 1. A primary perception P is a 5-ple, P = (OBJ,ATT,FUNC,REL,OBS)
where OBJ contains the types of objects considered in W, ATT denotes the types of
attributes of the objects, FUNC specifies a set of functions and REL is a set of relations
among object types.
Definition 2. Let P be a primary perception, A be an agent (who perceives W) and Oa
be a database. The objects with some abstract types in Oa are predefined by A. An
abstract perception of A is defined as P*=δa(P,Oa), where δa denotes an abstract per-
ception mapping.
Definition 3. Given a primary perception P, an abstract objects database Oa and an
abstract perception mapping δa, a general representation framework R* is a 4-ple
(P*,D*,L*,T*), where P*=δa(P,Oa) is an abstract perception, D* is a database, L* repre-
sents a language and T* specifies a theory.
 Formalizing the Modeling Process of Physical Systems in MBD 687

2.2 Fundamental Modeling

The primary perception can be taken as the fundamental model of W. Firstly we are
supposed to define the ontology of the model and identify what to represent of the real
system in the model according to the knowledge involved above. There exists repre-
sentational links and ontological links between the different fundamental models of the
same world [11]. Then we can choose some appropriate kinds of structural and be-
havioral knowledge to build the primary perception P within the framework R and
moreover construct the fundamental model of W as shown in figure 1 where L desig-
nates the union of representational links L1 and ontological links L2.

primary perception P1
P1(W)
W L
primary perception P2
P2(W)

Fig. 1. Fundamental Modeling

Fig. 2. Hydraulic system used as an example

In this section we take the diagnosis task as an example to show how to build up the
fundamental model of a particular physical system. We will consider the hydraulic
system reported in Figure 2, which is the same used in [12].
We here perceive four types of components, i.e., pipes, valves, three-way nodes, and
pumps. Moreover we treat ports that connect components with each other (internal
ports), or connect components to the environment(external ports), as the kind of entities
that the model must designate.
We assume the state of the valves has been given that V1 is open while V2 is closed
without considering the commands set on the valves. The volumetric pump PM1, that
we assume is active, is delivering a constant flow equal to Fk. We choose a qualitative
behavioral representation of the hydraulic system used in the example instead of its
quantitative version.
The possible behaviors of pipes, valves, pumps and three-ways are reported in [12].
＞
For example, pipes have two behavioral modes: OK(Fin=Fout) and LK(Fin Fout).
688 N. Wang et al.

Now we show how to construct the fundamental model of the system in Figure 2.
We use the representation in [13] with a few different details.
∪
-OBJ=COMP {PORT},COMP={PUMP, PIPE, VALVE, THREE-WAY}
-ATT={ObjType: OBJ->{pipe,pump,valve,three-way,port},
Direction: PORT->{in,out},THREE-WAY->{2wayOut,2wayIn},
Observable: PORT->{yes, no}, State : VALVE->{open,closed}}
-FUNC={Bpump : PUMP->{ok,uf,of,lk},Bpipe: PIPE->{ok,lk},
Bvalve: VALVE->{ok,so,sc},Bthree-way: THREE-WAY->{ok}}
-REL’={port-of ⊆ PORT×COMP,connected ⊆ PORT×PORT}
-OBS’={(PM1,P1,…,P6,V1,V2,TW1,TW2,t1,…,t13,t2’,…,t12),(ObjType(PM1)=pump,Ob-
jType(P1)=pipe,…,ObjType(P6)=pipe,ObjType(V1)=valve,…,ObjType(TW1)=three-way,
…,ObjType(t1)=port,…),(Direction(TW1)=2wayOut,…,Direction(t1)=in,…),(Observable(t
1)=yes,…),(State(V1)=open,State(V2)=closed),(port-of(t1,PM1),…),(connected(t2,t2’),…)}

For the sake of space, the contents of the structure/database D, the logical language
L and the theory T, which have been described in [13], will not be provided here.

2.3 Abstract Modeling

Two or more objects perceived in the fundamental modeling process may have the same
nature from some perspective so that we can take them individually as one more abstract
object. It is an automatically matching process, before which the abstract objects data-
base must be constructed manually directed by some kind of subject knowledge and at
the same time it is important to design abstract mapping in order to automatically op-
erate and then generate the more abstract model of the same world appropriate for some
particular reasoning tasks.
In the following section we will formalize the abstract modeling process within the
G-KRA model based on the fundamental model of the same system built up before.
We will give the formal definition of Abstract Objects Database and Abstract Map-
ping and take the functional knowledge for example to automatically construct the
model of functional roles which is based on an object-centered ontology.

2.4 Abstract Objects Database

We will represent the abstract objects using the concept of the framework R=(P,D,L,T)
proposed by Saitta and Zucker. We define P as abstract objects perception, denoted by
AOP, where the observation OBS has not been involved as it is not a particular system
that we perceive but a set of abstract objects. We will also introduce some assumptive
object representation for characterizing the behavior of the abstract objects. The rest of
the framework R will be automatically regulated to AOD, AOL, and AOT to form a
new framework AOR of different meaning with R.
Definition 4. An abstract objects perception is a 4-ple AOP =(A_OBJ,A_ATT,
A_FUNC,A_REL), where A_OBJ contains the types of abstract objects predefined
based not on fundamental knowledge, A_ATT denotes the types of attributes of the
abstract objects, A_FUNC specifies a set of functions, and A_REL is a set of relations
among object types. These sets can be expressed as follows:
 Formalizing the Modeling Process of Physical Systems in MBD 689

-A_OBJ = {A_TYPEi|1 ≤ i ≤ N}
- A_ATT= {A_Aj : A_TYPEj → Λj|1 ≤ j ≤ M}
- A_FUNC= {A_fk : A_TYPEik × A_TYPEjk × ... → Ck|1 ≤ k ≤ S}
- A_REL= {A_rh ⊆A_TYPEih × A_TYPEjh |1 ≤ h ≤ R}
The physical system comprises many individual components which have also been
assigned abstract roles in the functional role model. The functional role of a component
is an interpretation of the equations describing its behavior, aimed at characterizing how
the component contributes to the realization of the physical processes in which it takes
part. Here we introduce three kinds abstract functional objects involved in [14]: conduit,
which enables a generalized flow (effort, energy or information and so on) from one
point to another in the structure of a system; generator, which causes a generalized flow
from one point to another in the structure of a system; barrier, which prevents a gener-
alized flow from one point to another in the structure of a system.
Let AOP=( A_OBJ, A_ATT, A_FUNC, A_REL) be the abstract objects perception
based on functional knowledge specified as follows:

- A_OBJ= A_COMP ∪
A_PORT ∪
… where A_COMP=CONDUIT ∪
∪
GENERATOR BARRIER
- A_ATT={ A_ObjType: A_OBJ→{conduit,generator,barrier,port},
A_Direction: A_PORT→{in,out},A_MeasureType: A_PORT→{flow,pressure}}
- A_FUNC={ A_Fc: CONDUIT→{funcc},A_Fg: GENERATOR→{funcg},
A_Fb: BARRIER→{funcb}}
- A_REL={ A_portof ⊆ A_PORT×A_COMP,
A_connected ⊆ A_PORT×A_PORT }
Going back to the representation framework AOR, we will define three instantiated
abstract objects c(CONDUIT), g(GENERATOR) and b(BARRIER) and introduce
several ports(A_PORT) p1,p2,… to characterize the database AOD which contains the
tables described as follows.
-TableAObj=(A_obj, A_objtype, A_direction, A_measuretype), which describes
abstract components and their attributes.
-TableAFunc=(A_obj, A_func, A_funcmode), which describes the functional
modes of the abstract components.
-TableAPortOf=(A_port, A_comp), which describes what ports is attached to the
abstract components.

In the table TableAObj some of the entries can be set to N/A (not applicable), as not all
attributes are meaningful for all abstract objects. For the sake of simplification, we will
not discuss all of the generalized flow types but assume the measure type of the ports is
flow. We list the contents of the three tables as follows.
The definition of AOL and AOT are similar to L and T in the framework R so that
we can, for example, describe the structure relating a conduit using AOL:
∧ ∧ ∧
conduit(c,p1,p2) ⇔ A_comp(c) A_port(p1) A_portof(p1,c) in(p1) flow(p1) ∧
∧ ∧ ∧ ∧
A_port(p2) A_portof(p2,c) out(p2) flow(p2)
690 N. Wang et al.

Table 1. TableAObj

A_obj A_objtype A_direction A_measuretype

c Conduit N/A N/A
g Generator N/A N/A
b Barrier N/A N/A
p1 Ports in flow
p2 Ports out flow
… … … …

Table 2. TableAFun

A_obj A_func A_funmode

c funcc Pin=Pout
g funcg Pin=Pout=v≠0
b funcb Pin=Pout=0

Also the functional modes described with equations in Table 3 can be rewritten in
logical terms in the theory. For instance: conduit(c,p1,p2) ∧
funcc(c)→ △
FlowValue(p1,p2)=0.

Table 3. TableAPortOf

A_port A_comp A_port A_comp

p1 c p4 g
p2 c p5 b
p3 g p6 b

Structure Matching Procedure, SMP for short

We say the concrete component is matching the abstract one on structure when they
have exactly the same ports, including the number of input ports, the number of output
ports and the measure type of each port. Hence before identifying the corresponding
abstract object in the abstract objects database for some concrete component perceived
in the process of fundamental modeling, we should firstly make structural matching for
them. We can describe the matching procedure as follows.

Behavior Matching Procedure, BMP for short

The normal behavior of the considered component exhibits its function to the whole
system. For instance, when the pipes behave normally they can transport the flow from
one point to another just like the functional role CONDUIT works. So we need only to
see if the normal behavior of the component matches the functional mode of some
given abstract object. We can describe the matching procedure as follows: (we assume
either the behaviors of concrete components or the functional modes of the abstract
objects are represented by qualitative equations.
 Formalizing the Modeling Process of Physical Systems in MBD 691

------------------------------------------------------------------------------------
Procedure SMP COMP c1, A_COMP c2
in1=the set of all in ports of c1;
in2=the set of all in ports of c2;
out1=the set of all out ports of c1;
out2=the set of all out ports of c2;
If (|in1|<>|in2| or |out1|<>|out2|) return FAIL;
If the types of in1 and in2 can’t completely match, return FAIL;
If the types of out1 and out2 can’t completely match, return FAIL;
return SUCCEED;
------------------------------------------------------------------------------------
Fig. 3. Structure Matching Procedure

-------------------------------------------------------------------------------------------------------
Procedure BMP COMP c1, A_COMP c2
IN={in1,in2,…,inm}: the input ports of c1(or c2) individually;
OUT={out1,out2,…,outn}: the output ports of c1(or c2) individually;
E={E1,E2,…,Ej}: the equations representing c1’s normal behaviors;
F={F1,F2,…,Fk}: the equations representing c2’s functional modes;
Put the elements of IN and OUT to the corresponding equation of E and F, and receive the
new sets E’ and F’;
If E’ completely matches F’, return SUCCEED;
return FAIL;
-----------------------------------------------------------------------------------------------------
Fig. 4. Behavior Matching Procedure

-------------------------------------------------------------------------------------------------------
Procedure CRP ObjType type1, State S, A_ObjType type2
//replace the components having type type1 and state S with the abstract object having abstract
type type2 in the considered system

∪ ∪ ∪
IF S is arbitrary {
OBJ’=COMP’ PORT … where COMP’=COMP-type1 type2;
ATT’=ATT-{ATTname: type2→ ∧
j}-{ObjType} ∪
{ A_ObjType : COMP’ →
∧ ObjType-{type1}∪ {type2}}
∪
FUNC’=FUNC-{Btype1: type1→ck} {A_Ftype2: type2→functype2}
REL’=Keep REL unchanged except for replacing COMP in the relation of REL with
A_COMP.
Modify OBS to OBS’ based on OBJ’,ATT’, FUNC’,and REL’;
P’={OBJ’,ATT’, FUNC’, REL’, OBS’};
Modify the contents of D, L, and T accordingly ,based on P’, AOD, AOL,and AOT , to
D’, L’ and T’;
}
-------------------------------------------------------------------------------------------------------
Fig. 5. Components Replacement Procedure

·Components Replacement Procedure, CRP for short

After succeeding the matching in structure and behavior we think of both the consid-
ered component and all other ones of the same type (sometimes with the same states,
692 N. Wang et al.

like valves) are functionally equivalent to the given abstract object. So then we can
replace the components with the matched abstract type object through the procedure
CRP described as follows:
We have only considered in Procedure CRP the situation that the state of compo-
nents can be ignored (like pipes). While dealing with the components whose state needs
to be discussed we must do more works in the procedure. On the one hand, we are not
allowed to delete the object type type1 in the more abstract perception, because not all
components with type type1 stay in state S. Only, the new type type2 is added. On the
other hand, the replacement happens only on the components in state S, so we should
check the state of the considered component before we modify it.
Execute the three procedures repeatedly until there is no component matching any
abstract object of the abstract objects database, and then we can generate a more
abstract model based on functional knowledge with less object types in it. We use the
procedure GAMP (from Generate Abstract Model Procedure) to describe the process.
-------------------------------------------------------------------------------------------------------
Procedure GAMP
Add a flag ENABLE(initially ENABLE=YES) to each component of the considered
system;
//ENABLE=YES represents the component has not been checked or replaced, namely, it can
be dealt with. Whereas if ENABLE=NO, then it indicates either the component has been replaced
or it does not match any abstract object of the given abstract database.
While there exist components unsettled, do{
c=any component whose ENABLE value is YES;
type1=ObjType(c); S=State(c);
For each object oj of the given abstract database, do{
if (SMP(c, oj)=SUCCEED)
if(BMP(c, oj)=SUCCEED) {
type2=A_ObjType(oj);
CRP (type1, State S, type2);
Set ENABLE=NO to all the components of type type1 and state S;
}
Set ENABLE=NO to all the components of type type1 and state S;
}
}
-------------------------------------------------------------------------------------------------------

Fig. 6. Generate Abstract Model Procedure

In the rest of this section we give an example based on the hydraulic system de-
scribed in figure 2 to show how to automatically construct the abstract model from the
fundamental model using functional knowledge.
According to the procedure GAMP, we compare any component c unsettled to each
object of the abstract objects database to find an abstract object completely matching it
in both structure and behavior and then replace all components of the same type as c in
the considered system with the abstract object. We choose the components in such an
order as follows: PM1,P5,TW1,P1,V1,P2,P3,V2,P4,TW2,P6.
 Formalizing the Modeling Process of Physical Systems in MBD 693

·PM1 and conduit

PM1 and conduit can be seen as the same one on structure, namely, SMP(PM1,con-
duit)=SUCCEED, because they both have two ports (one input port and one output port)
and the two ports allow the same measure type. However, when comparing the PM1’s
normal behavior to the conduit’s functional mode, we find the equation Fin=Fout=Fk≠0(the
normal mode of pumps) is not completely equivalent to Pin=Pout, i.e.,
BMP(PM1,conduit)=FAIL, so we say PM1 is not completely matching conduit.
Here we stress two completely matching components to insure to avoid choosing the
inappropriate abstract object to do the replacing. For example, an active pump requires
the input flow value be equal to the output flow value and they both should be a nonzero
number which means an active pump delivers a constant flow equal to FK (FK≠0).
While the function of a conduit of the abstract type CONDUIT is to enable a general-
ized flow from one point to another, namely, the output flow value should be
equal to the input flow value which could be zero. So obviously matching the PUMP
type component with the CONDUIT type object, we say, is incomplete that will lead to
an incorrect abstract model based on functional knowledge.
·PM1 and generator
Similar to the process of comparing PM1 to conduit, we invoke SMP(PM1,generator)
and BMP(PM1,generator) individually and they both receive SUCCEED which means
PM1 completely matches the abstract object generator. Then we can replace all the
active pumps of the system with generators.

Fig. 7. Replace all the active pumps of the system with generators

Fig. 8. Replace all pipes with conduits

Fig. 9. Replace all open valves with conduits and closed valves with barriers
694 N. Wang et al.

-P’=(OBJ’,ATT’,FUNC’,REL’,OBS’) is presented as follows:

-OBJ’=COMP’ ∪ {PORT},COMP’={GENERATOR, PIPE, VALVE,
THREE-WAY}
-ATT’={ObjType: OBJ->{pipe,generator,valve,three-way,port},
Direction: PORT->{in,out},THREE-WAY->{2wayOut,2wayIn},
，
Observable: PORT->{yes, no} State : VALVE->{open,closed}}
-FUNC’={ A_Fg: GENERATOR→{funcg},
Bpipe: PIPE->{ok,lk},Bvalve: VALVE->{ok,so,sc},
Bthree-way: THREE-WAY->{ok}}
-REL’={port-of ⊆ PORT×COMP’,connected ⊆ PORT×PORT}
-OBS’={(PM1,P1,…,P6,V1,V2,TW1,TW2,t1,…,t13,t2’,…,t12),(ObjType(PM1)=gener
ator,ObjType(P1)=pipe,…,ObjType(P6)=pipe,ObjType(V1)=valve,…,ObjType(TW1)
=three-way,…,ObjType(t1)=port,…),(Direction(TW1)=2wayOut,…,Direction(t1)=in,
…),(Observable(t1)=yes,…),(State(V1)=open,State(V2)=closed),(port-of(t1,PM1),…),
(connected(t2,t2’),…)}
We adopt the representation to the three functional roles of [14] to show the trans-
formed system described as Figure 8.
～
The next replaced components are pipes(P1 P6), namely replacing all pipes with
the abstract objects of the type CONDUIT, described in Figure 9.
TW1 and TW2 can’t match any object of the abstract objects database, so they are not
replaced and there exists still the type of three-ways.
The last replaced components are valves(V1 staying open and V2 closed). V1 is re-
placed with an abstract object with type of CONDUIT while V2 is replaced with an
abstract object with type of BARRIER described in Figure 10.

3 Conclusion
Many researchers have explored how to use relevant knowledge to construct the
models of physical systems in MBD, for instance, structural and behavioral knowl-
edge(e.g.,[9]), functional and teleological knowledge(e.g.,[14]), multimodeling ap-
proach(e.g.,[11]). In recent years, new approaches have been proposed in MBD to
exploit the degree of observability for defining useful abstractions (e.g., [12],[15]).
While these approaches have developed some interesting solutions to the problem of
abstracting models for MBD, they have failed to investigate the relations between
proposed methods and general theories of abstraction. The present work represents a
step in the direction of filling this gap. In particular, the present paper introduces the
extended KRA model, i.e., G-KRA model, which can represent the world more general
and flexible. We have shown how to automatically construct an abstract model based
on the fundamental model which contributes to formalizing the process of transforming
models with fundamental knowledge to other models with abstract knowledge, such as
structural and behavioral models to functional role models described above. There still
exist some problems we will explore in future, e.g., how to deal with the situation that a
component has multiple roles about different domains to automatically build up mul-
tiabstract models; how to formalizing the realization of the links of two fundamental
models with different ontologies or representation assumption, and so on.
 Formalizing the Modeling Process of Physical Systems in MBD 695

Acknowledgment
The authors are grateful to the support of NSFC Major Research Program under Grant
Nos. 60496320 and 60496321, Basic Theory and Core Techniques of Non Canonical
Knowledge; NSFC under Grant Nos. 60773097 and 60873148, Program for New
Century Excellent Talents in University; Jilin Provine Science and Technology De-
velopment Plan under Grant Nos. 20060532 and 20080107, and European Commission
under Grant No. TH/Asia Link/010 (111084).

References
1. Holte, R., Mkadmi, T., Zimmer, R., MacDonald, A.: Speeding up problem-solving by ab-
straction: A graph-oriented approach. J. Art. Intelligence 85, 321–361 (1996)
2. Knoblock, C., Tenenberg, J., Qiang, Y.: A spectrum of abstraction hierarchies for planning.
In: Proc. AAAI WS on AGAA, pp. 24–35 (1990)
3. Mozetic, I.: Hierarchical model-based diagnosis. J. Int. Journal of Man-Machine Stud-
ies 35(3), 329–362 (1991)
4. Subramanian, D.: Automation of abstractions and approximations: Some challenges. In:
Proc. AAAI WS on AGAA, pp. 76–77 (1990)
5. Saitta, L., Zucker, J.: Semantic abstraction for concept representation and learning. In: Proc.
SARA, pp. 103–120 (1998)
6. Shan-wu, S., Nan, W., Dan-tong, O.Y.: General KRA Abstraction Model. J. Journal of Jilin
University (Science Edition) 47(3), 537–542 (2009)
7. Weld, D., De Kleer, J.: Readings in Qualitative Reasoning about Physical Systems. Morgan
Kaufmann, San Mateo (1990)
8. Bobrow, D.G. (ed.): Special Volume on Qualitative Reasoning about Physical Systems. J.
Artificial Intell. 24 (1984)
9. Davis, R.: Diagnostic reasoning based on structure and behavior. J. Artificial Intelli-
gence 24, 347–410 (1984)
10. Sticklen, J., Bond, E.: Functional reasoning and functional modeling. IEEE Expert 6(2),
20–21 (1991)
11. Chittaro, L., Guida, G., Tasso, C., Toppano, E.: Functional and teleological knowledge in
the multimodeling approach for reasoning about physical system:a case study in diagnosis.
IEEE Trans. Syst. Man, Cybern. 23(6), 1718–1751 (1993)
12. Chittaro, L., Ranon, R.: Hierarchical model-based diagnosis based on structural abstraction.
Art. Intell. 155(1-2), 147–182 (2004)
13. Saitta, L., Torasso, P., Torta, G.: Formalizing the abstraction process in model-based di-
agnosis. In: Tr cs, vol. 34, Univ. of Torino, Italy (2006)
14. Chittaro, L., Ranon, R.: Diagnosis of multiple faults with flow-based functional models:the
functional diagnosis with efforts and flows approach. Reliability Engineering and System
Safety 64, 137–150 (1999)
15. Torta, G., Torasso, P.: A Symbolic Approach for Component Abstraction in Model-Based
Diagnosis. In: Proceedings of the Model-Based Diagnosis International Workshop (2008)
 Study on Stochastic Programming Methods Based on
Synthesizing Effect

FaChao Li1,2, XianLei Liu2, and ChenXia Jin1,2

1
School of Economics and Management, Hebei University of Science and Technology,
050018, Shijiazhuang, China
2
School of Science, Hebei University of Science and Technology,
050018, Shijiazhuang, China
lifachao@tsinghua.org.cn, liuwoolong@163.com,
jinchenxia2005@126.com

Abstract. Stochastic programming is a well-known optimization problem in re-

source allocation, optimization decision etc. in this paper, by analyzing the essen-
tial characteristic of stochastic programming and the deficiencies of the existing
methods, we propose the concept of synthesizing effect function for processing
the objective function and constraints, and further we give an axiomatic system
for synthesizing effect function. Finally, we establish a general solution model
(denoted by BSE-SGM for short) based on synthesizing effect function for sto-
chastic programming problem, and analyze the model through an example. All the
results indicate that our method not only includes the existing methods for sto-
chastic programming, but also effectively merge the decision preferences into the
solution, so it can be widely used in many fields such as complicated system op-
timization and artificial intelligence etc.

Keywords: Stochastic programming, Stochastic decision-making, Stochastic

effect, Synthesizing effect function, Mathematical expectation, Variance,
Model.

1 Introduction
Randomness is a widespread phenomenon in the real world and is unavoidable in
many practical fields. How to establish an effective and workable method to process
random information is a widespread concern on production management, artificial
intelligence, optimization of complex systems research, and the stochastic programming
is generally involved in these research areas. As a random variable is a family of data
that satisfy some laws and dose not have clear order, the current programming methods
should not be directly applied to solving the problem of stochastic programming. At
present, there are three basic methods to solve the stochastic programming problem:
1) Expectation model, the basic idea is to use mathematical expectation to describe
random variables, then turn the stochastic programming into the general programming
problem. 2) Chance-constrained model, the basic idea is to convert stochastic constraints
and objective functions into ordinary constraints and objective functions through some
reliability principles. 3) Dependent-chance programming model, the basic idea is to

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 696–704, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 Study on Stochastic Programming Methods Based on Synthesizing Effect 697

regard objective functions and constraints as events under random environment and
solve the stochastic programming problem by maximizing the chances of all the
events happening. These methods have achieved much good applications, such as, in
[3], the authors studied measures program of oil field by using expectation model, and
paper [4] use chance-constrained model to size batteries for distributed power system,
then paper [5] use it to optimize allocation of harmonic filters on a distribution network.
But they could not solve the stochastic programming problem under complicated
environment effectively, and the deficiencies mostly are: 1) it is hard to use expectations
to describe and represent random variables effectively and it is difficult to ensure the
reliability of model when the randomness is larger by expectation model; 2) when the
stochastic characteristics are complex (that is to say, it is difficult to determine the
distribution of the stochastic environment), the computational complexities of the
chance-constrained model and dependent-chance programming model are so large to
get an analytic form of the program viable solution In response to these problems,
many scholars from different perspectives studied stochastic programming problem,
for example, papers[6-9] studied to solve chance-constrained model and dependent-
chance programming model by using the stochastic simulation technique and genetic
algorithm. But all of these methods have strongly points. Thus, so far, there are not
systematic and effective stochastic programming methods.
Based on the analyses above, this paper analyses the basic characteristics of stochastic
programming problems, combines with the inadequacy of existing methods, and have
the following contributions: a) we propose the concept of synthesizing effect function
for processing the objective function and constraints, and further we give some com-
monly-used models of synthesizing effect functions; b) we establish a general solution
model (denoted by BSE-SGM for short) based on synthesizing effect function for sto-
chastic programming problem; c) We analyze the characteristic of our model by an
example, and the results indicate that our methods are effective.
In this paper, let (Ω, B , P) be the probability space, and for every random variable
ξ in the (Ω, B , P), let E (ξ ) and D (ξ ) be the mathematical expectation and variation of
the random variable ξ respectively.

2 Overview of the Methods to Solve the Stochastic Programming

Stochastic programming problem is the core content in the production management,
resource allocation and other practical problems, its general form is:

⎧max f ( x, ξ ),
⎨ (1)
⎩s.t. g i ( x, ξ ) ≤ 0, i = 1, 2, L, m.
Where x = ( x1 , x2 ,L , xn ) is the decision vector, and ξ = (ξ1 , ξ 2 , L , ξ n ) is the given
random variable vector in the probability space (Ω, B , P ) , and f ( x, ξ ) , g j ( x, ξ ) ,
are random variable functions, j = 1, 2, L , m .
As there is no simple order between random variables and g j ( x, ξ ) ≤ 0 mostly
could not be completely satisfied, model (1) is just a model and can not be solved
698 F. Li, X. Liu, and C. Jin

directly. So, it is an essential way to solving stochastic programming problems that we

can convert the stochastic programming problems to normal programming problems
in some strategies.

2.1 Expectation Model

Mathematical expectation is a commonly used tool to describe the values of random

variable intently. It has some good theory characters. If we use the mathematical ex-
pectation of the random variable to take place of the random variable, the model (1)
becomes the following model (2):

⎧max E ( f ( x, ξ )),
⎨ . (2)
⎩s.t. E ( g j ( x, ξ )) ≤ 0, j = 1, 2, L , m.
Generally, we call the model (2) the Expectation model [3].
When the variation of the random variable is larger, the mathematical expectation
could not describe the variable effectively. So we could not get the optimum solution
of the stochastic programming by using the model (2).

2.2 Chance-Constrained Mode [1]

For the problem that the constraint of the stochastic programming often can not satisfy
absolutely, we can use the reliability to dispose of the constraints and objective functions
of the programming. Then the model (1) becomes the following model (3):
⎧max f ( x ),
⎪ .
⎨s.t. P( f ( x, ξ ) ≥ f ( x)) ≥ α , (3)
⎪ P( g j ( x, ξ ) ≤ 0) ≥ α j , j = 1, 2, L , m.
⎩
Generally, we call the model (3) the chance-constrained model. In the model,
α j , α ∈ [0, 1] are the reliabilities that the solutions satisfy constraints and objective
functions, and P( A) is the probability that the event A may happen. Compared to the
model (2), this model could control the quality of the decision, but we don’t know the
range of α j ’s value when the model (3) could be soluble.
As stochastic programming is an uncertain decision-making problem, the results of
its decision-making in general should not be bound to make the relevant constraint
established absolutely. So compared to bounding by the given reliability of constraints
and objectives in advance, it is more suitable for the characters of integrated decision-
making that considering the constraint satisfaction and the size characteristics of ob-
jective function. To establish a general solution model under this concept, we can
synthesize the objective function value and constraint satisfaction together through
some kind of strategy (called the synthesizing strategy the synthesizing effect func-
tion), then discuss the programming based on the synthesized value. Further we give
the following random effects multi-attribute axiomatic system of synthesizing effect
functions.
 Study on Stochastic Programming Methods Based on Synthesizing Effect 699

3 The Axiomatic System for Synthesizing Effect Function

As the minimum decision-making problems can be through some ways transfor-
mation into the maximum decision-making problems, in the following, we mainly
discuss the synthesizing strategy of objective function and the satisfaction of con-
straints on the maximization stochastic programming problem with single objective
(real function) and multi-constraints. According to the essential characteristic of the
optimal decision, we should obey the following rules in seeking for the decision
scheme of stochastic programming.
Principle 1. When the satisfaction degree of constraints is same, the greater the objec-
tive value function is, the better the effect is;
Principle 2. When the objective function value is same, the greater the satisfaction
degree of constraints is, the better the effect is;
Principle 3. When the constraints are absolutely satisfied, the decision only depends
on the objective value function;
Principle 4. When some constraints are absolutely dissatisfied, we can’t make a
decision.
If we abstractly regard the synthesizing strategy of objective function value and satis-
faction degree of constraints as a function S (u , v) = S (u, v1 , v 2 , L , v m ) (here, u is the
objective function value with the conversion interval Θ ; vi is the satisfaction degree
of the i-th constraint, with the conversion interval [0, 1], that is, S (u , v) is a map
on Θ × [0, 1]m → (−∞, + ∞) ), then the above principles can be accordingly stated as
follows:

Condition 1. For any given u ∈ Θ , S (u , v) is monotone non-decreasing on each vi ;

Condition 2. For any given v = (v1 , v2 , L, vn ) ∈ [0, 1]m , S (u , v) is monotone non-

decreasing on each u ;

Condition 3. S (u , 1, 1,L, 1) is strictly monotone increasing;

Condition 4. When ∏ nj=1 v j = 0 , S (u1 , v) = S (u2 , v) for any u1 , u 2 ∈ Θ . Obviously,

for uncertain decision problem, principles 1~3 are must be obeyed, while principle 4
can be loosed. For convenience, we call S (u , v) satisfying conditions 1~3 synthesizing
effect function on Θ , and S (u , v) satisfying conditions 1~4 uniform synthesizing
effect function on Θ .

Remark 1. Since the above four principles also must be obeyed for multi-objective deci-
sion, we can similarly establish the axiomatic system for synthesizing effect function
for multi-objective programming, and for the above symbolic system, we only con-
sider u as (u1 , u 2 ,L, u m ) , change that S (u , v) is monotone non-decreasing on u into
that S (u , v) is monotone non-decreasing on each u i , these changes will not have essen-
tial effect for the results.
700 F. Li, X. Liu, and C. Jin

According to the above definition, we have the following conclusions:

① For any a, b ∈ (−∞, + ∞), a < b , S (u, v) = (u − a)(b − a) ∧ v1 ∧ v2 ∧ L ∧ vn is a uni-
−1

form synthesizing effect function on [a, b] . Here, ∧ is min operation of real numbers.
② For any k ∈ (0, + ∞ ), c ∈ [0, + ∞ ), α j ∈ (0, + ∞ ) , S (u , v) = k (u + c)Π jn=1v j j is a
α

uniform synthesizing effect function on [0, + ∞) .

③ For any α ∈ (0, ∞), β j
β
∈ [1, ∞) , S (u , v) = u α Π jn=1v j j is a uniform synthesizing effect
β
function on [0, + ∞) ; S (u , v) = exp(α u )Π jn=1v j j is a uniform synthesizing effect func-
tion on (−∞, + ∞ ) .
④ For any α ∈ (0, ∞), λ ∈ (0, 1], j = 1, 2,L, n , S (u, v) = uΠ
j
n
j =1 δ (v j − λ j ) is a uniform
synthesizing effect function on [0, + ∞) ; S (u , v) = exp(αu ) × Π jn=1δ (v j − λ j ) is a uni-
form synthesizing effect function on (−∞, + ∞ ) ; S (u, v) = α u + ∑ jn=1 η (v j − λ j ) is a
uniform synthesizing effect function on (−∞, + ∞ ) . Here, δ (t ) = 0 for each t < 0 ,
δ (t ) = 1 for each t ≥ 0 , and η (t ) = −∞ for each t < 0 , η (t ) = 1 for each t ≥ 0 .
⑤ For any a, b, a ∈ (0, + ∞), j = 1, 2,L, n , S (u, v) = bu + a ∑
j
n
j =1
a j v j is a synthesizing
effect function on (−∞, + ∞) , but not a uniform synthesizing effect function.

4 The Stochastic Programming Model Based on Synthesizing

Effect
Using the synthesizing strategy in section 3, for single-objective (real value function)
and multi-attribute stochastic programming problem(1), if we regard u and v j in
S (u , v1 , v 2 , L , v n ) as the concentrated quantizing value C ( f ( x, ξ )) of the objective
function value f ( x, ξ ) and the satisfaction degree β j ( x) = Sat ( g j ( x, ξ ) ≤ 0) of the i th
constraint for the scheme x , then S (C ( f ( x, ξ )) , β1 , β 2 , L, β m ) considering both
objective and constraint, is a quantitative descriptive model measuring the quality of
the solution, and the model (1) can be converted into the following model (4):

⎧max S (C ( f ( x, ξ )), β 1 ( x), β 2 ( x ), L , β m ( x)),

⎨ . (4)
⎩s.t. x ∈ X .
Here, (inf f ( x), sup f ( x)) ⊂ Θ , S (u, v1 , v 2 , L , v n ) is a (uniform) synthesizing effect
function on Θ . For convenience, we call (4) the stochastic programming model based
on synthesizing effect, denoted it by BSE-SGM for short.
Remark 2. If we use E ( f ( x, ξ )) to describe the size characteristics of f ( x, ξ ) in
concentrated form, and use β j ( x) = δ ( E (− g j ( x, ξ ))) to represent the satisfaction of
g j ( x,ξ ) ≤ 0 , when S (u, v) = exp(u ) ⋅η (∏ mj=1 δ (v j − α j )) ,the model (4) is the expectation
 Study on Stochastic Programming Methods Based on Synthesizing Effect 701

model (2). Here, δ (t ) satisfies: δ (t ) = 0 for each t < 0 , δ (t ) = 1 for each t ≥ 0 ; and
η (t ) satisfies η (0) = −∞, η (1) = 1 .

Remark 3. If we use f (x) to describe the size characteristics of f ( x, ξ ) in concen-

trated form, and use β j ( x) = P ( g j ( x, ξ )) ≤ 0) to represent the satisfaction of
g j (x, ξ) ≤ 0 , when S (u, v) = exp(u ) ⋅η (∏ mj=1 δ (v j − α j )) The model(4) is the chance-
constrained model(3). Here, f (x) satisfies P( f ( x, ξ )) ≥ f ( x)) ≥ α , and δ (t ) satis-
fies: when t < 0 , δ (t ) = 0 , and when t ≥ 0 , δ (t ) = 1 ; η (t ) satisfies that when
η (0) = −∞ , η (1) = 1 .
Remark 4. If we use ( E ( f ( x, ξ )), D ( f ( x, ξ ))) to describe the size characteristic of
f ( x, ξ ) in a compound quantification mode, and use T ( E ( f ( x, ξ )), D( f ( x, ξ ))) to
represent C ( f ( x, ξ )) (Here, T ( E ( f ( x, ξ )), D( f ( x, ξ ))) is one synthesis of
E ( f ( x, ξ )) and D( f ( x, ξ )) . Obviously, T ( x, y ) should satisfy that: 1) T ( s, t ) mono-
tone non-decreasing on each s; 2) T ( s, t ) monotone non-increasing on each t), then
when S (u, v) = exp(u ) ⋅η (∏ mj=1 δ (v j − α j )) , the model (4) is the following model (5)
(Here, δ (t ) satisfies that δ (t ) = 0 for each t < 0 , and δ (t ) = 1 for each t ≥ 0 ;
η (t ) satisfies that η (0) = −∞ , η (1) = 1 ):

⎧max T ( E ( f ( x, ξ )), D( f ( x, ξ ))),

⎨ . (5)
⎩s.t. P( g j ( x, ξ ) ≤ 0) ≥ a j , j = 1, 2, L , m.

Remark 5. The above analysis indicate that, model (4) includes the existing stochastic
programming model, and also it has better structural characteristic and strong inter-
pretation, therefore model (4) provides a theoretical platform for solving stochastic
programming problem. For different problems, we can use different (uniform) synthe-
sizing effect functions to embody and describe different decision consciousness.

5 Example Analysis
In this part, we will further analyze the characteristic of stochastic programming model
(4) by an example.
Example 1. Consider the following programming problem.

⎧max z = −2ξ1 x12 − ξ 2 x 2 − ξ 3 x3 − ξ 4 sin x 4 ,

⎪⎪
⎨s.t. ξ 1 x1 + ξ 2 x 2 + ξ 3 cos x3 ≤ 5,
2
. (6)
⎪
⎪⎩ x 1 2 3 4 ≤
x x x 4 .
702 F. Li, X. Liu, and C. Jin

Here, ξ1 is a random variable with the uniform distribution on interval [2, 12]; ξ 2 is a
random variable with the normal distribution N(5, 10) ; ξ 3 is a random variable with
the normal distribution N (8, 20) ; ξ 4 is a random variable with the exponential distri-
bution that its parameters is 0.5, and ξ1 , ξ 2 , ξ3 , ξ 4 are independent with each other.
In order to facilitate the analysis of the characteristics of BSE-SGM, we only solve
the problem (6) by using the expectation model (2) and the BSE-SGM as follows.
I. By using the model (2), the problem (6) could be transformed into the general pro-
gramming problem (7):

⎧max z ∗ = −14 x12 − 5 x 2 − 8 x3 − 2 sin x 4 ,

⎪
⎨s.t. 7 x1 + 5 x2 + 8 cos x3 − 5 ≤ 0, .
2
(7)
⎪ x1 x 2 x3 x4 ≤ 4.
⎩
II. By using BSE-SGM to solve the problem (6): its parameters settings are: 1) use
( E( z ′), D( z ′)) to describe the size characteristic of z ′ in a compound quantification
mode, and use C ( z ′) = E ( z ′)[1 + k ( D( z ′)α ]−1 to represent the concentrated quantifying
value of z ′ ; 2) use S (u, v) = uv b to the stochastic synthesizing effect function (that
is S (C( z′), β ) = E(z′)[1 + k(D(z′)α ]−1 β b ), then problem (6) could be transformed into
the following general programming problem (8):

⎧ ∗ (−14 x12 − 5 x 2 − 8 x3 − 2 sin x 4 ) β b

⎪max z = ,
⎪ 1 + k (100 x14 + 10 x22 + 20 x32 + 4 sin 2 x4 )α
⎪⎪ 3
⎨ s.t. β = P (ξ x
1 1
2
+ ξ 2 x 2 + ξ 3 cos x3 ≤ 5), . (8)
⎪ x1 , x 2 , x3 , x4 ≤ 4.
⎪
⎪
⎩⎪

Obviously, (7) and (8) are nonlinear programming problems and they could not be
solved easily by an analytical way. Genetic Algorithm is used here (its parameters set
as follows: Code mode: binary; Mutation probability: 0.001; Crossover probability: 1;
Population size: 80; Evolutionary generations: 100). Then the results of the problem
under different methods are shown in Table 1 (Here, S.E.V. denotes the Synthesizing
Effect Value).
The above analysis and computation results indicate: 1) The variations of the de-
cision results for the same stochastic programming problem by using BSE-SGM are
smaller than the one by using expectation model, which shows that BSE-SGM is
much closer with the essence of the decision than expectation model; 2) For the
different synthesizing effect functions, the decision results are different, and even
the difference is great. Therefore, BSE-SGM can effectively merge uncertainty
 Study on Stochastic Programming Methods Based on Synthesizing Effect 703

Table 1. The results of solving the problem (6) under different methods

Solving Model Optimal Solution S.E.V. Expectation Variation Reliability

(-0.0098, -10.000, -
Expectation Model —— 131.9986 3004.0 0.7355
10.000, 4.7214)
B k=0.1,α=0.5, (-0.0098, -9.218, -
S 14.1813 114.1707 2224.1. 0.7100
E b=1 8.2796, -7.4585)
-
S k=0.1,α=0.5, (- 0.0098, -9.6676, - 11.6090 126.2389 2754.1 0.7580
G
b=2 9.5308, 5.3079)
M
k=0.1,α=0.3,b (- 0.0489, -9.9218, -
36.5984 129.4621 2887.7 0.7690
=2 9.7458, 4.4282)

k=0.1,α=0.5,b (0.0098, -0.2444, -

11.9354 67.3736 1327.7 0.9370
=3 8.1427, -6.8133)
k=0.1,α=0.3,b (-0.0098, -0.1466, -
35.6801 66.0624 1260.8 0.9970
=3 7.9277, 5.0147)
k=0.1,α=0.5,b (-0.0293, -9.9609, -
54.6945 131.7418 2996.0 0.7630
=5 10.0000, 4.4868)

process consciousness into decision process; 3) BSE-SGM has good structural

characteristics.

6 Conclusion
In this paper, for the solution of multi-attribute stochastic programming, by analyzing
the deficiencies of the existing methods, we propose the concept of multi-attribute
synthesizing effect function, give an axiomatic system for multi-attribute synthesizing
effect function, and establish a general solution model for stochastic programming
problem; we further analyze the characteristic of our model by an example. All the
results indicate that the multi-attribute synthesizing effect function is an effective tool
for processing decision preference, it can merge the processing thought of stochastic
information into the quantitative operation process, and it has theoretical systematiza-
tion and operational application.

Acknowledgements
This paper is supported by National Natural Science Foundation of China (70671034,
70871036, 70810107018) and the Natural Science Foundation of Hebei Province
(F2006000346) and the Ph. D. Foundation of Hebei Province (05547004D-2,
B2004509).
704 F. Li, X. Liu, and C. Jin

References
1. Charnes, A., Cooper, W.W.: Management Models and Industrial Applications of Linear
Programming. Johhn Wiley & Sons Inc., New York (1961)
2. Liu, B.: Dependent-chance programming: A class of stochastic programming. Computers
& Mathematics with Applications 34(12), 89–104 (1997)
3. Song, J.K., Zhang, Z.X., Zhang, Y.: A Stochastic expected value model for measures pro-
gram of oil field. Journal of Shandong University of Technology (Sci & Tech) 20(3), 9–12
(2006)
4. Sun, Y.J., Kang, L.Y., Shi, W.X., et al.: tudy on sizing of batteries for distributed power
system utilizing chance constrained programming. Journal of System Simulation 17(1),
41–44 (2005)
5. Zhao, Y., Deng, H.Y., Li, J.H., et al.: Chance-constrained programming based on optimal
allocation of harmonic filters on a distribution network. Proceedings of the CSEE 21(1),
12–17 (2001)
6. Iwamura, K., Liu, B.: A genetic algorithm for chance constrained programming. Journal of
Information & Optimization Sciences 17(2), 40–47 (1996)
7. Zhao, R., Iwamura, K., Liu, B.: Chance constrained interger programming and stochastic
simulation based genetic algorithms. Journal of Systems Science and Systems Engineer-
ing 7(1), 96–102 (1998)
8. Chen, J.M.: Solving order problems with genetic algorithms based on stochastic simula-
tion. J. Chongqing Technol. Business Univ (Nat. Sci. Ed.) 22(2), 179–181 (2005)
9. Liu, B.: Dependent-chance goal programming and its genetic algorithm based approach.
Mathematical and Computer Modelling 24(7), 43–52 (1996)
10. Holland, J.H.: Genetic algorithms and the optimal allocations of trials. SIAM J. of Com-
puting 2, 8–105 (1973)
 Rate Control, Routing Algorithm and Scheduling for
Multicast with Network Coding in Ad Hoc Networks

Xu Na Miao* and Xian Wei Zhou

Department of Communication Engineering, School of Information Engineering,

University of Science and Technology Beijing, Beijing, 100083, P.R. China
miaoxuna1227@126.com

Abstract. In this paper, we developed a distributed rate control and routing algo-
rithm for multicast session in ad hoc networks. We studied the case with dynamic
arrivals and departures of the users. With random network coding, the algorithm
can be implemented, and work at transport layer to adjust source rates and at net-
work layer to carry out network coding in a distributed manner. The scheduling
element of our algorithm is a dynamic scheduling policy. Numerical examples are
provided to complement our theoretical analysis. Modeling and solution algorithm
can be easily tuned according to a specific networking technology.

Keywords: Cross-layer design, Rate control, Network coding, Stochastic flow,

Multicast, Routing, Ad hoc networks.

1 Introduction
As a network layer problem, routing involves simply replicating and forwarding the
received packets by intermediate nodes in multi-hop networks. Network coding ex-
tends routing by allowing intermediate nodes to combine the information received from
multiple links in the subsequent transmissions and enables wired network connections
with rates that are higher than those achieved by routing only [1]. Subsequently, im-
portant progress has been made regarding the low-complexity construction of network
codes. Li et al. [2] showed that the maximum multicast capacity can be achieved by
performing linear network coding. Ho et al. [3], Jaggi et al. [4] and Sanders et al. [5]
showed that random linear network coding over a sufficiently large finite field can
(asymptotically) achieve the multicast capacity. Following these constructive theo-
retical results about network coding, Chou et al. [6] proposed a practical scheme for
performing network coding in real packet networks. Network coding has been extended
to wireless environments with distributed implementation ([6], [7], [8]).
In order to achieve high end-to-end throughout and efficient resource utilization, rate
control, routing and scheduling need to be jointly designed in ad hoc networks.
Cross-layer design is becoming increasingly important for improving the performance
of multihop wireless networks ([11], [12], [13], [14], [15]).
In this paper, we consider the problem of rate control and resource allocation
(through routing and scheduling) for multicast with network coding over a multi-hop
*
Corresponding author.

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 705–714, 2009.
© Springer-Verlag Berlin Heidelberg 2009
706 X.N. Miao and X.W. Zhou

wireless ad hoc network. Rate control is a key functionality in modern communication

networks to avoid congestion and to ensure fairness among the users ([9],[10]). Al-
though rate control has been studied extensively for wireline networks where links are
reliable and link capacities are usually well provisioned. However, these results cannot
be applied directly to multihop wireless networks where link is a shared medium and
interference-limited. Unlike in the wireline network where flows complete for trans-
mission resources only when they share the same link, here, network layer flows that do
not even share a wireless link in their paths can compete. Thus, in ad hoc wireless
networks the contention relations between link-layer flows provide fundamental con-
straints for resource allocation.

2 Related Work
There are several works on rate control of cross-layer design over wireless ad hoc
networks and works on rate control of multicast flows. In this section, we will only
briefly discuss several references that are directly relevant to this paper. To the best of
our knowledge, this paper is the first effort to study cross-layer design for network
coding based multicasting in ad hoc networks.
The work in [13], [15], [16], [17] provides a utility–based optimization framework
to study rate control over ad hoc networks, the authors study joint rate control and
media access control for ad hoc wireless network, and formulate rate allocation as a
utility maximization problem with the constraints that arise from contention for channel
access. This paper studies the case with dynamic arrivals and departures of the users,
and extends this work to include routing and to study cross-layer design for network
coding based multicast.
With network coding, the work that are most similar to our work are [18], [19], [20].
What differentiates our work from others are the following: First, we extend this model
to ad hoc wireless networks. Second, our rate control algorithm is a dual subgradient
algorithm whose dual variables admit concrete and meaningful interpretation as con-
gestion prices. Third, the session scheduling in our cross-layer design is a dynamic
scheduling over ad hoc networks.
The rest of the paper is structured as follows. The system model and problem for-
mulation is presented in Section 3. We present a fully distributed cross-layer rate con-
trol algorithm in Section 4.In section 5, we describe a scheduling policy over ad hoc
networks, and the simulation results are presented in Section 6, and the conclusion is
given in Section 7.

3 The System Model

3.1 System Model

We consider a set N of mobile nodes in an ad hoc network, labeled 1, 2, " N . We say

that link l = ij joining node i to node j exists if node i can successfully transmit to
G
node j . The set of links so formed is denoted by L . Let P = [ Pl , l ∈ L ] denote the
 Rate Control, Routing Algorithm and Scheduling Multicast with Network Coding 707

G
vector of global power assignments and let c = [ cl , l ∈ L ] denote the vector of data
G
G
( )
rates. We assume that c = u P the data rates are completely determined by the global

power assignment. The function u (.) is called the rate-power function of the system.
G
Note that the global power assignment P and the rate-power function u (.) summarize
the cross-layer control capability of the network at both the physical layer and the MAC
layer. Precisely, the global power assignment determines the Signal- to- Interference-
Ratio (SIR) at each link. Let M denote the number of multicast sessions. Each session
has one source s
m
∈ N and a set Dm ⊂ N of destinations. Network coding allows
flows for different destinations of a multicast session to share network capacity by
being coded together: for a single multicast session m of rate x m ，information must
m
flow at rate x to each destination. The arrival process of any multicast session m
according to a Poisson process with rate λ m and that each node brings with it a file for
transfer whose size is exponentially distributed with mean 1/ µm . Thus, the traffic in-
tensity brought by sessions of class m is ρm = λm / µm . Consider a graph
G = (V , E ) with capacity set c and a collection of multicast ses-
sions S m = ( s m , d , x m ) , m = 1, " M , d ∈ Dm as the end-to-end traffic demands.
Let nm ( t ) ( m = 1, 2,", M ) denote the number of multicast sessions that are pre-
sent in the system, In the rate assignment model that follows, the evolution of
nm ( t ) will be governed by a Markov process. Its transition rates are given by:
nm ( t ) → nm ( t ) + 1, with rate λm , and
nm ( t ) → nm ( t ) − 1 , with rate µm xm ( t ) nm ( t ) , if nm ( t ) > 0 .
Let p ( t ) denote congestion cost at link l in the network at time t . As in [17], we say
m
l

that the network is stable if

1 t
t ∫0 ⎧⎪⎨⎛⎜⎜ ∑ nm (t )+ ∑ plm (t ) ⎞⎟⎟ > H ⎫⎪⎬
lim sup 1 M L
dt → 0 ,
t →∞
⎪⎩⎝ m=1 l =1 ⎠ ⎪⎭

as H → ∞ . This means that the fraction of time that the amount of “unfinished work” in
the system exceeds a certain level H can be made arbitrary small as H → ∞ . In other
words, the number of sessions at each source node and the queues at each link must be
finite. The capacity region of the network is defined to be the set of user arrival rate
vectors for which the network can be stabilized by some scheduling policy. The ca-
pacity region of a constrained queuing system, such as a wireless network, is well
characterized in [27]. Let A = ⎡⎣ Alr ⎤⎦ denote the multicast matrix, i.e., Alm = 1 if
m

l ∈ L and Alm = 0 otherwise. For our model, the capacity region is given by the set
708 X.N. Miao and X.W. Zhou

⎧⎪ ⎡ M Am λ ⎤ ⎫⎪
Λ = ⎨λ : ⎢ ∑ l mm ⎥ ∈ Co ( S ) ⎬
⎪⎩ ⎣ m=1 µm cl ⎦ l∈L ⎪⎭
Where Co ( S ) represents the convex hull of all link schedules S .
With coding the actual physical flow on each link need only be the maximum of the
individual destination's flows [1]. For the case of multiple sessions sharing a network,
achieving optimal throughout requires in some cases coding across sessions. However,
designing such codes is a complex and largely open problem. In this paper, we use
intra-session coding approach to solve the multiple sessions problem similar to [26].
In this case, these constrains of the set of feasible flow vectors can be expressed as
⎧ xm if i = s m ;
⎪ (1)
∑ f md ( ij ) − ∑ f md ( ji ) = ⎨− x m if i = d ;
j:( ij )∈L j:( ji )∈L ⎪0 otherwise.
⎩

0 ≤ f md ( ij ) ≤ g md ( ij ) , ∀d ∈ Dm , ∀ij ∈ L (2)
M

∑ g ( ij ) ≤ c ( ij ) , ∀d ∈ D
m =1
md
m
, ∀ij ∈ L (3)

Where f md ( ij ) denote the information flow for destination d of session m , and

g md ( ij ) gives the physical flow for session m . The inequality (2) reflects the network
coding condition relating physical rate and information rate:
g md ( ij ) = max f md ( ij ) (4)
d

In practice, the network codes can be designed using the approach of distributed random
linear network coding ( [21]). If (1)-(2) holds, each sink receives with high probability a
set of packets with linearly independent coefficient vectors, allowing it or decode.
In this paper we consider the problem of rate control (congestion control) over a
multihop wireless ad hoc network. We take into among signals transmitted simulta-
neously by multiple nodes and the fact that a single node's transmission can be received
by multiple nodes, referred to as wireless multicast advantage. We model the conten-
tion relation between subflows as the interference set of the link, i.e., the links in the
interference set contend each other. Now we let S ( l ) = m ∈ Sm l ∈ L ( m ) , ∀l ∈ L be { }
the set of multicast sessions that use link l . Note that ij ∈ L ( m ) if and only if m ∈ S ( l ) .
We denote by I S ( l ) the interference set of link l = ij , including the link l itself. This set
indicates which groups of subflows interfere with the subflows which go through the
link l . Because links included in the interference set I S ( l ) share the same channel re-
source c ( l ) of the link l , only one of the subflows going through link

( )
k k ∈ I S ( l ) may transmit at any given time. The accurate interference set of the link
can be constructed based on the SIR model as proposed in [22].
 Rate Control, Routing Algorithm and Scheduling Multicast with Network Coding 709

3.2 Multicast with Network Coding

We assume each session m attains a utility function of the rate vector U m ( x m ) . We

assume U m x ( ) is continuously differentiable, increasing, and strictly concave. For
m

example, a utility that leads to proportional fairness is U m x m = wm log x m , ( )

where wm , m = 1, ", M are the weights. With the advantages of network coding [1],
we can model the routing problem at the network layer as follows:
m m ∑
max U m ( xm )
x ,y
m

s.t. Alm x m ≤ ylm , ∀m ∈ M (5)

∑ ∑
k∈I S ( l ) m∈S ( k )
ylm ≤ clm , ∀l ∈ L

The physical flow rate ylm for each multicast session m though
link l is y l
m
= m ax {A x m
lr
m
r }.
r
Note that network coding comes into action through the constraint (4). The utility
function is a concave function, and it is easy to prove that the network coding region
constraint (linear constraint) is a convex set. Therefore, solving the problem (5) is a
convex optimization problem. For a data with multiple multicast sessions, the maxi-
mum utility and its corresponding optimal routing strategy can be computed efficiently
and in a distributed fashion.

4 Cross-Layer Design with Network Coding

4.1 Distributed Rate Control Algorithm

At time t , given congestion price p ( t ) , the source s m adjusts its sending rate ac-
m
cording to aggregate congestion price over the multicast tree Tr . This rate control
mechanism has the desired price structure and is an end-to-end congestion control
mechanism.
The solution to the cross-layered rate control problem is of the following form
similar to [28].
⎛ ⎞
⎜ wm ⎟ (6)
x ( t ) = x ( kT ) = min ⎜ L
m m
, Vm ⎟
⎜ Am p m kT ⎟
⎜∑ l l ( ) ⎟
⎝ l =1 ⎠
For kT ≤ t < ( k + 1) T , where Vm is the maximum data rate for users of class m .The
congestion costs are updated as
710 X.N. Miao and X.W. Zhou

plm ( ( k + 1) T ) = [ plm ( kT ) +
(7)
⎛M ( k +1)T ⎞
α l ⎜ ∑ Alm ∫ nm ( t ) x m ( kT ) dt − cl ( t ) T ) ⎟]+
⎝ m =1 kT
⎠

[] [
+
where α l is a positive scalar stepsize. . is a projection on 0, ∞) .

4.2 Cross-Layer Design

G
• For each link l , in the beginning of period t allocates a capacity φ ( t ) over
link l such that
G L
φ ( t ) = argG max ∑ plm ( t ) cl (8)
φ ∈R l =1

The scheduling (8) is a difficult problem for ad hoc network. Our algorithm is similar to
a distributed variant of the sequential greedy algorithm presented in [24].
• Over link l , send an amount of coded packets for the session
G
ml ( t ) = arg max ∑ pl
m(t )
at rate φ ( t ) .
m
l

• Routing: over link l , a random linear combination of data of multicast ses-

sion ml to all destinations d is sent at an amount of bits according to the rate de-
termined by the above scheduling.
• Coding scheme: In wireless networks, nodes encode and transmit packets or re-
ceive and decode packets at different time instants. This coding scheme is a
random distributed coding scheme as the schemes in [21].
Directly applying the convergence results [17], we have the following result regarding
the convergence property of the distributed rate control algorithm for multicast.

5 Numerical Examples
In this section, we provide numerical examples to complement the analysis in the pre-
vious sections. We consider a simple ad hoc network with two multicast sessions
shown in Fig. 1. The network is assumed to undirected and a link has equal capacities in
1
both directions. And assume that session one with source node s and
destination x1 and x 2 , and session two with source node s 2 and
destination d and d with the same utility U m ( xm ) = log ( xm ) . We have chosen such a
1 2

small, simple topology to facilitate detailed discussion of the results.

We now simulate the distributed algorithm with dynamic scheduling and network cod-
Ⅳ
ing developed in .We assume the following link capacities: links ( s1 , s 2 ) and ( x 2 , d 1 ) have

( ) ( ) ( )
2 units of capacity, links s 2 , y , ( y, x1 ) , y, d 2 and y, x 2 have 3 units of capacity and
all other links have 1 units of capacity when active.
 Rate Control, Routing Algorithm and Scheduling Multicast with Network Coding 711

Fig. 1. A network topology with two multicast sessions

Fig. 2 and Fig. 3 show the evolution of source rate and congestion price of each
session with stepsize α l = 0.01 in (7). We see that the source rates converge quickly to a
neighborhood of the corresponding stable values .

4
session 1
3.5 session 2

3
Average Source Rates

2.5

1.5

0.5

0
0 20 40 60 80 100 120 140 160 180 200
Number of Iterations

Fig. 2. The average source rates with distributed algorithm

0.7
session 1
session 2
0.6

0.5
Average Congestion Prices

0.4

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160 180 200
Number of Iterations

Fig. 3. The congestion prices with distributed algorithm

712 X.N. Miao and X.W. Zhou

The simulation result also shows coding occurs over the multicast trees in Fig. 4 and
( 2
)
Fig. 5: 2 units of traffic of session one is coded over link s , y and 2 units of traffic of

(
session two is coded over link y , x
1
).

Fig. 4. Routing for session 1with distributed algorithm

Fig. 5. Routing for session 2 with distributed algorithm

Note that the routing depends not only on the network topology, which determines
the interference among links, but also link capacity configurations. So we can see that
( )
the link s 2 , x 2 is not used in Fig. 3. But the distributed dynamic scheduling always
picks the link with the locally heaviest weight, this feature such that the link ( s 2 , y ) is
always active due to it has a chance to be a locally heaviest link a lot of time.

6 Conclusions
We have presented a model for the joint design of distributed rate control algorithm,
routing and scheduling for multicast session with network coding in ad hoc networks.
With random network coding, the algorithm can be implemented in a distributed
manner, and work at transport layer to adjust source rates and at network layer to carry
out network coding. We study the case with dynamic arrivals and departures of the
users. The scheduling element of our algorithm is a dynamic scheduling policy. Nu-
merical examples are provided to complement our theoretical analysis. Modeling and
solution algorithm can be easily tuned according to a specific networking technology.
We will further study the practical implementation of our algorithm and extend the
results to networks with more general interference models. Solving this problem will
further facilitate the practical deployment of network coding in real networks.
 Rate Control, Routing Algorithm and Scheduling Multicast with Network Coding 713

Acknowledgement
This work is supported in part by the National Science Foundation of P.R. China under
Grant no.60773074 and by the National High Technology Research and Development
Program of P. R. China under Agreement no.2009AA01Z209.

References
[1] Ahlswede, R., Cai, N., Li, S.Y.R., Yeung, R.W.: Network information flow. IEEE Trans.
Inform. Theory 46, 1204–1216 (2000)
[2] Li, S.Y.R., Yeung, R.W., Cai, N.: Linear network coding. IEEE Trans. Inform. Theory 49,
371–381 (2003)
[3] Ho, T., Koetter, R., Karger, M.D.R., Effros, M.: The benefits of coding over routing in a
randomized setting. In: Proc. Int’l Symp. Information Theory, Yokohama, Japan. IEEE,
Los Alamitos (2003)
[4] Jaggi, S., Chou, P.A., Jain, K.: Low complexity optimal algebraic multicast codes. In: Proc.
Int’l Symp. Information Theory, Yokohama, Japan. IEEE, Los Alamitos (2003)
[5] Sander, P., Egner, S., Tolhuizen, L.: Polynomial time algorithms for network information
flow. In: Symposium on Parallel Algorithms and Architectures (SPAA), San Diego, CA,
pp. 286–294. ACM, New York (2003)
[6] Lun, D.S., Ratnakar, N., Médard, M., Koetter, R., Karger, D.R.: Minimum-cost multicast
over coded packet networks. IEEE Trans. Inform. Theory (2006)
[7] Sagduyu, Y.E., Ephremides, A.: Joint scheduling and wireless network coding. In: Proc.
WINMEE, RAWNET and NETCOD 2005 Workshops (2005)
[8] Wu, Y., Chou, P.A., Kung, S.-Y.: Minimum-energy multicast in mobile ad hoc networks
using network coding. IEEE Trans. Commun., 1906–1918 (2005)
[9] Deb, S., Srikant, R.: Congestion control for fair resource allocation in networks with ulti-
cast flows. IEEE Trans. on Networking, 274–285 (2004)
[10] Kelly, F.P., Maulloo, A., Tan, D.: Rate control in communication networks: Shadow
prices, proportional fairness and stability. Journal of the Operational Research Society,
37–252 (1998)
[11] Huang, X., Bensaou, B.: On Max-min Fairness and Scheduling in Wireless Ad-Hoc Net-
works: Analytical Framework and Implementation. In: Proceedings of IEEE/ACM Mo-
biHoc, Long Beach, CA, October 2001, pp. 221–231 (2001)
[12] Sarkar, S., Tassiulas, L.: End-to-end Bandwidth Guarantees Through Fair Local Spectrum
Share in Wireless Ad-hoc Networks. In: Proceedings of the IEEE Conference on Decision
and Control, Hawaii (2003)
[13] Yi, Y., Shakkottai, S.: Hop-by-hop Congestion Control over a Wireless Multi-hop Net-
work. In: Proceedings of IEEE INFOCOM, Hong Kong (2004)
[14] Qiu, Y., Marbach, P.: Bandwith Allocation in Ad-Hoc Networks: A Price-Based Ap-
proach. In: Proceedings of IEEE INFOCOM, San Francisco, CA (2003)
[15] Xue, Y., Li, B., Nahrstedt, K.: Price-based Resource Allocation in Wireless Ad hoc Net-
works. In: Jeffay, K., Stoica, I., Wehrle, K. (eds.) IWQoS 2003. LNCS, vol. 2707, pp.
79–96. Springer, Heidelberg (2003)
[16] Lin, X., Shroff, N.B.: Joint Rate Control and Scheduling in Multihop Wireless networks.
Technical Report, Purdue University (2004),
http://min.ecn.purdue.edu/_linx/papers.html
714 X.N. Miao and X.W. Zhou

[17] Lin, X., Shroff, N.: The impact of imperfect scheduling on cross-layer rate control in mul-
tihop wireless networks. In: Proc. IEEE Infocom (2005)
[18] Wu, Y., Chiang, M., Kung, S.Y.: Distributed utility maximization for network coding
based multicasting: A critical cut approach. In: Proc. IEEE NetCod (2006)
[19] Wu, Y., Kung, S.Y.: Distributed utility maximization for network coding based multi-
casting: A shortest path approach. IEEE Journal on Selected Areas in Communications
(2006)
[20] Ho, T., Viswanathan, H.: Dynamic algorithms for multicast with intrasession network
coding. In: Proc. Allerton Conference on Communication, Control and Computing (2005)
[21] Chou, P.A., Wu, Y., Jain, K.: Practical network coding. In: Proc. Allerton Conference on
Communication. Control and Computing (2003)
[22] Gupta, P., Kumar, P.R.: The capacity of wireless network. IEEE Trans. on Information
Theory 46(2), 388–404 (2000)
[23] Shor, N.Z.: Monimization Methods for Non-Differentiable Functions. Springer, Heidel-
berg (1985)
[24] Preis, R.: Linear time 1/2-approximation algorithm for maximum weighted matching in
general graphs. In: Meinel, C., Tison, S. (eds.) STACS 1999. LNCS, vol. 1563, p. 259.
Springer, Heidelberg (1999)
[25] Jain, K., Padhye, J., Padmanabhan, V.N., Qiu, L.: Impact of interference on multi-hop
wireless network performance. In: Proc. ACM Mobicom (2003)
[26] Ho, T., Viswanathan, H.: Dynamic algorithms for multicast with intra-session network
coding. In: Proc. 43rd Annual Allerton Conference on Communication. Control and
Computing (2005)
[27] Tassiulas, L., Ephremides, A.: Stability properties of constrained queuing systems and
scheduling policies for maximum throughout in multihop radio networks. IEEE Trans. on
Automatic Control 37(12), 1936–1948 (1992)
[28] Lin, X., Shroff, N.B.: On the stability region of congestion control. In: Proceedings of the
42nd Annual Allerton Conference on Communication. Control and Computing (2004)
 Design and Experimental Study on Spinning Solid
Rocket Motor

Heng Xue, Chunlan Jiang, and Zaicheng Wang

State Key Laboratory of Explosion Science and Technology, Beijing Institute of Technology,
Beijing, China
xh930@163.com, {jiangchuwh,wangskyshark}@bit.edu.cn

Abstract. The study on spinning solid rocket motor (SRM) which used as power
plant of twice throwing structure of aerial submunition was introduced. This kind
of SRM which with the structure of tangential multi-nozzle consists of a com-
bustion chamber, propellant charge, 4 tangential nozzles, ignition device, etc.
Grain design, structure design and prediction of interior ballistic performance
were described, and problem which need mainly considered in design were
analyzed comprehensively. Finally, in order to research working performance of
the SRM, measure pressure-time curve and its speed, static test and dynamic test
were conducted respectively. And then calculated values and experimental data
were compared and analyzed. The results indicate that the designed motor op-
erates normally, and the stable performance of interior ballistic meet demands.
And experimental results have the guidance meaning for the pre-research design
of SRM.

Keywords: Solid Rocket Motor SRM （ ）, Interior Ballistic Performance, Static

Test, Dynamic Test.

1 Introduction
Solid rocket motor (SRM) which has characteristics of small size, reliable operation,
simple structure, low cost and long-term preservation, is widely used in all kinds of
small-sized, short range military rocket and missile power plant. In recent years, a
series of studies on SRM have been made, and these studies are important for the
construction of structure design, internal flow field simulation and improvement of
interior ballistic performance [1], [2], [3], [4].
In this paper, spinning SRM is used as the power plant of twice throwing structure of
aerial submunition. Its structural characteristic is 4 tangential nozzles distributing
uniform on the circumference of the motor case, the thrust generated by nozzle acting
along tangential direction of the rocket body and then the motor rotates. Depending on
the centrifugal force generated by the rotation, bullets were dispersed in a large area.
Meanwhile, rotation ensures the flight stability of submunition.
Firstly, according to technical parameters the general scheme of SRM was deter-
mined, and three-dimensional overall structure design was achieved by using Inventor
software. Prediction of interior ballistic performance was carried out after the deter-
mination of overall structure design and grain design. Finally, static test and dynamic

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 715–723, 2009.
© Springer-Verlag Berlin Heidelberg 2009
716 H. Xue, C. Jiang, and Z. Wang

test were conducted, and the test results which were compared and analyzed under
different conditions, providing reference for similar SRM design.

2 Spinning Solid Rocket Motor Design

Unlike conventional tail nozzle, this article aims to present a motor with the structure
form of 4 tangential nozzles, the key of design is to determine a reasonable and feasible
design scheme. According to tactical and technical indexes of the SRM, general pa-
rameters were determined, and the main structural design was completed such as the
design of propellant charge, combustion chamber, nozzles, etc.

2.1 Grain Design

The size and geometry of grain determine gas producing rate and its change law of SRM,
thus determine thrust and the rule pressure with time. At the same time, volume and
weight of combustion chamber were determined by the volume of propellant charge.
Considering motor thrust is relatively large, while working time is not long, and
meanwhile in order to obtain a higher charge coefficient, tubular grain with seven
columns were used. As shown in Fig. 1. The grains with ends coated burn on the in-
ternal and external surface simultaneously. The greatest advantage of this program is
simple structure, easy to make, neutral burning and constant thrust can be obtained.
Diameter and length of the tubular grain were determined by parameters such as
combustion chamber pressure, working time, combustion rate of propellant etc.

Fig. 1. Tubular grain configuration

2.2 Structure Design

Combustion chamber is one of the most important components of SRM, and it provides
the place of storage and combustion for propellant charge. Furthermore, the combus-
tion chamber bears high temperature, high pressure and under the action of various
loads. At the same time, the shell of combustion chamber is opened holes to install
nozzles. In order to ensure combustion chamber still work reliably under the above bad
working conditions, the combustion chamber should have characteristics of sufficient
strength and rigidity, reliable connection and seal, light structural weight etc. Based on
the above considerations, shell material uses 35CrMnSi, connection of shell and noz-
zle, shell and closure head adopts screw connection.
 Design and Experimental Study on Spinning Solid Rocket Motor 717

Fig. 2. Structure of the spinning solid rocket motor

According to requirements of the general parameters, 4 tangential nozzles distrib-

uted uniform near the bottom of SRM, and its profile uses laval nozzle. In order to
convenient for processing and installation, the axial line of nozzle and the normal line
of rocket body inclined to one another at an angle of 45 degrees. Fig. 2 depicts the
architecture of the SRM.

3 Interior Ballistic Calculation

Interior ballistic calculation is an important step in the design of SRM, and combustion
chamber pressure is an important parameter. To begin with, pressure influences not
only the size of SRM thrust but also the combustion time of propellant grain, thus
affecting the motor’s working time. Secondly, from the perspective of structure design,
pressure directly determines the load bear by combustion chamber, thereby affects the
structural weight of SRM. From the above we can see that calculation of combustion
chamber pressure has vital significance to the prediction of interior ballistic perform-
ance, determination of SRM thrust as well as verification of structural strength.

3.1 Basic Assumptions

Through simplifying the working process of SRM, mass conservation equation and
state equation of the combustion chamber were established. Simplification is as fol-
lows: (1) Combustion chamber gas is ideal gas and heat loss is constant in combustion
process; (2) Propellant charge is complete combustion and the burning temperature is
unchanged; (3) The charge is ignited under the pressure of ignition pressure; (4) Do not
consider effects of erosive burning or the change of burning surface caused by scour-
ing. The charge obeys geometric burning law in combustion process.
Seriously, interior ballistic calculation is relevant to varying with time of flow pa-
rameters which distributing along the length of combustion chamber. Then
one-dimensional unsteady flow equations seemed to be used to determine the func-
tional relation between time and flow parameters such as gas pressure, temperature,
718 H. Xue, C. Jiang, and Z. Wang

density. However, a series of difficulties will be encountered in mathematics when

using this method. Therefore, methods of simplifying problem are usually adopted.
Calculation of average pressure and average temperature change with time, which
without considering flow parameters change along the length of combustion chamber,
is called zero-dimensional ballistic calculation. In addition, ignition pressure is re-
garded as the starting point of pressure calculation and ignition process is not consid-
ered. At this moment, motor working process can be divided into three stages: pressure
rising stage, pressure balance stage and after-effect stage.

3.2 Mathematical Model

Generally speaking, pressure rising stage is an unsteady state and it reflects motor’s
igniting process. The pressure rapidly increases to close to equilibrium state in a very
short time. Equation of this stage is as follows [5]:
1
⎧ ⎡ ⎛ p ⎞1− n ⎤ − (1− n )φ2 At Γ RTeq ⎫1− n
⎪ t ⎪
pc = peq ⎨1 − ⎢1 − ⎜ ⎟ ⎥e
(1)
ig
⎬
Vc

⎢
⎪⎩ ⎣ ⎝ peq ⎥
⎠ ⎦ ⎪⎭

Where pc is combustion chamber pressure, pa; peq is average pressure, pa; pig is igni-
tion pressure, pa; e is grain web thickness, mm; n is pressure index; φ2 is correction
coefficient; Γ is specific heat ratio function; t is working time, s; VC is initial chamber
volume, m3 ; R is molar gas constant, J ⋅ mol −1 ⋅ K −1 ; Teq is equilibrium temperature,
K ; At is nozzle exit area, m 2 .
Because of the designed tubular grain with ends coated burns on the internal and
external surface simultaneously, the burning surface is neutral burning. The stage of
pressure balance can be calculated as:

Vc a dpc
pc = peq (
1− n )
− (2)
xφ2 Γ At kC de
2 ∗

Where x is heat release coefficient; a is velocity coefficient; k is specific heat ratio;

C ∗ is characteristic velocity, m ⋅ s −1 .
After the end of combustion, combustion products which no longer producing are
continue to be exhausted from the nozzle. The combustion chamber pressure decreases
rapidly until reaching balance with the external pressure. The process of gas expansion
is after-effect stage and the pressure can be calculated as:

⎡ 2Vcf ⎤ 2k
pc = peq ⎢ ⎥ (3)
⎢⎣ 2Vcf + ( k − 1) RTeq φ2 At Γt ⎥⎦ k − 1

Where Vcf is combustion chamber volume, m3 .

 Design and Experimental Study on Spinning Solid Rocket Motor 719

k +1
⎛ 2 ⎞ 2( k −1)
Γ= k⎜ ⎟ (4)
⎝ k +1 ⎠

3.3 Calculation and Analysis

Basic parameters are as follows: combustion chamber diameter is Di = 132mm ; com-

bustion chamber length is L = 79mm ; density is ρ = 1.61g ⋅ cm −3 ; specific heat ratio
is k = 1.25 ; ignition pressure is pig = 4MPa ; characteristic velocity is C ∗ = 1341m ⋅ s −1 ;
initial chamber volume is Vc = 0.53 ×10−3 m3 .
The above differential equations can be solved by using fourth-order runge-kutta, and
then the law of combustion chamber pressure changes with time will be obtained. Table 2
presents calculation results, and the calculated pressure-time curve is shown in Fig. 3.
From the results and the pressure-time curve, we can see that the interior ballistic
performances of SRM meets design requirements. However, because of grains burn on
the internal and external surface simultaneously, the main problem existed in practical
combustion is erosive burning, and the actual curve will vary. Therefore, in order to
verify the correctness of theoretical model, experimental research should be carried out.

7
6
Pressure(MPa)

5
4
3
2
1
0
0.0 0.2 0.4 0.6 0.8 1.0
Time(s)

Fig. 3. Pressure-time calculating curve

4 Experimental Studies
In order to study working performance of the spinning SRM and measure pressure-time
curve and speed, static test and dynamic test were conducted respectively.

4.1 Static Test

4.1.1 Test Equipment

The static test of spinning SRM is designed to check working reliability of components,
rationality of charge structure and interior ballistic performance. As shown in Fig. 4,
720 H. Xue, C. Jiang, and Z. Wang

Fig. 4. Static test

the assembled spinning SRM is fixed with bolts in a test stand. By using pressure
sensor, pressure of the combustion process was converted into electrical signals. And
combustion chamber pressure and working time were measured by data acquisition
system.
The test motor consists of a combustion chamber, propellant charge, 4 tangential
nozzles, charge baffle plates, igniter cartridge, etc. Combustion chamber shell material
uses 35CrMnSi, connection of shell and nozzle, shell and closure head adopt screw
connection. Charge baffle plate is made of low carbon steel. As the motor’s working
time is short, no additional thermal protection measures are need. The motor uses the
ways of electric ignition, igniter charge bag is placed in front of the grain, and igniting
charge is 2# small powder grain. The charge is double base propellant grain, tubular
grain with seven columns are used.
The grains which two ends coated by 2 millimeter thick nitrocellulose lacquer, burn
on internal and external surface simultaneously. The ignition charge mass of first and
second test are 15gram and 20gram respectively.
Pressure sensor (model ZQ-Y1) was used, and the main technical specifications are
as follows, range: 20MPa, accuracy: 0.3%FS, power: 5-15VDC, output: 1.3672mV/V.

4.1.2 Test Results and Analyses

Pressure-time test curves of the first and second test are shown in Fig. 5 and Fig. 6
respectively. Because of the size of spinning SRM is relatively small, ignition charge
mass has a relatively high impact on interior ballistic performance. A comparison of
Fig. 5 and Fig. 6 indicates that less ignition charge mass make ignition delay too long,
and could not meet the needs. However, considering the restriction on motor size and
the affordability of solid rocket motor case, too much ignition charge mass is un-
suitability. Appropriate amount of ignition charge mass should be selected according
to the design requirements. And the ignition pressure can be improved and the prob-
lem of ignition delay at the preliminary stage was solved by adjusting ignition charge
mass.
 Design and Experimental Study on Spinning Solid Rocket Motor 721

8
7

Pressure(MPa)
6
5
4
3
2
1
0
0.0 0.2 0.4 0.6 0.8 1.0
Time(s)

Fig. 5. Pressure-time test curve of the first set

7
6
Pressure(MPa)

5
4
3
2
1
0
0.0 0.2 0.4 0.6 0.8 1.0
Time(s)

Fig. 6. Pressure-time test curve of the second set

The results also indicate that the designed motor work well, structure is reliable, the
thread connections have no loosening, combustion chamber and nozzles have no de-
formation, and the technical requirements are met.
Table 1 presents the first and second experimental data, and Table 2 presents a
comparison of calculated values and experimental data. As can be seen in Table 2, the
designed average thrust Fn and average pressure Pn is lower than experimental data, and
the burning time Tb is longer than experimental data. The main reason of this discrep-
ancy is that the actual burning rate is higher than theoretical value. And this causes
burning time shortened and pressure reduced.

Table 1. Experimental data of spinning SRM

Parameters First set Second set

pm (Mpa) 7.32 6.77
pn (Mpa) 7.09 6.55
Tb (s) 0.78 0.74
Tk (s) 0.85 0.81
722 H. Xue, C. Jiang, and Z. Wang

Table 2. Comparison of calculated values and experimental data

Parameters Calculated values Experimental data

pn (Mpa) 6.2 6.55
Fn (N) 1940.5 2058.2
Tb (s) 0.89 0.74
Tk (s) 0.93 0.81

4.2 Dynamic Test

In order to check dynamic performance of the spinning SRM, dynamic test was
conducted. As shown in Fig. 7, with bearings installed on the top, the motor was sus-
pended. By using high speed camera, dynamic working process of the motor was re-
corded and the speed was measured. According to high speed photographic records, the
speed is up to 15000r/min without load and meets the design requirements.
The results indicate that the spinning SRM work well, and have no abnormal sounds.
However, high-speed rotation has great influence on working performance of the
spinning SRM. Compared to static situation, the ablation of nozzle is more serious in
rotation conditions. Meanwhile, the centrifugal force generated by the rotation caused
metal oxides deposited on the motor case. And this may lead to thermal damage of head
structure.

Fig. 7. Dynamic test of the spinning SRM

5 Conclusions
Design and experimental methods of spinning SRM were preliminarily studied in the
paper. By comparing with calculated values and experimental data, the following re-
sults were obtained:
(1) The ignition charge mass can be of significant effect on the pressure-time curves
and internal ballistic performances of the spinning SRM. And in order to obtain satis-
fied internal ballistic performances, the ignition charge mass must be strictly
controlled.
 Design and Experimental Study on Spinning Solid Rocket Motor 723

(2) By comparing Fig. 3 with Fig. 6, and from Table.2 we can see that calculated
values accord with experimental data, and the calculation method presented in the
paper is simple and practicable.
(3) The experimental results indicate that the designed motor operated normally, and
the stable performance of interior ballistic meet demands.
(4) High-speed rotation has great influence on working performance of the spinning
SRM.

References
1. Kamm, Y., Gany, A.: Solid Rocket Motor Optimization. In: 44th AAIA/ASME/SAE/ASEE
Joint Propulsion Conference and Exhibit, AIAA-2008-4695, Hartford, CT (2008)
2. French, J., Flandro, G.: Linked Solid Rocket Motor Combustion Stability and Internal Bal-
listic Analysis. In: 41st AAIA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit,
AIAA-2005-3998, Tucson, Arizona (2005)
3. Shimada, T., Hanzawa, M.: Stability Analysis of Solid Rocket Motor Combustion by Com-
putational Fluid Dynamics. AIAA Journal 46(4), 947–957 (2008)
4. Willcox, M.A., Brewster, Q., Tang, K.C.: Solid Rocket Motor Internal Ballistics Simulation
Using Three-Dimensional Grain Burnback. Journal of Propulsion and Power 23(3), 575–584
(2007)
5. Zhang, P., Zhang, W.S., Gui, Y.: Principle of Solid Rocket Motor. Beijing Institute of Sci-
ence and Technology Press, Beijing (1992) (in Chinese)
An Energy-Based Method for Computing Radial Stiffness
of Single Archimedes Spiral Plane Supporting Spring

Enlai Zheng1, Fang Jia1, Changhui Lu1, He Chen1, and Xin Ji2
1
School of Mechanical Engineering, Southeast University, Nanjing, 211189, China
2
Nanjing Jienuo Environment Technology Co.Ltd, Nanjing, 210014, China
xx_xx1111@163.com, fangjia1988@yahoo.com, magiclu2007@126.com,
sjzyzch2222@yahoo.com.cn, njjn@Njjinuo.com

Abstract. With space-based adaptive performance of lower stiffness and greater

deformation energy, the plane supporting spring finds its wide application in
fields like aeronautics, astronautics, etc. In the current study, the radial stiffness
formula of a single Archimedes spiral plane supporting spring is derived by
means of energy approach, with three key parameters of the supporting spring as
independent variables. A series of the supporting spring FEA models are estab-
lished via APDL speedy modeling. According to the isolation requirements of
electronic equipment for a fighter, an example is presented in the form of finite
element analysis. The theoretical calculation and analysis data are studied and
fitted by MATLAB using the least-square method to obtain the discipline of the
radial stiffness of single spiral plane supporting spring with the changes of its
three key parameters. The validity of energy-based radial stiffness formula of the
spring is confirmed by the comparison between the theoretical calculation and
finite element analysis results.

Keywords: single archimedes spiral, plane supporting spring, energy approach,

radial stiffness.

1 Introduction
Precision instrument and system for aeronautics and astronautics usually endure
stronger impact or vibration under hostile environments. In order to obtain longer life
and higher reliability, it is essential to introduce effective measures for vibration re-
duction and isolation. Due to certain factors like constraints of assembly space, etc,
common isolation components usually fail to meet fully the requirements of vibration
isolation. Therefore, with its space-based adaptive performance of smaller volume,
lighter weight, lower stiffness, greater deformation energy and better static and dy-
namic characteristics, the plane supporting spring begins to attract attention recently.
The performance of plane supporting spring is mainly determined by its radial stiff-
ness, and environmental vibration varies from case to case in its requirements for radial
stiffness of springs, the study of which thus becomes an important issue. Currently, the
general way of obtaining the radial stiffness of plane supporting spring is to calculate
by the finite element simulation, or to measure by experiment. In our previous studies,
Jia Fang et al [1] give the relationship between turn number and radial stiffness by FEA;

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 724–733, 2009.
© Springer-Verlag Berlin Heidelberg 2009
 An Energy-Based Method for Computing Radial Stiffness 725

Chen Nan et al [2] [3] carry out the FEA and experimental research on the stiffness of the
spring; and Gao Weili et al [4] study the stiffness and stress performance of involutes of
circle applied to linear compressor by ANSYS. Besides, other researchers, like A. S.
Gaunekar et al[5], also make a study on a kind of structural supporting spring by FEA.
Nonetheless, finite element analysis is constrained by the need of hardware and soft-
ware support, and experimental method tends to have the deficiencies of long cycle,
high cost and complicated process, etc.
As far as we know, no attempts have ever been reported to have taken an energy-
based approach to calculating the radial stiffness of plane supporting spring until our
current study, in which the X dimensional radial stiffness formula of single spiral
plane supporting spring is derived by means of energy approach and its validity is
verified by finite element simulation.

2 Design of Plane Supporting Spring

The structure of a plane supporting spring studied in this paper is shown in Fig. 1.
Table 1 shows the independent parameters of the spring.
It uses Archimedes spiral as flexure spiral spring to connect the inner and outer
supports. The equation of Archimedes spiral in polar coordinates is
ρ = a + tθ (1)
where ρ stands for the polar radius corresponding to θ , mm; a for the polar radius
corresponding to θ = 0 , mm; t for the Archimedes spiral coefficient, mm/°; and θ
for the polar angle. And we have
t = ( Doi − Dio ) / 4π n2 (2)
where Doi is the outer support inner diameter; Dio the inner support outer diameter;
n2 the turn number of flexible spiral spring.

Fig. 1. The geometric model of plane supporting spring

726 E. Zheng et al.

Table 1. The independent parameters of plane supporting spring

No. Parameter Symbol

1 Thickness h
2 Outer support outer diameter Doo
3 Outer support inner diameter Doi
4 Inner support outer diameter Dio
5 Inner support inner diameter Dii
6 Length of inner connecting line Li
7 Length of outer connecting line Lo
8 Width of flexible spiral spring line b
9 Number of flexible spiral spring n1
10 Turn number of flexible spiral spring line n2

The flexible spiral spring line is composed of two Archimedes spirals with differ-
ent starting points. Width of flexible spiral spring line can be adjusted by regulating
the starting points.
The inner and outer connecting lines are designed to prevent stress concentration.
The outer support is used for spring location; the electronic instrument is tightened on
the inner support. The key parameters for radial stiffness include thickness, outer
support inner radius, inner support outer radius, width of flexible spiral spring line,
and number of flexible spiral spring. The influence of each parameter on radial stiff-
ness is analyzed by control variate method.

3 Formula Derivation of Radial Stiffness Based on Energy

Approach[6]
The analysis model of a single spiral/turning plane supporting spring is shown in
Fig.2.
As shown in Fig.3, taking the microsection of the plane supporting spring as focus,
the visual deformation of microsection can be decomposed into: the axial relative
displacement, angle, and dislocation of the two ends, of which only the internal force
at both ends works, its value being

dW = FN (θ )d ( ∆l ) * + M (θ )dθ * + FS (θ ) d λ * (3)

where FN (θ ) , M (θ ) and FS (θ ) stand for axial force, flexural moment and shearing
force under external force F respectively.
The total visual works obtained by the integration of Eq.3 is shown in the follow-
ing equation.
 An Energy-Based Method for Computing Radial Stiffness 727

Fig. 2. The analysis model of single spiral/turning plane supporting spring

Fig. 3. The visual deformation of microsection

W = ∫ FN (θ )d (∆l )* + ∫ M (θ )dθ * + ∫ FS (θ )d λ * (4)

Suppose the visual displacement of external force F is υ * ( x ) , as a result of which

the visual displacement remain unchanged under external force F and the total visual
works can be expressed in another form as follows:
W = Fυ *( x) (5)
According to the energy principle, the visual works under external force are equal to
those under internal force, as expressed below.

Fυ *( x) = ∫ FN (θ )d (∆l )* + ∫ M (θ )dθ * + ∫ FS (θ )dλ * (6)

728 E. Zheng et al.

With the displacement under external force F in X direction as the visual displace-
ment on condition that a unit force is added in X direction of the single helix plane
supporting spring, Eq.6 can also be reformulated into Eq.7.

∆ = ∫ FN (θ )d ( l ) + ∫ M (θ )dϕ + ∫ FS (θ )d λ (7)

where FN (θ ) ,M (θ ) and FS (θ ) stand for axial force, flexural moment and shear-
ing force under unit force respectively and d ( l ) , dϕ and d λ for relative axial
displacement, relative angle and relative dislocation, of which
M (θ ) = F isin θ i ρ

M (θ ) = sin θ i ρ
M (θ )
dϕ = i ρ iθ
EI
Being of Archimedes spiral, we have

ρ = Rio + tθ (8)

As the plane supporting spring mainly endures bending deformation, the strain energy
generated by axial tension and compression and shear can be ignored, unlike that
generated by bending deformation. Therefore,

∫F N (θ )d ( l ) =0

∫ F (θ )d λ =0
S

The displacement under F in X direction is

2 π − θ Lo F i sin α i ρ
∆= ∫θ Li
sin α i ρ
EI
i ρ idα

F (Rio + tθ )4 1
= ( − sin(2θ )(Rio + tθ )3 − (9)
2EI 4t 2
3t 3t 2 3t3
cos(2θ )(Rio + tθ )2 + sin(2θ )(Ri0 + tθ ) + cos(2θ )) θ2Liπ −θLo
4 4 8
Of which
t = ( Roi − Rio − b ) / 2π
θ Li = Li / t
θ Lo = b / t
 An Energy-Based Method for Computing Radial Stiffness 729

The stiffness of X direction is expressed in Eq.10.

F (R + tθ )4 1
KX = = 2EI /{( io − sin(2θ )(Rio + tθ )3 −
∆ 4t 2
2
(10)
3t 3t 3t3
cos(2θ )(Rio + tθ )2 + sin(2θ )(Ri0 + tθ ) + cos(2θ )) θ2Liπ −θLo }
4 4 8

4 Verification of Radial Stiffness Formula

The spring material used for finite element simulation is 65Mn, whose properties are
as follows:

E = 2.06 × 10 11 P a , u = 0.3,[σ ] = 475 M P ,

σ s = 800 M P , σ b = 980 M P
A group of independent parameters of plane supporting spring for a fighter are shown
in table 2.

Table 2. Lists independent parameter of a spring for a fighter

Value
No. Parameter Symbol
( mm )

1 Thickness h 1

2 Outer support outer diameter Doo 34

3 Outer support inner diameter Doi 30

4 Inner support outer diameter Dio 10

5 Inner support inner diameter Dii 8

6 Length of inner connecting line Li 0.5

7 Length of outer connecting line Lo 0.5

8 Width of flexible spiral spring line b 0.5

9 Number of flexible spiral spring n1 1

10 Turn number of flexible spiral spring n2 1

A. Modeling in APDL language[7] and analyzing

ANSYS creates the geometric model of plane supporting spring using APDL lan-
guage. And then the model will be divided into meshes. Automatic analysis will begin
after adding the load and boundary conditions.
730 E. Zheng et al.

Fig. 4. The finite element analysis model

Fig. 5. The displacement of finite element simulation

The finite element model and the displacement chart are demonstrated in Figures 4
and 5.
It can be seen from the Fig.5 that under the action of unit force in the X direction,
the value of displacement of finite element simulation is 1.4971mm, while the theo-
retical displacement by radial stiffness formula is 1.61334mm.

B. Analysis and Comparison between Theoretical Calculation and Finite

Element Simulation by Control Variate Method
1) Changing width of Flexible Spiral Spring Line from 0.4mm to1mm by the Step of
0.1mm
The radial stiffness curves of theoretical calculation and finite element simulation
with the changes of the width of flexible spiral spring line are shown in Fig.6.
 An Energy-Based Method for Computing Radial Stiffness 731

Fig. 6. The radial stiffness of theoretical calculation and finite element simulation varying with
changing width of flexible spiral spring line

It can be seen from the graph that under the action of radial unit force,the stiffness
goes up with the increasing width of flexible spiral spring line. Radial stiffness in
theory increases from 0.3125 N/mm to 0.4958 8 N/mm, and radial stiffness of finite
element simulation from 0.3368 N/mm to 0.48652 N/mm.The curves are cubic. The
results from theoretical calculation match those from finite element simulation.
2) Changing inner Support Outer Radius from 5mm to 11mm by the Step of 1mm
The radial stiffness curves of theoretical calculation and finite element simulation
with the changes of the inner support outer radius of plane supporting spring are
shown in Fig.7.
As revealed in Fig.7, the stiffness declines with the increasing inner support outer
radius of plane supporting spring under the action of X direction force. Radial stiff-
ness in theory decreases from 0.61985N/mm to 0.3269 N/mm, and radial stiffness of
finite element simulation from 0.6680N/mm to 0.4178 N/mm. The deviation between
the results of theoretical calculation and simulation turns narrower in that Eq.10 ig-
nores the impact of inner support energy in the calculation.
3) Changing thickness from 0.2mm to 0.8mm by the step of 0.1mm
The thickness-radial stiffness curves of theoretical calculation and finite element
simulation are shown in Fig.8.
This graph indicates that the radial stiffness ascends linearly with the increment of
thickness under the action of X direction force. Radial stiffness in theory increases
from 0.5N/mm to 0.8671 N/mm, and radial stiffness of finite element simulation from
0.5343N/mm to 0.9696 N/mm. The deviation between theoretical calculation and
simulation turns narrower due to the same reason mentioned above.
732 E. Zheng et al.

Fig. 7. The radial stiffness of theoretical calculation and finite element simulation varying with
changing inner support outer radius of plane supporting spring

Fig. 8. The radial stiffness of theoretical calculation and finite element simulation varying with
changing thickness of plane supporting spring

5 Conclusions
With wider application of the plane supporting spring, an ideal theoretical formula for
calculating radial stiffness grows more and more important in the process of its design.
This study has proposed a brand-new energy-based method for computing the radial
stiffness of single spiral plane supporting spring, with the radial stiffness formula well
 An Energy-Based Method for Computing Radial Stiffness 733

established theoretically. The results generated by the formula are basically consistent
with those from finite element simulation. Comparing both results makes it possible to
derive the general discipline of the radial stiffness of single spiral plane supporting
spring in relation to changes of the spring’s key parameters. And it may fairly be as-
sumed that our study has provided a solid theoretical basis for further research on the
multi-spiral/turn plane supporting spring as well.

Acknowledgement
We would like to acknowledge the contributions of a number of staff with Jienuo
Environment Technology. Special thanks go to their valuable work and, especially, to
their company, Nanjing Jienuo Environment Technology Co., Ltd., which provided us
with whatever we needed for the research and experiments. We are also greatly in-
debted to Professor Li Lingzhen for her successful organization and great support for
the project.

References
[1] Jia, F., Zhang, D., Zhang, Z.: Speedy stiffness modeling and designing of plane supporting
spring. In: The 15th International Conference on Mechatronics and Machine Vision in
Practice (M2VIP 2008), Auckland, New Zealand, pp. 209–214 (2009)
[2] Chen, N., Chen, X., Wu, Y.N., Yang, C.G., Xu, L.: Spiral profile design and parameter
analysis of flexure spring. Cryogenics (46), 409–419 (2006)
[3] Nan, C., Xi, C., Yinong, W., Lie, X., Chunguang, Y.: Performance analysis of spiral flex-
ure bearing. Cryogenics and Superconductivity 33(4), 5–8 (2005)
[4] Weili, G., Pengda, Y., Guobang, C.: Influence of geometry parameters on the performance
of flexure bearing. Cryogenics (6), 8–11 (2007)
[5] Gaunekar, A.S., Goddenhenrich, T., Heiden, C.: FiniteElement nalysis and Testing of
Flexure Bearing Elements. Cryogenics 36(5), 359–364 (1996)
[6] Liu, H.-w.: Mechanics of Materials (II). Higher Education Press, Beijing (2004)
[7] Jiang, S., Jia, F., Zhang, D., Wang, X.: Parameterized Modeling Technology of Plane Sup-
porting Spring Based on APDL. Equipment for Electronic Products Manufacturing 37(12),
46–49 (2008)
 Erratum to: Implementation of On/Off Controller for
Automation of Greenhouse Using LabVIEW

R. Alimardani1, P. Javadikia1, A. Tabatabaeefar1, M. Omid1, and M. Fathi2

1
Department of Agricultural Machinery Engineering, Faculty of Biosystems Engineering,
University of Tehran, 31587-77871 Karaj, Iran
2
Department of Electronic and Informatics, University of Siegen, 57074 Siegen, Germany
{Pjavadikia,atabfar,omid,rmardani}@ut.ac.ir
fathi@informatik.uni-siegen.de

H. Deng et al. (Eds.): AICI 2009, LNAI 5855, pp. 251–259, 2009.
© Springer-Verlag Berlin Heidelberg 2009
_______________________________________________

DOI 10.1007/978-3-642-05253-8_80

The paper “Implementation of On/Off Controller for Automation of Greenhouse Using

LabVIEW’’ in this volume has been withdrawn, because it was submitted by the
corresponding author, Payan Javadikia, without the knowledge of the other authors.

_______________________________________________
The original online version for this chapter can be found at
http://dx.doi.org/10.1007/978-3-642-05253-8_28
_______________________________________________
 Author Index

Alimardani, R. 251, E1 Hartmann, Benjamin 241

An, Liping 18 He, Jing 639
He, Yongyi 278
Belič, Aleš 241 He, Zhongshi 59
Breve, Fabricio 619 Ho, Zih-Ping 526
Honarvar, Ali Reza 86
Cai, Zhihua 28 Horng, Ming-Huwi 185
Cao, Cun Gen 479 Hu, Dike 231
Cao, Qixin 270 Hu, Jianming 658
Chang, Xueping 313 Hu, Wenpei 649
Chen, Chongchen 121 Huang, Liang-Tsung 579
Chen, Haizhu 59 Huang, Pei-Hwa 176
Chen, He 724 Huang, Rongfeng 425
Chen, Lifei 629 Hung, Nguyen Viet 330
Chen, Peihua 270
Chen, Shengbing 9 Igarashi, Shigeru 469
Chen, Si 629 Ikeda, Yasuwo 469
Chen, Yuquan 408
Cui, Yu 361 Javadikia, P. 251, E1
Ji, Hong Bing 322
Diao, Lu Hong 479 Ji, Xia 9
Dong, Min 141 Ji, Xin 724
Du, Lei 159 Jia, Fang 724
Duan, Bobo 677 Jiang, Chunlan 715
Duan, Zhongdong 221 Jiang, Lin-Hua 340
Jiang, Siwei 28
Fan, Guang-Zhong 176 Jiang, Weijin 379, 639
Fan, Jiancong 599 Jin, ChenXia 696
Fang, Gu 131 Jo, Hee-Won 507
Fang, Ji 103 Ju, Shi-Guang 361
Fard, Sepehr Meshkinfam 167
Fathi, M. 251, E1 Ko, Min-Suk 507
Feng, Xiaowei 278 Koo, Lock-Jo 507
Fu, He-fang 399 Krivosheev, Ivan 159
Fuangkhon, Piyabute 389 Kuo, Jung-Chih 579

Ge, Jinfa 1 Lai, Lien-Fu 579

Ghasem-Aghaee, Nasser 86 Lai, Yi-Chiuan 526
Gu, Tong-Yue 361 Lei, Jingsheng 555
Guo, Gongde 629 Lei, Kaiyou 38
Guo, Shuai 278 Lei, Xiujuan 206
Li, Bo 417
Hamzeh, Ali 167 Li, Dafa 463
Han, Fei 361 Li, FaChao 696
Han, Lei 298 Li, Hongjiao 96
736 Author Index

Li, Hui 668 Park, Sang-Chul 507

Li, Longshu 9 Peng, Daogang 668
Li, Qing-cheng 103 Perng, Chyuan 526
Li, Rong 399
Li, Shun-bao 306 Qigang, Shang 369
Li, Wei 340 Quiles, Marcos 619
Li, Weiming 435
Li, Xia 159 Ruan, Jiuhong 599
Li, Xianhua 278
Li, Yan 590 Shafiabady, Niusha 563
Liang, Yongquan 599 Shen, Hongyu 221
Liao, Daifa 568 Shen, Xizhong 497
Liu, Bingquan 417 Shi, Dejia 639
Liu, Chien-Heng 176 Shio, Masayuki 469
Liu, Hao 215 Shou, Zhou-xiang 535
Liu, Hui 408 Su, Yikang 668
Liu, Jin 322 Sun, Chengjie 417
Liu, Jun 590 Sun, Shanwu 685
Liu, Kaiyu 350 Sun, Tian 517
Liu, Lei 408, 479 Sun, Yu 649
Liu, Li-ying 399
Liu, Minna 517 Tabatabaeefar, A. 251, E1
Liu, Sheng 69 Tang, Xinggang 150
Liu, Wei 677 Tang, Zhong 568
Liu, Weixia 1 Tanprasert, Thitipong 389
Liu, XianLei 696 Tian, Xiuxia 96
Lo, Charles 270 Tong, Lingyun 18
Lu, Changhui 724 Tong, Zhiwei 215
Lu, Yang 658 Tseng, Ta-Hsiu 176

Ma, Qianzhi 206 Wang, Anhua 121

Ma, Yan 306 Wang, Baoxun 417
Ma, Zhanshan (Sam) 195, 608 Wang, Cong 103
MacDonald, Glen 131 Wang, Dawei 112
Mao, Wentao 231 Wang, Gi-Nam 507
Matsuhisa, Takashi 78 Wang, Hongjie 48
Mei, Yijun 544 Wang, Hongzhi 159
Meng, Guang 497 Wang, Jianmin 379
Miao, Xu Na 705 Wang, Li-dong 535
Mizutani, Tetsuya 469 Wang, Nan 685
Wang, Qin 1
Na, Lei 369 Wang, Shuning 658
Nekoui, Mohammad Ali 563 Wang, Shuqing 487
Nelles, Oliver 241 Wang, Xiaolong 417
Nie, Qing 435 Wang, Yadong 159
Wang, Yu 298
Omid, M. 251, E1 Wang, Yuanqing 544
Ou, Kairi 517 Wang, Zaicheng 715
Ouyang, Chin-Lun 526 Wei, Xueye 298
OuYang, Dantong 685 Wu, Chao-Chin 579
 Author Index 737

Wu, Haifeng 121 Zhang, Hao 668

Wu, Xiaojin 298 Zhang, Hongqiang 350
Wu, Yan 141 Zhang, Jin 103
Zhang, Ke 260
Xi, Liang 112 Zhang, Sen 479
Xia, Fei 668 Zhang, Shaojun 159
Xia, Feng 288 Zhang, Weihong 150
Xu, Jun 658 Zhang, Xianchao 288
Xu, Yi 9 Zhang, Xue 453
Xu, Yuhui 379, 639 Zhang, Yinhai 1
Xu, Yusheng 379, 639 Zhang, Zhen 270
Xu, Zhenzhen 288 Zhang, Zipeng 487
Xue, Heng 715 Zhao, Chengli 685
Zhao, Jinwei 231
Yan, Guirong 231
Yan, Shu Ying 479 Zhao, Liang 619
Yang, Bo 649 Zhao, Weidong 121
Yang, Chunyang 677 Zhao, Weigang 48
Yang, Jun 444 Zhao, Yuelong 517
Yang, Yang 270 Zheng, Enlai 724
Ying, Li 369 Zheng, Jia-heng 399
You, Xiaoming 69 Zheng, Zhonglong 313
Yuan, Huai 677 Zhong, ShangPing 453
Yuan, Xiaohui 487 Zhou, Xian Wei 705
Zhu, Fengxia 215
Zhan, Shouyi 435 Zhu, Jihong 150
Zhang, Caiqing 649 Zhu, Shi-jiao 444
Zhang, Fengbin 112 Ziarati, Koorush 167
Zhang, Guoxing 590 Zupančič–Božič, Damjana 241