LECTURE NOTES IN ELECTRICAL ENGINEERING 39

Sio-Iong Ao
Len Gelman
Editors

Advances in
Electrical
Engineering and
Computational
Science
1 23
Advances in Electrical Engineering
and Computational Science
Lecture Notes in Electrical Engineering
Volume 39

For further volumes:

http://www.springer.com/series/7818
Sio-Iong Ao • Len Gelman
Editors

Advances in Electrical
Engineering and
Computational Science

123
Editors
Dr. Sio-Iong Ao Dr. Len Gelman
Harvard School of Engineering Cranfield University
and Applied Sciences School of Engineering
Harvard University Dept. of Process and Systems
60 Oxford Street Engineering
Cambridge Cranfield, Beds
MA 02138, USA United Kingdom MK43 0AL
siao@seas.harvard.edu

ISBN 978-90-481-2310-0 e-ISBN 978-90-481-2311-7

DOI 10.1007/978-90-481-2311-7

Library of Congress Control Number: 2009926266

c Springer Science+Business Media B.V. 2009

No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work.

Printed on acid-free paper

9 8 7 6 5 4 3 2 1

springer.com
Preface

A large international conference on Advances in Electrical Engineering and

Computational Science was held in London, UK, July 2–4, 2008, under the World
Congress on Engineering (WCE 2008). The WCE 2008 was organized by the
International Association of Engineers (IAENG); the Congress details are available
at: http://www.iaeng.org/WCE2008. IAENG is a non-profit international associa-
tion for engineers and computer scientists, which was founded originally in 1968.
The World Congress on Engineering serves as good platforms for the engineering
community to meet with each other and exchange ideas. The conferences have also
struck a balance between theoretical and application development. The conference
committees have been formed with over 200 members who are mainly research
center heads, faculty deans, department heads, professors, and research scientists
from over 30 countries. The conferences are truly international meetings with a
high level of participation from many countries. The response to the Congress has
been excellent. There have been more than 1,000 manuscript submissions for the
WCE 2008. All submitted papers have gone through the peer review process, and
the overall acceptance rate is 57%.
This volume contains 61 revised and extended research articles written by promi-
nent researchers participating in the conference. Topics covered include Control
Engineering, Network Management, Wireless Networks, Biotechnology, Signal
Processing, Computational Intelligence, Computational Statistics, Internet Comput-
ing, High Performance Computing, and industrial applications. The book offers the
state of the art of tremendous advances in electrical engineering and computational
science and also serves as an excellent reference work for researchers and graduate
students working on electrical engineering and computational science.

Dr. Sio-Iong Ao
Prof. Len Gelman

v
Contents

1 Acoustic Behavior of Squirrel Cage Induction Motors : : : : : : : : : : : : : 1

C. Grabner
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Acoustic Noise Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Sound Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Sound Pressure Level in Decibel . . . . . . . . . . . . . . . . . . . . . 2
1.2.3 Sound Pressure Level and Human Hearing . . . . . . . . . . . . 3
1.2.4 Human Hearing Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.5 Loudness Level in Phon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.6 Weighing the Sound Pressure Level . . . . . . . . . . . . . . . . . . 5
1.3 Mutation of Stator and Rotor Parts . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Direct Comparison of Sound Pressure Levels for Different
Motor Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 Evaluation of Design Impacts to the Equivalent Sound
Pressure Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.2 Fast Fourier Sound Pressure Spectrum for Each Motor
Design at Rated-Load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Model Reduction of Weakly Nonlinear Systems : : : : : : : : : : : : : : : : : : : 13

Marissa Condon and Georgi G. Grahovski
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Perturbative Approximation of Nonlinear Systems . . . . . . . . . . . . . . 14
2.3 KRYLOV-Based Model Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Scale Invariance Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 KRYLOV Reduction of Perturbative Representation . . . . . . . . . . . . 19
2.6 Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

vii
viii Contents

3 Investigation of a Multizone Drift Doping Based Lateral Bipolar

Transistor on Buried Oxide Thick Step : : : : : : : : : : : : : : : : : : : : : : : : : : 23
Sajad A. Loan, S. Qureshi, and S. Sundar Kumar Iyer
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Ideal, Conventional and Proposed Devices . . . . . . . . . . . . . . . . . . . . 24
3.3 Simulation Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 Development of a Battery Charging Management Unit

for Renewable Power Generation : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 33
Yong Yin, Jihong Wang, Xing Luo, and Shen Guo
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Energy Conversion Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Lead Acid Battery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.4 Battery Charging and Discharging Strategies . . . . . . . . . . . . . . . . . . 37
4.4.1 Membership Function Database . . . . . . . . . . . . . . . . . . . . . 38
4.4.2 Fuzzification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4.3 Fuzzy Rule Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4.4 Defuzzification and Simulation . . . . . . . . . . . . . . . . . . . . . . 42
4.5 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.6 Conclusions and Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5 Optical CDMA Transceiver Architecture: Polarization Modulation

with Dual-Balanced Detection : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 47
Mohammad Massoud Karbassian and Hooshang Ghafouri-Shiraz
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 POLSK–OCDMA Transmitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3 Analysis of POLSK–OCDMA Receiver . . . . . . . . . . . . . . . . . . . . . . 50
5.4 Discussion of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

6 Template Based: A Novel STG Based Logic Synthesis

for Asynchronous Control Circuits : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 59
Sufian Sudeng and Arthit Thongtak
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.2 Petri-Net and Signal Transition Graph . . . . . . . . . . . . . . . . . . . . . . . . 61
6.3 Asynchronous DMA Controller Specification . . . . . . . . . . . . . . . . . . 62
6.4 Synthesis Using State Based Method . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.5 Synthesis Using Structural Encoding Method . . . . . . . . . . . . . . . . . . 66
6.6 Synthesis Using State Based Method . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.7 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Contents ix

7 A Comparison of Induction Motor Speed

Estimation Using Conventional MRAS
and an AI-Based MRAS Parallel System : : : : : : : : : : : : : : : : : : : : : : : : 75
Chao Yang and Professor John W. Finch
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.2 Speed Estimation Using Conventional Model Reference
Adaptive System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.3 Artificial Intelligence-Based Model Reference Adaptive System . . 77
7.4 MRAS Based Two-Layer ANN Speed Estimator
with Dynamic Reference Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7.5 Simulation Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
8 A New Fuzzy-Based Additive Noise Removal Filter : : : : : : : : : : : : : : : 87
M. Wilscy and Madhu S. Nair
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.2 Conventional Noise Removal Techniques . . . . . . . . . . . . . . . . . . . . . 89
8.3 Proposed Fuzzy Noise Removal Filter . . . . . . . . . . . . . . . . . . . . . . . . 90
8.3.1 Fuzzy Sub-filter I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
8.3.2 Fuzzy Sub-filter II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
8.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
9 Enhancement of Weather Degraded Color Images and Video
Sequences Using Wavelet Fusion : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 99
Jisha John and M. Wilscy
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.2 Atmospheric Scattering Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
9.3 Contrast Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.3.1 Normalized Brightness and Airlight . . . . . . . . . . . . . . . . . . 102
9.3.2 Correcting Contrast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.4 Intensity Based Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.5 Wavelet Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.6 Video Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.7 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
9.8 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
9.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data : : 111

L.M. San-José-Revuelta
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
10.2 Brief Tracking Algorithm Description . . . . . . . . . . . . . . . . . . . . . . . . 113
10.2.1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
x Contents

10.2.2 Anisotropy and Local Probability . . . . . . . . . . . . . . . . . . . . 114

10.2.3 Eigenvectors and Direction Considerations . . . . . . . . . . . . 114
10.2.4 Path Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
10.2.5 Final Criterion and Pool of “Future Seeds” . . . . . . . . . . . . 115
10.2.6 Parameters to Be Estimated . . . . . . . . . . . . . . . . . . . . . . . . . 115
10.3 Proposed GA for Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . 116
10.3.1 Estimation Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
10.3.2 GA Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
10.3.3 Genetic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
10.3.4 Elitism and Entropy Dependent Operators . . . . . . . . . . . . . 117
10.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
10.4.1 Synthetic Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
10.4.2 Real Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
10.4.3 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
10.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
11 A Bridge-Ship Collision Avoidance System Based on FLIR
Image Sequences : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 123
Jun Liu, Hong Wei, Xi-Yue Huang, Nai-Shuai He, and Ke Li
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
11.2 The FLIR Video Surveillance System . . . . . . . . . . . . . . . . . . . . . . . . 125
11.3 The Detection Algorithm for Moving Ships . . . . . . . . . . . . . . . . . . . 126
11.3.1 Extracting the ROS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
11.3.2 Calculating the Multi-scale Fractal Feature . . . . . . . . . . . . 126
11.3.3 Segmenting the Fractal Feature Image . . . . . . . . . . . . . . . . 129
11.3.4 Detecting Moving Ships . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
11.4 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 130
11.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

12 A Gray Level Feature Detector and Its Hardware Architecture : : : : : 135

Neeta Nain, Rajesh Kumar, and Bhavitavya Bhadviya
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
12.2 Feature Point Detection Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
12.3 Hardware Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
12.3.1 Architecture 1: A Complete Hardwired Approach . . . . . . 139
12.3.2 Architecture 2: A Hardware Software Co-design . . . . . . . 141
12.4 Synthesis Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
12.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Contents xi

13 SVD and DWT-SVD Domain Robust Watermarking using

Differential Evolution Algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 147
Veysel Aslantas
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
13.2 SVD Domain and DWT-SVD Domain Watermarking
Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
13.3 Optimal Watermarkings using DE . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
13.3.1 Basic Concepts of DE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
13.3.2 Optimisation of Scaling Factors . . . . . . . . . . . . . . . . . . . . . . 152
13.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
13.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

14 Design and Performance Evaluation of a Prototype Large Ring

PET Scanner : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 161
M. Monjur Ahasan and David J Parker
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
14.2 The Macropet Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
14.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
14.3.1 Spatial Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
14.3.2 Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
14.3.3 Scatter Fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
14.3.4 Count Rate Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
14.3.5 Imaging Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
14.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

15 Robust Wavelet-Based Video Watermarking Using Edge Detection : : 173

Tamás Polyák and Gábor Fehér
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
15.2 The Watermark Embedding Process . . . . . . . . . . . . . . . . . . . . . . . . . . 174
15.3 The Watermark Detection Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
15.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
15.4.1 Quality Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
15.4.2 Robustness of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 179
15.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

16 High Speed Soft Computing Based Circuit for Edges Detection

in Images : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 183
Nashaat M. Hussein and Angel Barriga
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
16.2 Edge Detection Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
16.2.1 The Filter Stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
16.2.2 The Threshold Stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
16.2.3 The Edge Detection Stage . . . . . . . . . . . . . . . . . . . . . . . . . . 187
xii Contents

16.3 Hardware Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

16.3.1 The Threshold Generation Circuit . . . . . . . . . . . . . . . . . . . . 189
16.3.2 Design of the Edge Detection System . . . . . . . . . . . . . . . . . 190
16.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

17 A Comparison Between 3D OSEM and FBP Image Reconstruction

Algorithms in SPECT : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 195
Khalid S. Alzimami, Salem A. Sassi, and Nicholas M. Spyrou
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
17.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
17.2.1 Analytical Reconstruction Algorithms: Filtered
Backprojection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
17.2.2 Iterative Reconstruction Algorithms . . . . . . . . . . . . . . . . . . 198
17.2.3 Physical Factors Affecting Quantitative and Qualitative
Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
17.3 A Comparison of 3D OSEM with FBP Image Reconstruction
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
17.3.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
17.3.2 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
17.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

18 Performance Improvement of Wireless MAC Using

Non-Cooperative Games : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 207
Debarshi Kumar Sanyal, Matangini Chattopadhyay, and Samiran
Chattopadhyay
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
18.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
18.3 Game-Theoretic Model of Medium Access Control . . . . . . . . . . . . . 209
18.4 Distributed Mechanisms to Achieve Nash Equilibrium . . . . . . . . . . 212
18.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
18.5.1 CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
18.5.2 AT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
18.5.3 CO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
18.5.4 SW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
18.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
18.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

19 Performance Evaluation of Mobile Ad Hoc Networking Protocols : : : 219

Nadia Qasim, Fatin Said, and Hamid Aghvami
19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
19.2 Mobile Ad Hoc Network’s Routing Protocols . . . . . . . . . . . . . . . . . . 221
19.3 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
19.4 Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
19.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Contents xiii

19.5.1 Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

19.5.2 End to End Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
19.5.3 Media Access Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
19.5.4 Packet Delivery Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
19.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

20 IEEE 802.11E Block Acknowledgement Policies : : : : : : : : : : : : : : : : : : 231

O. Cabral, A. Segarra, and F. J. Velez
20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
20.2 IEEE 802.11e and Block Acknowledgement Description . . . . . . . . 232
20.2.1 IEEE 802.11e User Priorities and Access Categories . . . . 232
20.2.2 Block Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
20.3 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
20.4 System, Scenario and Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . 235
20.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
20.5.1 Block Acknowledgement with Standalone Services . . . . . 237
20.5.2 Block Acknowledgement with Mixtures
of Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
20.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

21 Routing in a Custom-Made IEEE 802.11E Simulator : : : : : : : : : : : : : : 243

João Miguel Ferro and Fernando J. Velez
21.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
21.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
21.3 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
21.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
21.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
22 Web-Based Management of Distributed Services : : : : : : : : : : : : : : : : : : 255
George Oikonomou and Theodore Apostolopoulos
22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
22.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
22.3 WebDMF: A Web-Based Management Framework
for Distributed Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
22.3.1 Management and Service Nodes . . . . . . . . . . . . . . . . . . . . . 258
22.3.2 Management Representative . . . . . . . . . . . . . . . . . . . . . . . . 258
22.3.3 Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
22.3.4 CIM Schemas and Operations . . . . . . . . . . . . . . . . . . . . . . . 260
22.4 Implementation and Performance Evaluation . . . . . . . . . . . . . . . . . . 262
22.4.1 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
22.4.2 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
22.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
xiv Contents

23 Image Index Based Digital Watermarking Technique for Ownership

Claim and Buyer Fingerprinting : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 267
Sarabjeet S. Bedi, Rabia Bano, and Shekhar Verma
23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
23.2 Image Key and Buyer Fingerprint . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
23.2.1 Image Key . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
23.2.2 Buyer Fingerprint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
23.3 Proposed Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
23.3.1 Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
23.3.2 Generation of Watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
23.3.3 Insertion of Watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
23.3.4 Extraction of Watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
23.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
23.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276

24 Reverse Engineering: EDOWA Worm Analysis and Classification : : : 277

Madihah Mohd Saudi, Emran Mohd Tamil, Andrea J Cullen,
Mike E Woodward, and Mohd Yamani Idna Idris
24.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
24.2 Method of Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
24.2.1 Worm Analysis Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
24.2.2 Loading Specimen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
24.3 EDOWA Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
24.3.1 Infection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
24.3.2 Activation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
24.3.3 Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
24.3.4 Operating Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
24.3.5 Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
24.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
25 Reconfigurable Hardware Implementation
of a GPS-Based Vehicle Tracking System : : : : : : : : : : : : : : : : : : : : : : : : 289
Adnan Yaqzan, Issam Damaj, and Rached Zantout
25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
25.2 Upgrading the Aram Locator GPS System . . . . . . . . . . . . . . . . . . . . . 290
25.3 The FPGA-Based Aram System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
25.3.1 The Intersystem Process 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 293
25.3.2 The Intersystem Process 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 294
25.3.3 The Memory Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
25.3.4 Communication Protocols: I2C and UART . . . . . . . . . . . . 296
25.4 Performance Analysis and Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 296
25.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Contents xv

26 Unknown Malicious Identification : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 301

Ying-xu. Lai and Zeng-hui. Liu
26.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
26.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
26.3 Naı̈ve Bayesian and Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
26.3.1 String Extraction and Elimination . . . . . . . . . . . . . . . . . . . . 303
26.3.2 Naı̈ve Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
26.4 Increment Naı̈ve Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
26.4.1 Naı̈ve Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
26.4.2 Multi-naı̈ve Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
26.4.3 Half Increment Bayes (HIB) . . . . . . . . . . . . . . . . . . . . . . . . 306
26.4.4 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
26.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
26.5.1 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
26.5.2 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
26.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
27 Understanding Programming Language Semantics
for the Real World : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 313
Trong Wu
27.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
27.2 Extension of the Applications to a Real World . . . . . . . . . . . . . . . . . 315
27.3 Modularization and Communication in the Chaotic World . . . . . . . 316
27.4 Parallel or Concurrency Is the Nature of the World . . . . . . . . . . . . . 319
27.5 Accuracy and Efficiency Are Required by This Sophicificate
World . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
27.6 Exception Handling Is the Safe Guard in the Dangerous World . . . 324
27.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
28 Analysis of Overhead Control Mechanisms
in Mobile AD HOC Networks : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 329
S. Gowrishankar, T.G. Basavaraju, and SubirKumarSarkar
28.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
28.2 Theoretical Analysis of Overhead in Hierarchical
Routing Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
28.3 Theoretical Analysis and Overhead Minimizing Techniques for
AD HOC Networks Using Clustering Mechanisms . . . . . . . . . . . . . 332
28.4 Minimizing Overhead in AD HOC Networks by Header
Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
28.5 Minimizing Overhead for AD HOC Networks Connected
to Internet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
xvi Contents

29 Context Aware In-Vehicle Infotainment Interfaces : : : : : : : : : : : : : : : : 343

H. Sharma and A.K. Ramani
29.1 In-Vehicle Infotainment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
29.1.1 Infotainment System Interaction . . . . . . . . . . . . . . . . . . . . . 344
29.2 Context Aware Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
29.2.1 Context Aware Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
29.2.2 Context Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
29.3 Interaction Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
29.3.1 Layered Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
29.3.2 Component Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
29.3.3 Framework Core . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
29.3.4 Distribution and Logical Mobility . . . . . . . . . . . . . . . . . . . . 351
29.3.5 Application Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
29.4 Application Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
29.4.1 Application Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
29.4.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
29.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354

30 Facial Expression Analysis Using PCA : : : : : : : : : : : : : : : : : : : : : : : : : : 355

V. Praseeda Lekshmi, M. Sasikumar, Divya S. Vidyadharan,
and S. Naveen
30.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
30.2 Background and Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
30.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
30.3.1 Training stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
30.3.2 Face Recognition and Expression Classification . . . . . . . . 358
30.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
30.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
31 The Design of a USART IP Core : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 365
A. H. El-Mousa, N. Anssari, A. Al-Suyyagh, and H. Al-Zubi
31.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
31.2 USART Theory of Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
31.2.1 Modes of Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
31.3 Specifications of the Developed USART . . . . . . . . . . . . . . . . . . . . . . 368
31.4 USART System Design Methodology . . . . . . . . . . . . . . . . . . . . . . . . 368
31.4.1 Transmitter Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
31.4.2 Receiver Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
31.5 Testing and Verification Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . 373
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Contents xvii

32 Multilayer Perceptron Training Optimization for High Speed

Impacts Classification : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 377
Angel Garcia-Crespo, Belen Ruiz-Mezcua, Israel Gonzalez-Carrasco,
and Jose Luis Lopez-Cuadrado
32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
32.2 Ballistic Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
32.3 Use of ANN in Impacts Situations . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
32.3.1 Impact Scenario Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 381
32.3.2 Network Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
32.3.3 Input Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
32.4 Solution Proposed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
32.4.1 Randomness Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
32.4.2 Determination of the Influence of the Independent
Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
32.5 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
32.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
33 Developing Emotion-Based Pet Robots : : : : : : : : : : : : : : : : : : : : : : : : : : 389
W.-P. Lee, J.-W. Kuo, and P.-C. Lai
33.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
33.2 Building Emotion-Based Pet Robots . . . . . . . . . . . . . . . . . . . . . . . . . 390
33.2.1 A User-Oriented Framework . . . . . . . . . . . . . . . . . . . . . . . . 390
33.2.2 Control Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
33.2.3 User–Robot Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
33.2.4 Emotion Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
33.3 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
33.3.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
33.3.2 Building Emotion-Based Behavior Selector . . . . . . . . . . . . 396
33.4 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400

34 Designing Short Term Trading Systems with Artificial Neural

Networks : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 401
Bruce Vanstone, Gavin Finnie, and Tobias Hahn
34.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
34.2 Review of Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
34.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
34.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
34.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409

35 Reorganising Artificial Neural Network Topologies : : : : : : : : : : : : : : : : 411

Thomas D. Jorgensen, Barry Haynes, and Charlotte Norlund
35.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
35.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
xviii Contents

35.3 Neuroscientific Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413

35.4 Reorganising Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
35.4.1 Neural Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
35.4.2 Using the Complexity Measure . . . . . . . . . . . . . . . . . . . . . . 415
35.4.3 The Simulated Track and Robot . . . . . . . . . . . . . . . . . . . . . 415
35.5 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
35.5.1 The Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . 417
35.5.2 The Fitness Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
35.5.3 The Benchmark and the Reorganised Networks . . . . . . . . 417
35.5.4 Evaluation and Comparison of the Proposed Methods . . . 419
35.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421

36 Design and Implementation of an E-Learning Model

by Considering Learner’s Personality and Emotions : : : : : : : : : : : : : : 423
S. Fatahi, M. Kazemifard, and N. Ghasem-Aghaee
36.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
36.2 Previous Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
36.3 Psychological Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
36.4 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
36.5 Virtual Classmate Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
36.6 Simulation of Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
36.7 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
36.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
36.9 Conclusion and Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433

37 A Self Progressing Fuzzy Rule-Based System for Optimizing and

Predicting Machining Process : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 435
Asif Iqbal and Naeem U. Dar
37.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
37.2 System Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
37.3 The Self-Development Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
37.3.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
37.3.2 Self-Development of Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . 439
37.3.3 Self-Development of Prediction Rule-Base . . . . . . . . . . . . 439
37.3.4 Self-Development of Optimization Rule-Base . . . . . . . . . . 442
37.4 Optimization and Prediction of Machining Process . . . . . . . . . . . . . 442
37.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
38 Selection of Ambient Light for Laser Digitizing
of Quasi-Lambertian Surfaces : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 447
D. Blanco, P. Fernández, E. Cuesta, C. M. Suárez, and N. Beltrán
38.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
38.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Contents xix

38.3 Configuration of the Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449

38.4 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
38.5 Analysis Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
38.6 Results Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
38.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457

39 Ray-Tracing Techniques Applied to the Accessibility Analysis for

the Automatic Contact and Non Contact Inspection : : : : : : : : : : : : : : : 459
B. J. Álvarez, P. Fernández, J.C. Rico, and G.Valiño
39.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
39.2 Analysis Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
39.3 Analysis Considering Inspection Devices as Infinite Half-Lines . . . 460
39.4 Interference Analysis Considering Real Dimensions
of the Inspection Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
39.5 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
39.6 Application Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
39.6.1 Inspection Process by Means of a Touch-Trigger Probe . . 465
39.6.2 Scanning Process by Means of a Laser Stripe System . . . 466
39.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469

40 Detecting Session Boundaries to Personalize Search Using

a Conceptual User Context : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 471
Mariam Daoud, Mohand Boughanem, and Lynda Tamine-Lechani
40.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
40.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
40.3 Detecting Session Boundaries to Personalize Search . . . . . . . . . . . . 473
40.3.1 A Conceptual User Context Modelling for
Personalizing Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
40.3.2 How to Detect Session Boundaries . . . . . . . . . . . . . . . . . . . 475
40.4 Experimental Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
40.4.1 Experimental Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
40.4.2 Experimental Design and Results . . . . . . . . . . . . . . . . . . . . 477
40.5 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481

41 Mining Weather Information in Dengue Outbreak:

Predicting Future Cases Based on Wavelet, SVM and GA : : : : : : : : : : 483
Yan Wu, Gary Lee, Xiuju Fu, Harold Soh, and Terence Hung
41.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484
41.2 Model Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
41.2.1 Data Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
41.2.2 Wavelet Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
41.2.3 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
41.2.4 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
xx Contents

41.2.5 Regression Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488

41.2.6 Model Assessment Criteria . . . . . . . . . . . . . . . . . . . . . . . . . 488
41.3 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
41.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493

42 PC Tree: Prime-Based and Compressed Tree for Maximal

Frequent Patterns Mining : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 495
Mohammad Nadimi-Shahraki, Norwati Mustapha, Md Nasir B
Sulaiman, and Ali B Mamat
42.1 Problem of Maximal Frequent Patterns Mining . . . . . . . . . . . . . . . . 495
42.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
42.3 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
42.3.1 Database Encoding Technique . . . . . . . . . . . . . . . . . . . . . . . 497
42.3.2 PC Tree Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
42.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
42.5 Conclusion and Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504

43 Towards a New Generation of Conversational Agents Based

on Sentence Similarity : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 505
Karen O’Shea, Dr. Zuhair Bandar, and Dr. Keeley Crockett
43.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
43.2 Sentence Similarity Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
43.3 Traditional CA Scripting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
43.4 CA Scripting Using Sentence Similarity . . . . . . . . . . . . . . . . . . . . . . 510
43.5 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
43.5.1 Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
43.5.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
43.6 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
43.7 Conclusions and Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514

44 Direction-of-Change Financial Time Series Forecasting Using

Neural Networks: A Bayesian Approach : : : : : : : : : : : : : : : : : : : : : : : : : 515
Andrew A. Skabar
44.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
44.2 MLPs for Financial Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
44.3 Bayesian Methods for MLPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
44.4 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
44.4.1 Input Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
44.4.2 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520
44.4.3 Setting the Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
44.4.4 MCMC Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
44.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524
Contents xxi

45 A Dynamic Modeling of Stock Prices and Optimal Decision

Making Using MVP Theory : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 525
Ramin Rajabioun and Ashkan Rahimi-Kian
45.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
45.2 Modeling and Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
45.3 Virtual Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
45.4 The Market Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
45.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536

46 A Regularized Unconstrained Optimization in the Bond Portfolio

Valuation and Hedging : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 539
Yury Gryazin and Michael Landrigan
46.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 539
46.2 Option-Adjusted Spread Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
46.3 The Spread Variance Minimization Approach . . . . . . . . . . . . . . . . . . 542
46.4 Numerical Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
46.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
46.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
47 Approximation of Pareto Set in Multi Objective Portfolio
Optimization : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 551
I. Radziukyniene and A. Zilinskas
47.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
47.2 Multi-Objective Portfolio Optimization Problem . . . . . . . . . . . . . . . 552
47.3 Multi-Objective Optimization by the Method of Adjustable
Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
47.4 Evolutionary Methods for Multi-Objective Optimization . . . . . . . . 556
47.5 Description of Experimental Investigation . . . . . . . . . . . . . . . . . . . . . 558
47.6 Discussion on Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . 559
47.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561

48 The Scaling Approach for Credit Limit Management and Its

Application : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 563
M.Y. Konovalikhhin, D.O. Sergienko and O.V. Balaeva
48.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563
48.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
48.2.1 Model Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
48.2.2 B3 Coefficient Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . 566
48.2.3 B2 Coefficient Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . 569
48.2.4 B1 Coefficient Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . 570
48.3 Analysis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
48.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
xxii Contents

49 Expected Tail Loss Efficient Frontiers for CDOS of Bespoke

Portfolios Under One-Factor Copula Marginal Distributions : : : : : : : 575
Diresh Jewan, Renkuan Guo, and Gareth Witten
49.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
49.2 Bespoke CDO Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
49.3 Heavy-Tail Modelling and Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . 577
49.3.1 Stable Paretian Distributions . . . . . . . . . . . . . . . . . . . . . . . . 577
49.3.2 Copulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
49.4 Credit Risk Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
49.4.1 Convex and Coherent Risk Measures . . . . . . . . . . . . . . . . . 579
49.4.2 Expected-Tail-Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
49.4.3 Portfolio Optimisation Under Expected-Tail-Loss . . . . . . 580
49.5 Copula Marginal ETL Efficient Frontier for the CDO Collateral . . 581
49.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585

50 Data Mining for Retail Inventory Management : : : : : : : : : : : : : : : : : : : 587

Pradip Kumar Bala
50.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 587
50.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
50.3 Purchase Dependencies in Retail Sale in the Context
of Inventory Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
50.4 Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
50.5 Illustration with an Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
50.6 Case Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594
50.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
51 Economic Process Capability Index for Product Design and Process
Planning : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 599
Angus Jeang
51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
51.2 Process Capability Indices (PCI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
51.3 Proposed Process Capability Index . . . . . . . . . . . . . . . . . . . . . . . . . . . 603
51.3.1 Single Quality Characteristic . . . . . . . . . . . . . . . . . . . . . . . . 603
51.3.2 Multiple Quality Characteristics . . . . . . . . . . . . . . . . . . . . . 605
51.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608

52 Comparing Different Approaches for Design of Experiments (DoE) : : 611

Martı́n Tanco, Elisabeth Viles, and Lourdes Pozueta
52.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
52.2 Approaches to Design of Experiments . . . . . . . . . . . . . . . . . . . . . . . . 612
52.2.1 The Classical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
52.2.2 The Taguchi Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
52.2.3 The Shainin Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
Contents xxiii

52.3 Limitations of the Different Approaches . . . . . . . . . . . . . . . . . . . . . . 616

52.3.1 The Classical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
52.3.2 Taguchi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
52.3.3 Shainin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
52.4 Conclusions and Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . 618
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619

53 Prevention of Workpice Form Deviations in CNC Turning Based

on Tolerance Specifications : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 623
Gonzalo Valiño Riestra, David Blanco Fernandez, Braulio Jose Álvarez,
Sabino Mateos Diaz, and Natalia Beltrán Delgado
53.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
53.2 Form Errors in Turning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
53.3 Calculation of Radial Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
53.4 Relationship Between Deviations and Tolerances . . . . . . . . . . . . . . . 627
53.4.1 Length Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
53.4.2 Diametrical Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 629
53.4.3 Geometrical Tolerance of Flatness . . . . . . . . . . . . . . . . . . . 629
53.4.4 Geometrical Tolerance of Cylindricity . . . . . . . . . . . . . . . . 630
53.4.5 Geometrical Tolerance of Profile of a Surface . . . . . . . . . . 630
53.4.6 Geometrical Tolerance of Parallelism . . . . . . . . . . . . . . . . . 631
53.4.7 Geometrical Tolerance of Perpendicularity . . . . . . . . . . . . 631
53.4.8 Geometrical Tolerance of Coaxiality . . . . . . . . . . . . . . . . . . 631
53.4.9 Geometrical Tolerance of Circular Runout . . . . . . . . . . . . . 631
53.4.10 Geometrical Tolerance of Total Runout . . . . . . . . . . . . . . . 632
53.5 Deviations as Optimization Conditions . . . . . . . . . . . . . . . . . . . . . . . 632
53.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
54 Protein–Protein Interaction Prediction Using Homology
and Inter-domain Linker Region Information : : : : : : : : : : : : : : : : : : : : 635
Nazar Zaki
54.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
54.1.1 The Importance of Protein–Protein Interaction . . . . . . . . . 635
54.1.2 Current Methods to Predict PPI . . . . . . . . . . . . . . . . . . . . . . 636
54.1.3 Computational Approaches to Predict PPI . . . . . . . . . . . . . 636
54.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
54.2.1 Similarity Measures Between Protein Sequences . . . . . . . 637
54.2.2 Identify and Eliminate Inter-domain Linker Regions . . . . 639
54.2.3 Detecting Domain Matches and Associated Structural
Relationships in Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
54.3 Experimental Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
54.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
54.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
xxiv Contents

55 Cellular Computational Model of Biology : : : : : : : : : : : : : : : : : : : : : : : : 647

Alex H. Leo and Gang Sun
55.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647
55.2 List of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650
55.3 Determining mRNA Polymerases, mR(n + 1) . . . . . . . . . . . . . . . . . . 650
55.4 Determining ˇ-Galactosidases, G.n C 1/ . . . . . . . . . . . . . . . . . . . . . 651
55.5 Determining Lactoses with in and out the Cell, Lext(n C 1)
and L(n C 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
55.6 Determining ATP Molecules, A(n) . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
55.7 The System Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
55.8 In Comparison to Traditional Biological Experiment . . . . . . . . . . . . 657
55.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
55.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
56 Patient Monitoring: Wearable Device for Patient Monitoring : : : : : : : 659
Robert G. Lupu, Andrei Stan, and Florina Ungureanu
56.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659
56.2 Wearable Monitoring Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 661
56.3 Functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
56.4 Firmware Issue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
56.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668

57 Face Recognition and Expression Classification : : : : : : : : : : : : : : : : : : : 669

Praseeda Lekshmi.V, Dr.M.Sasikumar, and Divya S Vidyadharan
57.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
57.2 Background and Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
57.3 Discrete Cosine Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
57.4 Radial Basis Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
57.5 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
57.5.1 Facial Region Identification . . . . . . . . . . . . . . . . . . . . . . . . . 672
57.5.2 Face Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
57.5.3 Facial Expression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 676
57.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
57.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 678
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679
58 Spiking Neurons and Synaptic Stimuli: Neural Response
Comparison Using Coincidence-Factor : : : : : : : : : : : : : : : : : : : : : : : : : : 681
Mayur Sarangdhar and Chandrasekhar Kambhampati
58.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
58.2 Neuronal Model and Synapse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682
58.2.1 The Neuron Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682
Contents xxv

58.2.2 The Synaptic Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683

58.2.3 The Total External Current . . . . . . . . . . . . . . . . . . . . . . . . . . 684
58.3 Comparison of Two Spike Trains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
58.3.1 Responses of the Neuron . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
58.3.2 Comparison of Responses . . . . . . . . . . . . . . . . . . . . . . . . . . 685
58.3.3 Coincidence-Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
58.3.4 Two-Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 688
58.3.5 Binary Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688
58.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 690
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 690

59 Overcoming Neuro-Muscular Arm Impairment by Means

of Passive Devices : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 693
Federico Casolo, Simone Cinquemani, and Matteo Cocetta
59.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
59.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
59.3 Experimental Test Apparatus Set Up . . . . . . . . . . . . . . . . . . . . . . . . . 695
59.4 Experimental Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
59.5 Device Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 699
59.5.1 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 701
59.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 702
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704

60 EEG Classification of Mild and Severe Alzheimer’s Disease Using

Parallel Factor Analysis Method : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 705
Charles-Francois Vincent Latchoumane, Francois-Benois Vialatte,
Jaeseung Jeong, and Andrzej Cichocki
60.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
60.1.1 Diagnosis of Alzheimer’s Disease and EEGs . . . . . . . . . . . 705
60.1.2 Multi-way Array Decomposition . . . . . . . . . . . . . . . . . . . . . 706
60.2 Subjects and EEG Recordings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
60.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
60.3.1 Three-Way Tensor for Analysis . . . . . . . . . . . . . . . . . . . . . . 707
60.3.2 Classification in Divide and Conquer Scheme . . . . . . . . . . 708
60.3.3 The PARAllel FACtor Analysis (PARAFAC) . . . . . . . . . . . 709
60.3.4 Filter Estimation and Feature Extraction . . . . . . . . . . . . . . 710
60.3.5 From Filter to Feature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
60.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
60.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714
61 Feature Selection of Gene Expression Data Based on Fuzzy
Attribute Clustering : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 717
Elham Chitsaz, Mohammad Taheri, and Seraj D. Katebi
61.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
61.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 718
xxvi Contents

61.3 The Proposed Fuzzy Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 720

61.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
61.4.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
61.4.2 Classification Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724
61.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Chapter 1
Acoustic Behavior of Squirrel Cage Induction
Motors

C. Grabner

Abstract The evaluation of distinct influences to the acoustic noise emission of

industrial standard squirrel cage motors is fairly a very sensitive and complex topic.
Main interest is thereby given to the influence of the complete mechanical induction
motor design to the almost manifold acoustic response under sinusoidal power sup-
ply. This includes the investigation of unexpected effects even caused by passing
critical speed values which are corresponding to mechanical resonance frequen-
cies. To take remedial measures, totally-closed stator topologies and rotor skewing
opportunities, have been deeply investigated in a comparatively manner. The per-
formed comparatively measurement of the speed dependent equivalent continuous
A-weighted surface sound pressure level gives a good overview about the principal
design influences.

Keywords Acoustic Behavior  Motor Design  Induction Motor  Sound Pressure

Level

1.1 Introduction

As the motor design process has to consider multi-physical effects, such as, e.g.
electrical, thermal and mechanical aspects, a very deep and complex analysis has to
take place in order to find the best technical and commercial choice for the product.
Thus, extended investigations to the acoustic behavior of the complete speed vari-
able induction drive system, as it is schematically depicted in Fig. 1.1, have been
performed in the lab to obtain reproducible results with high accuracy.

C. Grabner
Research and Development, Siemens AG, Frauenauracherstrasse 80, D-91056 Erlangen,
E-mail: Grabner.Christian@SIEMENS.COM

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 1

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
2 C. Grabner

Fig. 1.1 A 750 W 4-pole squirrel cage induction motor in FS80

1.2 Acoustic Noise Physics

Sound is the mechanical vibration of a gaseous medium through which the energy is
transferred away from the source by progressive sound waves. Whenever an object
vibrates, a small portion of the energy involved is lost to the surrounding area as
unwanted sound – the acoustic noise [1, 2].

1.2.1 Sound Power

Any source of noise is commonly characterized by the emitted sound power,

measured in watts. This fundamental physical property of the source itself is often
known as absolute parameter for the acoustic output.
The sound power can be found conceptually by adding the product of the areas
times the acoustic intensities for the areas on any hypothetical surface that con-
tains the source. Since equipment for measuring of acoustic intensity is generally
not available, sound power inevitably has to be determined indirectly by taking
the spatial average of the sound pressure squared measured on equal areas on the
surrounding surface.

1.2.2 Sound Pressure Level in Decibel

As the acoustic intensity, the power passing through a unit area in space, is propor-
tional in the far field to the square of the sound pressure, a convenient scale for the
measurement can be defined as sound pressure level
1 Acoustic Behavior of Squirrel Cage Induction Motors 3
 2  
p .t/ p .t/
Lp .t/ D 10 log D 20 log ; (1.1)
p0 p0

in decibel, with p0 as the reference sound pressure of 20  Pa. The measured sound
pressure p .t/ in (1) depends on many insecure factors, such as e.g. orientation and
distance of receiver, temperature and velocity gradients inside the involved medium.
The analysis of (1) in the frequency domain is fortunately done with the discrete
fast Fourier-analysis. Therefore, the time-periodical signal (1) is sampled as Lp;n
and further processed at a distinct number of N samples as

X
N1
OLp; D Lp;n ej.2n=N/ ;  D 0; 1; 2 : : : N  1 (1.2)
nD0

in order to obtain the Fourier coefficients OLp; of the interested frequency-domain.

1.2.3 Sound Pressure Level and Human Hearing

Unfortunately, no simple relationship exists between the measured physical sound

pressure level and the human perception of the same sound. The treatment of
acoustic noise and its effects is therefore a complicated problem, which must take
a wide variety of parameters into account to achieve good correlations between
measurement and the resultant human reaction.

1.2.4 Human Hearing Mechanism

The quietest sound at 1,000 Hz which can be heard by the average person is found
in Fig. 1.2 to be about 20 μ Pa and this value has been standardized as the nomi-
nal hearing threshold for the purpose of sound level measuring. At the other end
of the scale the threshold of pain occurs at a sound pressure of approximately
100 Pa.
An inspection of Fig. 1.2 shows that a pure tone having a sound pressure level of,
e.g. 20 dB and a frequency of 1,000 Hz is plainly audible, whereas one having the
same sound pressure level but a frequency of 100 Hz is well below the threshold of
audibility and cannot be heard at all.

1.2.5 Loudness Level in Phon

The loudness of a pure tone of constant sound pressure level, perhaps the simplest
acoustic signal of all, varies with its frequency, even though the sound pressure may
be the same in every case.
4 C. Grabner

Fig. 1.2 Typical sound pressure levels in decibel and equal loudness levels in phon for different
noise sources

Although our hearing mechanism is not well-adapted for making quantity mea-
surements of the relative loudness of different sounds, there is a fair agreement
between observers, when two pure tones of different frequencies appear to be
equally load. It is therefore possible to establish in Fig. 1.2 contour plots of equal
loudness in phone. These subjective felled loudness curves are obtained by alter-
nately sounding a reference tone of 1,000 Hz and a second tone of some other
frequency. The intensity level of the second tone is then adjusted to the value that
makes the two tones appear equally load.
1 Acoustic Behavior of Squirrel Cage Induction Motors 5

Fig. 1.3 Standardized weighting curves for sound level measurement

A pure ton having a frequency of 100 Hz and a sound pressure level of about
35 dB sounds as load as a pure 1,000 Hz tone whose intensity level is 20 dB, and
hence the loudness level of the 100 Hz tone is by definition 20 phon. The 1,000 Hz
frequency is thus the reference for all loudness measurements, and all contours
of equal loudness expressed in phon have the same numerical value as the sound
pressure level in decibel at 1,000 Hz.

1.2.6 Weighing the Sound Pressure Level

Due to the human ears assessment of loudness, the defined purely physical sound
pressure terminus has to be modified by an implicit weighting process in a way
which corresponds to the more complex response of the human being. Therefore,
among several possibilities, distinguished by B, C, D, E or SI-weighting in Fig. 1.3,
the A-weighted level has be found to be the most suitable for modifying the fre-
quency response to follow approximately the equal loudness of 20 phon in Fig. 1.2.

1.3 Mutation of Stator and Rotor Parts

The stator part of a squirrel cage induction motor is typically manufactured so

far with semi-closed stator slots [3]. However, continual progress in the industrial
automation process allows the utilization of a novel stator construction with totally-
closed stator slots as shown in Fig. 1.4. The squirrel cage rotor could alternatively
be built up with a skewed or even non-skewed rotor as depicted in Fig. 1.5. Based on
the proposed different stator and rotor components, four motor designs, as listed in
6 C. Grabner

Fig. 1.4 Stator design with totally-closed (left) and semi-closed (right) stator slots

Fig. 1.5 Skewed (left) and un-skewed (right) rotor

Table 1.1 Test results with regard to their suitability for acoustic noise minimization
Combination Stator slot design Rotor design
Motor ① Totally-closed Skewed
Motor ② Totally-closed Un-skewed
Motor ③ Semi-closed Un-skewed
Motor ④ Semi-closed Skewed

Table 1.1, have been consecutively tested with regard to their suitability for acoustic
noise minimization.

1.4 Direct Comparison of Sound Pressure Levels for Different

Motor Designs

All investigated motor designs are operated at the rated electrical voltage level of
400 V. They have a rated power of approximately 750 W at the nominal speed of
about 1,395 rpm. As all listed samples in Table 1.1, are assisted with the same motor
fan, a direct comparison of the acoustic tests results is feasible.
1 Acoustic Behavior of Squirrel Cage Induction Motors 7

The noise measurement has been carried out for each motor design over a
wide speed range beginning from the lower value of 500 rpm up to rated-speed of
1,395 rpm at constant mechanical rated-torque of 5 Nm. Within increasing higher
speed ranges, the drive works in the field weakening range and so the load has been
continuously reduced.
After reaching the physically quasi-steady operational state for each adjusted
speed value, the measured data set has been stored. Thus, purely dynamic acoustic
noise in fact of a transient real-time run-up is herewith not considered.

1.4.1 Evaluation of Design Impacts to the Equivalent Sound

Pressure Level

The depicted sound pressure level courses in Figs. 1.6 and 1.7, which are repre-
senting the skewed rotor designs ① and ④, show a very smooth and continuous
track behavior, which fortunately avoids extensively noise peaks over the total speed
range. Thus, the utilization of the standard skewing technology avoids the generation
of undesired significant noise peak over the complete speed range.
As obviously from Figs. 1.6 and 1.7, the course of the proposed novel motor
combination ① has advantageously the lowest sound pressure level in comparison
to the usually design ④ at several speed values.
The introduced novel motor topology of the standard squirrel cage induction
motor with totally-closed slot tips is therefore suitable to reduce undesired noise
emission not only at the nominal speed of 1,395 rpm. The emitted noise of the pro-
posed motor design ① is reduced by the amount of 10 dB at no-load in Fig. 1.6 and
about 8 dB at rated-load in Fig. 1.7 in comparison to the state of the art design ④.

Fig. 1.6 Measured sound pressure level versus speed for the motor ① (blue), motor ② (red), motor
③ (green) and motor ④ (violet) at sinusoidal power supply and no-load
8 C. Grabner

Fig. 1.7 Measured sound pressure level versus speed for the motor ① (blue), motor ② (red), motor
③ (green) and motor ④ (violet) at sinusoidal power supply and rated-load

Contrarily, in case of both test objects ② and ③, a very discontinuous and even
speed sensitive sound level behaviors characterized by some very distinct noise
peaks is found from Figs. 1.6 and 1.7.
Especially the varying operational state from no-load to rated-load causes
extended magnetic saturation effects within the totally-closed stator tooth tip
regions. This fact has, in dependency on the motor design ② or ③ unexpected inverse
impacts to the emitted noise levels. There are some interesting secondary aspects
occurring at higher speed ranges. Considering the test object ③, the maximal occur-
ring peak of 81 dB in Fig. 1.6 at 1,842 rpm is slightly reduced at the same speed
value to 79 dB in Fig. 1.7. A completely contrarily observation could be for motor
②. The noise peak of 55 dB at 1,670 rpm at no-load is significantly shifted to much
higher levels with arising additional peaks of 62 dB at the same speed in case of the
rated-load condition in Fig. 1.7.

1.4.2 Fast Fourier Sound Pressure Spectrum for Each Motor

Design at Rated-Load

The description of the time-depended sound pressure in the frequency domain

allows a deeper inside into the fundamental aspects of noise generation. The har-
monic sound pressure frequency components of the motors ① to ④ are depicted in
Figs. 1.8–1.11 over the extended speed range.
The utilization of a skewed rotor type in motor ① obviously results in Fig. 1.8
in green shaded lower range spectral magnitudes. Contrarily, the test results of the
skewed design ④ given in Fig. 1.11 show a high 58 dB peak at the frequency of
1,900 Hz.
A distinct noise peak of electromagnetic origin of the un-skewed design ② is
visible in Fig. 1.9 at the speed 1,670 rpm with the frequency of 940 Hz.
1 Acoustic Behavior of Squirrel Cage Induction Motors 9

Fig. 1.8 FFT of the sound pressure level for motor ① at rated-load depicted for the speed range
500–3,000 rpm

Fig. 1.9 FFT of the sound pressure level for motor ② at rated-load depicted for the speed range
500–3,000 rpm

Fig. 1.10 FFT of the sound pressure level for motor ③ at rated-load depicted for the speed range
500–3,000 rpm
10 C. Grabner

Fig. 1.11 FFT of the sound pressure level for motor ④ at rated-load depicted for the speed range
500–3,000 rpm

Some very distinct spectral components due to stator slot harmonics of motor ③
at the speed of 1,900 rpm could be identified in Fig. 1.10 with magnitudes of 59 dB
at the surrounding of the single frequency of 2,300 Hz.

1.5 Conclusion

The evaluation of distinct influences to the acoustic noise emission of industrial

standard squirrel cage motors is fairly a very sensitive and complex topic.
Main interest is thereby given to the influence of the complete mechanical induc-
tion motor design to the almost manifold acoustic response under sinusoidal power
supply. This includes the investigation of unexpected effects even caused by passing
critical speed values which are corresponding to mechanical resonance frequen-
cies. To take remedial measures, totally-closed stator topologies and rotor skewing
opportunities, have been deeply investigated in a comparatively manner.
The performed comparatively measurement of the speed dependent equivalent
continuous A-weighted surface sound pressure level gives a good overview about
the principal design influences. The main focus at standard drives is thereby only
given to the occurring noise peaks at the interested nominal speed.
The modification of induction motors, which are commonly assisted by skewed
rotors, by applying novel totally-closed stator slot designs, is the most favorable
possibility for reducing the emitted acoustic noise level. Fortunately, the magnetic
utilization could almost be kept equal.
If the rotor is alternatively manufactured in an un-skewed kind, the acoustic
noise emission would reach the same level as with utilized semi-closed slots in
combination with a skewed rotor.
1 Acoustic Behavior of Squirrel Cage Induction Motors 11

References

1. L.E. Kinsler and A.R. Frey, Fundamentals of acoustics, Wiley:, New York/London, 1962.
2. M. Bruneau, Fundamentals of acoustics, ISTE, London, 2006.
3. H. Jordan, Der geräuscharme Elektromotor, Verlag W. Girardet, Essen, 1950.
Chapter 2
Model Reduction of Weakly Nonlinear Systems

Marissa Condon and Georgi G. Grahovski

Abstract In general, model reduction techniques fall into two categories – moment
–matching and Krylov techniques and balancing techniques. The present contribu-
tion is concerned with the former. The present contribution proposes the use of a
perturbative representation as an alternative to the bilinear representation [4]. While
for weakly nonlinear systems, either approximation is satisfactory, it will be seen
that the perturbative method has several advantages over the bilinear representation.
In this contribution, an improved reduction method is proposed. Illustrative exam-
ples are chosen, and the errors obtained from the different reduction strategies will
be compared.

Keywords Model Reduction  Weakly Nonlinear Systems  Perturbative Approxi-

mation  Nonlinear Circuit

2.1 Introduction

As the level of detail and intricacy of dynamical system models continues to grow,
the essential behaviour of such systems can be lost in a cloud of complexity. Fur-
thermore, numerical simulations of such systems can consume extensive memory
resources and take inordinate amounts of time. Consequently, to cope with the grow-
ing complexity and dimensionality of systems, model reduction has become a vital
aspect of modern system simulation. Model reduction techniques for linear systems
are well studied (e.g. [1, 2] and references therein). However, the study of nonlin-
ear systems is much more complicated and the development of model reduction
methods for large-scale nonlinear systems represents a formidable challenge.

M. Condon (B)
School of Electronic Engineering, Dublin City University, Glasnevin, Dublin 9, Ireland,
E-mail: Marissa.Condon@dcu.ie

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 13

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
14 M. Condon, G.G. Grahovski

In general, model reduction techniques fall into two categories – moment –

matching and Krylov techniques and balancing techniques. The present contribu-
tion is concerned with the former. The advantage of Krylov-based methods is that
matrix–vector multiplications are all that are involved in the formation of the projec-
tion matrices that are used to project the system onto a reduced system. Also sparse
linear solvers and iterative approaches can be used to improve the computational
efficiency [3].
To be amenable to the application of Krylov methods, the nonlinear function
describing the system must be approximated in a suitable manner. A widely used
approach is the utilisation of the bilinear representation. However, the present contri-
bution proposes the use of a perturbative representation as an alternative to the bilin-
ear representation [4]. While for weakly nonlinear systems, either approximation is
satisfactory, it will be seen that the perturbative method has several advantages over
the bilinear representation. The use of the perturbative representation in reduction
techniques based on Krylov methods has been addressed in [4]. In this contribution,
an improved reduction method is proposed. Illustrative examples are chosen, which
are typical of examples employed for comparing model reduction approaches [4].
The errors obtained from the different reduction strategies will be compared.

2.2 Perturbative Approximation of Nonlinear Systems

Let the nonlinear dynamical system under consideration be of the form:

P
x.t/ D f .x.t// C Bu.t/

y.t/ D C x.t/; (2.1)

where f W IRn ! IRn is a non-linear function, x 2 IRn are the state-space variables.
The initial condition is x.0/ D x0 and u.t/; y.t/ 2 IR. B; C 2 IRn are constant
vectors (C is a vector-row and B is a vector-column). It is assumed that x D 0 is
a stable equilibrium point of the system (2.1) and x0 D 0. Under this assumption,
f .x/ can be expanded in a generalised Taylor’s series about x D 0:

f .x/ D A1 x .1/ C A2 x .2/ C A3 x .3/ C : : : ; (2.2)

where x .1/ D x, x .2/ D x ˝ x, x .3/ D x ˝ x ˝ x, etc. and ˝ denotes the Kronecker

product. Since x D 0 is a stable equilibrium point, A1 is a stable matrix, i.e. all
of its eigenvalues have negative real parts. It is also assumed that each term in the
Taylor’s expansion is small compared to the previous one.
P
Now consider the case where a variational parameter ˛ is introduced, i.e. x.t/ D
f .x.t// C B˛u.t/ and let the response of the system x.t/ be perturbatively
expanded in a power series in ˛ [4]:

x.t/ D ˛x1 .t/ C ˛ 2 x2 .t/ C ˛ 3 x3 .t/ C : : : : (2.3)

2 Model Reduction of Weakly Nonlinear Systems 15

On comparing terms in the variational parameter ˛, the following set of

n-dimensional differential equations can be derived:

xP 1 D A1 x1 C Bu

xP 2 D A1 x2 C A2 .x1 ˝ x1 / (2.4)

xP 3 D A1 x3 C A2 .x1 ˝ x2 C x2 ˝ x1 / C A3 .x1 ˝ x1 ˝ x1 /
::
:

Each n-dimensional equation describes the time evolution of an xi , where xi rep-

resents the i th order perturbative term in the expansion (2.3). Defining a vector
x.t/:
2 3
x1
6 7
x.t/ D 4 x2 5 (2.5)
::
:

the system in (2.4) acquires the form:

xP D Ax C Bu
y D C x; (2.6)

where
2 3 2 3
A1 B 0 0 0 :::
6 A1 7 60 A2 0 0 7
6 7 6 7
AD6 A1 7 BD60 0 A2 A3 : : : 7 (2.7)
4 5 4 5
:: :: :: :: ::
: : : : :
2 3
u1
6 u2 7
6 7
u.t/ D 6 u3 7 C D ŒC; C; C; : : : ; u1 D u.t/; u2 D x1 ˝ x1 ;
4 5
::
:
u3 D x1 ˝ x2 C x2 ˝ x1 ; u4 D x1 ˝ x1 ˝ x1 ; : : :

The source u2 for the second equation in (2.4) depends only on the state vector x1
determined from the first equation and so on. Note that since A1 is a stable matrix,
A is also automatically stable. Now, u2 , u3 , etc. are not independent inputs like u1
and therefore, linear system theory cannot be applied directly to the representation
in (2.6). However, subsequent sections will show how the representation may be
adapted so that linear system theory and consequently, linear model reduction may
be applied to the representation in (2.6).
16 M. Condon, G.G. Grahovski

The well-known bilinear representation (Carleman bilinearization) [4–6] is an

alternative approach to approximation of (2.1) for weakly nonlinear systems:
PO
x.t/ D AOx.t/
O C NO x.t/ O C BO u.t/
O u.t/ O
O
y.t/ D C x.t/;
O (2.8)

where 2 3
x .1/
6 .2/ 7
O
x.t/ D 4x 5;
::
:

AO and NO are square matrices of dimension n C n2 C    C nK , BO and CO are vectors

with n C n2 C    C nK components if K terms in the Taylor’s series expansion are
taken into account. The matrices are defined in [4–6]. For example, for K D 2:
     
A1 A2 0 0 B
AO D ; NO D ; BO D ; CO D ŒC; 0;
0 A21 N 0 0

where A1 and A2 are the matrices in (2.2), A21 D A1 ˝ I C I ˝ A1 , N D B ˝

I C I ˝ B, where I is the n  n identity matrix.
However, the perturbative representation has several advantages over the bilinear
representation–namely:
1. The system (2.6) has a simple linear form unlike (2.8).
2. The size of the system (2.6) with K perturbative terms is nK. The size of (2.8)
with K terms in the series expansion is n C n2 C    C nK .
3. There is no need to restrict the input to the perturbative system to guarantee
stability. However, a restriction exists on the input to guarantee stability of the
bilinear system [5, 6]. A sufficient condition for stability on the interval Œ0; T  is
ju.t/j  Kc for all t 2 Œ0; T  where AO C NO is stable for all  2 ŒKc ; Kc .

2.3 KRYLOV-Based Model Reduction

The goal of any model reduction technique is to replace the n- dimensional system
(2.1) with a system of much smaller dimension k  n, such that the behaviour of
the reduced order system satisfactorily represents the behaviour of the full system.
In projection based reduction schemes, a projection matrix, V , is selected such that
its columns span the required subspace [4]. The reduced system is then formed
from approximating the state vector x with xO where xO D VO x. Consider a linear
state-space representation:

P
x.t/ D Ax.t/ C Bu.t/
y.t/ D C x.t/; (2.9)
2 Model Reduction of Weakly Nonlinear Systems 17

The state-space equations for the reduced system are then given as:
PO
x.t/ D AOx.t/
O C BO u.t/
O
O
y.t/ O
D C x.t/;
O (2.10)

where
AO D V t AV; BO D V t B; CO D C V:
In Krylov-based methods, the projection matrix is chosen to span the columns of
the Krylov subspace

Km ..s0 I  A/1 ; .s0 I  A/1 B/ D

D f.s0 I  A/1 B; .s0 I  A/2 B; : : : ; .s0 I  A/m Bg:

The rationale for selection of this subspace is that it results in matching the first m
moments of the original and reduced systems. Here s0 is the point in the complex
plane about which moments are matched. However, the Krylov-based reduction
methods are preferable to direct moment matching techniques as the methods avoid
the numerical problems arising in explicit moment matching.
For nonlinear systems, model reduction is not as straightforward as for linear
systems. In this contribution, we look at some of the properties of linear systems
with a view to adapting the perturbative representation of the nonlinear system so
that a reduction strategy similar to that for linear systems can be applied to it.

2.4 Scale Invariance Property

Consider the behaviour of a linear system when u ! ˛u. In this case, the output
also changes as y ! ˛y. We term this the scale invariance property which holds for
linear systems. The result is that the Krylov-based reduction method is unaffected
when u ! ˛u. Similarly, if x ! ˇx, the reduction process is unaffected. However,
nonlinear systems are not scale invariant. For example, consider the perturbative
system under a rescaling of the input. That is consider u ! ˛u. The Bu term of
(1.6) transforms as:
2 3
1
6 ˛ 7
6 7
Bu ! ˛ 6 ˛ 2 7 Bu: (2.11)
4 5
::
:

It is evident from (2.11) that the scale invariance property does not hold. To enable
application of linear theory to (2.6) would require that which is not the case as
evident from (2.11). Consequently, linear model reduction techniques may not be
applied directly to the perturbative representation and hence, a modification is
18 M. Condon, G.G. Grahovski

required. To this end, a parameter  is introduced with a view to explicitly account-

ing for the scale dependence of the nonlinear system. The role of  is to bear the
nonlinear properties of the system throughout the reduction process. Consider (2.6)
and (2.7). The Bu term can be rewritten as:

Bu D DBU ;

where
D D diag .1; ; 2 ; : : : /;
with
U D Œu1 ; 1 u2 ; 2 u3 ; 2 u4 ; : : : :
for any nonzero function . If when u1 .t/ ! ˛u1 .t/,  transforms as:  ! ˛,
then transforms as:

U ! ˛U : (2.12)

It transforms in the same manner as the input to a linear system. The property in
(2.12) is very important as it shows that to enable application of linear systems
theory to (2.6), then the proper input to (2.6) is actually U and not u.
An estimate for  may be determined as follows: If  D 0, then the system in
(2.6) is linear. Thus,  must be proportional to the output due to the nonlinearity
y  y1 , where y1 D C x1 is the output from the linear part of (2.6). For the purposes
of practical implementation of the reduction scheme, it is convenient to take  as a
constant parameter. Hence, the following is deemed an appropriate choice for :

jy  y1 j
D ; (2.13)
T juj
N
where the bar denotes the average value of the waveform over the time interval
Œ0; T  for which the behaviour of the system is under examination, provided uN ¤
0. An exact optimal value for  for a particular model parameterisation may be
chosen from computer simulations for a particular ‘test input’ that is close to the
inputs for which the system is designed.  is then determined by minimising an
error function using the Nelder–Mead algorithm. A suitable error function is the
following:
pP
.fex  fred /2
err D ; (2.14)
N
where fex is the output from the exact model and fred is the output from the reduced
model, N is the number of samples taken of fex to compute the error.
However, for most practical cases, the estimate in (2.13) suffices. Obviously, the
reduced model approximates the input–output behaviour of the system locally. No
conclusions however, can be drawn about its global behaviour.
2 Model Reduction of Weakly Nonlinear Systems 19

2.5 KRYLOV Reduction of Perturbative Representation

The reduction process for the perturbative representation proceeds as follows: Let
the approximation in (2.2) involve K terms. The dimension of the system represen-
tation in (2.6) is thus NK. Suppose it is required to reduce the system to dimension
k. The Krylov space for the first-order response x1 in (2.4) and (2.6) is formed as
K1 D .A1 1
1 ; A1 B/ (s0 is set to zero to simplify the explanation but this is not
necessary). An orthogonal projection matrix, V1 for the first-order system is formed
from K1 i.e. xP 1 D V1 x1 . Now, the second-order system in (2.4) and (2.6) is formed
as:

xP 2 D A1 x2 C A2 .xO 1 ˝ xO 1 / D A1 x2 C A2 .V1 ˝ V1 /.x1 ˝ x1 /

D A1 x2 C A2 .V1 ˝ V1 /uO 2 D A1 x2 C B2 uO 2 : (2.15)

This differs from the standard second-order system such as that presented by Phillips
[4]. In the standard version,  D 1. However, results in section 6 will show that
inclusion of the novel proposal for  achieves greater accuracy.
The Krylov space for the linear system in (2.15) is then formed as K2 D
.A1 1
2 ; A2 B2 /. An orthogonal projection matrix, V2 , is formed from K2 and this
matrix is used to reduce the second order system. The procedure for reducing the
higher order terms in (2.4) and (2.6) , i.e. x3 ; : : : , in the perturbative system is
similar.

2.6 Illustrative Example

The circuit employed is the nonlinear ladder shown in Fig. 2.1. The number of nodes
in the system is n D 30. The ladder in represents a heat flow model [7]. The voltage
at the mth node represents the temperature on a rod at a distance proportional to
m. The (input) voltage at node 1 represents the heat source. The nonlinearities re-
present the dependence of the conductivity on the temperature. The average voltage
at all nodes is taken as the output and this represents the average temperature of the
rod. Varying the circuit parameters corresponds to different spatial or environment

1 2 3 n

r r r

c c c c r
u(I)
r fref ( u ) ind ( u ) iref ( u ) inl ( u )

Fig. 2.1 Nonlinear circuit

20 M. Condon, G.G. Grahovski

conditions [7]. The nonlinear resistor introduces a quadratic nonlinearity at each

node:

inl D gv 2 ; (2.16)

for v > 0. The parameters are C D r D 1. The strength of the nonlinearity is varied
by varying g.
The dimension of the original state-space is n D 30. The perturbative representa-
tion (2.6) contains two terms, i.e. K D 2. The reduction process is performed from
the representation (2.6) of order nK D 60 to a representation of order k D 6. The
value is  D 1:6443. Fig. 2.2 shows the result for an exponential input et from
the reduced perturbative model for g D 0:1 superimposed on the result from a full
nonlinear model of Fig. 2.1. The root mean square error between this result and that
computed from a full model is 0:0026. The reduced model is equally accurate for
other input signals. In order to confirm the value of inclusion of , the root mean
square error is 0:0042 when  D 1.
As a second example, consider the 30-section nonlinear RC ladder shown in
Fig. 2.3.

0.015

0.01
v(t)

0.005

0
0 1 2 3 4 5
time s

Fig. 2.2 Reduced perturbative model g D 0:1 (solid line D full model, dashed line D reduced
model)

Fig. 2.3 Nonlinear RC ladder

2 Model Reduction of Weakly Nonlinear Systems 21

The reduction process described in Section 2.5 is applied. The system is reduced
from a dimension of 60 to a dimension of 6. The value of  is determined from
(1.13) as 0:6747. With this value of , the RMS error is 0:0037. With the standard
approach of [4], the RMS error is 0:0076.

2.7 Conclusions

Krylov model reduction of perturbative representations of weakly nonlinear systems

has been addressed (Fig. 2.4). The restriction to weak nonlinear systems arises as the
size of the perturbative representation would grow to impractical levels for highly
nonlinear systems. This restriction also applies to the bilinear representation and
indeed is even greater owing to the larger size of a bilinear representation of the same
order. This is the principal advantage of the perturbative representation compared to
the bilinear representation–it is the much smaller in size. It is of size (nK) compared
to the size of the bilinear representation (nK ). This results in reduced computa-
tional cost. The input–output mapping for nonlinear systems depends on inputs and
is not scale-independent. To explicitly account for this dependence, a parameter is
introduced into the perturbative representation. Results in this contribution indicate
that inclusion of the parameter leads to greater accuracy.
Previous work [5, 6] has shown that the same approach also leads to improved
accuracy in balanced truncation model reduction.

0.014

0.012

0.01

0.008
v(t)

0.006

0.004

0.002

0
0 2 4 6 8 10
time s

Fig. 2.4 Reduced perturbative model (solid line D full model, dashed line D reduced model)
22 M. Condon, G.G. Grahovski

Acknowledgements This material is based upon works supported by Science Foundation Ireland
Science Foundation Ireland under Principal Investigator Grant No. 05/IN.1/I18.

References

1. B. Moore (1981). Principal Component analysis in linear systems: Controllability, Observ-

ability and model reduction, IEEE Transactions on Automatic Control, AC-26, 1, 17–31.
2. A.C. Antoulas, D.C. Sorensen and S. Gugercin (2001). A survey of model reduction methods
for large-scale systems, Contemporary Mathematics, AMS Publications, Providence, RI.
3. C.A. Beattie and S. Gugercin (2007). Krylov-based minimization for optimal H2 model
reduction, 46th IEEE Conference on Decision and Control, New Orleans.
4. J.R. Phillips (2003). Projection-based approaches for model reduction of weakly nonlinear,
time-varying systems, IEEE Transactions on computer-aided design of integrated circuits and
systems 22, 2, 171–187.
5. M. Condon and R. Ivanov (2005a). Nonlinear systems-Algebraic Gramians and Model Reduc-
tion, COMPEL: The International Journal for Computation and Mathematics in Electrical and
Electronic Engineering 24, 1, 202–219.
6. M. Condon and R. Ivanov (2005b). Balanced Model Reduction from a Perturbative Represen-
tation of Weakly Nonlinear Systems, NOLTA 2005, Bruges.
7. T. Veijola and L. Costa (1998). Combined electrical and thermal circuit simulation using
APLAC, Circuit Theory Laboratory Report Series, No. CT-34, Helsinki University of Tech-
nology.
Chapter 3
Investigation of a Multizone Drift Doping
Based Lateral Bipolar Transistor on Buried
Oxide Thick Step

Sajad A. Loan, S. Qureshi, and S. Sundar Kumar Iyer

Abstract A 2D numerical simulation study of a multizone step doped lateral bipo-

lar junction transistor (LBJT) on buried oxide thick step (BOTS) is performed. The
combination of multizone doping and BOTS has been used for increasing the break-
down voltage of the device. The steps in doping and step in oxide result in the
creation of additional electric field peaks in the collector drift region which increases
the uniformity of lateral surface electric field and hence the breakdown voltage.
Numerical simulations have revealed that a LBJT with two doping zones and a thick
buried oxide results in 200% increase in breakdown voltage than the conventional
device. Increasing the number of zones to three from two makes the breakdown volt-
age 260% higher than the conventional one. An improvement in tradeoff between
the breakdown voltage and the on-resistance has been observed in the device. The
multizone doping also reduces the kinks in the device characteristics.

Keywords Multizone step  multizone doping  lateral bipolar junction transistor 

buried oxide thick step  breakdown voltage

3.1 Introduction

The SOI-BiCMOS is emerging as a promising technology for realization of wireless

system-on-chip. This technology offers advantages in terms of reduction of par-
asitic capacitance, high quality isolation and reduction in crosstalk [1]. However,
the problem of compatibility lies in the integration of vertical bipolar device with
SOI-CMOS [2]. This problem has been reduced by using lateral bipolar device as an
alternative to the vertical device. The lateral bipolar junction transistor (LBJT) offers
low parasitic capacitance, promises low power consumption and high breakdown
voltage.

S.A. Loan (B)

Electrical Engineering Department, Indian Institute of Technology, Kanpur, India,
E-mail: sajad@iitk.ac.in

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 23

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
24 S.A. Loan et al.

To obtain the large breakdown voltage, the lateral surface electric field
distribution along the silicon surface must be uniform [3]. Several ideas have been
used to enhance the breakdown voltage of the lateral bipolar devices. These include
Reduced Surface Field (RESURF) principle [4], fully depleted collector drift region
[5], graded drift region [6, 7], semi insulating polysilicon (SIPOS) passivation lay-
ers [8]. The concept of extended box has been used to increase the breakdown
voltage of lateral Schottky transistor [9]. By using linearly varying doping con-
centration and linearly varying oxide thickness, the lateral surface component of the
electric field in the device can be made perfectly uniform and the maximum break-
down voltage can be obtained. But it is extremely difficult to fabricate a device with
linearly varying doping concentration and simultaneously having linearly varying
oxide thickness [10, 12].
In this paper, numerical simulation of LBJT with multizone step doped drift
region on buried oxide thick step (BOTS) has been performed. we have studied
and simulated three types of the devices in this work. In type one, the device is
having two doping zones but without BOTS. This is the conventional device. The
second type again uses two doping zones but is having BOTS, called as two zone
proposed (2ZP) device. The third type uses three doping zones and BOTS, called
as three zone proposed (3ZP) device. To increase the breakdown voltage and at the
same time retain the ease of fabrication of the device, the linearly varying doping
and the linearly varying oxide thickness has been replaced with step profile of both.
The multizone step doping and thick step oxide results in increased uniformity in the
lateral surface electric field in the drift region, reduction of base-collector junction
electric field by using lower doping concentration near the junction and enhance-
ment of collector breakdown voltage. The simulation results using MEDICI [13]
has shown that the 2ZP device has a breakdown voltage of 200% higher than the
conventional device. A 3ZP device possesses breakdown voltage 260% higher than
the conventional one. It has been observed that increasing number of zones fur-
ther increases breakdown voltage marginally but increases complexity of the device
significantly. The high breakdown voltage in the proposed devices can be obtained
even at high drift doping concentration. This improves the tradeoff between the
breakdown voltage and the on-resistance of the device.

3.2 Ideal, Conventional and Proposed Devices

The four structures of lateral bipolar transistor employed to obtain maximum

possible breakdown voltage is shown in Fig. 3.1. Figure 3.1a employs a linearly
varying drift doping concentration and a linearly varying buried oxide thickness in
a lateral bipolar device. This is an ideal structure for obtaining highest breakdown
voltage and best performance in terms of on-resistance and speed of the operation of
the device. A highest breakdown voltage is obtained in such a structure even at high
drift doping concentration. But, the problem associated with the ideal structure is its
fabrication complexity. The designing and fabrication of such a device is extremely
3 Lateral Bipolar Transistor 25

Fig. 3.1 (a) An ideal lateral bipolar transistor (b) Conventional LBJT (c) Two zone proposed (2ZP)
LBJT (d) Three zone proposed (3ZP) device

difficult as it needs large number of masking and lithographic steps, which renders
its fabrication practically impossible. A simple and conventionally used lateral BJT
on SOI is shown in Fig. 1b. This structure is easier to fabricate in comparison to
the ideal one, but at the cost of increase in on-resistance and the decrease in break-
down voltage. The breakdown voltage is also poor in the conventional structure.
The proposed practical alterations to the structure in Fig. 3.1a, that is, 2ZP and 3ZP
are shown in Fig. 3.1c and d respectively. These structure retains the advantages of
the ideal structure, that is, high breakdown voltage and improved tradeoff between
breakdown voltage and on-resistance. The proposed structures are easier to fabricate
than the ideal one, but are more complex than the conventional one.

3.3 Simulation Results and Discussion

Extensive device simulations were carried out using the device simulator MEDICI.
The various models activated in the simulation are analytic, prpmob, fldmob, consrh,
auger and BGN [13]. The mobility models prpmob and fldmob specify that the
perpendicular and parallel electrical field components are being used. The concen-
tration dependent Shockley-Read-Hall model (consrh) and Auger recombination
model (auger) are activated. Since the device is bipolar, having PN junctions, the
band gap narrowing effect has been considered by using BGN model. The various
device and simulation parameter used are shown in Table 3.1. The current gain and
Gummel plot curves for the conventional and 3ZP devices are shown in Fig 3.2.
The common emitter current gain of the all devices is chosen to be identical 50)
by an appropriate choice of various doping concentrations. This is being done for
better comparison of the breakdown voltage. The Gummel plot for the proposed and
26 S.A. Loan et al.

Table 3.1 Device and simulation parameters

Parameters Conv. devices 2ZP devices 3ZP devices

Si film thickness 1 μm 1 μm 1 μm
Oxide thickness 1.5 μm 1.5 μm 1.5 μm
Emitter length 6.4 μm 6.4 μm 6.4 μm
Emitter doping conc.(cm3 )(n type) 5  1019 5  1019 5  1019
Substrate doping conc.(cm3 )(p type) 5  1013 5  1013 5  1013
Collector doping conc.(n type)(cm3 / nC = 5  1019 nC = 5  1019 nC = 5  1019
n1 = 3  1015 n1 = 3  1015 n1 = 3  1015
n2 = 5  1015 n2 = 5  1015 n2 = 5  1015
n3 = 7  1015
Drift region length LDN1 = 15 μm LDN1 = 15 μm LDN1 = 10 μm
LDN 2 = 19 μm LDN 2 = 19 μm LDN 2 = 11 μm
LDN 3 = 13 μm
TAUN0, TAUP0 (SRH electron and hole 5  106 (s) 5  106 (s) 5  106 (s)
carrier life time coefficients)
VSURF at poly base contact (Surface 3  106 cm/s 3  106 cm/s 3  106 cm/s
recombination velocity)
Metal (Al) work function 4.17 eV 4.17 eV 4.17 eV

Fig. 3.2 (a) Beta versus collector current (b) Gummel plots for the the conventional and 3ZP
device)

conventional devices are more or less similar, that is why they are overlapping for
most of the base voltage range. Since the oxide is fully covering the base region in
both the cases, the area of the base remains constant in all cases and hence the base
current. The output characteristics of the conventional, 2ZP and 3ZP devices are
shown in Fig. 3.3a, b, c respectively. The various parameters used in obtaining these
characteristics are shown in Table 3.1. In all structures the buried oxide is 1.5 μm
thick and BOTS thickness is varied from 1 to 9 μm.
After comparing the simulation results we have found that the common emitter
breakdown voltage with base open (BVCEO ) is significantly higher for three zone
proposed (3ZP) and two zone proposed (2ZP) devices than the conventional device.
The BVCEO of the conventional device is 124 V. Using same device parameters
3 Lateral Bipolar Transistor 27

Fig. 3.3 Output characteristics of (a) conventional device (b) 2ZP device (c) 3ZP device

as for the conventional device, the 2ZP device with BOTS length of 34 μm and
thickness of 6 μm, the BVCEO is 370 V. Enhancing the number of zones to three,
that is, the 3ZP device, the breakdown voltage increases to 440 V. This shows that
the breakdown voltage in 3ZP device is 260% higher than the conventional device
and more than 20% higher than 2ZP device.
The reason behind this enhancement in breakdown voltage is explained in
Fig. 3.4, which shows that the lateral surface electric field is more uniform in the
proposed devices than in the conventional device. The conventional device is having
least uniformity in the electric field and the 3ZP device is having highest uniformity
in the electric field. The enhanced uniformity in surface electric field is due to the
presence of steps in doping and step in oxide [10]. These steps result in the gener-
ation of additional electric peaks in the collector drift region, which are shown in
Fig. 3.4. These generated peaks pull down the peaks at the edges of the collector drift
region. Further, the buried oxide thick step (BOTS) reduces the peak electric field
at n2 -nC junction in 2ZP device and n3 -nC in 3ZP device by redistributing applied
potential in the collector drift region and in thick oxide. The analytical approach
deduced in [12] explains the creation of the electric field component due to BOTS
and increase in breakdown voltage.
28 S.A. Loan et al.

Fig. 3.4 Electric field profile in proposed and conventional devices

Fig. 3.5 (a) Potential con-

tours in the conventional
device. (b) Potential contours
in the proposed device

The potential contours of the conventional and the proposed devices are shown
in Fig. 3.5a and b respectively. The crowding of the potential contours along the
oxide layer as shown in Fig. 3.5a results in lowering the breakdown voltage in the
conventional device. In the proposed device, BOTS helps in reducing the crowding
of the potential contours as shown in Fig. 5b. This reduction in crowding of the
potential contours in BOTS, makes it possible to apply more voltage to the device
before it breaks down.
Figure 6a and b show the electric field profiles at top and bottom of the SOI
film respectively. The electric field profile in the oxide layer is shown in Fig. 6c.
These figures also show how the location of electric field peaks vary with BOTS
length. The increase in breakdown voltage can be attributed to how the electric field
peaks are distributed in the device for different values of BOTS length. As can be
seen from these figures that there is an optimum length which results in maximum
uniformity of the lateral surface electric field and hence the maximum breakdown
3 Lateral Bipolar Transistor 29

Fig. 3.6 Electric field profile in the proposed device at (a) top (b) bottom of silicon film (c) inside
the oxide layer for different values of LBOTS

Fig. 3.7 Effect of variation in (a) BODS length and (b) BODS thickness on breakdown voltage in
proposed devices

voltage. It is observed that for LBOT S = 34 μm and TBOT S = 6 μm, a maximum

breakdown voltage of 440 and 370 V is obtained in 3ZP and 2ZP devices respec-
tively for specific values of doping concentration in different regions. Figure 3.7
gives the impact of BOTS length and thickness on the breakdown voltage of the
proposed device. It is quite obvious from the Fig. 7a that an optimum length of
the BOTS is needed to get the maximum breakdown voltage of the device. From
30 S.A. Loan et al.

this plot it is clear that for the BOTS length of 30–40 μm, the breakdown voltage
2ZP device is >200% higher than that of the conventional device. For 3ZP device,
the breakdown voltage is > 260% higher than that of the conventional device. At
the optimum length, the lateral surface electric field is more uniform in the pro-
posed devices. The impact of oxide thickness on breakdown voltage is shown in
Fig. 7b. It shows that the breakdown voltage first increases with increase in BOTS
thickness, then subsequently saturates. This is because thick BOTS helps to sustain
high electric field to a value lower than the field causing breakdown in the oxide.
Since both horizontal and vertical components of the electric field contribute to the
avalanche process, increasing BOTS thickness further does not improve breakdown
voltage [9]. The simulation results have shown that for TBOT S = 8 μm and LBOT S
= 36 μm, the breakdown voltage of the 3ZP device is 445 V, 2ZP device is 380 V
and that of the conventional device is 124 V. This shows that the breakdown voltage
in the 3ZP device is enhanced by 260% and in 2ZP by 206%.
Figure 3.8 shows the effect of substrate doping on the breakdown voltage. It is
observed that there is an optimum doping concentration which gives the maximum
breakdown voltage. The effect is explained in Fig. 8b, which shows electric field
profile at specified values of substrate doping concentration in a 2ZP and in conven-
tional device. At high substrate doping concentration, the electric field builds up at
the n2 -nC junction. This electric field peak breaks down the device at low voltage.
At low substrate doping, the electric field builds up at the collector-base junction,
and ultimately breakdown occurs due to this electric field peak.
Figure 3.9a shows the effect of varying doping concentration in two zones of drift
region on breakdown voltage in the two zone proposed device. It is clear that there
are optimum values of drift region doping concentration in the two zones, which
gives maximum breakdown voltage. The optimum drift region doping concentra-
tion in N1 and N2 is about 6  1015 cm3 . The electric field profile at specified
values of doping concentration in two zones N1 and N2 is shown in Fig. 3.9b. For
n1 = 3 1016 cm3 , the electric field at collector-base junction is high and results

Fig. 3.8 (a) Breakdown voltage versus substrate concentration in conventional and proposed
devices. (b) Electric field profile at specified values of substrate doping concentration in 2ZP and
in conventional devices
3 Lateral Bipolar Transistor 31

Fig. 3.9 (a) Breakdown voltage versus drift doping concentration in 2ZP device. (b) Electric field
profile at specified values of drift region doping concentration in 2ZP device

Fig. 3.10 Thermal anal-

ysis of the propo-
sed and conventional devices

in breaking down the device at 40 V. Similarly, at doping concentration of n2 D

7  1016 cm3 , the electric field peak shifts to n2 -nC junction and hence breaks
down the device at 180 V. The electric field profile is more uniform for optimum
values of n1 = n2 = 6 1015 cm3 .
The thermal analysis of the proposed and the conventional structures has been
performed for a base current of Ib = 11016 A/μm, LBOT S = 34 μm and TBOT S
= 6 μm, as shown in Fig. 3.10. The peak temperature of the conventional device
at its breakdown voltage of 124 V is 320 K. In the 2ZP device it rises to 420 K,
when operated at its breakdown voltage (BV) of 380 V. The temperature in the
3ZP device is slightly higher than 2ZP device, although having higher breakdown
voltage than 2ZP device. The reason behind this temperature enhancement is the
presence of thick oxide under collector and operation at high voltage. However,
as the breakdown voltage does not increase continuously with increase in oxide
thickness, temperature of the device can be kept in safe region by choosing optimum
oxide thickness.
32 S.A. Loan et al.

3.4 Conclusion

Numerical simulation of a novel lateral bipolar junction transistor on SOI has been
performed. The novelty in the device is the combination of two/three zone step
doping with the thick step oxide. It is known that the linearly varying drift doping
and linearly varying oxide thickness results in maximum breakdown voltage. How-
ever, it is very difficult to fabricate such a device. In this paper a device with high
breakdown voltage and relatively easy to fabricate has been proposed. The BVCEO
has been maximized by choosing appropriate values of TBOT S and LBOT S . The
breakdown voltage has been found to be dependent on substrate doping and drift
region doping concentration. In this simulation study a more than 260% increase in
breakdown voltage has been observed in the three zone proposed device.

References

1. Ning T. H., “Why BiCMOS and SOI BiCMOS” IBM Journal of Res and Dev, vol. 46, no. 2/3,
pp. 181–186, Mar./May 2002.
2. I-Shun Michael Sun et al., “Lateral High speed Bipolar transistors on SOI for RF SoC
applications” IEEE Transactions on Electron Decives, vol. 52, no. 7, pp. 1376–1383, July
2005.
3. Roy S. D. and Kumar M. J., “Realizing High Voltage Thin Film Lateral Bipolar Transistors
on SOI with a Collector Tub” Microelectronics International, 22/1, pp. 3–9, 2005.
4. Huang Y. S. and Baliga B. J., “Extention of resurf principle to dielectrically isolated power
devices,” Proceedings of the 3rd International Symposium on Power Semiconductor Devices
and ICs, pp. 27–30, April 1991.
5. Kumar M. J. and Bhat K. N., “Collector recombination lifetime from the quasi-saturation
analysis of high voltage bipolar transistors,” IEEE Trans. Electron Devices, vol. 37, no. 11,
pp. 172–177, 1995.
6. Zhang S., Sin J. K. O., Lai T. M. L. and KO P. K., “Numerical modeling of linear doping
profiles for high-voltage think film SOI devices,” IEEE Trans. Electron Devices, vol. 46, no.
5, pp. 1036–1041, 1995.
7. Cao G. J., De Soza M. M. and Narayanan E. M. S., “Resurfed Lateral Bipolar Transis-
tor for High-Voltage, High-Frequency Applications” Proceedings of the 12th International
Symposium on Power Semiconductor Devices and ICs. Toulouse, France, pp. 185–187, 2000.
8. Jaume D., Charitat G., Reynes J. M. and Rossel P., “High-Voltage planar devices using
filed plate and semi-resistive layers,” IEEE Transactions on Electron Devices, vol. 38, no. 7,
pp. 1478–1483, July 1991.
9. Kumar M. J. and Roy S. D., “A New High Breakdown Voltage Lateral Schottkey Collector
Bipolar transistor on SOI: Design and Analysis” IEEE Transactions Electron Device, vol. 52,
no. 11, Nov. 2005.
10. Sunkavalli R et al., “Step drift doping profile for high voltage DI lateral power devices”
Proceedings of IEEE International SOI Conference, pp. 139–140, Oct. 1995, 2005.
11. Luo J et al., “A high performance RF LDMOSFET in thin film SOI technology with step drift
profile”, Solid State Electronics, vol. 47, pp. 1937–1941, 2003.
12. Kim I. J et al., “Breakdown Voltage improvement for thin film SOI Power MOSFET’s by a
Buried Oxide Step Structures” IEEE Electron Device Letters, vol. 15, no. 5, pp. 148–150, May
1994.
13. TMA MEDICI 4.2 . Technology Modeling Associates Inc. Palo Alto, US 2006.
Chapter 4
Development of a Battery Charging
Management Unit for Renewable Power
Generation

Yong Yin, Jihong Wang, Xing Luo, and Shen Guo

Abstract To alleviate the intermittences in renewable power generation, use of

energy storage is an evitable route of choice. Batteries are nowadays the most real-
istic means in energy storage. The charging and discharging of batteries affect the
battery life time and energy efficiency greatly. At present, batteries have been used to
provide electricity for lighting, radios, TVs and communications. The use of a small
size wind generators with the assistance of batteries could provide local alternative
means to enable more households to have access to electricity. The work reported in
the chapter mainly focuses on the development of a battery control unit. The energy
conversion modes are analyzed. Then the working principles of chargeable batter-
ies are probed into. Control strategies based on fuzzy logic have been suggested to
ensure the optimal performance in battery charging and discharging. A charging/
discharging control module using embedded system technology is designed to
charge the battery by following the optimal charging/discharging curve of the bat-
tery. The whole unit is controlled and monitored by an embedded microprocessor
system.

Keywords Renewable power generation  Battery Charging  battery life time 

control module  embedded system

4.1 Introduction

Renewable energy refers to sustainable natural resources such as sunlight, wind,

wave and tides, which are naturally replenished [1]. Renewable power generation
systems have been paid a great attention recently due to the pressure from the global

J. Wang (B)
School of Electronic, Electrical and Computer Engineering, University of Birmingham,
Birmingham, B15 2TT, UK,
E-mail: j.h.wang@bham.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 33

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
34 Y. Yin et al.

warming. China has abundant inland and offshore wind energy resources, providing
potential conditions for various types of capacity, in-grid or off-grid wind stations.
According to a new study from an environmental group, China’s southern province
of Guangdong could support 20 GW of wind generating capacity by 2020, provid-
ing as much as 35,000 GWh of clean electrical energy annually, which is equivalent
to 17% of Guangdong’s total current power demand [2]. By end of year 2005, China
had built 59 wind farms with 1,854 wind turbine generators and a 1,266 MW wind
power installed capacity [2]. In China, a large number of remote rural or mountain-
ous inhabitants have no access to electricity, but in these areas there are abundant
natural energy resources such as wind or solar energy. To alleviate the intermittences
in renewable power generation, use of energy storage is an evitable route of choice.
Batteries are nowadays the most realistic means in energy storage. The charging and
discharging of batteries affect the battery life time and energy efficiency greatly. At
present, batteries have been used to provide electricity for lighting, radios, TVs and
communications. The use of a small size wind generators with the assistance of
batteries could provide local alternative means to enable more households to have
access to electricity. Generally, the framework of the energy conversion process in
a small-scale renewable energy generation system can be described by Fig. 4.1. The
energy conversion process illustrated in Fig. 4.1 can be divided into eight steps.
Many research works such as the modeling of the wind turbine, the control strate-
gies for the three-phase generator have been conducted and the research results
were reported in many publications [3, 4]. The work reported in the chapter mainly
focuses on the development of a battery control unit. The energy conversion modes
are analyzed. Then the working principles of chargeable batteries are probed into.
Control strategies based on fuzzy logic have been suggested to ensure the optimal
performance in battery charging and discharging. A charging/discharging control
module using embedded system technology is designed to charge the battery by
following the optimal charging/discharging curve of the battery. The whole unit is
controlled and monitored by an embedded microprocessor system.

Embedded
microprocessor

Resource Wind or Synchronous Uncontrollable DC-DC Chargeable

(Wind or tide) water turbine generator rectifier bridge converter battery
1 2 3 4 5
8 6
Control and /or monitoring DC-AC
Loads inverter
Energy flow
7

Power
grid

Fig. 4.1 Framework of the renewable energy generation system

4 Development of a Battery Charging 35

4.2 Energy Conversion Modes

If the output electrical power is over the load demand, the surplus is used to charge
the battery. Provided that the load does not need all the output power, and the battery
is fully charged, the superfluous power is then sent to the main grid. In the system,
the output electrical power is provided to the loads with priority. There exist five
possibilities with the relationship among Poutput , Pload , Pbattery and PGrid , and five
working modes of energy conversion in the system is formed accordingly [3], in
which Poutput represents the electrical power output, Pload the power consumption
of the loads, Pbattery the power charged into or discharged from the battery, PGrid the
power supplied to the main grid. The five modes are:
Mode1: Poutput D 0; Pload D 0; Pbattery D 0; PGrid D 0 – the system is idle.
Mode 2: Poutput > Pload ; Poutput  Pload < Pbattery ; PGrid D 0 – the generated
electricity is more than the load demand and the battery is charged to less than
its full capacity.
Mode 3: Poutput > Pload ; Poutput  Pload > Pbattery ; PGrid > 0 – the generated
electrical power is more than the load demand and the battery is over charged so
the extra power is sent to the grid.
Mode 4: Poutput < Pload ; Pload < 0; PGrid D 0 – the generated electrical power
is less than the load demand and the battery is discharged.
Mode 5: Poutput < Pload ; Pload D 0; PGrid D 0 – the generated electrical power
is much less than the demanded energy supply and the battery is fully discharged.
The battery must be disconnected from the system to avoid over discharge.
From the discussion, it can be seen that the battery works in three statuses: dis-
connected from the system, charged by the renewable power or discharged to supply
power to the loads, as shown in Fig. 4.2. The status of the battery is depended on
the working modes of the system, and shifts according to different modes [4].

4.3 Lead Acid Battery

The most commonly used batteries are divided into three categories, alkali bat-
tery, nickel iron battery and lead acid battery. The lead acid battery is chosen for
our system. The working principles of the lead acid battery can be found in many

Generator Loads Generator Loads Generator Loads

Being charged by Disconnected from

Discharged to the loads
the system the system
Battery Battery Battery

Fig. 4.2 Three working status of the battery

36 Y. Yin et al.

literatures [5, 6]. In 1960s, an America scientist Mass has put forward an optimal
charging curve for the battery based on the lowest output gas rate, as shown in
Fig. 4.3. If the charging current keeps to the track of the optimal curve, the charging
hours can be sharply cut down without any side effect on the capacity and life-span
of the battery. The Ampere-hour rule for charging the Lead-acid batteries can be
considered as the most efficient charging approach, considering the charging time
and the excellent performance provided by this method to maintain the life-span
of the battery. One example of the application of the Ampere-hour rule to charge
battery is shown in Fig. 4.4. More details in charging method for lead-acid batteries
was discussed in reference [5].
When the battery is charged, the charging current Ic , the charging voltage (port
voltage) Uc , the potential difference Eb between the positive plate and the negative
plate of the battery and the internal resistor Rb of the battery has the following

i(A)

80 90Ah

60Ah
Y

60

40 40Ah

20

0 t(h)
Fig. 4.3 Optimal charging 1 2 3 4
curve of a battery X

U,I
y (V,A)

2.23V U
Per cell

Under charging
t(h)
O
x

First stage Second stage

Fig. 4.4 Application of the ampere-hour rule to charge battery

4 Development of a Battery Charging 37

relationship:
Uc  Eb
Ic D (4.1)
Rb
The set point of the charging current is closely related with the capacity of the bat-
tery [7]. With the potential difference of the battery gradually increasing, the
charging voltage rises with it so that the battery can be charged according to the
formula ( 4.1), but it kept below a certain maximum level as shown in the part one
of Fig. 4.4. The second stage starts when the battery voltage reaches a certain level,
which is defined by the characteristics of the battery. During this phase, the control
variable switches over from current to voltage. The voltage is maintained almost
unchanged at a given value, while the current is gradually decreasing until it drops
to a value set at the initial stage [8].

4.4 Battery Charging and Discharging Strategies

In general, frequently charging and discharging, overcharging or insufficiently

charging a battery will shorten the battery’s life time. The problem associated with
wind power is when and how the battery should be charged to provide the best
energy efficiency for the system and to prolong the battery life time. If the output
electrical power is excessive for the consumption of the loads, the surplus will be
provided to charge the battery. As shown in Fig. 4.5, from the period t1 to t2 and
t7 to t8, the wind power is more than the load consumption, and it lasts a relatively
long period of time, so the battery may be charged, while from t3 to t4 or t5 to t6,
even though the load can’t use up the wind power, it should not change the battery
to avoid insufficient or too frequent charging.
To achieve the optimal charging and discharging performance, fuzzy logic con-
trol algorithms can be a suitable choice for the management of charging and
discharging process. The main advantages of Fuzzy control presents in its feature –
including human experiences and decision making into the control logic ( [9, 10]).

Wind power

Load consumption

X
0
t3 t4t5 t6
t1 t2 t7 t8

Fig. 4.5 Relationship between the wind power and the load consumption
38 Y. Yin et al.

Fuzzy logic control can be well implemented to a low-cost hardware system and
easily updated by amending rules to improve performance or adding new features.

4.4.1 Membership Function Database

Referred to the Ampere-hour charging curve of the battery in Fig. 4.4, the charging
process seems infinite. At the beginning of charging process, the charging current
is quite large, while it drops quickly with time elapses. During this period, most
electrical power has been converted into chemical energy. At the end of the charg-
ing process, the current is close to zero and hardly changes. In general, when the
charging capacity of the battery reaches to 90% of its rated value, the charging
process is considered to complete.
In this system, there are four input variables. P D P  P n is the difference of
the wind power P and the load consumption Pn; P 0 D dP =dt 0 represents the
changing rate of P ; T D T n  T the relative temperature of the battery to the
surrounding temperature; T 0 D dT =dt the changing rate of T . The output
variable is the charging voltage U . So, the control function can be described as:

U D F .P; P 0 ; T; T 0 / (4.2)

The general fuzzy rules can be represented by (3), which is illustrated in Fig. 4.6.
If (P is : : : and P 0 is : : : and T is : : : and T 0 is : : :)

Then .U is : : :/: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: : : (4.3)

To reduce the complexity of the controller, an improvement on organization of

the fuzzy rules has been considered. According to the control function in (4.2) and
(4.3), the fuzzy rules can be organized into two separate parts in formulae (4.4) and
(4.5).
U 1 D F 1.P; P 0 / (4.4)
U 2 D F 2.T; T 0 / (4.5)
The fuzzy rules for them can be described as follows.

If (P is : : : and P 0 is : : :) then .U1 i s: : :/: : :: : :: : :: : :: : :: : :: : :: : :: : :: : : (4.6)

If (T is : : : and T 0 is : : :) then .U2 i s: : :/: : :: : :: : :: : :: : :: : :: : :: : :: : :: : :: (4.7)

U1 ˝ U 2 D U (4.8)
The improved fuzzy rules can be demonstrated with the block diagram in Fig. 4.7.
The improved version is much simpler, which can be implemented with two con-
trol loops. The outer loop is based on the U and U 0 , while the inner loop is on
the basis of the T and T 0 . The controller designed by following the above fuzzy
rules can be implemented with the structure as shown in Fig. 4.8.
4 Development of a Battery Charging 39

Ñ
IF U is very high
Ñ
AND U' is large
Very high Ñ
AND T is very high
Ñ High Ñ
AND T' is large
U
Ok THEN U is Pn
Low Ñ
Very low IF U is high
Ñ
AND U' is large
Ñ
AND T is very high
Ñ
Large AND T' is largeTHEN
Ñ U is Pn-1
U' Medium
Small Ñ
IF U is ok
Zero Ñ
AND U' is large
Ñ
AND T is very high
Ñ
AND T' is large Defuzzifier U
Very high THEN U is Pn-2
Ñ High
T Ñ
Ok IF U is low
Ñ
Low AND U' is large
Ñ
Very low AND T is very high
Ñ
AND T' is large
THEN U is Pn-3
Ñ Large Ñ
IF U is very low
T' Medium Ñ
AND U' is large
Small Ñ
AND T is very high
Zero Ñ
AND T' is largeTHEN
U is Pn-4
…...

Fig. 4.6 Block diagram of the fuzzy rules

Very high Ñ
IF P is very high
Ñ
P High Ñ
AND P' is large
Ok THEN U1 is P 24
Low
Very low Ñ
IF P is high
Ñ
AND P' is large Defuzzifier U1
THEN U1 is P 23
Large
Ñ Medium Ñ
P' IF P is ok
Ok Ñ
AND P' is large
Small THEN U1 is P 22
Very small …...

Very high Ñ
IF T is very high
Ñ
T High Ñ
AND T' is large
Ok THEN U 2 is P 23
Low
Very low Ñ
IF T is high
Ñ U2
AND T' is large Defuzzifier
THEN U 2 is P 23
Large Ñ
Ñ
T' Medium IF T is ok
Ñ
Small AND T' is large
Zero THEN U 2 is P 22
…...

Fig. 4.7 Block diagram of the improved fuzzy rules

40 Y. Yin et al.

P Ñ
P U
+
Ñ F1 + + Battery T
P'
Pn Ñ
T +
F2 Ñ
T'

Tn

Fig. 4.8 Fuzzy control based on improved rules

4.4.2 Fuzzification

The fuzzification of the system has two separate parts. One is for the outer loop
and the other is for the inner loop. As the fuzzification of the outer control loop is
much the same as that of the inner loop, we only present the discussion of the outer
loop. The input variables for the outer loop is the P and P 0 . P is positive
in the whole charging process. At the same time, it drops gradually in the whole
process. As P 0 is the differential coefficient of P  to t, it can be positive or
negative in the charging process. Given a value set of X in Œ0; C1, P is labeled
by:

PV PH PO PL PZ

Positive Positive Positive Positive Positive

very high high ok low but close to zero

While P 0 is labeled with the following values, defined a value set of X in

[1, C1].

PL PM PZ NM NL

Positive large Positive Positive Negative Negative

medium but close to zero medium large

The output variable U1 can be positive or negative labeled in five possibilities.

LP SP ZE SN LN

Large Small Zero Small Large

positive positive negative negative

The fuzzification of P and P 0 is shown in Fig. 4.9. The fuzzication of the

output U1 can be demonstrated in Fig. 4.10.
4 Development of a Battery Charging 41

PV PH PO PL PZ PL PM PZ NM NL
Y Y
Ñ Ñ
U U'
1

0 1 -1 0 1
X X

Fig. 4.9 Input variables fuzzification

LN SN ZE SP LP

U1
1

0
0
Min Max

Fig. 4.10 Output variable fuzzification

Table 4.1 Rule base for outer loop

P 0 =P PV PH PO PL PZ

PL P24 P23 P22 P21 P20

PM P19 P18 P17 P16 P15
PZ P14 P13 P12 P11 P10
NM P9 P8 P7 P6 P5
NL P4 P3 P2 P1 P0

Another approach is to label them with the minimum and maximum values of the
analogue to digital converter, such as the set of [0, 255] or [255, 0] with an 8-bit
converter (or more details, see [11]).

4.4.3 Fuzzy Rule Evaluation

Fuzzy rule evaluation is the key step for a fuzzy logic control system. This step
processes a list of rules from the knowledge base using input values from RAM to
produce a list of fuzzy outputs to RAM. There are many mutual methods to evaluate
fuzzy rules, among which State Space method is the most popular one due to its
close relationship with the transference function of the system [11]. The rule base
for outer loop of the system is summarized in Table 4.1.
These rules are expressed in the form of IF–THEN, as generally given in formula
(4.3). The fuzzy rules are specified as follows:
42 Y. Yin et al.

PL
Y
Ñ
P
1
0.4

0 X Y
U1
1
Y
0.4 AND 0.6 = 0.4
Ñ 0.4
P'
PZ 1 X
Min 0 Max
0.6
X
0

Fig. 4.11 Output variable fuzzification

Rule 1: IF P D PV AND P0 D NL THEN U1 = ZE

Rule 2: IF P D PV AND P0 D NM THEN U1 D SN
Rule 3: IF P D PL AND P0 D PZ THEN U1 D LP
Rule 4: IF P D PV OR P0 D PM THEN U1 D LN
: : :: : :
Rule n: IF P D : : : OR/AND P0 D : : : THEN U1 D : : :
To work out the values of the output P0 to P19 in Table 4.1, the Min–Max method
was selected in the rule evaluation process [11]. In the Min–Max rule, the mini-
mum of the two inputs is used as the combined truth-value. Fig. 4.11 displays this
method.
Combining with rule3, we can solve out that P11 equals 0.4 but belongs to LP.

4.4.4 Defuzzification and Simulation

The final step in the fuzzy logic controller is to defuzzify the output, combining the
fuzzy output into a final systems output. The Centroid method is adopted, which
favours the rule with the output of greatest area [12, 13]. The rule for outputs can be
defuzzified using the discrete Centroid computation described in formula (4.9).

P
4
S i F i
1
Uout D
P
4
Si
1
D SUM.i D 1 to 4 of . S.i / F .i ///=SUM.i D 1 to 4 of S.i //
S.1/ F .1/ C S.2/ F .2/ C S.3/  F .3/ C S.4/ F .4/
D (4.9)
S.1/ C S.2/ C S.3/ C S.4/

In (4.9), S.i / is the truth value of the result membership function for rule i , and F .i /
represents the value while the result membership function (ZE) is maximum over the
4 Development of a Battery Charging 43

Fig. 4.12 Battery charging current and output port voltage

output variable fuzzy set range. With the fuzzy control rules discussed above, the
typical battery charging current and voltage curves are shown in Fig. 4.12.
In the simulation, the setting point for the charging voltage is 14.5 V and that
of the charging current to 6 A. The charging process can be separated into two
stages. During the first stage, the fuzzy control strategy is implemented to deter-
mine the proper start charging time and to prevent it from being over or insufficiently
charged. At the beginning, the port voltage Uc starts with a small value, while the
current keeps constant to the set value, so the battery will be fully charged. During
the second stage, a normal charging strategy is used. The control variable switches
over from current to voltage. The voltage is maintained almost unchanged at a given
value, while the current is gradually decreasing until it drops to a value set at the
initial stage. As can be seen from Fig. 4.12, the performance of the fuzzy controller
proves to be very good, and the charging curve is much close to the Ampere-hour
curve in Fig. 4.5. Charging time in this method has been reduced to about 30%,
comparing to the classical methods. For a fully discharged battery, the required
charging time is approximately 2.5 h (9,000 s).

4.5 System Implementation

The performance of the renewable energy generation system, including the battery
charging and discharging unit, is controlled and monitored by an embedded micro-
processor. In our system, a 32-bit RISC microprocessor from Samsung Company,
S3C2410X, is employed. The S3C2410X was developed using an ARM920T core,
0.18 um CMOS standard cells and a memory complier. The S3C2410X provides a
44 Y. Yin et al.

Remote server

TCP/IP

Fuzzy control
algorithm Driver circuit Battery
D/A
S3C2410 A/D Sensors
X circuit
Disk USB

Keyboard, Mouse, LCD...

Fig. 4.13 System functional architecture

complete set of common system peripherals, including16 KB Instruction and 16 KB

Data Cache, MMU to handle virtual memory management, LCD Controller (STN &
TFT), NAND Flash Boot Loader, System Manager (chip select logic and SDRAM
Controller), 3-ch UART, 4-ch DMA, 4-ch Timers with PWM, I/O Ports, RTC,
8-ch 10-bit ADC and Touch Screen Interface, IIC-BUS Interface, IIS-BUS Inter-
face, USB Host, USB Device, SD Host & Multi-Media Card Interface, 2-ch SPI
and PLL for clock generation and so on [14].
In the battery charging and discharging control unit, the S3C2410X mainly
takes charge of three essential functions: charging and discharging control, status
surveying and monitoring and data communication, as shown in Fig. 4.13.
The working processes of the system as shown in Fig. 4.13 is: The embed-
ded microprocessor S3C2410X compares the control parameters inputted from
Keyboard and those fed back from sensors of the battery. Then the results are pro-
cessed with fuzzy control algorithm by the embedded microprocessor. After that,
corresponding signals, which have been converted from digital instructions into
those of analogue form and amplified by the driver, are sent out to control the
charging and discharging process. Functions of remote networked monitoring are
provided to monitor and survey the status of the battery through networked level,
often based on web technology with TCP/IP protocol stack. Components of fault
diagnosis expert system can be adopted to employ professional strategies and mea-
sures to analyze or judge the existed or potential faults of the battery. Also in the
system, history records of the control and monitoring of the battery can be stored to
removable disks with USB interface.
The main module is the S3C2410X, which has abundant and powerful peripheral
resources and communication ability with very high operational and data processing
speed, so various kinds of advanced digital control algorithms can be implanted or
emendated into it.
As the charging and discharging process have important relationship with the
temperature of the battery, the relationship between the temperature and Uout can
4 Development of a Battery Charging 45

Communication program
Fuzzy control program
Data management program
…...
Application program

Drivers for Network, USB, LCD...

Drivers

File system Network interface

Memory management Multi-task management…...

Operating system environment

Hardware

Fig. 4.14 System software architecture

be described with formula (4.10) [15].


Uout 106
T .ı C / D Iout  273 D   (4.10)
R3 1 C R5
R4

The software architecture of the renewable energy generation system can be sep-
arated into three parts, namely the operating system environment, hardware driver
and application program, as shown in Fig. 4.14.
The operating system environment is based on Embedded Linux, which is
transplanted into the S3C2410X microcontroller, providing the basic software envi-
ronment for file system, network interface, memory management and multi-task
schedule etc. to the system. Embedded Linux is a free and open source real-time
operating system, and it is programmed with standard C language. It can be tailed
according to specific system requirements with some redundant tasks removed and
certain enhanced functions added, which guarantees the system’s reliability and
real-time and powerful performance.

4.6 Conclusions and Future Works

The control process of the battery charging and discharging is non-linear, time-
varying with pure time delay, multiple variables and many external disturbances.
Many parameters such as the charging rate, the permitted maximum charging cur-
rent, the internal resistor, the port voltage, the temperature and moisture, change
during the charging and discharging process can not be obtained directly, and it
is impossible to use traditional control system. A fuzzy control unit for battery
46 Y. Yin et al.

charging and discharging in a renewable energy generation system is studied in

detail. Simulation based on fuzzy strategies introduced in the paper shows that
the control unit has excellent performance in a laboratory environment. Analysis
involved in the paper ignores the temperature element in the inner loop of the con-
trol unit, but to utilize this unit in the system, changes in the temperature must
be taken into account. Future works may involve genetic algorithms or neural net-
work based methods, combined with fuzzy algorithm, to get quicker and accurate
system response. Also remote online monitoring and alarming functions should be
developed to prevent the battery from destruction.

Acknowledgements The authors would pay their appreciations to The Science Foundation for
New Teachers of Ministry of Education of China (No. 20070497059) and The Open Research
Projects Supported by The Project Fund (No. 2007A18) of The Hubei Province Key Laboratory
of Mechanical Transmission and Manufacturing Engineering, Wuhan University of Science and
Technology and The National Science Foundation of China (No. 50675166).

References

1. Renewable energy, last modified on February 2008, Available: http:// en.wikipedia.org/

wiki/Renewable energy#Market development of renewable heat energy. Accessed on Febru-
ary 2008.
2. Li, Z. J., China’s Wind energy potential appears vast, November, 2005, Available: http://www.
worldwatch.org/node/57. Accessed on February 2008.
3. Zhao, Q., Integrated control of load power tracking and battery charging for stand-alone small
scale wind power generating systems, University of Inner Mongolia, China, Master’s thesis.
June 2006.
4. Oneyama, N., Air compressor and its operation control, The Journal of Energy Conservation
Center, Vol. 50, No. 2, pp. 65–73, 1998.
5. Northern Arizona Wind & Sun, Deep cycle battery FAQ, Available: http://www.windsun.
com/Batteries/Battery FAQ.htm. Accessed on February 2008.
6. Buchmann, I., The lead-acid battery, created on April 2003, last edited on February 2007,
Available: http://www.batteryuniversity.com/partone-13.htm. Accessed on February 2008.
7. Kirchev, A., D., Pavlov and B. Monahov., Gas-diffusion approach to the kinetics of oxygen
recombination in lead-acid batteries. Journal of Power Sources, Vol. 113, pp. 245–254, 2003.
8. Osaka, T. S., M. Rajamäki and T. Momma. Influence of capacity fading on commercial
lithium-ion battery impedance. Journal of Power Sources. Vol. 119, pp. 929–933, 2003.
9. Fuzzy control system, From Wikipedia, the free encyclopedia, 2007, Available:
http://en.wikipedia.org/wiki/Fuzzy system. Accessed on February 2008.
10. Lee, C. C., Fuzzy logic in control systems: Fuzzy logic controller – part I and II, IEEE
Transactions on Systems, Man and Cybernetics, Vol. 20, No. 2, pp. 404–435, 1990.
11. Lee, H. D. and S. K. Sul, Fuzzy Logic based Torque control strategy for parallel type HEV,
IEEE Transactions on Industrial Electronics, Vol. 45, No. 4, pp. 625–632, 1998.
12. Berardinis, L. A., A clear thinking on fuzzy logic, Machine Design, Vol. 64, No. 8, pp. 46–52,
23 April 1992.
13. Cai, B. and D. Konik, Intelligent vehicle active suspension control using fuzzy logic, IFAC
World Congress, Sydney, Vol. 2, pp. 231–236, 1993.
14. Samsung Electronics, S3C2410X 32-Bit RISC Microprocessor User’s Manual, Revision 1.2.
Publication Number: 21.2-S3-C2410X-052003, 2003.
15. Analog Devices, Inc., Two-terminal IC temperature transducer, AD590, Available: http://
www.analog.com, 1997.
Chapter 5
Optical CDMA Transceiver Architecture:
Polarization Modulation with Dual-Balanced
Detection

Mohammad Massoud Karbassian and Hooshang Ghafouri-Shiraz

Abstract In this chapter, we propose a novel OCDMA architecture design based on

the polarization modulation and dual-balanced detection. The application of optical
tapped-delay line (OTDL) as the incoherent OCDMA decoder is also investigated.
It is assumed that the signal is degraded by (i) fiber amplifier spontaneous emission
(ASE) noise, (ii) electronic receiver noise (iii) photo detectors (PD) shot-noise (iv)
and mainly MAI. For optical signature sequences, we have used a coding scheme
based on double-padded modified prime code (DPMPC).

Keywords Optical CDMA Transceiver  Polarization Modulation  Dual-Balanced

Detection  OCDMA architecture  optical tapped-delay line

5.1 Introduction

Bandwidth-hunger world demands higher bit-rate and ultra-fast services such

as video-on-demand and streaming over the Internet protocol e.g. IPTV. Due
to tremendous resources of bandwidth and extremely low loss, fiber-optic can
be the best physical transmission medium for telecommunications and com-
puter networks. Unlike conventional time- division multiple-access (TDMA) and
wavelength-division multiple-access (WDMA) techniques, code-division multiple-
access (CDMA) can make full use of the huge bandwidth in the fiber-optic.
Meanwhile, it has the potential to support random access protocol, different services
with different data-rates and bursty traffics.
Optical CDMA (OCDMA) has been proposed as an access protocol that takes
advantage of the excess bandwidth particularly in single-mode fiber-optic (SMF). It
provides multi-channel communications over a single frequency band. Accordingly,

M.M. Karbassian (B)

University of Birmingham, Edgbaston, B15 2TT, UK,
E-mails: mmk404@bham.ac.uk,h.ghafouri-shiraz@bham.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 47

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
48 M.M. Karbassian, H. Ghafouri-Shiraz

signals must be designed to reduce mutual interferences. This can be achieved by

subdividing each data into a number of binary chips. The chip sequences constitute
a code that permits a bit-stream broadcast on a network to be selected by means of
a correlation process at the receiver destination. A large number of chip sequence
signatures can be assigned to different users, in the process of which the set of
optical signatures essentially becomes a set of address codes for the network users.
Incoherent OCDMA technique inherently suffers from multiple-access interfer-
ence (MAI) that requires estimation and removal through cancellation techniques
or higher order modulations [1–4]. In the past few years, advances in photonics
technology made it possible for incoherent scheme to approach the sensitivity per-
formance of coherent ones. Hence, there has been a considerable shift of interest
from coherent to incoherent scheme on the part of both research and industry, due
to its simple architecture and cost effectiveness.
Polarization shift keying (PolSK) is the only modulation that takes advantage
of vector characteristic of lightwave. It has been experimentally reported [5] that
PolSK has insensitive behavior to (i) phase noise, (ii) polarization fluctuations, (iii)
self- and cross-phase modulations due to the constant-power envelop of the light-
wave. A comprehensive set of results [5, 6] showed that the performance of the
binary PolSK is approximately 3 dB better than intensity modulation (on the peak
optical power) of other PolSK or equivalent phase shift keying (PSK) modulations.
It has very low sensitivity to fiber dispersions in which elevates the system per-
formance accordingly. PolSK encodes the information on a constellation of signal
points in the space of the Stokes parameters which are the coordinates of the sig-
nal’s states of polarization (SOP) over the Poincaré sphere [6]. In general, each
signal point corresponds to a given SOP and a given optical power. To perform
PolSK detection avoiding optical birefringence compensation, it is necessary to
use a receiver that extracts the Stokes parameters of the incoming lightwaves [6].
Additionally, the advantages of PolSK over OCDMA have been reported [7, 8].
In this chapter, we propose a novel OCDMA architecture design based on the
polarization modulation and dual-balanced detection. The application of optical
tapped-delay line (OTDL) as the incoherent OCDMA decoder is also investigated.
It is assumed that the signal is degraded by (i) fiber amplifier spontaneous emission
(ASE) noise, (ii) electronic receiver noise (iii) photo detectors (PD) shot-noise (iv)
and mainly MAI. For optical signature sequences, we have used a coding scheme
based on double-padded modified prime code (DPMPC). It has been extensively
introduced and employed in pulse-position modulation (PPM), [2] in overlapping
PPM, [4] in coherent heterodyne, [9] and homodyne [10, 11] schemes, and also
in frequency modulated OCDMA system, [12] with even novel MAI cancellation
technique [13].

5.2 POLSK–OCDMA Transmitter

A generic transformation of the SOP of a fully polarized lightwave, propagating

along the z-axis which preserves the degree of polarization, is explained as fol-
N
lows. Let E.t/ and EN 0 .t/ be the electromagnetic field vectors before and after the
5 Optical CDMA Transceiver Architecture 49

transformation (i.e. modulation) respectively, and given as follows:

N
E.t/ D .Ex .t/xN C Ey .t/y/e
N j! t ; EN 0 .t/ D .Ex0 .t/xN 0 C Ey0 .t/yN 0 /e j! t (5.1)

where ! is the optical angular frequency, where Ex .t/ and Ey .t/ are .x; y/-
components of electric filed before transformation and Ex0 .t/; Ey0 .t/ are .x; y/-
components of electric filed after. Thus, we have:
 0   
Ex .t/ Ex .t/
D Q (5.2)
Ey0 .t/ Ey .t/

where Q is a complex Jones matrix with unit determinant. A subset of Jones matri-
ces, called the set of matrices of birefringence or optical activity that not only
preserves the degree of polarizations, but also has the additional feature of preserv-
ing two orthogonal fields (according to the Hermitian scalar product) [14] which
were orthogonal before the transformation. Matrices of this kind are complex uni-
tary matrices with unit determinant. Throughout this chapter we strictly refer to the
Jj as subset of Q D ŒJ0 : : : Jj : : : Jk1 .
By using the Jones representation, the field can be represented by the vec-
tor, J D ŒEx Ey T and the intensity of the beam can be normalized so that
ˇ ˇ2
jEx j2 C ˇEy ˇ D 1. Two SOPs represented by J1 and J2 are orthogonal if their
inner product is zero. Any SOP can be transformed into another by multiplying it by
a Mueller matrix [14] which are required for SOP processing e.g. polarizers, rota-
tors and retarders. In PolSK, the angle of one polarization component is switched
relative to the other between two angles; therefore, binary data bits are mapped into
two Jones vectors. A block diagram of the proposed PolSK–OCDMA transmitter
is illustrated in Fig. 5.1. The light source is a highly coherent laser with a fully
polarized SOP. If a non-polarized source is used, then a polarizer can be inserted
after the laser source. The light beam first passes through the polarization controller
that sets the polarization to an angle of 45ı for simplicity. Then, the lightwave gets
divided through polarization beam splitter (PBS) to become SOP-encoded in PolSK
modulator which switches the SOP of the input beam between two orthogonal states

Fig. 5.1 Proposed architecture of incoherent PolSK–OCDMA transmitter

50 M.M. Karbassian, H. Ghafouri-Shiraz

Fig. 5.2 Proposed architecture of incoherent PolSK–OCDMA receiver

(i.e. 0ı and 180ı at the phase modulator in Fig. 5.1 N times per bit according to an
externally supplied code (i.e. DPMPC) that spreads the optical signal into CDMA
format. Thereafter, the PolSK–OCDMA modulated signals are combined through
polarization beam combiner (PBC) and broadcasted. It is also displayed in Fig. 5.1
that for a K-user system with the first user as the desired one (for example), the ith
user SOP-encoded signal can be written as:
(
J0 if di .t/ ˚ ci .t/ D 0
Ji .t/ D (5.3)
J1 if di .t/ ˚ ci .t/ D 1

where di .t/ is the data signal with symbol duration of Ts , ci .t/ is the N -chip code
sequences of DPMPC signal with chip duration of Tc and di .t/; ci .t/ 2 f0; 1g; ˚
denotes the signal correlation. As the emitted light is initially (linearly) polarized at
an angle of 45ı , therefore J0 D p1 Œ1 1T and J1 D p1 Œ1 1T. In other words,
2 2
we have [8]:  
1 1 1
Q D ŒJ0 J1  D p (5.4)
2 1 1
Therefore, the polarization-modulated signal travels a distance of L km through an
optical SMF. Consequently, the SOP-encoded signal undergoes several impairments
such as attenuation, dispersion, polarization rotation and fiber nonlinearity. At the
receiver end shown in Fig. 5.2, the SOP rotation is compensated by the polariza-
tion controller whose function is to insure that the received signal and the optical
components at the receiver have the same SOP reference axis.

5.3 Analysis of POLSK–OCDMA Receiver

We previously discussed the configuration of the transmitter. Now we consider

the alignment and analysis of the received optical signal. The electric field of the
received polarization-modulated lightwave for K number of users can be expressed
as [7]: ( )
XK  
d .t/
EN .t/ D Re E.t/
0 N Q i
u .t  iTs /:ci .t/ (5.5)
1  di .t/ T
i D1
5 Optical CDMA Transceiver Architecture 51

The channel is represented by the Jones matrix Q and Ref:g refers to the real part
of complex EN 0 .t/. Since equipower signal constellations have been considered, both
orthogonal components are assumed to be equally attenuated. Thus, these terms can
be included in the constant amplitude of the electric field E.t/, N which neglects a
loss of orthogonality on the channel. While switching time in SOP (i.e. bit-rate) is
much slower than the chip-rate, the elements of the Jones matrix can be understood
as time-independent (i.e. Tc << Ts ). The x-component of the received electric field
vector based on Q D ŒJ0 J1  (see Eq. (5.4)) is:
( )
XK
0
Ex .t/ D Re E.t/ N ŒJ0 di .t/ C J1 .1  di .t// uT .t  iTs /:ci .t/ (5.6)
i D1

Thus, orthogonal components of the ith user are given as Exi .t/ D J0 di .t/ci .t/
N and Eyi .t/ D J1 .1di .t//ci .t/E.t/
E.t/ N where the .x; y/-components of received
modulated signal are [7]:
!
Exi .t/CEyi .t/ X
K
0 Exi .t/Eyi .t/
Exi .t/D C ci .t/di .t/ uT .t  iTs / cos.'xi /
2 2
i D1
!
E xi .t/E yi .t/ X K
E xi .t/CE yi .t/
0
Eyi .t/D C ci .t/di .t/ uT .t  iTs / cos.'yi /
2 2
i D1
(5.7)
where 'xi and 'yi describe the frequencies and phases of transmitting lasers in a
general form of ' D !t C . Based on the concept of CDMA, the field vectors of
all K transmitters are combined and multiplexed over the same channel. Thus, the
overall channel field vector can be expressed as:

X
K
EN C hannel D EN i0 .t/ (5.8)
i D1

Figure 5.2 illustrates the application of the OTDLs used as the optical correlator
in this incoherent PolSK–OCDMA configuration. The delay coefficients in OTDLs
are designed in such a way to make them perform as a CDMA chip-decoder in both
branches. Additionally, OTDL in lower branch is set up with complement of code
used in upper branch (i.e. shown by OTDL/ to decode other symbol (i.e. ‘1’). It
can be observed from Fig. 5.2 that OTDLs output contain N chip pulses that can
be assumed as a parallel circuit of many single PDs so that their currents are added
and no interference between the OTDL pulses is possible. The signals are photo-
detected in the balanced-detector arrangement to generate the differential electrical
current (Idiff D I1  I2 ) ready for data-extraction in decision processor unit. The
total upper branch current (i.e. x-component) considering all chip currents after
photo-detection is then obtained as:
52 M.M. Karbassian, H. Ghafouri-Shiraz

ZTs X ( K 
c.nTc / C 1 X Exi .t/ C Eyi .t/
N
0
I D<
2 2
nD1
t D0 i D1

 2 )
Exi .t/  Eyi .t/
C di .t/ci .t  nTc / cos.'xi / dt (5.9)
2

where < is the PD responsivity, ci .t  nTc / is the nth chip of assigned spreading
code of the ith user.
By further simplification, it can be modified as:

ZTs X (
c.nTc / C 1 X  2
N K
<
I D
0
Exi .t/ C Eyi
2
.t/ C di .t/ci .t  nTc /
4 2
t D0 nD1 i D1
o
2 2
 .Exi .t/  Eyi .t// .1  cos 2'xi / dt
8 2
ZTs XN ˆ< c.nTc / C 1 6X X n
K K
<
C 4 Exi .t/ C Eyi .t/
8 nD1 :̂
2
t D0 i D1 j D1
j ¤i

C di .t/ci .t  nTc /

 .Exi .t/  Eyi .t// Exj .t/ C Eyj .t/ C dj .t/cj .t  nTc /
) #)

 .Exj .t/  Eyj .t/// cos .'xi C 'xj / C cos .'xi  'xj / dt

(5.10)

Since PD frequency response behaves similar to a low-pass filter, both terms of

cos 2'xi in the first element and cos.'xi C 'xj / in the second element of Eq. (5.10)
are filtered out as they are outside of the PD frequency range. Furthermore, the term
cos.'xi  'xj / can also be removed provided that 'xi  'xj >> !c where !c is
the cut-off frequency of the PD. Therefore, the total current of upper branch can be
expressed as:

< X c.nTc / C 1
N
I0 D
4 nD1 2
!)
X
K
 2 
Exi .t/ C Eyi
2
.t/ C di .t/ci .t  nTc /.Exi
2
.t/  Eyi
2
.t// (5.11)
i D1

Then the Stokes parameters are defined as:

Si0 D Exi
2 2
.t/ C Eyi .t/
(5.12)
Si1 D Exi
2 2
.t/  Eyi .t/
5 Optical CDMA Transceiver Architecture 53

where Si0 refers to signal intensity part, generated in upper branch of polarization
modulator at the transmitter while Si1 refers to the linear polarized part, generated
in lower branch containing data (see Fig. 5.1. Thus, Eq. (5.11) can be rewritten as:
( !)
< X c.nTc / C 1 X  0
N K

I D
0
Si C di .t/ci .t  nTc /Si
1
(5.13)
4 nD1 2
i D1

Similarly the total current of the lower branch (i.e. y-component) can be
derived as:
( !)
< XN
1  c.nT c / XK
 
I1 D Si0 C di .t/ci .t  nTc /Si1 (5.14)
4 nD1 2
i D1

Thus, the balanced-detector output (I D I 0  I 1 / is then derived as:

<X X
N K
 0 
I D c.nTc / Si C di .t/ci .t  nTc /Si1 C n.t/ (5.15)
4 nD1
i D1

2
where n.t/ represents the total filtered Gaussian noise with variance ˝ 2 ˛ of  that
includes: (i) the PD shot-noise with electric current variance of iav D 2eiBo
where iav is the average photo-current; (ii) optically filtered ASE noise with vari-
2
ance of ASE D 2N0 B0 where N0 is the (unilateral) power spectral density (PSD)
of the white ASE noise arriving on each polarization and B0 is the optical filter
bandwidth; (iii) electronic receiver noise current (i.e. thermal noise) at the low-pass
2
filter with variance of LP D 2kb TBel =R: where R is the filter direct-current (dc)
equivalent impedance, T the absolute temperature and kb the Boltzmann constant,
Bel the filter bandwidth. Thus the overall variance of additive noise of n.t/ can be
represented as: ˝ 2˛
/ D iav C ASE C LP
2 2 2
n.t (5.16)
By considering the first user as the intended user, then we can modify the
differential output current, i.e. Eq. (5.15), as:

<X 0 <X
N N
I D S1 c.nTc / C c.nTc /c1 .t  nTc /d1 .t/:S11
4 nD1 4 nD1
(5.17)
< XX
K N
C c.nTc /ci .t  nTc /di .t/Si1 C n.t/
4
i D2
nD1

The first element in Eq. (5.17) is a dc current that needs estimation and removal
in the balanced-detector. The second element represents the intended data mixed
with its assigned spreading code auto-correlation and polarization while the third
element assumes the interference (i.e. MAI) caused by other transmitters and the
last one is the noise. Thus, the system SNR can be expressed as:
54 M.M. Karbassian, H. Ghafouri-Shiraz
!2
< P
N
4
c.nTc /c1 .t  nTc /d1 .t/:S11
nD1
SNR D !2 (5.18)
< P
K P
N
4 c.nTc /ci .t  nTc /di .t/Si1 C n.t
2
/
i D2 nD1

Both the auto- and cross-correlation of the DPMPC can be expressed respectively
as [9–13]:
XN
c.nTc /c1 .t  nTc / D P C 2 (5.19)
nD1

X
N
Xli D cl .nTc /ci .t  nTc / (5.20)
nD1

where P is the prime number, which is set to generate DPMPC sequences.

The cross-correlation, i.e. Eq. (5.20), probability density function (PDF) can be
obtained from the independent values of random variable Xli . The in-phase cross-
correlation value of DPMPC is either zero or one depending on whether the codes
are in the same group or from the different groups [2]. Obviously, the zero value
does not cause the interference due to perfectly orthogonal sequences, while the
one value causes the interference which is only among intended user and (P 2  P /
users from the different groups (i.e. P 2 whole sequences and P sequences from the
same group of intended user which are orthogonal) [4]. The cross-correlation values
are uniformly distributed among interfering users, since the users are independently
active. Thus the PDF of w, realization of Xli , is:
i
P .w D i / D (5.21)
P2  P
where P .w D i / is the probability that w assumes the value i (the number of
actively involved users in the transmission). Therefore, by substituting Eqs. (5.19)
and (5.21) into Eq. (5.18) and a little further calculation, the system SNR can be
simplified as:
1
SNR.K/ D  2 2
(5.22)
.KC2/.K1/ 16n.t
2.P 2 P /.P C2/
C 2
/
<2 d12 .t /:S11 .P C2/2

2
.
Note that SNR.1/ D <2 d12 .t/:S11 .P C 2/2 16n.t 2
/ D Eb =N0 , where Eb is the
energy of one bit, N0 is the noise PSD, denotes the single-user SNR. Equation (5.22)
is one of the main results of this study as it represents the SNR of polarization-
modulated optical CDMA system.
5 Optical CDMA Transceiver Architecture 55

5.4 Discussion of Results

BER estimation of binary PolSK modulation has already been evaluated [6, 14].
Here, the numerical results of the BER performance of the proposed transceiver
based on the above detailed analysis, resulted the OCDMA system SNR, are
demonstrated and discussed.
Figure 5.3 shows the BER of this architecture against the single-user SNR (shown
on Figs. by Sdb /. Different trends like 10%, 15%, 20% and 25% of full-load (i.e.
P 2  P interfering users) [2] as the number of simultaneous active users where
P D 19 have been evaluated in this investigation. As illustrated in Fig. 5.3, the
system that can manage 25% of full-load is able to provide BER D 109 with
Sdb D 16:5dB; whereas for Sdb D 8:5dB the system can support 20% load which
is still superior enough to deliver the network services. Furthermore, the system can
tolerate 15% load with only Sdb D 7 dB. It is indicated that the delivering network
services under these conditions is very power efficient. Although for supporting
greater number of users, higher values for P and Sdb are recommended.
Figure 5.4 also indicates the BER performance against the number of simul-
taneous users (K/ for the discussed architecture. It is observable from Fig. 5.4,
when the number of transmissions (i.e. users) increases, the BER also becomes
higher due to the growing interferences. The system employed Sdb D 14dB can
tolerate 80 simultaneous users where P D 19 which is equal to 24% of full-load.
While 73 users (21% of full-load) are guaranteed very consistent communication

Fig. 5.3 BER performance of the transceiver against single-user SNR, Sdb
56 M.M. Karbassian, H. Ghafouri-Shiraz

Fig. 5.4 BER performance of the transceiver against the number of simultaneous users, K

link (BER  109 / with only Sdb D 10dB, which refers to cost-effective design
with less power consumption.
As a comparison [7, 8], the proposed architecture can tolerate greater number of
users with less Sdb . On the other hand, the results achieved in this analysis are with
the code-length of 399 (i.e. for P D 19) [2] which is much shorter than analyzed
Gold sequences [7, 8]. That implies the proposed structure can provide even higher
throughput as the code-length is smaller.

5.5 Conclusion

This chapter has proposed and evaluated novel transceiver architecture of incoher-
ent PolSK–OCDMA with dual-balanced detection. The application of OTDL as the
CDMA-decoder has also been investigated. The BER performance of PolSK over
OCDMA in cooperation with DPMPC as the spreading code has been demonstrated
taking into account the effects of optical ASE noise, electronic receiver noise, PDs
shot-noise and mainly the MAI. The results indicated that the architecture can reli-
ably and power-efficiently accommodate great number of simultaneous users, and
can be promising for high-data-rate long-haul optical transmission.
5 Optical CDMA Transceiver Architecture 57

References

1. F. Liu, M. M. Karbassian and H. Ghafouri-Shiraz, “Novel family of prime codes for synch-
ronous optical CDMA”, J. Opt. Quant. Electron., vol. 39, no. 1, pp. 79–90, 2007
2. M. M. Karbassian and H. Ghafouri-Shiraz, “Fresh prime codes evaluation for synchronous
PPM and OPPM signaling for optical CDMA networks”, J. Lightw. Tech., vol. 25, no. 6,
pp. 1422–1430, 2007
3. H. Ghafouri-Shiraz, M. M. Karbassian, F. Lui, “Multiple access interference cancellation
in Manchester-coded synchronous optical PPM-CDMA network”, J. Opt. Quant. Electron.,
vol. 39, no. 9, pp. 723–734, 2007
4. M. M. Karbassian and H. Ghafouri-Shiraz, “Capacity enhancement in synchronous optical
overlapping PPM-CDMA network by a novel spreading code”, in Proceedings of GlobeCom,
pp. 2407–2411, 2007
5. S. Benedetto et al., “Coherent and direct-detection polarization modulation system experi-
ments,” in Proceedings of ECOC, 1994
6. S. Benedetto, R. Gaudino and P. Poggiolini, “Direct detection of optical digital transmission
based on polarization shift keying modulation”, IEEE J. Selected Areas Comms., vol. 13,
no. 3, pp. 531–542, 1995
7. K. Iversen, J. Mueckenheim and D. Junghanns, “Performance evaluation of optical CDMA
using PolSK-DD to improve bipolar capacity”, in SPIE Proceedings, vol. 2450 (Amsterdam),
pp. 319–329, 1995
8. N. Tarhuni, T. O. Korhonen and M. Elmusrati, “State of polarization encoding for optical code
division multiple access networks”, J. Electromagnet. WavesAppl. (JEMWA), vol. 21, no. 10,
pp. 1313–1321, 2007
9. M. M. Karbassian and H. Ghafouri-Shiraz, “Performance analysis of heterodyne detected
coherent optical CDMA using a novel prime code family”, J. Lightw. Technol., vol. 25, no. 10,
pp. 3028–3034, 2007
10. M. M. Karbassian and H. Ghafouri-Shiraz, “Phase-modulations analyses in coherent homo-
dyne optical CDMA network using a novel prime code family”, in Proceedings of WCE-
ICEEE (IAENG), pp. 358–362, 2007
11. M. M. Karbassian and H. Ghafouri-Shiraz, “Performance analysis of unipolar code in different
phase modulations in coherent homodyne optical CDMA”, J. Eng. Lett. (IAENG), vol. 16,
no. 1, pp. 50–55, 2008
12. M. M. Karbassian and H. Ghafouri-Shiraz, “Novel channel interference reduction in optical
synchronous FSK-CDMA networks using a data-free reference”, J. Lightw. Technol., vol. 26,
no. 8, pp. 977–985, 2008
13. M. M. Karbassian and H. Ghafouri-Shiraz, “Frequency-shift keying optical code-division
multiple-access system with novel interference cancellation”, J. Microw. Opt. Techno. Lett.,
vol. 50, no. 4, pp. 883–885, 2008
14. S. Benedetto and P. Poggiolini, “Theory of polarization shift keying modulation”, IEEE Trans.
Comms., vol. 40, no. 4, pp. 708–721, 1992
Chapter 6
Template Based: A Novel STG Based Logic
Synthesis for Asynchronous Control Circuits

Sufian Sudeng and Arthit Thongtak

Abstract This article introduces a template based technique in state encoding

process to insert additional signals to STG with a small number. According to our
method, complete state coding (CSC) property can be satisfied without using state
graph tracing. Our method is useful for complicated asynchronous controllers, and
also it can guarantee the other relevant properties, such as persistency and consis-
tency. Our process begins with an encoding STG using Petri-net level in order to
form a template STG. Then the projection to each non-input signals from an orig-
inal STG is done. After that, we trace the projection to smaller state space. If the
small state space shows conflicts, we have to insert balance signals from template
STG. Unbalance signals are inserted after in case the state space still shows con-
flicts. Finally, we can get the STG with appropriate insertion points which is used
to be projected for CSC support on each non-input signals. Asynchronous DMA
controller is an example of our proposed method. The final part of this paper is con-
cluded with a complexity comparison between our template based method with state
based method and structural encoding method. It shows that the number of iterative
signal removal according to our method is less than others.

Keywords Signal Transition Graph (STG)  logic synthesis  asynchronous control

circuits  asynchronous DMA controller  template based technique

6.1 Introduction

State encoding is a process in logic synthesis for asynchronous control circuits. It

is defined as a given specification to interact between input and non-input signals.
The target is to find an encoding for every state of non-input signals in order to

S. Sudeng and A. Thongtak (B)

Digital System Engineering Laboratory (DSEL), Department of Computer Engineering Faculty
of Engineering, Chulalongkorn University, 254 Phyathai Road, Pathumwan, Bangkok, Thailand
10330, E-mails: sovanyy@msn.com,Arthit@cp.eng.chula.ac.th

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 59

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
60 S. Sudeng, A. Thongtak

implement the circuits with hazard free. To find a state encoding, several authors
have concerned with different ways, such as a logic synthesis using Petri-net
unfolding synthesis using state based technique [1–3], synthesis using structural
encoding technique, and etc. State based technique is involved several problems.
The major problem is the detection of state encoding conflicts. It causes exponential
complexity of conflict states with size of specification.
Another problem is the insertion of additional signals to solve the conflict states.
Signal insertion must be done in such a way that the new signals are consistent
(rising and falling transitions alternate) and their behaviors are hazard free. This
method is insufficient for the specification such large scale of signals. Structural
encoding method is more suitable than classical state based technique. The structural
encoding scheme is Petri-net level encoding. This method encodes with two silent
transitions [3, 5] to generate encode STG. Signal removal is begun after encoded
STG has been generated. Iteratively removing signal is done with greedy removal
from encode STG. In the end, it is produced a reduced STG. The next is to start
projection for CSC support for each non-input signals. This method is available
for large state encoding. However, it suffers from complexity reduction in greedy
removal phase according to the number of iterative signal reduction.
This article proposes a template based method for synthesizing complicated
STGs, and it also useful for some large scale STGs. The template based technique
not only be useful for complicated and large scale STGs but also solves the reduction
complexity of structural encoding technique.
Our template based technique was explained by one complicated example called
asynchronous DMA controller that has been used in our asynchronous processor
implementation [5].
We introduced three types of synthesis scheme in our explanation. Firstly, the
synthesis of asynchronous DMA controller using state based technique. We derived
the DMA’s STG to state graph and showed the limitation of synthesizable properties
which is limited by STG’s rule called complete state coding (CSC). The complete
state coding meant the state graph must be unique on state space; the asynchronous
DMA controller specification is unsatisfied while state based was applied. From
the state graph, if we try to insert an additional signal, the complication level is
increased and suffers the state explosion.
Secondly, structural encoding technique is applied. The asynchronous DMA con-
troller can be synthesized with this technique, but it shows some complexity in
signal removal phase. Finally, template based is introduced. We generate the tem-
plate STG from original STG using Petri-net level then we trace all non-input signals
to small state space and derive it to Karnaugh map, if it shows conflicts, we have
to insert some signals from template STG repeatedly, and we can get the final STG
called STG with appropriate insertion point.
The final part of this paper is concluded with the complexity comparison between
our template based technique, state based technique and structural encoding
technique.
6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 61

6.2 Petri-Net and Signal Transition Graph

To be able to introduce the methods of synthesis of asynchronous circuits in sub-

sequent sections, we will need a more formal definition of an STG. STGs are a
particular type of labeled Petri-Nets, while the definition of Petri-net is a net is a
triple Ndf D .PI TI F/ such that P and T are disjoint sets of respectively places and
transitions, and F .P  T/Œ.T  P/ is a flow relation. A marking of N is a multi
set M of places, i.e., M W P > f0; 1; 2: : :g. The transitions are associated with the
changes in the values of binary variables. These variables can be associated with
wires, when modeling interfaces between blocks, or with input, output and internal
signals in a control circuit [4].
Signal Transition Graph (STG) is a quadruple I 0 D .NI M0I ZI /,
where
P
D .NI M0 / is a Petri net (PN) based on a net N D .PI TI F/
Z˙ is a finite set of binary signals, which generates a finite alphabet Z˙ D Z˙ 
fC; g of signal transitions
 W T > Z˙ is a labeling function
An STG inherits the basic operation semantics from the behaviour underlying Petri-
Net. In particular, this includes: the rules for transition enabling and firing, the
notions of reachable markings, traces, and the temporal relation between transi-
tions. STGs also inherit the various structural and behavioural properties and the
corresponding classification of Petri-Nets Namely:
Choice place: A place is called a choice or conflict place if it has more than one
output transition.
Marked graph and State machine: A Petri-Net is called a marked graph if each
place has exactly one input and one output transition. A Petri-Net is called a state
machine if each transition has exactly one input and one output place. Marked
graph has no choice. Save state machine has no concurrency.
Free-choice: A choice place is called free-choice if every its output transition
has only one input place. A Petri-Net is free-choice if all its choice places are
free-choice.
Persistency: A transition t E T is called non-persistence if some reachable mark-
ing enables t together with another transition t’. Non-persistency of t with respect
to t’ is also called a direct conflict between t and t’. A Petri-Net is persistence if
it does not contain any non-persistence transition.
Boundless and safeness: A Petri-Net k-bounded if for every reachable mark-
ing the number of tokens in any place is not greater than k (a place is called
k-bounded if for every reachable marking the number of tokens in it is not
greater than k). A Petri-Net is bounded, if there is a finite k for which it is
k-bounded. A Petri-Net is safe if it is 1-bounded (1 bounded place is called a safe
place).
Liveness: A Petri-Net is live if for every transition t and every reachable marking
M there is a firing sequence that leads to a marking M’ enabling t.
62 S. Sudeng, A. Thongtak

The properties for synthesizable STGs:

Persistency – Excited signal fire, namely they cannot disable by other transition.
Consistency – Signal strictly alternate between C and .
Complete state coding – Different marking must represent different states.

6.3 Asynchronous DMA Controller Specification

The Asynchronous processor is a master device on the asynchronous system bus

[5], that it is able to initiate read and write transfers to all devices on the bus. The
other peripherals are slave devices, which they are only able to response to read or
write requests. The DMA controller is required to initiate read and write transfers to
memory and also other I/O on the bus, DMA controller also required to be a master
device (Fig. 6.1).
As there will be more than one master device in the system, asynchronous sys-
tem bus specification required the presence of a bus arbiter. The bus arbiter selects
which master is to have right to the bus at any one instant in time. An arbiter also
exits for the asynchronous processor. Initially the DMA controller is programmed

Setup
Transfer

Request
I/O not ready
I/O

Count ≠ 0 I/O ready

Transfer
data

Count = 0

Fig. 6.1 DMA basic

6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 63

Hold Dev_Req
Hold_Ack DMA WT_Req
CPU_INT WT_ack

(a)
TC_Req
TC_Ack

Fig. 6.2 Asynchronous DMA controller specification

addr
Bus data
Interface

Intr CPU Hold

Hold_Ack
DMA Dev_Req
CPU_INT
WT_Req
I/O
WT_Ack Module

TC_Req TC_Ack Intr_Ack

Base Addr
INC/DEC
Count

(a)

Fig. 6.3 Asynchronous DMA controller architecture

by asynchronous processor. This information is stored in internal DMA registers.

Include following information: base address for transfer source, base address in
memory to where the data is transferred and size of data transfer. The DMA con-
troller also be slave on the asynchronous system bus in order the asynchronous
processor access these registers.
The DMA controller waits for Dev req line to assert and take over the asyn-
chronous system bus from asynchronous processor as shown in Figs. 6.2 and 6.3. It
checks its internal registers to obtained details of the transfer. The DMA controller
read data from the source, and then writes it out to the memory, until the transfer is
complete.
64 S. Sudeng, A. Thongtak

WT_ack+

TC_Req+ WT_Req−

TC_Ack+ WT_Ack−

TC_Req– CPU_INT+

TC_Ack− Hold−
WT_Req+

Hold_ack−

Hold_Ack+

Dev_Req−

Hold+
CPU_INT−

(b)
Dev_Req+

Fig. 6.4 Signal transition graph

The internal architecture is very closely based on the classic DMA controller.
The architecture proposed of the DMA controller split into above functional
units. The most complex of these units is timing and control unit, which consist
of a large and complicated STG which described in the diagram on the Fig. 6.4.
The DMA controller works follow on STG as shown in Fig. 6.4. Finally, DMA
releases the usage of asynchronous system bus and activate the CPU INT line in
order to indicate the asynchronous processor that transfer was complete.

6.4 Synthesis Using State Based Method

The purpose of this section is to present the synthesis of DMA controller using
state-based technique. The key steps in this method are the generation of a state
graph, which is a binary encoded reach ability graph of the underlying Petri-net,
6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 65

11111000
TC_Req+ WT_Req-

WT_Ack+
11111100 11101000
TC_Ack+ TC_Req+ WT_Ack-
WT_Req-

11111110 11101100 11100000

TC_Ack+ TC_Req+
WT_Req- WT_Ack-

TC_Req- CPU_INT+ 11101110 11100100

TC_Req- CPU_INT+ 11110000
TC_Ack+
WT_Ack-
11111010 11111111
11100010 11100111
TC_Ack- TC_Req- Hold- 11100110
CPU_INT+ TC_Ack- TC_Req- Hold-
CPU_INT+
11111000 11111011 10110000 WT_req+
11100000 11100011 10100111
TC_Ack- TC_Req-
CPU_INT+ Hold- TC_Ack- TC_Req-
CPU_INT+ Hold-
11111001 10111011
TC_Req- CPU_INT+ 11100001 10100011
TC_Ack-
Hold- TC_Ack-
Hold- 11100000
10111001 11101010 11101111
TC_Ack- TC_Req- 10100001
Hold-
CPU_INT+

11101000 11101011 10101111

TC_Ack- TC_Req-
Hold- Hold_Ack+
Hold-
Hold_ack– Hold_ack–
11101001 10101011
Hold- TC_Ack-

10101001 11000000

Hold_ack-
01

01

WT_req-
01
10

WT_Ack- Hold+
10

00
01

00

00
10

10

10

Dev_req- Dev_req-
Dev_req-
01

01
10

01
10
01

00

00000000 10000000
00

WT_req-
00

WT_Ack-
00

CPU_INT– Dev_req+
00

00

Fig. 6.5 State graph

and deriving boolean equations for the output signals via their next state functions
obtained from the state-graph.
The state graph shown in Fig. 6.5 is a state space of DMA controller specification
that was derived from Fig. 6.4, the shadowed states are the conflict states, it clearly
suffers to state explosion problem of state based method, to solve the explosion
problem, an additional states are inserted, from state graph, it hardly finds an inser-
tion of the internal signals due to many of redundant states, it’s the major problem
of state based technique.
State encoding, at this point, we must have noticed a peculiar situation for the
value of the next-state function for signal that in two states with the same binary
encoding. This binary encoding is assigned to the shadowed states in Fig. 6.5.
The binary encoding of the state graph signals alone cannot determine the future
behavior. Hence, an ambiguity arises when trying to define the next-state function.
This ambiguity is illustrated in the Karnaugh map. When this occurs, the system is
said to violate the Complete State Coding (CSC) property. Enforcing CSC is one of
the most difficult problems in the synthesis of asynchronous circuits.
66 S. Sudeng, A. Thongtak

This section shows a limitation and problem of state based synthesis. However,
the solution of state based design is also possible; with inserting internal signals to
solve the two conflicting states are disambiguated by the value of CSC, which is the
last value in the binary vectors. Now Boolean minimization can be performed and
logic equations can be obtained.

6.5 Synthesis Using Structural Encoding Method

The structural encoding scheme provides a way to avoid the state space explosion
problem. The main benefit of using structural method is the ability to deal with large
highly concurrent specifications, which cannot be tracked by state base method, the
structural method for synthesis of asynchronous circuits is the class of mark graph.
By transform the STG specification to Petri net, and then encode it to be well form
specification as shown in Figs. 6.6 and 6.7.

WT_Ack+

Sp7+ Sp17+

WT_ack+
Sp6-
Sp16-

TC_Req+

WT_Req-

Sp13+ Sp8+
TC_Req+ WT_Req−
Sp18+
Sp7-

Sp17-
TC_Ack+

TC_Ack+ WT_Ack− Sp9+ Sp9+ WT_Ack-

Sp8- Sp14- Sp5-

Encoding Sp16+ Sp6+

TC_Req− CPU_INT+ TC_Req- Sp13-

WT_Req+

Sp10+ CPU_INT+

Sp4-

TC_Ack− Hold− WT_Req+ Sp9- Sp15+

Sp5+

TC_Ack- Sp14-

Hold_Ack+

Sp11+ Hold-
Hold_ack−
Sp3-
Sp10- Sp15-

Hold_Ack+ Sp4+
Hold_ack-

Sp12+
Dev_Req− Hold+

Sp11-

Sp2-

Dev_Req-

Hold+ Sp1+ Sp3+

CPU_INT− Sp18-
Sp2-
Dev_Req+

(b) CPU_INT-
Sp1-

Sp2+
Dev_Req+

Fig. 6.6 Encoded STG

6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 67

Sp2+ Sp11 Dev_req+

P1 P2 P1 P2 P1 P2

Sp6+ Sp5-
P6 P6
P6
WT_Ack+
P5
P5 P5
Sp5-
Sp16+
P16 P16 P16

Fig. 6.7 Encoding scheme (Dev Req+,WT Ack+)

The main idea of structural method is insertion of new signals in the initial
specification in a way that unique encoding is guaranteed in the transform speci-
fication.
Structural encoding technique, working at level of Petri-net can synthesize a big
size of specification. State based is used in the final state of this method, when
specification has been decomposing to smaller ones [4].
The synthesis flow is given a consistent STG, encode for all signals resulting an
STG contains a new set of signals that ensure unique state and complete state cod-
ing property. Depending encoding technique applied, since many of encoding signal
may unnecessary to guarantee unique and complete state coding, they are iteratively
removed each signal using greedy heuristics until no more signal can be removed
without violating unique and complete state coding property. The reduced STG next
projected onto different sets of signals to implement each individual output signal.
One the reduced STG is reached, it must be computed the complete state support
for each non-input signal, applying CSC support algorithm as shown in next section.
Afterward the projection of the STG into the CSC support for each non-input signal
is performed, finally speed independent synthesis of each projection is performed,
provide the small size of the projection, state based techniques can be applied for
performing the synthesis. When synthesizing the projection for each non-input sig-
nal a, every signal is the projection but a is considered as an input signal. These
prevent the logic synthesis to try to solve possible conflicts for the rest of signals
(Fig. 6.8).

6.6 Synthesis Using State Based Method

Structural encoding can be synthesizing a big size of specification; state based is

used in the final state of this method, when specification has been decomposed to
smaller ones. To design with structural encoding method, Petri-net level is applied,
begins with encode original STG with two silent transition, the encoding ele-
ment composes of two transitions and one place, then iterative signals reduction
immediately with complete state checking, finally, STG with remaining signals are
68 S. Sudeng, A. Thongtak

Initial STG

Encoded STG

Signal Removal
Transformation
+
CSC Checking
Reduced STG

Projection for a1 Projection for an

Synthesis
Synthesis

Circuit for a1 Circuit for an

Fig. 6.8 Structural encoding synthesis

produced, and project for all non-inputs signal to get the small size CSC support,
apply state based method for each CSC support to get a circuits for each non-input
signals, However, structural based still consumed more wasting time in the design
phase, especially signal reduction phase, each time of reduction, it must verify the
persistence, consistence and complete state coding properties before reached the
final reduction.
This work focuses on complexity reduction; our process begins with an encoding
STG using Petri-net level in order to form a template STG. Then the projection
to each non-input signals from an original STG is done. After that, we trace the
projection to smaller state space. If the small state space shows conflicts, we have
to insert balance signals from template STG. Unbalance signals are inserted after in
case the space still shows conflicts. Finally, we can get the STG with appropriate
insertion points which are used to be projected for CSC support on each non-input
signal as shown in Fig. 6.9.

Template based method (insertion point calculation)

Let:
a = non input signal
b = balance signal
ub = unbalance signal
Original STG Projection (a)
6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 69

Initial STG

Template STG

Original STG
Conflict = Yes Projection for s1 Projection for sn Projection

Conflict = No

STG with Insertion

Points

Projection for a1 Projection for an STG with

Insertion Point
Synthesis Synthesis Projection

Circuit for a1 Circuit for an

Fig. 6.9 Template based synthesis

While not Conflict do

Trace a;
If Conflict Then
Insert b from template STG;
Else If Still Conflict Then
Insert ub from template STG;
Else
Trace a;
End If;

Figure 6.10 shows the template STG that encoded with the encoding scheme in
Fig. 6.7.
After the template based technique was applied. We can get the STG with
appropriate insertion signals as shown in Fig. 6.11.

CSC Support calculation:

CSC Support (STG S, a) return CSC support of a
CSC support a = Trigger (a) U fag
While it’s still conflict do
Let b an unbalance signal in z
CSC Support for a = CSC Support for a U fbg
70 S. Sudeng, A. Thongtak

WT_Ack+

Sp7+ Sp17+

Sp6−
Sp16−

+
TC_Req

WT_Req−

Sp 13+ Sp8+

Sp18+
Sp7−

Sp17−
TC_Ack+

Sp9+ Sp9+ WT_Ack−

Sp8− Sp14− Sp5−

Sp16+ Sp6+

TC_Req− Sp13−

WT_Req+

Sp10+ CPU_INT+

Sp4−

Sp9− Sp15+

Sp5+

TC_Ack− Sp14−

Hold_Ack+

Sp11+ Hold−

Sp3−
Sp10− Sp15−

Hold_ack− Sp4+

Sp12+
Hold+

Sp11−

Sp2−

Dev_Req−

Sp1+ Sp3+

Sp2− Sp18−
Dev_Req+

CPU_INT−
Sp1−

Sp2+

Fig. 6.10 Template STG

6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 71

WT_ack+

Sp5+
TC_Req+ WT_Req−

TC_Ack+ WT_Ack−
WT_Req+

Sp9+ CPU_INT+

Hold_Ack+
Sp9− Hold−

TC_Ack−
Sp5−

Hold_ack−

Sp1+
Hold+
Dev_Req−

Dev_Req+

CPU_INT−

Sp1−

Fig. 6.11 STG with appropriate insertion signals

CSC support calculation for each non-input (output and internal signals) signal
is calculated. Finally we got the small state graph which guarantee that it save
from state explosion and also satisfied the synthesizing properties (Figs. 6.12
and 6.13).
72 S. Sudeng, A. Thongtak

TC_Req+ TC_Ack–

Sp9–
TC_Ack+

WT_ack+ Sp9–
TC_Req– WT_Req+
Sp9+ Sp5+

TC_Req–
TC_req+ SP5–
CSC Support WT_Ack+
(Sp9)
Sp9+
WT_ack– WT_Ack–
WT_Req+

CSC Support CSC Support

(TC_Req) WT_ack+
(WT_Req)

Sp5+
TC_Req WT_Req–
+

TC_Ack+ WT_Ack
-
TC_Ack+ WT_Req+
Sp1+

Sp9+ CPU_INT+
Sp5+ Hold_ack+
Sp9+ CPU_INT+
Dev_Req+ Hold_Ack+
Sp9– Hold–
SP5–
TC_Ack– CSC Support WT_Req-
Sp1– (CPU_INT) TC_Ack– CSC Support
Sp9–
Sp5– (Sp5) WT_Req+
Hold_Ack-
CPU_INT–
Hold_ack–

Dev_req– Sp1+
Hold+
Dev_Req–

Dev_Req+
CPU_INT–

CSC Support
Sp1–
(Hold)

CSC Support
WT_Req+
(Sp1) WT_Ack+

Hold+ Dev_Req+
CPU_INT+ WT_Req–
Dev_Req+
CPU_INT–
CPU_INT+
Dev_Req– Sp1+

Dev_req– WT_Ack–
Hold–
Sp1–

CPU_INT–

Fig. 6.12 STG with Insertion Point and CSC support calculation

6.7 Result

This section presents complexity comparison between structural encoding method

and template based method. The conventional state based method is not shown on a
comparison table, it clearly suffers its exponential complexity of conflict state with
the size of specification, and others difficult problem. Due to absence of synthesis
6 Template Based: A Novel STG Based Logic Synthesis for Asynchronous 73

2. CPU_INT 3. TC_Req
Sp1
1. Hold Dev_Req Sp9
Dev_Req
Hold CPU_INT TC_Req
CPU_INT TC_Ack WT_Ack
5. Sp1

WT_Ack
4. WT_Req
CPU_INT Sp1
6. Sp5
WT_Ack
Sp5 Dev_Req Hold_Ack
Sp5
Hold_Ack WT_Req WT_Req
7. Sp9

TC_Req
Sp9
TC_Ack

Fig. 6.13 Synthesized circuits from CSC support of Fig.6.12

Table 6.1 Result

Circuits Signals Structural encoding Template based
Iteration Inserted Iteration Inserted
signals signals
VME bus 6 10 2 6 2
Asyn DMA 8 10 6 7 6
Pipeline-Scheduler 14 26 2 18 2
QDI-bus 16 28 4 20 3
Pipeline control block 13 23 3 18 3
Classical DMA 11 13 9 8 6
PPArbCSC (2,3) 10 18 2 8 2

tools for asynchronous controller synthesis, all circuits in Table 6.1 were manually
synthesis.

6.8 Concluding Remarks

This paper proposes template based technique for reducing additional signals in
signal transition graph (STG) based logic synthesis, and solves the reduction com-
plexity of structural encoding method; our method is explained by asynchronous
DMA controller specification. The final section showed the comparison between
our template based method with state based and structural encoding method. The
number of iterative signal removal according to our method is less than others,
Roughly speaking, structural encoding is better than classical state based method,
and template based method is less complexity than structural encoding.
74 S. Sudeng, A. Thongtak

Acknowledgement The research is supported by TJTTP-OECF (Thailand – Japan Technology

Transfer Project – Japanese Overseas Economic Cooperation Fund).

References

1. S.B. Park. Synthesis of Asynchronous VLSI circuit from Signal Transition Graph Specifica-
tions, Ph.D. thesis, Tokyo Institute of Technology, 1996.
2. J. Carmona, J. Cortadella. ILP models for the synthesis of asynchronous control circuits. In
Proceedings of the International Conference Computer-Aided Design (ICCAD), pp. 818–825,
November 2003.
3. J. Carmona and J. Cortadella. State encoding of large asynchronous controllers. In Proceedings
of the ACM/IEEE Design Automation Conference, pp. 939–944, July 2006.
4. J. Carmona, J. Cortadella, and E. Pastor. A structural encoding technique for the synthesis of
asynchronous circuits. In International Conference on Application of Concurrency to System
Design, pp. 157–166, June 2001.
5. S. Sudeng and A. Thongtak FPGA Implementation of Quasi-Delay Insensitive Microproces-
sor. The World Congress on Engineering and Computer Science (WCECS2007), Clark Kerr
Campus, University of California Berkeley, San Francisco, CA, USA on 24–26 October 2007.
Chapter 7
A Comparison of Induction Motor Speed
Estimation Using Conventional MRAS
and an AI-Based MRAS Parallel System

Chao Yang and Professor John W. Finch

Abstract The Model Reference Adaptive System (MRAS) is probably the most
widely applied speed sensorless drive control scheme. This chapter compares induc-
tion motor speed estimation using conventional MRAS and AI-based MRAS with
Stator Resistance Compensation. A conventional mathematical model based MRAS
speed estimation scheme can give a relatively precise speed estimation result, but
errors will occur during low frequency operation. Furthermore, it is also very sen-
sitive to machine parameter variations. An AI-based MRAS-based system with a
Stator Resistance Compensation model can improve the speed estimation accuracy
and is relatively robust to parameter variations even at an extremely low frequency.
Simulation results using a validated machine model are used to demonstrate the
improved behaviour.

Keywords Dynamic reference model  Model Reference Adaptive System

(MRAS)  neural networks  induction motor control

7.1 Introduction

Much effort has been devoted to speed-sensorless induction machine drive schemes,
with the Model Reference Adaptive System (MRAS) being the most popular [1].
In a conventional mathematical-model-based MRAS, some state variables are esti-
mated in a reference model (e.g. rotor flux linkage components, rd , rq , or
back e.m.f. components, ed , eq , etc.) of the induction machine obtained by using
measured quantities (e.g. stator currents and perhaps voltages).
These reference model components are then compared with state variables esti-
mated using an adaptive model. The difference between these state variables is then

Prof. J. W. Finch (B)

School of Electrical, Electronic and Computer Engineering, Newcastle University, Newcastle upon
Tyne, NE1 7RU, UK,
E-mail: J.W.Finch@ncl.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 75

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
76 C. Yang, Prof. J. Finch

used in an adaptation mechanism, which, for example, outputs the estimated value
of the rotor speed .!O r / and adjusts the adaptive model until satisfactory performance
is obtained [2–6].
Greater accuracy and robustness can be achieved if the mathematical model is
not used at all and instead an AI-based non-linear adaptive model is employed. It is
then also possible to eliminate the need of the separate PI controller, since this can
be integrated into the tuning mechanism of the AI-based model [7].
However, both the conventional MRAS and AI-based MRAS scheme are easily
affected by machine parameter variations, which happen during practical opera-
tion [8, 9]. Hence an online stator resistance estimator is applied to the AI-based
MRAS scheme which is shown to make the whole scheme more robust using a com-
puter simulation and could make the scheme usable for practical operation [10, 11].
The comparison of schemes presented here is felt to be valuable since much of the
literature presents results for the novel approach alone [1].

7.2 Speed Estimation Using Conventional Model Reference

Adaptive System

It is possible to estimate the rotor speed by using two estimators (a reference-model-

based estimator and an adaptive-model-based one), which independently estimate
the rotor flux-linkage components in the stator reference frame rd , rq /, and,
using the difference between these flux-linkage estimates, drive the speed of the
adaptive model to that of the actual speed [12]. The expressions for the rotor flux
linkages in the stationary reference frame can be obtained by using the stator volt-
age equations of the induction machine (in the stationary reference frame). These
give (7.1) and (7.2), which are now rearranged for the rotor flux linkages:
Z 
0
rd D .L r =Lm / .u sD  R i
s sD /dt  L i
s sD (7.1)
Z 
rq D .Lr =Lm / .usQ  Rs isQ /dt  L0s isQ (7.2)

These two equations represent a so-called stator voltage model, which does not con-
tain the rotor speed and is therefore a reference model. However, when the rotor
voltage equations of the induction machine are expressed in the stationary reference
frame, they contain the rotor fluxes and the speed as well. These are the equations
of the adaptive model:
Z 
O rd D .1=Tr / Lm isD  O rd  !O r Tr O rq dt (7.3)
Z 
O rq D .1=Tr / Lm isQ  O rq  !O r Tr O rd dt (7.4)
7 A Comparison of Induction Motor Speed Estimation 77

Flux estimator ψ′ ε
Speed tuning signal ω
reference model
ψ ′ × ψ ′ =Im ψ ′ψ ′

Flux estimator ψ′
adaptive model
ω

Kp+K ¤p

Fig. 7.1 MRAS-based rotor speed observer using rotor flux linkages for the speed tuning signal

The reference and adaptive models are used to estimate the rotor flux linkages and
the angular difference of the outputs of the two estimators is used as the speed
tuning signal. This tuning signal is the input to a linear controller (PI controller)
which outputs the estimated rotor speed as shown in Fig. 7.1. The estimated speed
can be expressed as (7.5)
Z
!O r D Kp "! C Ki "! dt (7.5)

7.3 Artificial Intelligence-Based Model Reference Adaptive

System

The MRAS-based schemes described in the previous section contain a reference

model and an adaptive model. Greater accuracy and robustness can be achieved if
the mathematical model is partially replaced by a neural network. It is then also pos-
sible to eliminate the need of the separate PI controller, since this can be integrated
into the tuning mechanism of the neural network-based model.
The neural network-based model can take various forms: it can be an artificial
neural network (ANN) or a fuzzy-neural network etc. and there is also the possibility
of using different types of speed tuning signals. It is believed that some of these
solutions can give high accuracy and are relatively robust to parameter variations
even at extremely low stator frequency.
One specific implementation of the ANN-based MRAS speed estimator system
which is popular in academic work, shown in Fig. 7.2, which is similar to the con-
ventional MRAS system. In this new model the adaptive model is replaced by a
78 C. Yang, Prof. J. Finch

Fig. 7.2 MRAS-based rotor speed estimator containing a two-layer ANN

simple two layer neural network, which gives the whole system a fast response and
better accuracy than the conventional MRAS [13, 14].

7.4 MRAS Based Two-Layer ANN Speed Estimator

with Dynamic Reference Model

Compared to the conventional MRAS scheme, the MRAS based rotor speed esti-
mator containing a two-layer ANN could give more accurate estimation result
and be relatively robust to parameters variations. The two-layer ANN replaces the
adjustable model and adaptive mechanism in the conventional MRAS, but the ref-
erence model is still necessary for estimation the rotor flux which is used as speed
tuning signal. Several machine parameters are used to build the conventional refer-
ence model, such as stator resistance .Rs / and stator reluctance .Ls /. These parame-
ters may change during the different periods of motor operation. The values of these
parameters are fixed in the reference model. So the ANN speed estimator is still sen-
sitive to parameter variations especially during the motor low speed running period.
To solve this problem and make this scheme more independent to the machine
parameters a stator resistance estimator is built in the new reference model, in which
the stator resistance Rs value could be estimated online. Figure 7.3 shows the total
scheme of this neural network based MRAS with a dynamic reference model.
In this new system, both the reference model and adaptive model of the conven-
tional MRAS system are modified for better performance. The whole system can
be divided into two main parts, the dynamic reference model part and the neural
network part. The dynamic reference part consists of the dynamic reference model
derived from Eqs. (7.1) and (7.2), in which the stator resistance Rs is replaced by
7 A Comparison of Induction Motor Speed Estimation 79

ψrd +
usD
usQ Dynamic Reference
ψrq - +
isD Model
isQ
-

Adaptive
Z-1 Z-1 Rs Tuning
Mechanism
PI Controller Signal

Two-layer ANN

W2 Weight Weight
-1
Z Z-1 Adjustment

Fig. 7.3 MRAS based ANN speed estimator with dynamic reference model

the online estimated value RO s coming from Eqs. (7.6) and (7.7),
 
O Ki
Rs D Kp C eRs (7.6)
p
eRs D isD . rd  O rd / C isQ . rq  O rq / (7.7)

The neural network part contains a simple two-layer network, with only an input
layer and an output layer. Adjustable and constant weights are built in the neural net-
work, and the adjustable weights are proportional to the rotor speed. The adjustable
weights are changed by using the error between the outputs of the reference model
and the adjustable model, since any mismatch between the actual rotor speed and
the estimated rotor speed results in an error between the outputs of the reference and
adaptive estimators.
To obtain the required weight adjustments in the ANN, the sampled data forms
of Eqs. (7.3) and (7.4) are considered. using the backward difference method the
sampled data forms of the equations for the rotor flux linkages can be written as
(7.8) and (7.9), where T is the sampling time.

ΠO rd .k/  O rd .k  1/=T D  O rd .k  1/=Tr

(7.8)
!r O rq .k  1/ C .Lm =Tr /isD .k  1/
ΠO rq .k/  O rq .k  1/=T D  O rq .k  1/=Tr
(7.9)
!r O rd .k  1/ C .Lm =Tr /isQ .k  1/
80 C. Yang, Prof. J. Finch

Thus the rotor flux linkages at the kth sampling instant can be obtained from the
previous .k  1/th values as

O rd .k/ D O rd .k  1/.1  T =Tr /

(7.10)
!r T O rq .k  1/ C .Lm T =Tr /isD .k  1/
O rq .k/ D O rq .k  1/.1  T =Tr /
(7.11)
!r T O rd .k  1/ C .Lm T =Tr /isQ .k  1/

Introducing c D T =Tr , the following weights are given:

w1 D 1  c
w2 D !r cTr D !r T (7.12)
w3 D cLm

It can be seen that w1 and w3 are constant weights, but w2 is a variable weight and
is proportional to the speed. Thus Eqs. (7.10) and (7.11) take the following forms:

O rd .k/ D w1 O rd .k  1/  w2 O rq .k  1/
(7.13)
C w3 isD .k  1/

O rq .k/ D w1 O rq .k  1/ C w2 O rd .k  1/
(7.14)
C w3 isQ .k  1/
These equations can be visualized by the very simple two-layer ANN shown in
Fig. 7.4.
The neural network is training by the backpropagation method, the estimated
rotor speed can be obtained from:

!O r .k/ D !O r .k  1/ C w2 .k/=T C .˛=T /w2 .k  1/

D !O r .k  1/ C fΠrd .k/  O rd .k/ O rq .k  1/ (7.15)
C Œ rq .k/  O rq .k/ O rd .k  1/g=T C .˛=T /w2 .k  1/

Fig. 7.4 Neural network representation for estimated rotor flux linkages
7 A Comparison of Induction Motor Speed Estimation 81

where is the learning rate and ˛ is a positive constant called the momentum con-
stant. The inclusion of the momentum term into the weight adjustment mechanism
can significantly increase the convergence, which is extremely useful when the ANN
shown in Fig. 7.4 is used to estimate in real time the speed of the induction machine.

7.5 Simulation Results and Discussion

To compare the conventional MRAS and the AI-based MRAS with dynamic refer-
ence model, simulations are established by using Matlab-Simulink software, based
on the standard well established validated 2-axis machine model [6].
Speed estimation results using conventional MRAS and neural network based
MRAS are shown in Figs. 7.5 and 7.6 respectively. These results assume the
machine parameters are correctly measured and unchanged during operation. Both
of the two schemes can give good speed tracking results.
Further simulation were carried out with changed stator resistance to test how
much changing parameters would affect the speed estimation results.
In Figs. 7.7 and 7.8, simulations are carried out with the stator resistance changed
by a small amount, 2%. Obviously both schemes are still sensitive to parameter
variations.
A final simulation for AI-based MRAS with the dynamic reference model is
shown in Fig. 7.9. The online estimated stator resistance is displayed in Fig. 7.10.
This simulation result shows the effect caused by the stator resistance variation has
been considerably improved.

Fig. 7.5 Speed estimation using Conventional MRAS

82 C. Yang, Prof. J. Finch

Fig. 7.6 Speed estimation using two-layer ANN MRAS

Fig. 7.7 Speed estimation by using conventional MRAS (with stator resistance Rs changed 2%)

Comparing all the above simulation results shows that the conventional MRAS
scheme works well when the parameters are precisely measured and do not change
during operation. The MRAS with adaptive model replaced by the two-layer neural
network can slightly improve the performance when working in the same situation.
But both schemes can still be easily affected by parameters variations, which do
occur during practical operation. Introducing the stator resistance online estimator
gives much improved performance which should enable the scheme to be usable for
practical operation.
7 A Comparison of Induction Motor Speed Estimation 83

Fig. 7.8 Speed estimation using two-layer ANN MRAS (with stator resistance Rs changed 2%)

200
Estimated Speed
150 Rotor Speed

100

50
Speed

-50

-100

-150

-200
0 1 2 3 4 5 6
Time

Fig. 7.9 Speed estimation using two-layer MRAS with dynamic reference model

7.6 Conclusion

The main objective of this chapter is to compare conventional MRAS and AI-based
MRAS parallel system for induction motor sensorless speed estimation. The con-
ventional MRAS can give good speed estimation in most of the operation period,
but errors will occur during low frequency operation mainly caused by machine
parameter variations. An AI-based MRAS system can give improved accuracy and
84 C. Yang, Prof. J. Finch

2.5
Estimated Rs
Rs
2

1.5
Rs

0.5

-0.5
0 1 2 3 4 5 6
Time

Fig. 7.10 Estimated Rs in the dynamic reference model

bypasses the PI controller tuning problems. The simple structure of the two-layer
neural network shown in Fig. 7.4 yields a speed estimation system working online
with a fast response. Also the simple two-layer neural network does not require a
separate learning stage, since the learning takes place during the on-line speed esti-
mation process. This is mainly due to the fact that the development time of such an
estimator is short and the estimator can be made robust to parameter variations and
noise. Furthermore, in contrast to most conventional schemes, it can avoid the direct
use of a speed-dependent mathematical model of the machine.
However, the Two-layer neural network MRAS lies more in the realm of adaptive
control then neural networks. The speed value is not obtained at the output, but as
one of the weights. Moreover, only one weight is adjusted in the training. Therefore,
it would still be sensitive to parameter variations and system noise.
In the new approach, an online stator resistance estimator is used to compen-
sate the parameter variations. The computer simulation results show that this new
approach makes the whole scheme more robust to parameter variations, enhancing
the possibility of practical use of the neural network based MRAS scheme. The sta-
tor resistance estimator works under an adaptive mechanism (PI controller). Further
study could be carried out to replace the PI controller with another simple neural
network which could also estimate more machine parameters.

References

1. Finch, J.W. and Giaouris, D., Controlled AC Electrical Drives, IEEE Transactions on
Industrial Electronics, Feb. 2008, 55, 1, 1–11.
2. Landau, Y.D., Adaptive Control the Model Reference Approach, 1979: Marcel Dekker, New
York.
7 A Comparison of Induction Motor Speed Estimation 85

3. Vas, P., Sensorless Vector and Direct Torque Control, 1998: Oxford University Press, Oxford.
4. Shauder, C., Adaptive Speed Identification for Vector Control of Induction Motors without
Rotational Transducers. IEEE Transactions on Industry Applications, 1992, 28, 1054–1062.
5. Yang, G. and T. Chin, Adaptive-Speed Identification Scheme for a Vector-Controlled Speed
Sensorless Inverter-Induction Motors. IEEE Transactions on Industry Applications, 1993, 29,
820–825.
6. Fitzgerald, A.E., C. Kingsley, and S.D. Umans, Electric Machinery. 6th ed., 2003: McGraw-
Hill International Edition, New York.
7. Vas, P., Artificial-Intelligence-Based Electrical Machines and Drives. 1999: Oxford University
Press, Oxford.
8. Kumara, I.N.S., Speed Sensorless Field Oriented Control for Induction Motor Drive. Ph.D.
thesis, 2006, University of Newcastle upon Tyne.
9. Leonhard, W., Controlled AC Drives, a Successful Transition from Ideas to Industrial Practice,
1996: Elsevier, Amsterdam, Netherlands.
10. Zhen, L. and L. Xu, Sensorless Field Orientation Control of Induction Machines Based on a
Mutual MRAS Scheme. IEEE Transactions on Industrial Electronics, 1998, 45, 824–831.
11. Holtz, J. and J. Quan, Drift-and Parameter-Compensated Flux Estimator for Persistent Zero-
Stator-Frequency Operation of Sensorless-Controlled Induction Motors. IEEE Transactions
on Industry Applications, 2003, 39, 1052–1060.
12. Ohtani, T., N. Takada, and K. Tanaka, Vector Control of Induction Motor without Shaft
Encoder. IEEE Transactions on Industry Applications, 1992, 28, 157–164.
13. Peng, F.Z. and T. Fukao, Robust Speed Identification for Speed-Sensorless Vector Control of
Induction Motors. IEEE Transactions on Industry Applications, 1994, 30, 945–953.
14. Vasic, V. and S. Vukosavic, Robust MRAS-Based Algorithm for Stator Resistance and Rotor
Speed Identification. IEEE Power Engineering Review, 2001, 21, 39–41.
Chapter 8
A New Fuzzy-Based Additive Noise Removal
Filter

M. Wilscy and Madhu S. Nair

Abstract A digital color image C can be represented in different color space such
as RGB, HSV, L a b etc. In the proposed method, RGB space is used as the basic
color space. Different proportions of red, green and blue light gives a wide range
of colors. Colors in RGB space are represented by a 3-D vector with first element
being red, the second being green and third being blue, respectively. The general
idea in this method is to take into account the fine details of the image such as edges
and color component distances, which will be preserved by the filter. The goal of
the first filter is to distinguish between local variations due to image structures such
as edges. The goal is accomplished by using Euclidean distances between color
components instead of differences between the components as done in most of the
existing filters. The proposed method uses 2-D distances instead of 3-D distances,
and uses three fuzzy rules to calculate weights for the Takagi-Sugeno fuzzy model.

Keywords Additive Noise Removal Filter  Fuzzy-Based  digital color image 

fuzzy rules  Image restoration

8.1 Introduction

Images are often degraded by random noise. Noise can occur during image capture,
transmission or processing, and may be dependent on or independent of image con-
tent. Noise is usually described by its probabilistic characteristics. Gaussian noise
is a very good approximation of noise that occurs in many practical cases [1].
The ultimate goal of restoration techniques is to improve an image in some pre-
defined sense. Although there are areas of overlap, image enhancement is largely

M. S. Nair (B)
Rajagiri School of Computer Science, Rajagiri College of Social Sciences, Kalamassery,
Kochi 683104, Kerala, India,
E-mail: madhu s nair2001@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 87

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
88 M. Wilscy, M.S. Nair

a subjective process, while image restoration is for the most part an objective
process. Restoration attempts to reconstruct or recover an image that has been
degraded by using a priori knowledge of the degradation phenomenon [2]. Thus
restoration techniques are oriented toward modeling the degradation and applying
the inverse process in order to recover the original image. This approach usually
involves formulating a criterion of goodness that will yield an optimal estimate
of the desired result. By contrast, enhancement techniques basically are heuris-
tic procedures designed to manipulate an image in order to take advantage of the
psychophysical aspects of human visual system. For example, histogram equaliza-
tion is considered an enhancement technique because it is primarily on the pleasing
aspects it might present to the viewer, whereas removal of image blur by applying a
deblurring function is considered a restoration technique.
Image restoration differs from image enhancement in that the latter is con-
cerned more with accentuation or extraction of image features rather than restoration
of degradations. Image restoration problems can be quantified precisely, whereas
enhancement criteria are difficult to represent mathematically. Consequently, restor-
ation techniques often depend only on the class or ensemble properties of a data
set, whereas image enhancement techniques are much more image dependent. The
degradation process is usually modeled as a degradation function that, together with
an additive noise term, operates on an input image f .x; y/ to produce a degraded
image g.x; y/. Given g.x; y/, some knowledge about the degradation function H ,
and some knowledge about the additive noise term .x; y/, the objective of restora-
tion is to obtain an estimate fO.x; y/ of the original image. The estimate needs to
be as close as possible to the original input image and, in general, the more about
H and is known, the closer fO.x; y/ will be to f .x; y/.
If H is a linear, position-invariant process, then the degraded image is given in
the spatial domain by

g.x; y/ D h.x; y/ f .x; y/ C .x; y/

where h.x; y/ is the spatial representation of the degradation function and the
symbol “*” indicates convolution [2].
Image noise reduction has come to specifically mean a process of smoothing
noise that has somehow corrupted the image. During image transmission, noise
which is usually independent of the image signal occurs. Noise may be additive,
where noise and image signal g are independent.

f .x; y/ D g.x; y/ C v.x; y/

where f .x; y/ is the noisy image signal, g.x; y/ is the original image signal and
v.x; y/ is the noise signal which is independent of g [3]. The additive noise image v
models an undesirable, unpredictable corruption of g. The process v is called a two-
dimensional random process or a random field. The goal of restoration is to recover
an image h that resembles g as closely as possible by reducing v. If there is an ade-
quate model for the noise, then the problem of finding h can be posed as the image
8 A New Fuzzy-Based Additive Noise Removal Filter 89

estimation problem, where h is found as the solution to a statistical optimization

problem. The detailed statistics of the noise process v may are unknown. In such
cases, a simple linear filter approach can yield acceptable results, if the noise satis-
fies certain simple assumptions such as zero-mean additive white noise model [4].
Noise may be impulse noise, which is usually characterized by some portion of
image pixels that are corrupted, leaving the remaining pixels unchanged. The most
common example of the impulse noise is the salt-and-pepper noise removal.

8.2 Conventional Noise Removal Techniques

Noise reduction is the process of removing noise from a signal. Noise reduction
techniques are conceptually very similar regardless of the signal being processed,
however a priori knowledge of the characteristics of an expected signal can mean the
implementations of these techniques vary greatly depending on the type of signal.
Although linear image enhancement tools are often adequate in many applications,
significant advantages in image enhancement can be attained if non-linear tech-
niques are applied [3]. Non-linear methods effectively preserve edges and details
of images, whereas methods using linear operators tend to blur and distort them.
Additionally, non-linear image enhancement tools are less susceptible to noise.
One method to remove noise is to use linear filters by convolving the original
image with a mask. The Gaussian mask comprises elements determined by a Gaus-
sian function. It gives the image a blurred appearance if the standard deviation of
the mask is high, and has the effect of smearing out the value of a single pixel over
an area of the image. Averaging sets each pixel to the average value of itself and its
nearby neighbors. Averaging tends to blur an image, because pixel intensity values
which are significantly higher or lower than the surrounding neighborhood would
smear across the area. Conservative smoothing is another noise reduction technique
that is explicitly designed to remove noise spikes (e.g. salt and pepper noise) and
is, therefore, less effective at removing additive noise (e.g. Gaussian noise) from an
image [5].
Additive noise is generally more difficult to remove from images than impulse
noise because a value from a certain distribution is added to each image pixel, for
example, a Gaussian distribution. A huge amount of wavelet based methods [6]
are available to achieve a good noise reduction (for the additive noise type), while
preserving the significant image details.. The wavelet denoising procedure usually
consists of shrinking the wavelet coefficients, that is, the coefficients that contain
primarily noise should be reduced to negligible values, while the ones containing
a significant noise-free component should be reduced less. A common shrinkage
approach is the application of simple thresholding nonlinearities to the empirical
wavelet coefficients [7, 8]. Shrinkage estimators can also result from a Bayesian
approach, in which a prior distribution of the noise-free data (e.g., Laplacian [4],
generalized Gaussian [9–11]) is integrated in the denoising scheme.
90 M. Wilscy, M.S. Nair

Fuzzy set theory and fuzzy logic offer us powerful tools to represent and process
human knowledge represented as fuzzy if-then rules. Several fuzzy filters for noise
reduction have already been developed, e.g., the iterative fuzzy control based filters
from [12], the GOA filter [13, 14], and so on. Most of these state-of-the-art methods
are mainly developed for the reduction of fat-tailed noise like impulse noise. Never-
theless, most of the current fuzzy techniques do not produce convincing results for
additive noise [15, 16]. Another shortcoming of the current methods is that most of
these filters are especially developed for grayscale images. It is, of course, possible
to extend these filters to color images by applying them on each color component
separately, independent of the other components. However, this introduces many
artifacts, especially on edge or texture elements.
A new fuzzy method proposed by Stefan Schulte, Valérie De Witte, and Etienne
E. Kerre, is a simple fuzzy technique [17] for filtering color images corrupted with
narrow-tailed and medium narrow-tailed noise (e.g., Gaussian noise) without intro-
ducing the above mentioned artifacts. Their method outperforms the conventional
filter as well as other fuzzy noise filters. In this paper, we are presenting a modified
version of the fuzzy approach proposed by Stefan Schulte, et al. [17], which uses a
Gaussian combination membership function to yield a better result, compared to the
conventional filters as well as the recently developed advanced fuzzy filters.

8.3 Proposed Fuzzy Noise Removal Filter

A digital color image C can be represented in different color space such as RGB,
HSV, L*a*b etc. In the proposed method, RGB space is used as the basic color
space. Different proportions of red, green and blue light gives a wide range of colors.
Colors in RGB space are represented by a 3-D vector with first element being red,
the second being green and third being blue, respectively. These three primary color
components are quantized in the range 0 to 2m  1, where m D 8. A color image
C can be represented by a 2-D array of vectors where (i, j) defines a position in C
called pixel and Ci;j;1 ; Ci;j;2 , and Ci;j;3 , denotes the red, green and blue components,
respectively.

8.3.1 Fuzzy Sub-filter I

The general idea in this method is to take into account the fine details of the image
such as edges and color component distances, which will be preserved by the filter.
The goal of the first filter is to distinguish between local variations due to image
structures such as edges. The goal is accomplished by using Euclidean distances
between color components instead of differences between the components as done
in most of the existing filters. The proposed method uses 2-D distances instead of
3-D distances (distance between three color components red, green and blue), that is,
the distance between red-green (rg) and red-blue (rb) of the neighbourhood centered
8 A New Fuzzy-Based Additive Noise Removal Filter 91

at (i, j) is used to filter the red component [4]. Similarly, the distance between RG
and green-blue (gb) is used to filter the green component and the distance between
rb and gb is used to filter the blue component, respectively. The method uses three
fuzzy rules to calculate weights for the Takagi-Sugeno fuzzy model [10].
The current image pixel at position (i, j) is processed using a window size of
.2K C 1/  .2K C 1/ to obtain the modified color components. To each of the
pixels in the window certain weights are then assigned namely Wk;l , where k; l 2
f1; 0; 1g. WiCk;jCl;1 , WiCk;jCl;2 , and WiCk;jCl;3 denotes the weights for the red,
green and blue component at position (i + k, j + l), respectively. These weights are
assigned according to the following three fuzzy rules. Let DT (a, b) represents the
distance between the parameters a and b, and NG(y) represents the neighbourhood
of the parameter y. In this case, y represents a pixel with the neighbourhood given
by a 3  3 window. The three fuzzy rules can be represented as follows:
1. If DT(rg, NG(rg)) is SMALL AND DT(rb, NG(rb)) is SMALL THEN the weight
Wk;l;1 is LARGE
2. IF DT(rg, NG(rg)) is SMALL AND DT(gb, NG(gb)) is SMALL THEN the weight
Wk;l;2 is LARGE
3. IF DT(rb, NG(rb)) is SMALL AND DT(gb, NG(gb)) is SMALL THEN the weight
Wk;l;3 is LARGE
In the above fuzzy rules DIST represents the Euclidean distance.

DT.rg; NG.rg// D Œ.CiCk;jCl;1  Ci;j;1 /2 C .CiCk;jCl;2  Ci;j;2 /2 1=2

DT.rb; NG.rb// D Œ.CiCk;jCl;1  Ci;j;1 /2 C .CiCk;jCl;3  Ci;j;3 /2 1=2
DT.gb; NG.gb// D Œ.CiCk;jCl;2  Ci;j;2 /2 C .CiCk;jCl;3  Ci;j;3 /2 1=2

Fuzzy sets are commonly represented by membership functions from which the cor-
responding membership degrees are derived. Membership degrees between zero and
one indicate the uncertainty that whether the distance is small or not. In the proposed
approach, a new membership function SMALL has been used which incorporates a
two-sided composite of two different Gaussian curves. The Gaussian membership
function depends on two parameters  and c as given by

.x  c/2
f .x; ; c/ D e 2 2

The membership function gauss2mf () is a combination of two of these two parame-

ters. The first function, specified by 1 and c1, determines the shape of the leftmost
curve. The second function specified by 2 and c2 determines the shape of the right-
most curve. Whenever c1 < c2, the gauss2mf () reaches a maximum value of 1.
Otherwise, the maximum value is less than one. The membership function SMALL
is then defined as

μS .x/ D gauss2mf.x; Œx ; Cx ; x ; 0/

where x is the standard deviation of the distance measure and Cx is the mean of the
distance measure, respectively.
92 M. Wilscy, M.S. Nair

In the above fuzzy rules, the intersection of two fuzzy sets is involved. The inter-
section of two fuzzy sets A and B is generally specified by a binary mapping T,
which aggregates two membership functions as follows:
μA\B .y/ D T.μA .y/; μB .y//, where μA and μB are the membership functions
for the two fuzzy sets A and B, respectively. The fuzzy intersection operator, known
as triangular norms (T-norms), used in this paper is the algebraic product T-norms.
For example, the antecedent of Fuzzy rule 1 is:

μSMALL .DT.rg; NG.rg///: μSMALL .DT.rb; NG.rb///

The above obtained value, called the activation degree of the fuzzy rule 1, is used
to obtain the corresponding weight. So the weights WiCk;jCl;1 , WiCk;jCl;2 , and
WiCk;jCl;3 are calculated as follows:

WiCk;jCl;1 D μSMALL .DT.rg; NG.rg///: μSMALL .DT.rb; NG.rb///

WiCk;jCl;2 D μSMALL .DT.rg; NG.rg///: μSMALL .DT.gb; NG.gb///
WiCk;jCl;3 D μSMALL .DT.rb; NG.rb///: μSMALL .DT.gb; NG.gb///

The output of the Fuzzy Sub-filter I, denoted as F1, is then given by:

P
CK P
CK
Wi Ck;j C1;1 :Ci Ck;j C1;1
kDK lDK
F1i;j;1 D
P CK
CK P
Wi Ck;j C1;1
kDK lDK
P
CK P
CK
Wi Ck;j C1;2 :Ci Ck;j C1;2
kDK lDK
F1i;j;2 D
P CK
CK P
Wi Ck;j C1;2
kDK lDK
P
CK P
CK
Wi Ck;j C1;3 :Ci Ck;j C1;3
kDK lDK
F1i;j;3 D
P CK
CK P
Wi Ck;j C1;3
kDK lDK

where F1i;j;1 , F1i;j;2 and F1i;j;3 denotes the red, green and blue components of the
Fuzzy sub-filter I output image respectively.

8.3.2 Fuzzy Sub-filter II

The second sub-filter is used as a complementary filter to the first one. The goal
of this sub-filter is to improve the first method by reducing the noise in the color
components differences without destroying the fine details of the image. In this
step, the local differences in the red, green and blue environment are calculated
8 A New Fuzzy-Based Additive Noise Removal Filter 93

separately. These differences are then combined to calculate the local estimation of
the central pixel. In this step also, a window of size .2L C 1/  .2L C 1/ is used
centered at (i, j) to filter the current image pixel at that position. The local differ-
ences for each element of the window for the three color components are calculated
as follows:

DRk;l D F1iCk;jCl;1  F1i;j;1

DGk;l D F1iCk;jCl;2  F1i;j;2
DBk;l D F1iCk;jCl;3  F1i;j;3

where k; l 2 f1; 0; C1g. The correction term " is calculated as follows:

"k;l D .1=3/:.DRk;l C DGk;l C DBk;l /

for k; l 2 fL; : : : :; 0; : : : :; CLg:

The output of the Fuzzy sub-filter 2, denoted as FS2, is then given by

P
CL P
CL
.F 1i Ck;j C1;1  "k;l /
kDL lDL
F2i;j;1 D
.2L C 1/2
P
CL P
CL
.F 1i Ck;j C1;2  "k;l /
kDL lDL
F2i;j;2 D
.2L C 1/2
P
CL P
CL
.F 1i Ck;j C1;3  "k;l /
kDL lDL
F2i;j;3 D
.2L C 1/2
where F2i;j;1 , F2i;j;2 and F2i;j;3 denotes the red, green and blue components of the
output image respectively.

8.4 Results and Discussion

The performance of the discussed filter has been evaluated and compared with con-
ventional filters dealing with additive noise, using MATLAB software tool. As a
measure of objective similarity between a filtered image and the original one, we
use the peak signal-to-noise ratio (PSNR) in decibels (dB).

PSNR.i mg; org/ D 10 log10 .S2 =MSE.i mg; org//

This similarity measure is based on another measure, namely the mean-square error
(MSE).
94 M. Wilscy, M.S. Nair

P
3 P
N P
M
Œorg .i;j;c/  i mg .i;j;c/2
cD1 i D1 j D1
MSE .i mg; org/ D
3:N:M
where org is the original color image, img is the filtered color image of size
N:M , and S is the maximum possible intensity value (with m-bit integer values,
S will be 2m  1). The standard color images used in this paper are House and
Gantrycrane images. The original image, the noisy image (original image corrupted
with Gaussian noise with a selected  value) and restored images using mean filter,
median filter, fuzzy method of [17] and the modified fuzzy method of the above
mentioned standard color images along with their corresponding PSNR values are
shown in Figs. 8.1 and 8.2. From experimental results, it has been found that our

(a) (b) PSNR = 22.10

(c) PSNR = 26.57 (d) PSNR = 27.66

(e) PSNR = 28.38 (f) PSNR = 29.33

Fig. 8.1 (a) Original House image (256  256). (b) Noisy image (Gaussian noise,  D
20). (c) After applying mean filter (3  3 window). (d) After applying median filter (3  3
window). (e) After applying fuzzy filter of [17] with K D 3 .7  7 window/ and L D
2 .5  5 window/. (f) After applying proposed fuzzy filter with K D 3 .7  7 window/ and
L D 2 .5  5 window/
8 A New Fuzzy-Based Additive Noise Removal Filter 95

(a) (b) PSNR = 16.10

(c) PSNR = 22.03 (d) PSNR = 21.52

(e) PSNR = 22.75 (f) PSNR = 23.61

Fig. 8.2 (a) Original Gantrycrane image (400  264). (b) Noisy image (Gaussian noise,  D
40). (c) After applying mean filter (3  3 window). (d) After applying median filter (3  3
window). (e) After applying fuzzy filter of [17] with K D 3 .7  7 window/ and L D
2 .5  5 window/. (f) After applying proposed fuzzy filter with K D 3 .7  7 window/ and
L D 2 .5  5 window/

proposed method receives the best numerical and visual performance for low lev-
els and higher levels of additive noise, by appropriately selecting window size for
the two fuzzy sub-filters. Numerical results that illustrate the denoising capability
of the proposed method, modified method and conventional methods are pictured
in Table 8.1. Table 8.1 shows the PSNRs for the colored House image that were
corrupted with Gaussian noise for  D 5; 10; 20; 30 and 40. The window size for
different filters is appropriately chosen to give better PSNR value. The PSNR value
of the noisy image and the best performing filter were shown bold.
96 M. Wilscy, M.S. Nair

Table 8.1 Comparative results in PSNR of different filtering methods for various distortions of
Gaussian noise for the (256  256) colored House image
PSNR (dB)

 D5  D 10  D 20  D 30  D 40
Noisy 34.13 28.12 22.10 18.57 16.08
Mean 28.05 27.70 26.57 25.13 23.72
Median 32.31 30.81 27.66 25.02 22.95
Proposed fuzzy method 34.12 31.79 28.38 25.85 23.76
Modified fuzzy method 34.22 32.77 29.33 26.51 24.18

8.5 Conclusion

A fuzzy filter for restoring color images corrupted with additive noise is proposed
in this paper. The proposed filter is efficient and produces better restoration of color
images compared to other filters. Numerical measures such as PSNR and visual
observation have shown convincing results. Further work can be focused on the con-
struction of other fuzzy filtering methods for color images to suppress multiplicative
noise such as speckle noise.

References

1. A. K. Jain, “Fundamentals of Digital Image Processing,” Pearson Education, Prentice Hall,

Englewood Cliffs, NJ, 1989, pp. 233–356.
2. R. C. Gonzalez and R. E. Woods, “Digital Image Processing,” 2nd Ed., Pearson Education,
Prentice-Hall, Englewood Cliffs, NJ, 2002, pp. 147–163.
3. G. R. Arce, J. L. Paredes and J. Mullan, “Nonlinear Filtering for Image Enhancement,”
Handbook of Image and Video Processing, Academic, London, 2006, pp. 81–100.
4. M. Hansen and B. Yu, “Wavelet thresholding via mdl for natural images,” IEEE Transactions
on Inference Theory, Vol. 46, No. 5, pp. 1778–1788, August 2000.
5. http://homepages.inf.ed.ac.uk/rbf/HIPR2/csmooth.htm
6. H. L. Resnikoff and R. O. Wells, “Wavelet Analysis: The Scalable Structure of Information,”
Springer-Verlag, New York, 1998.
7. D. Donoho, “Denoising by soft-thresholding,” IEEE Transactions on Inference Theory, Vol.
41, No. 3, pp. 613–627, May 1995.
8. S. Chang, B. Yu, and M. Vetterli, “Adaptive wavelet thresholding for image denoising
and compression,” IEEE Transactions on Image Processing, Vol. 9, No. 9, pp. 1532–1546,
September 2000.
9. E. Simoncelli and E. Adelson, “Noise removal via Bayesian wavelet coring,” in Proceedings
of IEEE International Conference on Image Processing, 1996, pp. 379–382.
10. P. Moulin and J. Liu, “Analysis of multi-resolution image denoising schemes using generalized
Gaussian and complexity priors,” IEEE Transactions on Inference Theory, Vol. 45, No. 3, pp.
909–919, April 1999.
11. A. PiLzurica, W. Philips, I. Lemahieu, and M. Acheroy, “A joint inter and intra-scale statistical
model for Bayesian wavelet based image denoising,” IEEE Transactions on Image Processing,
Vol. 11, No. 5, pp. 545–557, May 2002.
8 A New Fuzzy-Based Additive Noise Removal Filter 97

12. F. Farbiz and M. B. Menhaj, “A fuzzy logic control based approach for image filtering,” in
Fuzzy Technical Image Processing, E. E. Kerre and M. Nachtegael, Eds., 1st ed. Physica
Verlag, Heidelberg, Germany, 2000, Vol. 52, pp. 194–221.
13. D. Van De Ville, M. Nachtegael, D. Van der Weken, W. Philips, I. Lemahieu, and E. E.
Kerre, “A new fuzzy filter for Gaussian noise reduction,” in Proceedings of SPIE Visual
Communications and Image Processing, 2001, pp. 1–9.
14. D. Van De Ville, M. Nachtegael, D. Van der Weken, E. E. Kerre, and W. Philips, “Noise
reduction by fuzzy image filtering,” IEEE Transactions on Fuzzy Systems, Vol. 11, No. 8, pp.
429–436, August 2003.
15. M. Nachtegael, S. Schulte, D. Van der Weken, V. De Witte, and E. E. Kerre, “Fuzzy fil-
ters for noise reduction: The case of Gaussian noise,” in Proceedings of IEEE International
Conference on Fuzzy Systems, 2005, pp. 201–206.
16. S. Schulte, B. Huysmans, A. PiLzurica, E. E. Kerre, and W. Philips, “A new fuzzy-based wavelet
shrinkage image denoising technique,” Lecture Notes in Computer Science, Vol. 4179, pp.
12–23, 2006.
17. S. Schulte, V. De Witte, and E. E. Kerre, “A Fuzzy noise reduction method for color images,”
IEEE Transaction on Image Processing, Vol. 16, No. 5, May 2007, pp. 1425–1436.
18. A. C. Bovik and S. T. Acton, “Basic Linear Filtering with Application to Image Enhancement,”
Handbook of Image and Video Processing, Academic, New York, 2006, pp. 71–79.
19. B. Vidakovic, “Non-linear wavelet shrinkage with Bayes rules and Bayes factors,” Journal of
the American Statistical Association, Vol. 93, pp. 173–179, 1998.
20. H. Chipman, E. Kolaczyk, and R. McCulloch, “Adaptive Bayesian wavelet shrinkage,” Journal
of the American Statistical Association, Vol. 92, pp. 1413–1421, 1997.
21. L. Sendur and I. W. Selesnick, “Bivariate shrinkage functions for wavelet-based denoising
exploiting inter-scale dependency,” IEEE Transactions on Signal Processing, Vol. 50, No. 11,
pp. 2744–2756, November 2002.
22. M. Crouse, R. Nowak, and R. Baranuik, “Wavelet-based statistical signal processing using
hidden Markov models,” IEEE Transactions on Signal Processing, Vol. 46, No. 4, pp. 886–
902, April 1998.
23. J. Romberg, H. Choi, and R. Baraniuk, “Bayesian tree-structured image modeling using
wavelet-domain hidden Markov models,” IEEE Transactions on Image Processing, Vol. 10,
No. 7, pp. 1056–1068, July 2001.
24. M. Malfait and D. Roose, “Wavelet-based image denoising using a Markov random field
a priori models,” IEEE Transactions on Image Processing, Vol. 6, No. 4, pp. 549–565,
April 1997.
Chapter 9
Enhancement of Weather Degraded Color
Images and Video Sequences Using Wavelet
Fusion

Jisha John and M. Wilscy

Abstract The image in outdoor scene is degraded by optical scattering of light

which produces additional lightness present in some parts of the image. In this work
a method for enhancing visibility and maintaining color fidelity is proposed using
wavelet fusion. This method mainly consists of three phases. Given an input image,
first phase is to apply a contrast correction using the depth information. Here we
compute the value of airlight present in the image by optimizing a cost function. The
second phase consists of finding an approximate airlight value by using the intensity
information of YIQ color model. Contrast restoration in the first and second phase
is performed by removing the airlight from the image and applying depth informa-
tion. The third and final phase of the proposed method consists of a wavelet fusion
method to get a resultant image which has considerable visibility improvement and
also maintains the color fidelity.

Keywords Weather Degraded  Color Images  Video Sequences  Wavelet Fusion 

Contrast restoration  Enhancing visibility

9.1 Introduction

One of the major reasons for accidents in air, on sea and on the road is the poor
visibility due to presence of fog or mist in the atmosphere. During winter, visibility
is worse, sometimes up to few feet only. Under such conditions, light reaching the
human eye is extremely scattered by constituents of atmosphere like fog, haze and
aerosols and the image is severely degraded. Images taken under such bad weather
conditions suffer from degradation and severe contrast loss. The loss of image qual-
ity is a nuisance in many imaging applications. For example, in underwater imaging

J. John (B)
Mtech student, Department of Computer Science, University of Kerala, Karyavattom – 695581,
Trivandrum, Kerala, India,
E-mail: jisha.json@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 99

Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
100 J. John, M. Wilscy

in murky water, the detection of artifacts becomes difficult due to poor image qual-
ity. Hence, imaging must be performed at close range and this usually results in a
long time required to inspect a small area. Another example is in the navigation of
surface ships and aircraft in bad weather. In weather conditions such as fog, visibility
is low and navigation is more difficult, dangerous and slow.
The image in outdoor scene is degraded by optical scattering of light which pro-
duces additional lightness present in some parts of the image, an effect that has
been referred to as “atmospheric background radiation” [1, 2] or “airlight” [3, 4].
This results in degradation of image contrast, as well as alteration of scene color,
which finally leads to a poor visual perception of the image. Contrast enhancement
methods fall into two groups: non-model-based and model-based.
Non-model-based methods analyze and process the image based solely on the
information from the image. The most commonly used non-model-based methods
are histogram equalization and its variations [5–8]. For color images, histogram
equalization can be applied to R, G, B color channels separately but this leads to
undesirable change in hue. Better results are obtained by first converting the image
to the Hue, Saturation, Intensity color space and then applying histogram equaliza-
tion to the intensity component only [9]. However, even this method does not fully
maintain color fidelity.
There are also other non-model-based methods like unsharp masking [10],
approaches based on the Retinex theory [11–13] and wavelet-based methods [14,
15]. Generally, all non-model-based methods have a problem with maintaining
color fidelity. They also distort clear images, which is an important limitation for
fully automatic operation. Model-based methods use physical models to predict
the pattern of image degradation and then restore image contrast with appropri-
ate compensations. They provide better image rendition but usually require extra
information about the imaging system or the imaging environment.
In [16] John P Oakley and Hong Bu have suggested a method of enhancement
by correcting contrast loss by maintaining the color fidelity. In this method it is
assumed that if the distance between a camera position and all points of a scene rep-
resented by an image generated by the camera is approximately constant, the airlight
will be uniform within the image. But in most real-time situations this assumption
is not valid. This method gives good contrast restoration but does not provide much
visibility enhancement. To enhance the visibility R.T. Tan et al. [17] have proposed
a visibility enhancement method which makes use of color and intensity informa-
tion. Visibility is greatly improved in the resulting images but color fidelity is not
maintained. Hence in situations were the naturality of the image is important this
method cannot be used.
In this work a method for enhancing visibility and maintaining color fidelity is
proposed using wavelet fusion. This method mainly consists of three phases. Given
an input image, first phase is to apply a contrast correction using the depth infor-
mation. Here we compute the value of airlight present in the image by optimizing
a cost function. The second phase consists of finding an approximate airlight value
by using the intensity information of YIQ color model. Contrast restoration in the
first and second phase is performed by removing the airlight from the image and
9 Enhancement of Weather Degraded Color Images 101

applying depth information. The third and final phase of the proposed method con-
sists of a wavelet fusion method to get a resultant image which has considerable
visibility improvement and also maintains the color fidelity.
The rest of the paper is organized as follows: In Section 9.2 we will discuss the
atmospheric scattering models concentrating on the airlight model which forms the
basis of this method. In Section 9.3, the contrast correction is given which forms
the first phase of this work. In Section 9.4 the approximate airlight estimation is
discussed. Section 9.5 introduces the third phase were a wavelet fusion is applied
to get the enhanced image. Section 9.6 explains how the method can be applied on
video sequences. To show the effectiveness of the proposed method, performance
analysis is done in Section 9.7 with the help of a contrast improvement index and
sharpness measure. Section 9.8 includes the experimental results and discussion.

9.2 Atmospheric Scattering Models

Scattering of light by physical media has been one of the main topics of research
in the atmospheric optics and astronomy communities. In general, the exact nature
of scattering is highly complex and depends on the types, orientations, sizes, and
distributions of particles constituting the media, as well as wavelengths, polarization
states, and directions of the incident light [1,2]. Here, we focus on one of the airlight
model, which form the basis of our work.
While observing an extensive landscape, we quickly notice that the scene points
appear progressively lighter as our attention shifts from the foreground toward the
horizon. This phenomenon, known as airlight results from the scattering of environ-
mental light toward the observer, by the atmospheric particles within the observer’s
cone of vision. The environmental illumination can have several sources, including,
direct sunlight, diffuse skylight and light reflected by the ground. While attenua-
tion causes scene radiance to decrease with path length, airlight increases with path
length. It therefore causes the apparent brightness of a scene point to increase with
depth. The, the irradiance due to airlight is given by

E.x/ D I1 .x/eˇ d.x/ C I1 1  eˇ d.x/ : (9.1)

The first term in the equation represents the direct transmission, while the second
term represents the airlight. E is the image intensity, x is the spatial location, I1 is
the atmospheric environmental light, which is assumed to be globally constant and
is the normalized radiance of a scene point, which is the function of the scene
point reflectance, normalized sky illumination spectrum, and the spectral response
of the camera. ˇ is the atmospheric attenuation coefficient. d is the distance between
the object and the observer. It is the second term in Eq. (9.1) or airlight that causes
degradation in the image taken under bad weather conditions and hence all con-
trast restoration methods are aimed at removing this additional lightness from the
image.
102 J. John, M. Wilscy

9.3 Contrast Correction

In Simple Contrast Loss, the degradation can be described by the applying

y D m.x – / (9.2)

to each pixel of the image, where

‘x’ is the distorted pixel value
‘’ is an estimate of the airlight-contributed part of the pixel value
‘m’ is a scaling parameter
and ‘y’ is a modified pixel value.
To estimate the airlight in an image the normalized brightness value needs to be
known

9.3.1 Normalized Brightness and Airlight

The normalized brightness, BK , is defined by:

k
Bk D ; k D 1; 2; : : : : : : :K (9.3)
Nk
were k is the pixel position, k is the value of the image at pixel position k, and Nk
is the output of a spatial low-pass filter at pixel position k. K is the total number of
pixels in the image. The type of spatial low-pass filter is not critical. The Gaussian-
shaped kernel is used here but other shapes provide similar results. The Gaussian
shape is preferred since it introduces the least spurious structure. Also, an efficient
recursive implementation of the Gaussian kernel may be used to reduce compu-
tational effort. For natural images, Bk has a near-symmetric and near-Gaussian
distribution with a mean close to unity.

9.3.2 Correcting Contrast

In order to perform contrast correction, an airlight estimate is required. An algorithm

is used for estimating the level of this airlight given the assumption that it is constant
throughout the image. Airlight estimation is done by finding the minimum value of
a cost function Eq. (9.4) that is a scaled version of the standard deviation of the
normalized brightness is given by;

pk0  
Sg m ./ D S TD W k D 1; 2; ::::K :GM fpNk   W k D 1; 2; ::::Kg
pNk0  
(9.4)
9 Enhancement of Weather Degraded Color Images 103

Where GM f:g denotes the geometric mean. The geometric mean can also be
written as
( )
1X
K
GM fxk W k D 1; 2::::::K g D exp ln.xk / (9.5)
k
kD1

Another possible variation on the cost function is to use sample variance in Eq.
(9.4) rather than sample standard deviation, in which case the scaling factor must be
squared.
K  2 ( )
1X 1X
K
k  Nk
S./ D : exp ln. Nk  /2
(9.6)
k Nk   k
kD1 kD1

Obtain the optimum value of  which minimizes the cost function by calculating,
˚
cO D arg min Sg m ./g (9.7)

This is done using a standard optimization algorithm. This estimated value of 

represents the airlight present in the image. From this computation we can rewrite
Eq. (9.1) as
E.x/ D I1 .x/eˇ d.x/ C  (9.8)
Hence the enhanced image

I1 .x/ D .E.x/  /eˇ d.x/ (9.9)

From Eq. (9.1) we can again write


 D I1 1  eˇ d.x/ (9.10)

So, in order to estimate eˇ d.x/ we can rewrite Eq. (9.10) as

 r 
 D I1 C I1g
C I1r
1  eˇ d.x/ ; Hence

1
eˇ d.x/ D (9.11)
1 g

r/
r
.I1 CI1 CI1
g
Where .I1 r
C I1 C I1 r
/, namely, the environmental light, is assumed to be the
largest intensity in the image.  is found by optimizing the cost function Eq. (9.4)
and depth information is obtained from Eq. (9.11). Thus Eq. (9.9) gives the contrast
corrected image. Section IV and V describes how visibility enhancement can be
achieved.
104 J. John, M. Wilscy

9.4 Intensity Based Enhancement

To accomplish the goal of intensity enhancement, airlight ./ is computed based on

the intensity value of YIQ color model [17] which is defined as

Y D 0:257  Er C 0:504  Eg C 0:098  Eb (9.12)

where Er ; Eg ; Eb represents the r, g and b color channels respectively.

It is assumed that the value of Y is the value of . However the values of 
are approximated values and thus to create a better approximation, these values are
diffused by using Gaussian blur. Depth information, eˇ d.x/ is computed as described
in the earlier section by Eq. (9.11). Enhanced image is obtained as in Eq. (9.9).This
resultant image contains all the detailed information present in the image.

9.5 Wavelet Fusion

The first phase described in Section 9.3 results in an image maintaining the color
fidelity but the visibility is not enhanced particularly in scenes were the distribution
of airlight is not uniform. The second phase uses an approximate airlight estimation
method which results in an image with enhanced visibility but the color fidelity
is not maintained. In the third phase a novel fusion method is used which helps
in extracting the useful information from both images and hence obtaining an
image with enhanced visibility at the same time maintaining the color fidelity. The
daubechies wavelet is used here. The two images obtained as described in Sec-
tion 9.3 and 9.4 are decomposed by using Daubechies wavelet method. The wavelet
decomposition is done using shift invariant wavelet transform. The four images
obtained per image after decomposition are coefficients extracted from the given
image.
The first image is actually approximate coefficients displayed while the sec-
ond image is formed when horizontal coefficients are displayed. The third image
is formed when vertical coefficients are displayed. And the final image comes when
diagonal coefficients are displayed. These coefficients are obtained by the following
process. The image is actually passed through some sets of filters then these images
are obtained. The image is passed through two low pass filters one aligned vertically
and one aligned horizontally.
If image is passed through two filters, one low pass aligned horizontally and
a high pass filter aligned vertically the vertical coefficients are obtained. Vertical
coefficients and are obtained from high pass filter aligned horizontally and low pass
filter aligned vertically. And the final image is for diagonal coefficients which are
obtained with both high pass filters aligned horizontally and vertically. After obtain-
ing the wavelet bands of the two images merge the coefficients by obtaining the
mean value of the approximate coefficients and maximum value from the detailed
coefficients. The resultant image is an enhanced image which contains the maximum
details and also maintains the color fidelity. video enhancement.
9 Enhancement of Weather Degraded Color Images 105

9.6 Video Enhancement

There are only few existing methods to enhance weather degraded video sequences.
In order to apply the enhancement process on video sequence, two methods can
be adopted. Each of the frames can be separately processed or the method can be
applied to the background and foreground pixels separately.
In the first method the computation time is more since the whole method is
applied on each frame. In the approach followed in this paper the background and
foreground pixels are separated and processed separately. The computation time is
much less and the visibility of the resultant video sequence is found to increase
considerably while color fidelity is maintained. In this method, first a static back-
ground frame is estimated. Then, analyze each of the frames to detect the motion
pixels using any standard methodology [7] and apply morphological operations to
remove all noise pixels. The motion pixels are then bounded by boxes so that each
of the frames will be divided into sub images containing motion pixels. The back-
ground frame is processed using the wavelet fusion method. The estimated lightness
parameter in step 1 is utilized globally for the entire video sequence and hence need
not be computed for the sub images. Wavelet fusion is applied to each of the sub
images using the estimated lightness and approximate lightness parameters. Finally
the sub images are merged with the background image to form the enhanced video
sequence. This method saves a lot of computation time since the processing is done
only on the sub images of each frame and the estimated lightness parameter in step
1 of wavelet fusion method is computed only for the background image.

9.7 Performance Analysis

There is a lack of methodology to assess the performances of the methods or to com-

pare them with one another. Unlike image quality assessment or image restoration
areas, there is no easy way to have a reference image, which makes the problem not
straightforward to solve.
For performance analysis a contrast improvement index is used here as proposed
in [18]. This measure helps in comparing the contrast of foggy and restored image
and hence analyzing the efficiency of the proposed method. However, it does not
rate the fidelity of the contrast restoration method. To achieve such an objective, the
same scene without fog must be grabbed and compared with restored image. The
contrast improvement index is given by
Cprocessed
CI D (9.13)
COriginal

where C is the average value of the local contrast measured with a 3*3 window as:
max  min
CD (9.14)
max C min
106 J. John, M. Wilscy

In order to evaluate the effectiveness of the resultant image a well-known

benchmark-image sharpness measure, the tenengrad criterion [19, 20] can be used.
The tenengrad criterion is based on gradient rI.x; y/ at each pixel .x; y/; were the
partial derivatives are obtained by a high-pass filter, eg., the sobel operator, with the
convolution kernels ix and iy . The gradient magnitude is given by
q
S.x; y/ D .ix  I.x; y//2 C .iy  I.x; y//2 (9.15)

And the tenengrad criteria is formulated as

XX
TEN D S.x; y/2 ; for S(x, y > T) (9.16)
x y

were T is the threshold. The image quality is usually considered higher if its
tenengrad value is larger.
The tenengrad values (TEN) of all images given below has been calculated and
listed in corresponding figure, captions. It is noted that images processed using the
wavelet fusion method described above gives significantly larger tenengrad values,
which indicate the effectiveness of this method. This result agrees with the visual
evaluation of human eye.

9.8 Results and Discussion

The performance of the proposed method has been evaluated and compared with
conventional methods of contrast enhancement using MATLAB software tool. The
performance is analyzed using the measures described above. As a measure of
objective similarity between a contrast restored image and the original one, the
mean-squared error (MSE) is used.

P
3 P
N P
M
Œorg .i; j; c/  img .i; j; c/2
cD1 i D1 j D1
MSE.img,org/ D (9.17)
3:N:M
were org is the original color image, img is the restored color image of size N:M
The color images used in this paper are Trees, Traffic and Aerialview images
of size 256  256. The original image, the noisy image (original image degraded
by fog) and restored images using visibility enhancement method and proposed
wavelet fusion method along with their corresponding tenengrad, contrast improve-
ment index values and mean-squared error values are shown in Fig. 9.1. From
experimental results, it has been found that the proposed method has good contrast
improvement and visibility (measured by the sharpness measure-tenengrad). It also
maintains color fidelity which is shown using the low mean-squared error compared
to other method.
9 Enhancement of Weather Degraded Color Images 107

(a) (b) TEN = 224626 (c)TEN = 225000,CI = 2.6356 (d)TEN = 225000,CI = 1.1555
MSE = 2.0401 e +004 MSE = 9.4580 e+ 003

Fig. 9.1 (a) Original Trees image (256  256). (b) Foggy image. (c) After applying method of
visibility enhancement. (d) After applying proposed method

(a) TEN = 225017 (b) TEN = 2 30371, CI = 1.4743

Fig. 9.2 (a) Frame #1 Traffic video sequence. (b) After applying proposed method

(a) TEN = 173397 (b) TEN = 184135 CI = 4.7925

Fig. 9.3 (a) Degraded Aerialview image. (b) After applying proposed method

Figures 9.2 and 9.3 gives the foggy Traffic image and the Aerialview image
degraded by mist. The method proposed is not restricted to uniform suspension of
airlight and hence is applicable for all real-time scenarios.

9.9 Conclusion

In this paper the problem of restoring the contrast of atmospherically degraded

images is addressed. This method uses the shift invariant discrete wavelet transform
to decompose the images into various levels and then the frames are reconstructed
108 J. John, M. Wilscy

by coefficient based fusion. It provides excellent color fidelity as well as visibility

enhancement and hence is applicable for all real time operations. The method can
be efficiently applied on video sequences also. Future research will be focused on
contrast restoration of images degraded by dense fog or mist.

References

1. S. Chandrasekhar, Radiative Transfer. Dover, New York, 1960. Raefal C. Gonzalez, Richard
E. Woods, “Digital Image Processing,” 2nd Ed., Pearson Education, 2002, pp. 147–163.
2. H. C. Van De Hulst, Light Scattering by Small Particles. Wiley, New York, 1957. Alan C.
Bovik and Scott T. Acton, “Basic Linear Filtering with Application to Image Enhancement,”
Handbook of Image and Video Processing, Academic, 2006, pp. 71–79:
3. J. P. Oakley and B. L. Satherley, “Improving image quality in poor visibility conditions using
a physical model for contrast degradation,” IEEE Transactions on Image Processing, vol. 7,
no. 2, pp. 167–179, February 1998.
4. S. K. Nayar and S. G. Narasimhan, “Vision in bad weather,” in Proceedings of IEEE
International Conference Computer Vision, 1999, vol. 2, pp. 820–827.
5. R. C. Gonzalez and R. E. Woods, Digital Image Processing. Reading, MA: Addison-Wesley,
1993.
6. S. M. Pizer et al:, “Adaptive histogram equalization and its variations,” Computer Vision,
Graphics, and Image Processing, vol. 39, pp. 355–368, 1987.
7. K. Zuiderveld, “Contrast limited adaptive histogram equalization,” in Graphics Gems IV, P.
Heckbert, Ed. New York: Academic, 1994, ch. VIII.5, pp. 474–485.
8. J. A. Stark, “Adaptive image contrast enhancement using generalizations of histogram
equalization,” IEEE Transactions on Image Processing, vol. 9, no. 5, pp. 889–896, May 2000.
9. J. J. Rodriguez and C. C. Yang, “High-resolution histogram modification of color images,”
Graphical Models and Image Processing, vol. 57, no. 5, pp. 432–440, September 1995.
10. A. Polesel, G. Ramponi, and V. J. Mathews, “Image enhancement via adaptive unsharp
masking,” IEEE Transactions on Image Processing, vol. 9, no. 3, pp. 505–510, March 2000.
11. D. J. Jobson, Z. Rahman, and G. A. Woodell, “Properties and performance of a center/surround
retinex,” IEEE Transactions on Image Processing, vol. 6, no. 3, pp. 451–462, March 1997.
12. D. J. Jobson, Z. Rahman, and G. A. Woodell, “A multi-scale retinex for bridging the gap
between color images and the human observation of scenes,” IEEE Transactions on Image
Processing, vol. 6, no. 7, pp. 965–976, July 1997.
13. Z. Rahman, D. J. Jobson, and G. A. Woodell, “Retinex processing for automatic image
enhancement,” Journal of Electronic Imaging, vol. 13, no. 1, pp. 100–110, January 2004.
14. P. Scheunders, “A multivalued image wavelet representation based on multiscale funda-
mental forms,” IEEE Transactions on Image Processing, vol. 10, no. 5, pp. 568–575, May
2002.
15. L. Grewe, and R. R. Brooks, “Atmospheric attenuation reduction through multi-sensor fusion,”
Proceedings of SPIE, vol. 3376, pp. 102–109, 1998.
16. J. P. Oakley, and H. Bu, “Correction of simple contrast loss in color images,” IEEE
Transactions on Image Processing, vol. 16, no. 2, February 2007.
17. R. T. Tan, N. Pettersson Lars Petersson, “Visibility enhancement of roads with foggy or hazy
scenes,” Proceedings of the 2007 IEEE Intelligent Vehicles Symposium Istanbul, Turkey, June
13–15, 2007.
9 Enhancement of Weather Degraded Color Images 109

18. P. Zeng, H. Dong, J. Chi, X. Xu, School of Information Science and Engineering, Northeast-
ern University. Shenyang, China, Proceedings of the 2004 IEEE International Conference on
Robotics and Biomimetics, August 22–26, 2004, Shenyang, China.
19. E. P. Krotkov, Active Computer vision by Cooperative focus and stereo. Springer-Verlag, New
York, 1989.
20. A. Buerkle, F. Schmoeckel, M. Kiefer, B. P. Amavasai, F. Caparrelli, A. N. Selvan, and J. R.
Travis, “Vision-based closed-loop control of mobile micro robots for micro handling tasks,”
Proceedings of SPIE, vol. 4568, Microrobotics and Micro assembly 111, pp. 187–198, 2001.
Chapter 10
A GA-Assisted Brain Fiber Tracking Algorithm
for DT-MRI Data

L.M. San-José-Revuelta

Abstract This work deals with the problem of fiber tracking in diffusion tensor
(DT) fields acquired via magnetic resonance (MR) imaging. Specifically, we focus
on tuning-up a previously developed probabilistic tracking algorithm by making use
of a genetic algorithm which helps to optimize most of the adjustable parameters of
the tracking algorithm. Since the adjustment of these parameters constitutes a hard
NP-complete problem, traditionally, this task has been heuristically approached. In
previous work, we have already developed a multilayer neural network that was suc-
cessfully applied to this issue. Though robustness was one of its major advantages,
the complexity of the whole algorithm constituted its main drawback. In order to
avoid this situation, here we have explored the possibility of using a computation-
ally simpler method based on a micro-genetic algorithm. This strategy is shown
to outperform the NN-based scheme, leading to a more robust, efficient and human
independent tracking scheme. The tracking of white matter fibers in the human brain
will improve the diagnosis and treatment of many neuronal diseases.

Keywords Diffusion tensor magnetic resonance imaging  fiber tracking  genetic

algorithm

10.1 Introduction

The Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) technique measures

the diffusion of hydrogen atoms within water molecules in 3D space. Since in
cerebral white matter most random motion of water molecules are restricted by
axonal membranes and myelin sheets, diffusion anisotropy allows depiction of
directional anisotropy within neural fiber structures, [1, 2]. The DT-MRI technique

L.M. San-José-Revuelta
E.T.S.I. Telecomunicación, University of Valladolid, 47011 Valladolid, Spain,
E-mail: lsanjose@tel.uva.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 111
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
112 L.M. San-José-Revuelta

has raised great interest in the neuro-science community for a better understanding
of the fiber tract anatomy of the human brain. Among the many applications that
arise from tractography we find: brain surgery (knowing the extension of the fiber
bundles could minimize the functional damage to the patient), white matter visu-
alization using fiber pathways (for a better understanding of brain anatomy) and
inference of connectivity between different parts of the brain (useful for functional
and morphological research of the brain).
Most of DT-MRI visualization techniques focus on the integration of sample
points along fiber trajectories, [3], using only the principal eigenvector of the dif-
fusion ellipsoid as an estimate of the predominant direction of water diffusion, [2].
Several approaches have already been developed, some of them include the Runge-
Kutta approach, the multiple diffusion tensor approach, the tensorline approach and
the exhaustive search approach, to name a few.
Though, the in vivo visualization of fiber tracts opens up new perspectives for
neurological research, these algorithms may depict some fiber tracts which do not
exist in reality or miss to visualize important connectivity features (e.g. crossing
or branching structures) mainly due to both some deficiencies in these methods
and several shortcomings inherent in datasets. In order to avoid misinterpretations,
the viewer must be provided with some information on the uncertainty of every
depicted fiber and of its presence in a certain location. In [4], we proposed an esti-
mation algorithm that takes into account the whole information provided by the
diffusion matrix, i.e., it does not only consider the principal eigenvector direction
but the complete 3D information about the certainty of continuing the path through
every possible future direction. An improved version of this algorithm was devel-
oped in [5]. This article included two main aspects: (i) a procedure that on-line
adapts the number of offspring paths emerging from the actual voxel, to the degree
of anisotropy observed in its proximity (this strategy was proved to enhance the
estimation robustness in areas where multiple fibers cross while keeping complex-
ity to a moderate level), and (ii) an initial version of a neural network (NN) for
adjusting the parameters of the algorithm in a user-directed training stage. Subse-
quent work, [6], studied with more detailed the architecture of the neural network
and numerically evaluated its tracking capability, robustness and computational load
when being used with both synthetic and real DT-MR images. This work showed
that in many cases, such as real images with low SNR, a huge computational load
was required.
In this work we propose to use an evolutionary computation-based approach
for tuning-up the parameters of the tracking algorithm instead of using the neural
network scheme. The main aim is to adjust the parameters with a less complex pro-
cedure and to obtain a robust and efficient tracking algorithm. The required human
intervention time should also be reduced. Specifically, we propose a genetic algo-
rithm for this optimization task. Numerical results will prove that this approach
leads to similar and even better convergence results while offering much lower
computational requirements.
10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data 113

10.2 Brief Tracking Algorithm Description

The basic version of the algorithm here used is described in [4]. Thus, this section
just presents a summary of the method, with special emphasis on the new aspects.
The algorithm uses probabilistic criteria and iterates over several points in the ana-
lyzed volume (the points given by the highest probabilities in the previous iteration).
The process starts in a user-selected seed voxel, V0 , and, at every iteration, it eval-
uates a set of parameters related to the central voxel of a cubic structure consisting
of 3  3  3 D 27 voxels, similar to that shown in Fig. 32.3, left.
The central point, Vc , (No. 14 in the figure) represents the last point of the tract
being analyzed. In the first iteration, Vc D V0 . Obviously, there exist 26 possible
directions to take for the next iteration in order to select the next point of the tract.
Once Vc is selected, the previous point and all those points exceeding the limits of
the MR volume are also removed from the list of possible destination points (valid
points).

10.2.1 Basic Concepts

First, a measure Pi , i 2 fvalidpointsg, is evaluated based on the probability of going

from voxel Vc to voxel Vi . This probability takes into account the eigenvalues and
eigenvectors available at point Vc from the DT-MR image diffusion matrix. In order
to calculate this probability, the information shown in Fig. 32.3, right, is used.
The table shows, for every voxel shown in Fig. 32.3, left, the changes that
must occur in indices (m; n; p), when a pathway goes from voxel Vc to voxel Vi .
For instance: when going from point No. 14 to point No. 17, coordinates m and
n increase by 1, and p remains the same. This is represented in the table with
“ m n p D .C C 0/”. With this information, the probability of each possible
destination Vi can be calculated taking into account the projection of each of the
eigenvectors to each of the directions defined in triplet m n p . Besides, each
projection is weighted by the corresponding eigenvalue . Thus, in the previous
example, Pi should be calculated as Pi D V1y 1 C V2y 2 C V3y 3 C V1z 1 C

Fig. 10.1 Modifications of indices (m; n; p) when moving from Vc to the neighboring voxel Vi ,
1  i  27, i ¤ 14
114 L.M. San-José-Revuelta

V2z 2 CV3z 3 , where Vj˛ represents the ˛-component of eigenvector j , 1  j  3,

˛ 2 fx; y; zg. In the general case we have,

X X
3
Pi D ˛ Vj;˛ j (10.1)
˛2fx;y;zg j D1

with x ; y ; z being zero if m ; n ; p are zero, respectively, and equal to 1

otherwise.
The axes reference criterion for the (x; y; z) vector components is also shown in
Fig. 32.3, left. Note that, for this calculus, the sign “” in the triplet is equivalent to
sign “C”. In order to properly calculate Pi , it must be weighed by 0.33 if there are
no zeros in triplet i , and by 0.5 if there is one zero.

10.2.2 Anisotropy and Local Probability

The following anisotropy index is used in the algorithm:

s
.1  2 /2 C .2  3 /2 C .1  3 /2
fa D ; (10.2)
2.21 C 22 C 23 /

where 1 2 3 . When both fa.Vc / and fa.Vi / do not exceed a certain

threshold, then point Vi is eliminated as a possible destination point.
Taking into account both Pi and the anisotropy given by Eq. (10.2), the local
probability of voxel i is defined as

Pi0 D a  1  fa.Vi / C .1  a/  2  Pi ; 0<a<1 (10.3)

where parameter a allows the user to give a higher relative weight to either the
anisotropy or the local probability, and 1 and 2 are scaling factors (normally, 1
and 1,000, respectively). The set of values Pi0 is properly normalized so that they
can be interpreted as probabilities.

10.2.3 Eigenvectors and Direction Considerations

Besides these considerations, the final probability of voxel i makes also use of the
so-called smoothness parameters (described in [7]) which judge the coherence of
fiber directions among the trajectories passing through voxel Vc . The mathematical
expressions of these four parameters, fspi g4iD1 , as well as their geometrical mean-
ing, is explained in [6]. They measure the angles between the directions that join
successive path points, as well as the angles between these directions and the eigen-
vectors associated to the largest eigenvalues found in those voxels. sp2 ; sp3 and
sp4 are used to maintain the local directional coherence of the estimated tract and
10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data 115

avoid the trajectory to follow unlikely pathways, [7]. The threshold for sp1 is set
such that the tracking direction could be moved forward consistently and smoothly,
preventing the computed path from sharp transitions.
Next, the following parameter is calculated for every valid point whose smooth-
ness parameters satisfy the four corresponding threshold conditions,
00
Pi D b. 1 sp1 C 2 sp2 C 3 sp3 C 4 sp4 / C .1  b/Pi0 (10.4)

where, 1 , 2 , 3 and 4 are the corresponding weights of the smoothness parameters

(normally, 0.25), and b stands for a weighting factor.

10.2.4 Path Probabilities

Probabilities Pi00 can be recursively accumulated, yielding the probability of the path
generated by the successive values of Vc ,
000
Pp .k/ D Pi  Pp .k  1/ (10.5)
000 00 P 00
with k being the iteration number, and Pi D Pi = i Pi .
At the end of the visualization stage, every estimated path is plotted with a color
that depends on its probability Pp .

10.2.5 Final Criterion and Pool of “Future Seeds”

A pool of voxels is formed by selecting, at the end of each iteration, the s best voxels
according to Eq. (10.4). The first voxel of the pool becomes the central voxel Vc at
next iteration, expanding, this way, the current pathway.
As proposed in [6], the value of s is adjusted depending on the degree of
anisotropy found in current voxel Vc and its surroundings. When this anisotropy
is high, it means that a high directivity exists in that zone, and the probability that
Vc belongs to a region where fibers cross is really low. Consequently, s takes a small
value (1, 2 or 3). On the other hand, if Vc is found to be situated in a region of high
anisotropy, the probabilities of having fibers crossing or branching is higher. In this
case, it is interesting to explore various paths starting in Vc . This can be achieved by
setting parameter s to a higher value.

10.2.6 Parameters to Be Estimated

In this work we propose to use an genetic algorithm for adjusting the parameters
of the algorithm (a; b; 1 ; 2 ; 1 ; 2 ; 3 ; 4 ), instead of using the complex and time
consuming NN proposed in [5, 6]. This adjustment is necessary when the algorithm
116 L.M. San-José-Revuelta

is applied to a different area of the brain (fiber bundles) or even to the same portion
but having been scanned with under different conditions. In these cases, the volume
of interest will have a different smoothness and anisotropy characterization.

10.3 Proposed GA for Parameter Estimation

Genetic algorithms (GAs) represent a set of potential solutions (population) with

a predetermined encoding rule. At every iteration, each potential solution (chro-
mosome) is associated to a figure of merit, or fitness value, in accordance to its
proximity to the optimal solution. Considering the tracking problem described in
Section 10.2, the goal is to estimate the set of algorithm’s parameters and thresholds
 D .a; b; 1 ; 2 ; 1 ; 2 ; 3 ; 4 ).

10.3.1 Estimation Procedure

When the proposed estimation strategy is used, the user is requested to manually
draw a sample fiber path as well as to compare this fiber path to those estimated by
the GA during its first stages. Specifically, the steps for the estimation of the param-
eters are: (i) the user manually draws a sample fiber path, ru , (ii) the GA starts with
np
a randomly generated population of np D 10 individuals fui gi D1 , each of them
being a possible binary representation of parameters , (iii) the tracking algorithm
of Section 10.2 is applied np times, each of them with the set of parameters repre-
sented by each GA’s individual. This way, np different paths ri are obtained, (iv)
every path ri is compared with ru and given a fitness value i , (v) iterate the GA
during ng D 25 generations and then go to step (ii).
Every time that the fiber paths are obtained at step (ii) the user must compare
them to his sample ru and, in case he finds that a tract rj , 1  j  np , is better
than his first estimation ru , then rj becomes the new reference path ru . At the end,
solution  is obtained from the encoding of the fittest individual.
Though this scheme seems initially complicated, experiments show that a few
iterations lead to sets  that allow to obtain good tracking results. The user doesn’t
have to assign too many fitness values or to perform many comparisons. The
extremely reduced size of the population and the low number of generations per
GA execution, lead to moderately short training periods.

10.3.2 GA Description

The initial set of potential solutions, PŒ0 (population at k D 0), is randomly gen-
erated – in our specific application, the user drawn tract could be included in the
np
initial population. Let us denote PŒk D fui gi D1 to the population at iteration k.
10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data 117

As previously mentioned, the population size np has been fixed to a low value
exploiting, this way, the properties of the so called micro genetic algorithms
(-GAs).

10.3.3 Genetic Operators

In order to perform step (v) in subsection estimation procedure, individuals are

modified by making use of the genetic operators, mainly mutation and crossover,
to a subset of individuals selected using the roulette wheel selection scheme, [8].
This selection procedure is based on a biased random selection, where the fittest
individuals have a higher probability of being selected than their weaker coun-
terparts.PThis way, the probability of any individual to be selected is Pi .k/ D
n
i .k/= j pD1 j .k/, with i .k/ being the fitness of the i th individual in the popu-
lation during the kth iteration of the algorithm.
The mutation operator modifies specific individuals with probability pm , chang-
ing the value of some concrete positions in the encoding of ui . Both the specific
positions and the new values are randomly generated, and mutation is effectively
performed with probability pm . Notice that this genetic operator promotes the
exploration of different areas of the solutions space.
On the other hand, the crossover operator requires two operands (parents) to
produce two new individuals (offspring). These new individuals are created when
merging parents by crossing them at specific internal points. This operation is per-
formed with probability pc . Since parents are selected from those individuals having
a higher fitness, the small variations introduced within these individuals are intended
to also generate high fit individuals.

10.3.4 Elitism and Entropy Dependent Operators

The proposed GA also implements an elitism strategy: The best individual in the
population is preserved and directly introduced in the new population. Nevertheless,
a duplicate of it can also be used as input for the genetic operators.
This elitist model of the genetic algorithm presents some convergence advantages
over the standard GA. In fact, using Markov chain modeling, it has been proved that
GAs are guaranteed to asymptotically converge to the global optimum – with any
choice of the initial population – if an elitist strategy is used, where at least the best
chromosome at each generation is always maintained in the population, [9]. How-
ever, Bhandari et al. [9] provided the proof that no finite stopping time can guarantee
the optimal solution, though, in practice, the GA process must terminate after a finite
number of iterations with a high probability that the process has achieved the global
optimal solution.
Following the ideas described in [10], the crossover and mutation probabilities
depend on the Shannon entropy of the population (excluding the elite) fitness, which
118 L.M. San-José-Revuelta

is calculated as
X
np
H.PŒk/ D  i .k/ log i .k/ (10.6)
i D1

with i .k/ being the normalized fitness of individual ui , i.e., i .k/ D i .k/=
Pnp
i D1 i .k/. When all the fitness values are very similar, with small dispersion,
H.PŒk/ becomes high and pc is decreased – it is not worthwhile wasting time
merging very similar individuals. This way, exploration is boosted, while, con-
versely, exploitation decreases. On the other hand, when this entropy is small, there
exists a high diversity within the population, a fact that can be exploited in order to
increase the horizontal sense of search. Following a similar reasoning, the prob-
ability of mutation is increased when the entropy is high, so as to augment the
diversity of the population and escape from local suboptimal solutions (exploita-
tion decreases, exploration becomes higher). Therefore, we have that probabilities
pm and pc are directly/inverselly proportional to the population fitness entropy,
respectively.
Some exponentially dependence on time k must also be included in the model –
making use of exponential functions – in order to relax (decrease), along time,
the degree of dependence of the genetic operators’ probabilities with the disper-
sion measure. This avoids abandoning good solution estimates when very low
fitness individuals are sporadically created, specially when most of the population
individuals have converged to the global optimum.
As a consequence of these entropy dependent genetic operators, the resulting
complexity of the GA is notably decreased since crossover is applied with a very
low probability (and only on individuals not belonging to the elite), and the diver-
sity control allows the algorithm to work properly with a much smaller population
size, [10].

10.4 Numerical Results

In order to evaluate the proposed algorithm (parameter tuning + tracking), both syn-
thetic and real DT-MR images have been used. For the sake of comparison, we have
used the same test images as in [6].

10.4.1 Synthetic Images

Figure 6 in [6] shows four different synthetic DT-MRI data defined in a 50  50  50

grid (we will refer to them as “cross”, “earth”, “log” and “star” as in previous work).
To make the simulated field more realistic, Rician noise – [11] – was added in the
diffusion weighted images which were calculated from the Stejskal-Tanner equation
using the gradient sequence in [12], and a b-value of 1,000.
10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data 119

Fig. 10.2 Tracking results for the “star” synthetic DT-MR image. Black: seed points. Blue: fiber
paths obtained using adjustment with NN, Green: paths using estimation with the proposed AG.
Red: extrinsic voxels. Initial seeds V0 D fS1 ; S2 ; S3 g. Top left: original synthetic image

Satisfactory tracing results for the first three cases can be found in [4], where a
simpler algorithm was used. For the sake of brevity, in this paper we have worked
with the most complex case, the star (Fig. 10.2, top left). This image consists of
six orthogonal sine half-waves, each of them with an arbitrary radius. Under this
scenario the diffusion field experiments variations with the three coordinate axes and
there exists a crossing region. Three different tracking results are shown in Fig. 10.2,
each of them for a different seed V0 D fS1 ; S2 ; S3 g. Blue tracts were obtained with
an algorithm were parameters were estimated with a NN – [6] – while green ones
correspond to the estimation using the proposed GA.
It can be seen that, in both cases, the path estimates pass through isotropic zones
where different fiber bundles cross. It is also appreciated how both methods dif-
ferentiate between the totally isotropic zones extrinsic to the tracts and the fiber
bundles.
The differentiation between voxels belonging to a fiber or to a very isotropic area,
respectively, is attained by mapping the path probabilities given by Eq. (10.5) into
a color scale and classifying them according to some fixed thresholds. Notice that
seeds S1 and S2 belong to the intrinsic volume (voxels with a very high anisotropy).
In this case both methods move through the most probable direction following the
main direction of the star in each situation. When extrinsic point S3 is selected as
seed, the algorithms explore in the neighboring voxels until they find a voxel with a
high anisotropy value (point P1 ). Once P1 is found, the tracking algorithm proceeds
as in the case of S1 and S2 . Fig. 10.2 shows how the algorithm finds the proper
120 L.M. San-José-Revuelta

Table 10.1 Convergence performance for different SNRs values. Cell values represent percentage
of right convergence for two configurations of the algorithm: s D 1=s D 4. Each cell shows: top:
NN-estimation, middle: GA-estimation, bottom: Bayesian tracking, [13]
SNR (dB)
Image Method 5 10 15 20 25 30
NN 78.3/82.8 89.7/93.6 92.1/94.3 98.3/98.7 99.0/99.0 100/100
Cross AG 81.0/84.9 93.1/94.1 94.2/95.7 98.4/98.3 99.0/100 100/100
Friman, 2005 76.8 89.0 90.7 97.0 100 100
NN 77.7/76.2 88.6/87.5 89.9/89.0 98.2/98.2 99.0/99.0 100/100
Earth AG 81.5/79.0 88.9/91.8 92.8/93.3 98.4/98.5 99.0/100 100/100
Friman, 2005 74.4 83.2 85.0 97.3 99.2 100
NN 71.0/69.7 82.1/81.0 86.1/85.5 96.0/95.8 98.0/97.8 100/100
Log AG 73.5/74.2 84.6/84.2 89.1/87.0 96.7/97.0 97.8/98.0 100/100
Friman 2005 68.8 78.3 85.2 96.0 98.0 100

fiber path whatever (extrinsic or intrinsic) seed voxel is chosen, for both methods of
parameters’ estimation.
Next, the robustness of the tracking algorithm with both parameter estimation
methods is now studied. For the sake of brevity, these experiments were run with
parameter s kept constant during the fiber tract estimation (see Section 10.2).
The convergence performance for different SNRs is shown in Table 10.1. The
first row in each cell corresponds to tracking results when parameters where esti-
mated using the NN, the second contains the results when the proposed GA is
used for this estimation, and the third one shows the values obtained with a slightly
modified version of the Bayesian method proposed in [13].
It can be seen that both algorithms (with NN- and AG-based adjustment) con-
verge properly within a wide range of SNRs, with the AG version showing a
convergence gain of about 3–6% in all cases. The percentage values for the “cross”
and the “earth” test images are very close, while for the “log” case both algorithms
exhibit a slightly lower convergence. Comparing our methods with the Bayesian
approach, we see that the proposed tracking algorithm performs slightly better when
the SNR is low, while the three methods tend to similar results with high SNRs.
Analyzing the simulations of the synthetic images considered, it is seen that con-
vergence results improve whenever the MR image contains branching or crossing
areas – as it is the case in real DT-MR images. This is the case of our “cross” image.
For this image, the convergence results are improved  5% when parameter s is
modified according to the anisotropy. Besides, for these studied cases, we see that
the influence of the procedure that adapts s is higher for low SNRs.

10.4.2 Real Images

The proposed tracking algorithm has also been applied to real DT-MR images.
Specifically, we have selected the corpus callosum of the brain (see Fig. 10.3).
10 A GA-Assisted Brain Fiber Tracking Algorithm for DT-MRI Data 121

20
40
60
80
100
120 NN
140
160
AG 180

40 60 80 100 120 140

Fig. 10.3 Tracking results for the corpus callosum area of the human brain. Left: tracts obtained
with the tracking algorithm tuned-up with the proposed GA, Right: parameter estimation with NN

Simulation results show that whichever parameters tuning-up method is used,

the algorithm is able to follow the main fiber bundle directions without getting out
of the area of interest. Fig. 10.3 shows some bundles of properly estimated tracts.
Red/green color indicates high/low certainty.

10.4.3 Final Remarks

The proposed parameter estimation procedure is useful when the volume being
analyzed varies. For instance, with just 5–10 training iterations (repetitions of the
procedure described in 10.3.1), in synthetic images, or 8–16, in real images, the
parameters of the algorithm are fine-tuned so as to get satisfactory results. Note
that these training times are: (i) always inferior to those required by the NN-
based method proposed in [5, 6], (ii) always greatly inferior to the time required to
heuristically adjust the parameters, (iii) only required when the scanning conditions
vary.

10.5 Conclusions

The work here presented expands upon previous work of the author on fiber tracking
algorithms to be used with DT-MR images. Previous papers presented and improved
the basic probabilistic tracking algorithm [4] and developed a novel multilayer
122 L.M. San-José-Revuelta

neural network that helps to tune-up the tracking method, [5]. In this paper we
have presented an Evolutionary Computation-based algorithm that outperforms the
neural network approach.
Numerical simulations have shown that the tracking algorithm that has been
tuned-up using the proposed GA-based method is capable of estimating fiber tracts
both in synthetic and real images. The robustness and convergence have been stud-
ied for different image qualities (SNRs). Results show a convergence gain of about
3–6% with respect to our previous work in [5, 6].
The experiments carried out show that an efficient parameter adjustment in con-
junction with precise rules to manage and update the pool of future seeds lead to:
(i) a better use of computational resources, (ii) a better performance in regions with
crossing or branching fibers, and (iii) a minimization of the required human inter-
vention time. The method has been tested with synthetic and real DT-MR images
with satisfactory results, showing better computational and convergence properties
than already existing Bayesian methods.

References

1. Bjornemo, M. and Brun, A. (2002). White matter fiber tracking diffusion tensor MRI. Master’s
Thesis, Linkoping University.
2. Ehricke, H. H., Klose, U. and Grodd, U. (2006). Visualizing MR diffusion tensor fields by
dynamic fiber tracking and uncertainty mapping. Computers & Graphics, 30:255–264.
3. Mori, S., van Zijl, P.C.M. (2002). Fiber tracking: principles and strategies – a technical review.
Nuclear Magnetic Resonance in Biomedicine, 15:468–480.
4. San-José-Revuelta, L. M., Martı́n-Fernández, M., and Alberola-López, C. (2007). A new
proposal for 3D fiber tracking in synthetic diffusion tensor magnetic resonance images. Pro-
ceedings IEEE International Symposium on Signal Processing and Its Applications, Sharjah,
United Arab Emirates.
5. San-José-Revuelta, L. M., Martı́n-Fernández, M., and Alberola-López, C. (2007). Neural-
network assisted fiber tracking of synthetic and white matter DT-MR images. Proceedings
International Conference on Signal and Image Engineering, ICSIE 2007, London, United
Kingdom I:618–623.
6. San-José-Revuelta, L. M., Martı́n-Fernández, M., and Alberola-López, C. (2008). Efficient
tracking of MR tensor fields using a multilayer neural network. IAENG International Journal
of Computer Science 35:129–139.
7. Kang, N. et al. (2005). White matter fiber tractography via anisotropic diffusion simulation in
the human brain. IEEE Trans. on Medical Imaging, 24:1127–1137.
8. Mitchell, M. (1996). An Introduction to Genetic Algorithms. MIT, Cambridge, MA.
9. Bhandari, D., Murthy, C.A, and Pal, S.K. Genetic algorithm with elitist model and its conver-
gence. International Journal on Pattern Recognition and Artificial Intelligence, 10:731–747.
10. San-José-Revuelta, L. M. (2005). Entropy-guided micro-genetic algorithm for multiuser
detection in CDMA communications. Signal Processing, 85:1572–1587.
11. Gudbjartsson, H., Patz, S. (1995). The Rician Distribution of Noisy MRI Data. Magnetic
Resonance in Medicine, 34:910–914.
12. Westin, C.-F. et al. (2002). Processing and visualization for diffusion tensor MRI. Medical
Image Analysis, 6:93–108.
13. Friman, O., Westin, C.-F. (2005). Uncertainty in white matter fiber tractography Proceedings
of the MICCAI 2005, LNCS 3749, 107–114.
Chapter 11
A Bridge-Ship Collision Avoidance System
Based on FLIR Image Sequences

Jun Liu, Hong Wei, Xi-Yue Huang, Nai-Shuai He, and Ke Li

Abstract In this paper, a forward-looking infrared (FLIR) video surveillance system

is presented for collision avoidance of moving ships to bridge piers. An image pre-
processing algorithm is proposed to reduce clutter background by multi-scale fractal
analysis, in which the blanket method is used for fractal feature computation. Then,
the moving ship detection algorithm is developed from image differentials of the
fractal feature in the region of surveillance between regularly interval frames. When
the moving ships are detected in region of surveillance, the device for safety alert is
triggered. Experimental results have shown that the approach is feasible and effec-
tive. It has achieved real-time and reliable alert to avoid collisions of moving ships
to bridge piers.

Keywords Bridge-ship collision avoidance  infrared image  fractal  target detec-

tion

11.1 Introduction

The fact that more and more ships are built while their size becomes bigger and
bigger is introducing the high risk of collision between bridge and ship in inland
waterways. Incidences of ship-bridge collision mainly cause six types of results,
i.e. damage of bridge, people casualty, damage of ship and goods, economical loss,
social loss and environmental loss. A large amount of statistical analysis indicates
that one of main reasons resulting in ship-bridge collision is the execrable natural
environment such as poorly visible conditions, floods, etc. [1, 2].

J. Liu (B)
Navigation & Guidance Lab, College of Automatic, University of Chongqing,
Chongqing, 400030, China,
E-mail: h.wei@reading.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 123
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
124 J. Liu et al.

Mainly, there are two existing strategies to avoid bridge-ship collision at present
[3, 4]. One is a passive strategy in which fixed islands or safeguard surroundings
are built around bridge piers. The shortages of the passive method are: it could not
avoid ship damage from a collision; the costs are normally high; and it becomes
less effective with constant increase of ship size. The other is an active strategy
that uses radar or video images to monitor moving ships by measuring their course
for collision estimation. Compared with the passive method, the active method
avoids damage of both bridge and ship and its costs are low. However radar is
difficult to detect course changes immediately due to its low short-term accuracy,
and the high noise level makes radar sometimes hardly detect any objects from a
clutter background. Sensors for visible light do not work well under poorly illu-
minated conditions such as fog, mist and night. In contrast, infrared sensors are
capable of adapting weather and light changes during a day. Moreover, the FLIR
images overcome the problems that radar has, i.e. they have high short-term angle
accuracy.
In design, the first consideration of the FLIR surveillance system is its robust-
ness for detecting moving ships. The main difficulties are: (1) low thermal contrast
between the detected object and its surroundings; (2) relatively low signal to noise
ratio (SNR) under the weak thermal contrast; and (3) insufficient geometric, spatial
distribution and statistical information for small targets [5].
Motion detection in a surveillance video sequence captured by a fixed camera can
be achieved by many existing algorithms, e.g. frame difference, background estima-
tion, optical flow method, and statistical learning method [6]. The most common
method is the frame difference method for the reason that it has a great detec-
tion speed and low computation cost. However the detection accuracy by using
this method is strongly affected by background lighting variation between frames.
The most complex algorithm is the optical flow method, by which the computa-
tion cost is high. The statistical learning method needs training samples which may
not be available in most cases, and its computation cost is also relatively high. The
background estimation method is extremely sensitive to the changes of the lighting
condition in which the background is established.
In the FLIR surveillance video sequence used for moving ship detection, a back-
ground normally consists of various information, such as sky, the surface of river,
water waves, large floating objects (non-detected objects) in a flooding season, etc.
In many cases, ships and background in FLIR images are visually merged together.
It is very difficult to detect the targets (moving ships) using normal methods
mentioned above.
The new FLIR video surveillance system for bridge-ship collision avoidance
is proposed in Section 11.2. Section 11.3 presents the novel infrared image pre-
processing algorithm using multi-scale fractal analysis based on the blanket method.
The moving ship detection algorithm in the region of surveillance is also developed
in Section 11.3. Section 11.4 demonstrates the experimental results with detailed
discussion and analysis. Finally, conclusions and future work are presented in
Section 11.5.
11 A Bridge-Ship Collision Avoidance System Based on FLIR Image Sequences 125

11.2 The FLIR Video Surveillance System

In the system, a pan-tilt is fixed on bridge pier, and FLIR camera is installed on the
pan-tilt. The visual region of FLIR i.e. the region of surveillance, can be adjusted by
the pan-tilt, and is configured according to real conditions. The FLIR camera links
to a personal computer (PC) through a frame grabber. When images are captured,
the image processing program in PC’s memory is used to detect the moving ships.
When the moving ships are detected in the region of surveillance, the device for
safety alert is started. The ship driver could be alarmed if necessary, and he/she
would take maneuvers to avoid ship-bridge collision. The flowchart of the system
and sketch map of installation is depicted in Fig. 11.1.
Large amount of experiments carried out in the Yangtse River have proved that
the minimum pre-warning distance between bridge pier and ship to avoid collision
is 500 m in inland waterway, and the valid distance for moving ship detection is
from 800 to 2,000 m when the uncooled infrared FPA (focal plane arrays) ther-
mal imaging camera is used. Therefore, this type of camera is suitable for the
application. The camera resolution is 320  240 pixels. There are three ways
designed to trigger the pre-warning signal, i.e. automatically broadcast the pre-
recorded voice through very high frequency (VHF), automatically broadcast the
pre-recorded voice through loudspeaker, and automatically turn on the assistant
lighting system.

FLIR

Infrared camera

Pan-tilt

Frame grabber

Data acquisition to PC
Bridge Pier
Ship detection algorithms

No
Are there ships?
FOV
Yes
Safety alert

Fig. 11.1 Flowchart of the system framework and sketch of installation

126 J. Liu et al.

11.3 The Detection Algorithm for Moving Ships

In order to detect moving ships in an FLIR image from complicated background

along the inland waterway, a novel detection algorithm is developed. It consists
of four main stages: extracting the region of interest, i.e. the region of surveillance
(ROS); calculating the multi-scale fractal feature; generating the binary image based
on the fractal feature; detecting moving ships by a frame difference method. The
algorithm is schematically demonstrated in Fig. 11.2.

11.3.1 Extracting the ROS

The ROS is defined based on various conditions in real parts of the inland waterway.
Consequently, the ROS appears as a part of region in the original image. The image
analysis and processing is focused on this region only. This excludes unwanted
regions to reduce computation cost in further processing.

11.3.2 Calculating the Multi-scale Fractal Feature

In practice, ships in FLIR images are treated as man-made objects in contrast to

natural background. A fractal model can well describe complex surface structure
characteristics for natural objects, but not for man-made objects [7, 8]. To some
extent, fractal features of natural background keep relatively stable, but fractal fea-
tures of man-made objects behave obviously variety. Therefore, the fluctuation of
fractal features distinguishes natural and man-made objects with the variation of
scale. The multi-scale fractal feature is proposed to reduce interference of natural
background and enhance the intensity of ships.
Many researchers [9, 10] adopted Mandelbrot’s idea and extended it to surface
area calculation. For a grey image surface, the fractal dimension can be estimated
as in Eq. (11.1).
A."/ D K"2D (11.1)
where D is the fractal dimension, K is a constant, " is the scale with value of
0; 1;    ; "max , and A."/ is the surface area of image in the scale ".
Let f .x; y/a grey image, the image could be viewed as a hilly terrain surface
where height from the normal ground is proportional to the image grey value. Then
all points at distance " from the surface on both sides create a blanket of thickness
2". The estimated surface area is the volume of the blanket divided by 2". For differ-
ent scale "." 0/, the blanket area can be iteratively estimated. The upper surface
u.x; y; "/ and the lower surface b.x; y; "/ are expressed in Eqs. (11.3) and (11.4),
respectively.
11 A Bridge-Ship Collision Avoidance System Based on FLIR Image Sequences 127

Start

The ith Frame The (i+s)th Frame

Region of Surveillance Region of Surveillance

(ROS) Extraction (ROS) Extraction

Multi-scale Fractal Multi-scale Fractal

Feature Calculation Feature Calculation

The ith Frame The (i+s)th Frame

Segementation Segementation
Binary Image Binary Image

½Frame i - Frame (i+s)½

Difference Image

Count the Number of Non-zero Pixels (Num)

No
Num >Th (threshold)?

Yes
The Difference Number of ROS
N_dif = N_dif+1

No
N_dif > N_set?

Yes
There are moving ships
in ROS; N_dif=0

Safety alert

Fig. 11.2 The detection algorithm

128 J. Liu et al.

u.x; y; 0/ D b.x; y; 0/ D f .x; y/ "D0 (11.2)

u.x; y; "/ D maxfu.x; y; "  1/; max u.m; n; "  1/g (11.3)
j.m;n/.x;y/j1

b.x; y; "/ D minfb.x; y; "  1/; min b.m; n; "  1/g (11.4)
j.m;n/.x;y/j1

where, " D 0; 1;    ; "max , the image point .m; n/ with distance less than one from
.x; y/ is the four neighbors of .x; y/. To assure the blanket of the surface for scale
" includes all the points of the blanket for scale "  1, .m; n/ are chosen to be the
eight neighbors of pixel .x; y/ in the experiments.
"max is the maximum scale when the fractal features are calculated. "max 2 N;
"max 2:
At scale ", the volume between the upper surface and the lower surface is
calculated by Eq. (11.5):

X
xC" X
yC"
V .x; y; "/ D Œu.k; m; "/  b.k; m; "/ (11.5)
kDx" mDy"

Then, the estimate of the surface area at .x; y/ and " can be obtained by Eq. (11.6):

V .x; y; "/
A.x; y; "/ D (11.6)
2"
Taking the logarithm of both sides in Eq. (11.1), we have:

log A.x; y; "/ D .2  D/ log."/ C log K (11.7)

Linearly fitting log A.x; y; "/ against scale " in Eq. (11.7), fractal dimension D can
be obtained as a constant for all scales ".
For the constant K in Eq. (11.1), also named D-dimensional area [12], which
characterizes the roughness of surface, i.e. different surfaces have different K val-
ues. From this point of view, K is not a constant for the variation of scale ". When
two scales "1 , "2 are used in Eq. (11.7), we have:

log A.x; y; "1 / D .2  D/ log."1 / C log K (11.8)

log A.x; y; "2 / D .2  D/ log."2 / C log K (11.9)

In Eqs. (11.8) and (11.9), fractal dimension D is a constant for scale "1 and "2 , and
let "1 D "; "2 D " C 1, " D 0; 1; 2;    "max , K.x; y; "/ is derived as
 
log A.x; y; "/ log." C 1/  log A.x; y; " C 1/ log."/
K.x; y; "/ D exp
log." C 1/  log."/
(11.10)
11 A Bridge-Ship Collision Avoidance System Based on FLIR Image Sequences 129

From Eq. (11.10), the D-dimension area K.x; y; "/ can be calculated from surface
area A.x; y; "/ at point .x; y/ along with scale ". We use a new function C.x; y/ to
measure the deviation of K.x; y; "/ against scale ", as presented in Eq. (11.11).
"max h
X X
"max i2
1
C.x; y/ D K.x; y; "/  K.x; y; "/ (11.11)
"D2
"max  1 "D2

C.x; y/ is the fractal feature used for ship detection in the algorithm. C.x; y/
appears high value for man-made objects, and much lower value for natural back-
ground.

11.3.3 Segmenting the Fractal Feature Image

C.x; y/ provides sufficient information to discriminate natural background and

ships. The simplest OSTU segmentation method [11] is used to segment the C.x; y/
image. In the resulting binary image, pixel value 255 represents ships and other
man-made objects.

11.3.4 Detecting Moving Ships

In the process of moving ship detection, a difference between two binary images
generated from segmentation of C.x; y/ is used. A group of pixels with non-zero
values represent the difference. Based on the fact that the FLIR camera is fixed on a
bridge pier, the process is summarized as follows.
1. Generate two binary images Ci .x; y/ and Ci Cs .x; y/ by segmentation of two
frames with interval of s, fi .x; y/ and fi Cs .x; y/. In practice, the interval s is
chosen as five to ten frames.
2. Calculate the difference between image Ci .x; y/ and Ci Cs .x; y/, and obtain
D.x; y/.
3. To count the number of non-zero pixels in D.x; y/, and record it as Num.
4. If Num is larger than the threshold (th) which is experimentally obtained, denote
the number of pixels in ROS have changed once, then the difference number of
ROS (N-dif ) add one.
5. If N-dif is larger than a pre-set value N set from experiments, denote that moving
ships are detected in the ROS. In practice, the value N set is set as 2 or 3, which
is effective in reducing the false alarm ratio.
130 J. Liu et al.

11.4 Experimental Results and Discussion

The testing experiments were carried out in the Yangtse River in Chongqing city,
China. An FLIR camera was mounted on a bridge pier. A Celeron1.5 Ghz PC was
connected with the camera through a frame grabber, the frame size was 320  240,
and the frame rate was 30 fps. The parameter settings in the algorithm were, the
frame interval as ten frames, the value of threshold (th) as 5, the value of N set as 2,
and the value of "max as 4. A group of testing results is demonstrated in Fig. 11.3.
The average processing time for each step in the algorithm is shown in Table 11.1.
Observations have indicated that the speed of moving ships is from 20 to 30
km/h. The ROS defines the distance between moving ships and bridge pier as 800–
2,000 m. Therefore the time during which a ship driver takes action to avoid collision
to a bridge pier after altered is between 96 and 360 s. From Table 11.1, it is clearly
shown that the FLIR surveillance system takes about one s to complete a process,
due to the value N set as 2 or 3. It is satisfactory to the application with a real-time
manner.
Comparative experiments were also carried out for system performance analysis
in terms of reliability and effectiveness. The frame difference method was imple-
mented to be compared with the proposed method. FLIR frames were carefully
selected for this comparison. Weather conditions and time of a day were taken into
account when 400 frames with 286 moving ships involved were chosen as the testing
set. Two parameters were introduced as the criterion for the performance, i.e. false
alarm ratio (FAR) and missed alarm ratio (MAR). The comparative results are shown
in Table 11.2. Some typical experimental results are demonstrated in Fig. 11.4.
From the results in Table 11.2, it can be seen that the proposed method for bridge-
ship collision avoidance is superior to the frame difference method in the criterion
of both false alarm ratio and missed alarm ratio. From the results of the testing set
and Fig. 11.4, the system is capable of adapting weather and light changes during a
day.
It is worth to mention that while the FLIR system is mounted on bridge deck,
the performance of surveillance system is impaired by vibration caused by moving
vehicles on the bridge. Therefore, the FLIR system is mounted on a bridge pier in
practice.

11.5 Conclusion

This paper presented a novel FLIR video surveillance system for bridge-ship col-
lision avoidance by using the multi-scale fractal feature, by which moving ships
have successfully been separated from the complex background in inland waterway
images. The proposed algorithm for moving ship detection has achieved the real-
time performance within the ROS in FLIR video sequences. Experimental results
have proved that the developed FLIR video surveillance system is efficient in detect-
ing moving ships to alert possible bridge-ship collisions. Its wide adaptability and
11 A Bridge-Ship Collision Avoidance System Based on FLIR Image Sequences 131

Fig. 11.3 Testing results. (a) The 1st original frame. (b) The 11th original frame. (c) The ROS
image extracted from (a). (d) The ROS image extracted from (b). (e) C.x; y/ of the 1st frame. (f)
C.x; y/ of the 11th frame. (g) Segmenting binary image from (e). (h) Segmenting binary from
(f). (i) D.x; y/ between the 11th and the 1st frames, Num = 32. (j) between the 21st and the 11th
frames, Num = 25

Table 11.1 Average processing time for each step

Steps Time (ms)
Extracting the region of surveillance 5
Calculating C.x; y/ 331
Segmenting C.x; y/ image 22
Detecting moving ships based on frame difference 25
Other steps 10
Total 393
132 J. Liu et al.

Table 11.2 Comparative performance analysis

Frame difference method The proposed method

FAR 23.1% Lower than 1%

MAR 17.6% Lower than 1%

Original image C(x,y) Segmenting binary image

(a) Date - Jul 12, 2006; Time - 00:50; Cloudy

(b) Date - May 26, 2007; Time - 02:30; Moderate fog

(c) Date - Jul 11, 2006; Time - 21:20; Shower

Fig. 11.4 Testing results

high reliability fro weather and light changes have been proved in long period testing
in Yangtse River, Chongqing city, China.
Our future work will be extended to two venues. First, the investigation for track-
ing multiple ships will be carried out based on the ship detection results. Relative
tracking algorithms will be developed to address the issue in a clutter background.
Second, a data fusion of radar and infrared sensors may be explored to improve
system performance in making use of the complement of data from different
sensors.
11 A Bridge-Ship Collision Avoidance System Based on FLIR Image Sequences 133

Acknowledgement The authors would like to thank the Maritime Safety Administration of
Chongqing for providing experimental sites. This research is partially supported by the China
Scholarship Council (CSC).

References

1. Dai, T. D., Lie, W., and Liu, W. L., 1993, The analysis of ship-bridge collision in main
waterway of the Yangtze River, Navigation of China, 4:44–47.
2. Van Manen, S. E., 2001, Ship collisions due to the presence of bridge, International Navigatio
Association (PIANC), Brussels, Report of WG 19.
3. Zhu, Q. Y., 2006, Pier anticollision system based on image processing, Master’s thesis,
Department of Information Engineering, Wuhan University of Technology, Wuhan.
4. Wu, J., 2004, Development of ship-bridge collision analysis, Journal of Guangdong Commu-
nication Polytechnic, 4:60–64.
5. Liu, J., Huang, X. Y., Chen, Y., and He, N. S., 2007, Target recognition of FLIR images on
radial basis function neural network, in Proceedings of Advances in Neural Networks, ISNN
2007, 4th International Symposium on Neural Networks, Nanjing, pp. 772–777.
6. Zhan, C. H., Duan, X. H., Xu, S. Y., Song, Z., and Luo, M., 2007, An improved moving
object detection algorithm based on frame difference and edge detection, in Proceedings of
the Fourth International Conference on Image and Graphics, Chengdu, pp. 519–523.
7. Mandelbrot, B. B., 1982, The fractal geometry of nature. New York: W.H. Freeman, pp. 1–24.
8. Pentland, A., 1984, Fractal-based description of natural scenes, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 6:661–674.
9. Peleg, S., Naor, J., Hartley, R., and Avnir, D., 1984, Multiple resolution texture analysis and
classification, IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:518–523.
10. Zhang, H., Liu, X. L., Li, J. W., and Zhu, Z. F., 2006, The study of detecting for IR weak and
small targets based on fractal features, In Lecture Notes in Computer Science, Advances in
Multimedia Modeling. Springer, Heidelberg/Berlin, pp. 296–303.
11. Ostu, N., 1979, A threshold selection method from gray-level histograms, IEEE Transactions
System, Man and Cybernetics, SMC-9:62–66.
12. Li, J. and Zhang, T. X., 1996, Target detection based on multiscale fractal parameter change,
Journal of Data Acquisition & Processing, 11(3):218–221.
Chapter 12
A Gray Level Feature Detector and Its
Hardware Architecture

Neeta Nain, Rajesh Kumar, and Bhavitavya Bhadviya

Abstract This chapter describes a fast real time gray scale based feature point detec-
tor and its hardware architecture for FPGA based realization. The implementation
is based on a new efficient technique which is fusion of both affine transformation
invariance and robustness to noise. The novelty of the proposed approach lies in
its highly accurate localization and realization only in terms of addition, subtraction
and logic operations. The algorithm is designed to keep the high throughput require-
ments of today’s feature point detectors and applications in Silicon. The proposed
implementation is highly modular with custom scalability to fit devices like FPGAs
etc, with different resource capacity. The implementation can be ported to any real
time vision processing systems where power and speed are of utmost concern.

Keywords Feature Detector  gray scale  Hardware Architecture  affine transfor-

mation invariance  FPGA

12.1 Introduction

A one dimensional change in intensity can be classified as an edge and it can be

approximated using gradient or first order derivative of the image. Similarly feature
points (FP s) are the sudden changes in two dimensions or second order proper-
ties of a surface, which can be approximated with second order derivatives. Since
second order derivatives are noise sensitive, and hence the proposed FP detector
uses only first order derivatives to approximate the second order derivative. FP s
are locations in the image where the signal changes significantly in two-dimensions.
Examples include corners and junctions as well as locations where the texture varies

N. Nain (B)
Associate Professor, Department of Computer Engineering,
Malaviya National Institute of Technology Jaipur-302017, India,
E-mail: neetanain@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 135
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
136 N. Nain et al.

significantly. A large number of feature point detectors like [1–7] etc. are reported
in literature which are mainly classified into two categories: template based and
geometry based. Template based approach is to build a set of corner templates and
determine the similarity between the templates and all the sub windows of the gray
level image. Though simple these techniques are time consuming due to multiple
iterations. Geometry based relies on measuring differential geometry features of
corners. They are of two kinds: boundary based and gray level based. Boundary
based, first extract the boundary as a chain code, and then search for significant
turnings at boundary. These techniques suffer from high algorithm complexity, as
multiple steps are needed also the corner accuracy is dependent on the bound-
ary extraction technique used. Gray-level based methods, directly operate on the
gray level image. The feature detector proposed in this work belongs to this last
category.
The chapter is organized as Section 12.2 describes the proposed feature point
detection algorithm. Its hardware realization is explained in Section 12.3. The syn-
thesis results and its analysis is given in Section 12.4 followed by related discussion
and conclusions in Section 12.5.

12.2 Feature Point Detection Model

A true FP can be defined as a sudden and significant 2D change in intensity gen-

erally classified as corners [2, 3]. Alternatively FP can also be defined as a point
where multiple regions of different intensity meet and convey significant informa-
tion change in 2D. As the number of regions meeting at a point increases, the
measure of entropy also increases at that point which is defined as Information
Content.IC / of the FP in this chapter. Such points are commonly referred as inter-
est points [5, 8]. The proposed technique detects both corners and interest points as
FP s. With the aim of satisfying the universal criterion of localization, consistency,
accuracy and noise immunity in a real time environment a short but effective five
step algorithm is proposed. Where first 1D changes in intensity are detected using
a very simple 2  2 difference mask unlike the usual convolution masks of sizes
varying from 3  3 to 7  7, the complete algorithm is as follows:
1. Apply Difference Mask – The complex convolution is replaced with simple
differences between the pixel intensity values. The use of the difference masks
gives rise to four orthogonal difference values: H1 , H2 , V1 and V2 as shown in
Fig. 12.1. These are the four critical difference values that responds positively to
intensity changes. Each of these can be defined as:

H1 D j I.i;j /  I.i;j C1/ j (12.1)

H2 D j I.i C1;j /  I.i C1;j C1/ j (12.2)
V1 D j I.i;j /  I.i C1;j / j (12.3)
V2 D j I.i;j C1/  I.i C1;j C1/ j (12.4)
12 A Gray Level Feature Detector and Its Hardware Architecture 137

Fig. 12.1 The 2  2 differ-

ence operator mask (a) and
(b) are 90ı rotated to each
other

Where I.i; j / is the pixel under consideration. The response of the 2  2 differ-
ence operator can be classified as, If (H1 D H2 D V1 D V2 D 0) then it is
a constant region; If ((H1 AND H2 > P1 ) OR (V1 AND V2 > P1 /) then it is
a region of only 1D change; If ((H1 OR H2 AND V1 OR V2 / > P1 ) then the
region contains FP s (detected as 2D change). Here P1 is the threshold value
for FP detection which can be varied as per the interest of feature detection. For
example, P1 D 1 will detect all intensity variations as FP s, and P1 D 255 will
detect only step changes (black-to-white or white-to-black) as FP s. The differ-
ence operator detects candidate points for feature detection. As we are interested
only in the 3rd category of responses for FP detection hence each pixel satisfy-
ing this criteria is further processed to determine whether it is a true FP or not.

2. Apply the Pseudo Gaussian Mask to the Points Satisfying the First Differ-
ence Operator – In real case scenario we receive images which are blurred due
to several types of noise and thus FP s and also edges are not well defined. To
overcome these problems and increase the noise immunity of our algorithm we
propose a Pseudo-Gaussian kernel which is derived from a 5  5 Gaussian kernel
having  D 1:3. It is further modified and normalized so as to have all factors
some exponent of 2 to accomplish operations like multiplication or division(in
the Gaussian) by simple shift of the bits during calculations. Due to constrained
time and performance intensive requirement for real time applications the Gaus-
sian smoothing is applied only around those pixels which are part of the region
containing FP s. Additionally due to the nature of our difference kernel we apply
only a partial of the entire Gaussian kernel. This partial implementation as shown
in Fig. 12.2 reduces the Gaussian averaging overhead by 75% but produces the
desired noise smoothing on the set of 2  2 pixels under consideration. The new
Gaussian averaged values of the four pixels under consideration (where we get
the third category of responses) are calculated as:
0 0
I.i;j / D Gk  I.i;j / I I.i;j C1/ D Gl  I.i;j C1/ (12.5)
0 0
I.i C1;j / D Gm  I.i C1;j / I I.i C1;j C1/ D Gn  I.i C1;j C1/ (12.6)

where Gk ; Gl ; Gm ; Gn are as shown in Fig. 12.2.

3. Calculate the Second Difference Operator with Threshold Parameter P2 –
To approximate second order differences the difference operator is re-applied
on the above new Gaussian averaged values of the four pixels under considera-
tion and the four orthogonal differences H1 , H2 , V1 and V2 are updated using a
138 N. Nain et al.

Fig. 12.2 Pseudo-Gaussian

mask. Each of the four central
pixel(weight = 64) is aver-
aged with the neighboring
similar colored pixels with
their respective weights(in
top right corner). Note the
weights are all a power of
2 and that the total weight
of the partial mask is also
64C216C16C.28/ D
128 (a power of 2)

second threshold parameter P2 . The use of P2 is encouraged to avoid missing

weak FP s in the presence of noise. A low P1 will detect most of the FP s and
P2 gives control over the noise interference with the desired FP s. The pseudo
Gaussian kernel is significant in many ways and is one the principle part of the
proposed technique. In addition to noise removal it is amplifying the change in
intensity at the meeting point of regions, in contrast to smoothing the meeting
point. Consider a set of pixels where four different gray scale regions meet. As
the values of each pixel become averaged by pixels having the same pixel inten-
sity the difference mask in the next step responds very strongly to this point.
This is unlike applying a standard kernel wherein the pixel is averaged with all
neighboring pixels and may lead to a very low or no response at all if the nearby
regions are close to each other in gray scale values.
4. Compute Information Content of the FP – The measure of entropy or distinc-
tiveness of the detected FP is computed as the Information Content.IC / of the
FP . For the set of four pixels that form a part of the zero crossings (satisfy the
second difference operator) the IC.i;j / of that FP is increased by a constant. For
example, we can add a 12 to its IC , a 7 to the IC of its 4-connected neighbors
and a 5 to the IC of its diagonally connected neighbors (any multiple of 12; 7
and 5 can be used). This results in a averaged and normalized IC as 60, when
all 8-connected neighbors satisfy the second difference operator including itself
as maximum entropy at the current pixel. The more significantly distinctive the
FP is, the larger is the average IC . For the best accuracy we pin-point all FP s
with IC > 52 in the nxn neighborhood (where n depends upon the localization
accuracy required) as the most prominent FP s (as shown in red), where at least
6 pixels from different regions in its neighborhood are contributing to this FP .
Similarly the less dominant are shown in blue with 44 < IC < 52, in green for
32 < IC < 44, and in white for 22 < IC < 32, where at least 5, 4, and 3 pixels
from different regions are contributing to the FP respectively.
12 A Gray Level Feature Detector and Its Hardware Architecture 139

Fig. 12.3 Mask used for 2

removing the false positives
1 9 8

3 11 12 5

7 10 4

5. Eliminate False Positives – The orthogonal difference masks will respond

strongly to all diagonal edges (false positive FP s). To avoid these we elimi-
nate all the candidate pixels that are part of the diagonal edges and are only
2-pixel connected in the direction of the diagonal. A 4  4 mask as shown
in Fig. 12.3 is used to eliminate the false positives. To understand this, we
have defined two logical functions: I nrange2.P x1 ; P x2 / where in P x1 and
P x2 are the points of concern as shown in the mask (numbered 1  12).
I nrange2.P x1 ; P x2 / is true if and only If j.P x1  P x2 /j < P2 . Similarly we
define I nrange3.P x1 ; P x2 ; P x3 /, which is true if and only If j.P x1  P x2 /j
AND j.P x2  P x3 /j < P2 . We eliminate all pixels that return true for only
I nrange3.x; y; z/ in the same direction and no I nrange2./ is true other than
I nrange2.x; y/ and I nrange2.y; z/. These are the non maxima FP s which
are part of a diagonal edge and needs to be eliminated.

12.3 Hardware Implementation

The feature detector is implemented on FPGA. Two different implementation appro-

aches have been used for realizing the proposed hardware as explained in the
following sub Sections 12.3.1 and 12.3.2. In the first approach the entire algorithm is
completely implemented in hardware with custom scalability to fit FPGA’s with dif-
ferent resource capacity. In the second approach it is implemented using the concept
of SOPC.

12.3.1 Architecture 1: A Complete Hardwired Approach

The overall architecture is depicted in the block diagram as shown in Fig. 12.4. Two
RAM blocks are used where RAM block 1 stores the original image with each pixel
represented in gray scale value of 8 bits. Whereas RAM block 2 stores the resultant
image after false positive removal and also stores the IC of each pixel. RAM block
2 is twice the size of the input image. RAM block 2 also feeds the Non-Maxima
suppression module with the resultant IC for suppression of weak features. It is
best to have the fastest possible memory for reducing the processing time of the
140 N. Nain et al.

Fig. 12.4 Block diagram of the Proposed Architecture 1

algorithm as reading and writing from the RAM is the most expensive operation in
our entire algorithm.
Address Generator module generates the memory addressing sequence for both
the RAM modules. It is capable of generating individual addresses depending upon
the control signals from the Control Block.
Control Block is the most important module as it synchronizes the various blocks.
It is made of a State Machine converted to its VHDL entity. It controls the read/write
sequences necessary to proceed further. It takes as input the results of each block
compares them to its present state and decides where to proceed. It is capable
of clearing the module data and assigning the threshold values P1 and P2 for
comparison operation.
Gradient Module 1 uses 4 comparators and 4 subtractors along with other logic
elements to calculate the Gradient mask. This module is again reused as Gradient
Module 2, the only difference being in the threshold value which is P2 instead of P1
as used in the 1st module. This requires 4 read cycles to obtain the pixel information
necessary for calculations.
Pseudo Gaussian Kernel Module is called only if a pixel satisfies the Gradi-
ent Module 1, otherwise processing of all the following modules are skipped and
the address generator moves to the next pixel. This step requires another 20 read
cycles to obtain the total of 24 pixel values compulsory for its mask calculation and
obtaining the updated values for the next gradient module. It has been determined
experimentally (on real images) that on an average only 40% of the entire image
12 A Gray Level Feature Detector and Its Hardware Architecture 141

pixels actually enters this module. This module uses only left shift operations and
20 adders(8 bit) to approximates the convolution and multiplication.
This is followed by the Gradient Module 2 in which apart from mask calculations
the IC is also computed. It is a background module which adds specific value to the
IC map and writes it into the RAM module 2 taking a total of 9 write cycles.
All the pixels that satisfy the Gradient Module 2 are then processed into the False
Positive Removal Module which requires another 7 read cycles, to obtain the unread
pixels required for its processing. It further requires around 16 comparators and 16
subtractors to formulate the input for its Conditional Block.
Once the entire image is over, the Control Unit starts the Non-Maxima Suppres-
sion module which reads the IC of each pixel and determines its significance. It
uses 3 comparators which compares the IC of the current pixel with 3 constant
values in a step manner and colors the pixel accordingly. It uses 2 Latches(8 bit
wide) that prevents unwanted comparing operations. The module first works on all
the column pixels of the current row and then increments the row number for fur-
ther processing. Maximum column and row number for a module can be specified
in the Control Unit Registers. We have used a common shared register having 31
8-bit values, this register stores the pixel values as and when they are read from the
RAM for the current iteration. Once the iteration is completed, instead of clearing all
the values, we shift the values to new positions for the next iteration. This saves the
common read values for the next iteration. This on an average reduces the memory
overhead by 50%. Figure 12.5 shows the Low level block diagram of two of the
various modules depicting the actual mapping of the VHDL into schematic entities.
The top level module is a schematic and hence each VHDL module was converted
into its schematic before adding to the top level module.

12.3.2 Architecture 2: A Hardware Software Co-design

The complete hardwired approach is suitable for use in ASIC design or a system
with a vision co-processor. It would be difficult to be used in an existing embedded

Fig. 12.5 Internal structure of module 1 (a) Gradient module, (b) partial non-maxima suppression
module
142 N. Nain et al.

applications. For full filling this usability we took another approach. The system was
designed as an SOPC with NIOS II as the core processor, and then the algorithm’s
logic was added by means of Custom Instructions. The NIOS II processor was then
programmed in C code for executing on the [11] and [12] systems.

12.4 Synthesis Results and Analysis

The Compilation, verification and Simulation of the architecture is done for both
Altera’s Cyclone II and FPGA device (EP2C20). The entire verification process was
carried on 480 x 640 sized images. The timing analysis post synthesis and simulation
resulted in a throughput speed of about 500 images/s (assuming a minimum average
delay of 6:69 ns). The worst case scenario is evaluated by testing an image with
alternate black and white pixels and assigning P1 and P2 both to 1, thus ensuring
that each and every pixel enters all the modules and time taken is maximum. This
resulted in the worst case timing delay of 19 ns, making our implementation execute
200 images/s. Further the complete implementation requires a tightly coupled (or a
dedicated) RAM having a total of 1;500 KB of free space. Its synthesis resulted
in the implementation summary as enlisted in Table 12.1. Table 12.1 also depicts
the minimum timing latency for each module. Assuming the serial nature of the
iteration for single feature detection module the minimum latency post synthesis
turns out to be 18:2 ns. Thus a 480  640 size image would take 5:3 ns to execute.
Although in real time applications it never takes that much as only 40% of the total
pixels reach the second module, from which only 20% reach the Gradient module
2 while only 10% of the total image pixels actually reach the false positive removal
module. Further this results in only 10% of all the pixels being finally fed into the
Non-maxima suppression module. Re-calculating the modified minimum delay it
comes out as:

Delay D 3:08  4=21 C 2:02 C 0:4  .3:08  5=21 C 4:12 C 2:02/

C 0:2  .4:98/ C 0:1  .3:99/ D 6:69 ns

Table 12.1 Resource utilization summary and timing analysis for each of the individual modules
Module name Utilization Utilization Latency
(slices) percentage (ns)
(out of 18742)

Address generator 16 0.00 3.08

Control unit 16 0.00 0
Gradient module 112 0.60 2.02
Gaussian module 133 0.71 4.12
False positive removal module 168 0.90 4.98
Non maxima suppression module 48 1.32 3.99
Complete architecture 493 2.63 20.21
12 A Gray Level Feature Detector and Its Hardware Architecture 143

Table 12.2 Compilation report of the SOPC system

Resource type Utilized Available Percentage

LUTs 5,278 18,752 28

Total pins 287 315 91
Memory bits 79,360 239,616 33
PLLs 2 4 50
NIOS II system code 72 KB 8,192 KB 0.9
Run time memory usage 1,872 KB 8,704 KB 21

(a) (b) (c)

Highest information Content Lower information Content
Medium information Content Some information Content

Fig. 12.6 Output on two of the test images after (a) First difference operator, (b) Second difference
operator (c) False positive removal

The SOPC system with the updated NIOS II core (Supplemented with our Custom
Instructions) was compiled and executed on the DE1 board. Table 12.2 depicts a
detailed description of the system’s resource utilization. The simulated result of the
architecture on two of the test images is shown in Fig. 12.6. Comparing our imple-
mentation results with two of the most popular FP detection algorithms by [5, 8]
as shown in Table 12.3 ( [9] implementation is yet to be completed and thus we
have only included its estimated resource utilization) we can conclude that our
architecture can be regarded as one of the most resource effective and high per-
forming implementation of a feature detector in hardware. Further a frequency of
50 MHz ensures that the power requirement for this implementation is very compet-
itive. Thus for real time applications, we can use Architecture 1 for high throughput
requirements, as it is suitable for the highest resolution video and vision processing
systems. Whereas Architecture 2 can be added to an existing embedded system and
can process QCIF and CIF frames at a rate of upto 25 frames/s.
144 N. Nain et al.

Table 12.3 Comparison with two other popular feature detectors implemented in hardware
Implementation Resource Maximum Throughput
name utilization- frequency- (images/s)
slices MHz
[9] 78,728 — —
[10] 685 58.72 120
Our Algorithm-
Architecture 1 493 50 500
Our Algorithm-
Architecture 2 5,278 100 30

12.5 Conclusions

We proposed a true real time FP detector and successfully depicted the possibility
of embedding a high performance feature detector in an embedded Vision System.
The proposed architecture effectively shows the hardware porting of our own novel
algorithm for feature detection. Our architecture uses only integer add, subtract and
shift operations leading to high performance throughput. It is also modular in its
entirety and can be scaled depending upon the resource availability of the FPGA
device. The above resource utilization numbers can be used for estimating the uti-
lization for the scaled version of the above algorithm. The algorithm incurs at most
28 subtractions(or additions) at any pixel, also assuming the worst case scenario of
highest noise content in the images only 80% of all the pixels will pass the first
step for further processing, giving an upper limit of average subtractions per pixel
to 22. No expensive multiplication or division operations are involved in our algo-
rithm making it one of the most suitable contender for real time implementations.
Further emphasizing the claim that it is an excellent contender for use in hardware
implementations and embedded systems alike. Comparing the algorithms code-size
to [7], it is a mere 5 KB (100 lines of Code) instead of a 145 KB (3;500 lines of
code) of Rostenś. Thus it will find direct use in embedded systems having a very
low memory footprint. Apart from the performance gain, as every pixel operation is
dependent only on its surrounding 5  5 pixels independent of the rest of the image,
so coarse level parallelization is very much possible. As a future work the proposed
technique can be parallelized and pipelined to obtain higher throughput.

References

1. Moravec, H P (1977). In Towards Automatic Visual Obstacle Avoidance, page 584. Proceed-
ings of the 5th International Joint Conference on Artificial Intelligence.
2. Harris, C. and Stephens, M. (1988). In A Combined Corner and Edge Detector, volume 23,
pages 147–151. Proceedings of 4th Alvey Vision Conference, Manchester.
12 A Gray Level Feature Detector and Its Hardware Architecture 145

3. Mokhtarian, F and Soumela, R (1998). In Robust Image Corner Detection Through Curva-
ture Scale Space, volume 20, pages 1376–1381. IEEE Transcations on pattern Analysis and
Machine Intelligence.
4. Lowe, D G (2004). In Distinctive Image Features from Scale-Invariant Keypoints, volume 60,
pages 91–110. International Journal of Computer Vision.
5. Smith, Stephen M. and Brady, J. Michael (1997). Susan – a new approach to low level image
processing. International Journal for Computer Vision, 23(1):45–78.
6. Trajkovic, M and Hedley, M (1998). Fast corner detection. International Journal for Image
and Vision, 16(2):75–87.
7. Rosten, Edward and Drummond, Tom (2006). In Machine Learning for high-speed Corner
Detection, volume 17, pages 211–224. ECCV.
8. C Schmid, Roger Mohr and Bauckhage, C (2004). In Comparing and Evaluating Interest
Points, volume 655, pages 63–86. INRIA Rhone, Montbonnot, France, Europe.
9. Cabani, Cristina and Maclean, W. James (2006). In A Proposed Pipelined Architecture
for FPGA Based Affine Invariant Feature Detectors, pages 112–116. IEEE Conference on
Computer Vision and Pattern Recognition.
10. Cesar Torris-Huitzil, Miguel Arias-Estrada (2000). In An FPGA Architecture for High Speed
Edge and Corner Detection, pages 112–116. IEEE Transactions in Computing.
Chapter 13
SVD and DWT-SVD Domain Robust
Watermarking using Differential Evolution
Algorithm

Veysel Aslantas

Abstract This study aims to develop two optimal watermarking techniques based
on SVD and DWT-SVD domain for grey-scale images. The first one embeds the
watermark by modifying the singular values of the host image with multiple SFs.
In the second approach, a hybrid algorithm is proposed based on DWT and SVD.
Having decomposed the host image into four bands, SVD is applied to each band.
Then, the same watermark is embedded by modifying the singular values of each
band with different SFs. Various combinations of SFs are possible, and it is dif-
ficult to obtain optimal solutions by trial and error. Thus, in order to achieve the
highest possible transparency and robustness, optimization of the scaling factors
is necessary. This work employs DE to obtain optimum SFs. DE can search for
multiple solutions simultaneously over a wide range, and an optimum solution can
be gained by combining the obtained results appropriately. Experimental results of
developed techniques show both the significant improvement in transparency and
the robustness under attacks.

Keywords Watermarking  Differential Evolution Algorithm  grey-scale image 

singular values  hybrid algorithm

13.1 Introduction

In the past decade, there have been great developments in computer and communica-
tion technology that have enabled digital multimedia contents such as audio, image
and video to be reproduced and distributed with ease. Besides these advantages are
provided by the technology, it also enables illegal operations in these materials such
as duplication, modification and forgery. Due to that, protection of multimedia data

V. Aslantas
Erciyes University, Department of Computer Engineering, 38039 Melikgazi, Kayseri, Turkey,
E-mail: aslantas@erciyes.edu.tr

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 147
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
148 V. Aslantas

has become a very crucial and urgent matter for data owners and service providers.
Several methods have been proposed in order to provide protection against copyright
forgery, misuse or violation. Among the methods, digital watermarking techniques
are the most commonly used.
The principal objective of digital watermarking is to embed data called a water-
mark (tag, label or digital signal) into a multimedia object with the aim of broadcast
monitoring, access control, copyright protection etc. The object may be an image
or video or audio. Watermarking techniques can be divided into various categories
according to visibility, permanency and domain [1, 2].
In relation to the visibility, digital watermarks can be of two types: visible or
invisible watermarks. Visible watermarks can easily be detected by observation such
as an embossed logo superimposed upon the image indicating the ownership of
the image. Although the owner of the multimedia object can be identified with no
computation, watermarks can be removed or destroyed with ease. On the contrary,
invisible watermarks are designed to be imperceptible (transparent) and are inserted
on the unknown places in the host data. The watermarked data should look similar
to the original one, and should not cause suspicion by others. If it is used in an
illegal way, the embedded watermark can then be utilized as a proof for showing the
ownership of the multimedia object. The invisible watermarks are more secure and
robust than the visible watermarks and are the subject of this chapter.
In the class of invisible watermarks, one may further categorise techniques
according to permanency as fragile, semi-fragile and robust. Fragile watermarks
are embedded in such a way that slight changes to the watermarked image would
destroy the watermark. Hence, they are mainly used with the purpose of authentica-
tion. Semi-fragile watermarks are designed to tolerate some degree of modification
to the watermarked image, for instance, the addition of quantization noise from
lossy compression. Robust watermarks are designed to resist intentional or unin-
tentional image modifications called attacks that attempt to destroy or remove the
watermark. Such attacks include filtering, geometric transformations, noise addi-
tion, lossy compression, etc. In general, this kind of watermarks is employed for
copyrights protection and ownership verification.
Depending upon the domain in which the watermark is embedded, the techniques
can be grouped into two classes: spatial and frequency-domain techniques. In spa-
tial domain, the pixel values of the cover image are modified for embedding the
watermark. On the other hand, the watermark is embedded in transform domain by
means of modulating the coefficients of the transformed host image according to
watermark. Most of the transform domain watermarking techniques developed with
the use of Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT),
and Discrete Wavelet Transform (DWT). Although spatial domain methods are less
complex as no transform is used, transform domain watermarking techniques are
more robust in comparison to spatial domain methods against attacks.
Two important and conflicting requirements in watermarking are perceptual
transparency and robustness. If significant modifications are introduced to the host
image either in spatial or transform domain, greater robustness, in general, can
be obtained. However, such modifications are distinguishable and hence do not
13 SVD and DWT-SVD Domain Robust Watermarking 149

satisfy the requirement of transparency (invisibility). The design of an optimal

watermarking for a given application always involves a trade-off between these
competing criteria. Thus, image watermarking can be formulated as an optimiza-
tion task. Artificial intelligence techniques such as genetic algorithm, differential
evolution, particle swarm, Fuzzy logic, neural networks, etc. have been employed
to solve the problem optimally in spatial domain [3], DCT [4–9], DWT [10–12] and
SVD [12–14].
In general, a watermarked image can be obtained by adding the watermark multi-
plied by a scaling factor (SF) to the host image either in spatial or transform domain.
SF is used to control the watermark strength. The smaller the SF, the better the image
quality (transparency) but the weaker the robustness. In contrast, the larger the SF,
the stronger the robustness but the more the degradation of the quality of the host
image. Moreover, a single SF may not be applicable for modifying all the values of
host image because different spectral components may exhibit different tolerance to
modification. Therefore, multiple SFs need to be used for adapting to them to reduce
visual artifacts [13].
Singular value decomposition is one of the most popular methods of linear alge-
bra with various applications in signal processing including watermarking [12–23].
This study aims to develop two optimal watermarking techniques based on SVD
and DWT-SVD domain for grey-scale images. The first one embeds the watermark
by modifying the singular values of the host image with multiple SFs. In the second
approach, a hybrid algorithm is proposed based on DWT and SVD. Having decom-
posed the host image into four bands, SVD is applied to each band. Then, the same
watermark is embedded by modifying the singular values of each band with different
SFs. Various combinations of SFs are possible, and it is difficult to obtain optimal
solutions by trial and error. Thus, in order to achieve the highest possible trans-
parency and robustness, optimization of the scaling factors is necessary. This work
employs DE to obtain optimum SFs. DE can search for multiple solutions simulta-
neously over a wide range, and an optimum solution can be gained by combining the
obtained results appropriately. Experimental results of developed techniques show
both the significant improvement in transparency and the robustness under attacks.

13.2 SVD Domain and DWT-SVD Domain Watermarking

Techniques

SVD-Domain Watermarking: SVD decomposes an N  N matrix A into a product

of 3 matrices A D US V T where U and V T are N  N orthogonal matrices. S is an
N  N diagonal matrix. The elements of S are only nonzero on the diagonal and are
called the SVs of A. The steps of SVD based watermarking algorithm is described
in Table 13.1 [15]:
DWT-SVD Domain watermarking: DWT transforms a signal from spatial
domain to frequency domain. The transformation is done using wavelet filters
such as Haar Wavelet Filter, Daubechies Orthogonal Filters and Daubechies Bi-
Orthogonal Filters. Each of these filters decomposes the image into several
150 V. Aslantas

Table 13.1 SVD based watermarking scheme

Watermark embedding Watermark extracting
1. Apply SVD to the host image (I ): 1. Apply SVD to the watermarked (possibly

I D US V T distorted) image (IW ):
   T
IW D U S W V

2. Modify the S with the watermark 2. Compute possibly corrupted SM :
SM D S C kW 
SM D UW SW VW  T

3. Apply SVD to the SM : 3. Extract the watermark (possibly distorted)


SM D UW SW VW T
image (W  ): W  D .SM  S /=k
4. Compute watermarked image (IW /:
IW D USW V T

Table 13.2 DWT-SVD based watermarking scheme

Watermark embedding Watermark extracting
1. Apply DWT to decompose the host image 1. Decompose the watermarked (possibly dis-
I into 4 subbands torted) image IdW into 4 subbands (LL, HL,
LH and HH) using DWT
2. Compute SVD of each subband I k D 2. Compute SVD of each subband image:
U k S k V kT, kD 1; 2; 3; 4 where k denotes IdkW D
subbands, and k i , i D 1; : : :; N are the SVs Ud W SdkW VdkW , k D 1; 2; 3; 4
k

of S k
3. Apply SVD to the watermark: W D 3. Obtain possible corrupted singular val-
UW SW VW where W i i D 1; : : :; N are ues of the watermark from each subband:
the singular values of SW d W i k D .d i k  i k /=˛k ,
4. Modify the SVs of each subband with the 4. Using the SVs, calculate the four watermarks
SVs of the watermark: k i D k i C (possibly distorted): Wk D UW Sd W k VW T
,
˛k wi k D 1; 2; 3; 4
5. Obtain the modified coefficients of the each
subbands: I k D U k S k V kT
6. Compute the watermarked image by apply-
ing the inverse DWT using

frequencies. In two dimensional DWT, each level of decomposition produces four

subbands of the image identified as LL1, HL1, LH1 and HH1 (L: i.e., low, H: i.e.,
high). The LL1 band is termed as approximate subband while the others i.e. HL1,
LH1, HH1 are termed as detailed subbands. The low pass band (LL1) can further
be decomposed to obtain another level of decomposition. The process continues
until the desired number of levels determined by the application is reached. The
watermarking algorithm initially decomposes the host image into subbands. After-
wards, it determines the SVs of each subband and modifies these SVs with the same
watermark by scaling with different scaling factors. The steps of DWT-SVD domain
watermarking scheme is given in Table 13.2 [16].
13 SVD and DWT-SVD Domain Robust Watermarking 151

13.3 Optimal Watermarkings using DE

The reminder of this section explains the way that DE is employed to optimize
watermark embedding process with respect to the two conflicting requirements:
transparency and robustness to different attacks.
An N xN host image (or subband) can have N SVs that may reveal different
tolerance to modification. Prior knowledge may not be available about the sensi-
tivity of the image with respect to various values of the SFs so an algorithm needs
to be employed to compute the optimum scaling factors that produce maximum
robustness and transparency. Thereby, DE is employed in this work to achieve this
objective.

13.3.1 Basic Concepts of DE

Differential Evolution (DE) [24] is a novel population-based heuristic related to

evolutionary computation. It works with real-valued parameters and is fairly fast
and reasonably robust. It is capable of handling nonlinear, non-differentiable and
multimodal continuous space functions. Like other evolutionary algorithms, DE
also utilizes similar operators such as the initial population generation, crossover,
mutation and selection but differs from those algorithms such as Genetic Algo-
rithms and Evolutionary Strategies, where perturbation occurs in accordance with
a random quantity, DE employs weighted differences between solution vectors to
perturb the population. For each generation, the operators are continually repeated
until the best vector representing the optimum solution is produced or some prede-
fined termination criterion, such as maximum generation number or when there is
no improvement in the fitness function is reached. The operators of DE algorithm
are briefly described as follows:
Initial Population: DE starts with a population of NP solution vectors as is the
case with other evolutionary algorithms. The population consists of real valued vec-
tors with dimension D which equals to the number of design parameters. Each
vector forms a candidate solution to optimization problem. If a priori knowledge
is not available about the problem, the first population of solutions can be generated
randomly as follows:
 
xi;G D xi.L/ C randi Œ0; 1 xi.H /  xi.L/ (13.1)

where xi;.L/ and xi;.H / are the lower and higher boundaries of D-dimensional
vector xi D fxj;i g D fx1;i ; x2;i ; : : :; xD;i gT , respectively.
Mutation: The main operator of DE is the mutation operator that is employed to
expand the search space. For each target vector xi;G , a mutant vector vi;G C 1 is gen-
erated by the combination of vectors randomly chosen from the current population
at generation G as:
vi;GC1 D xr1;G C F .xr2;G  xr3;G / (13.2)
152 V. Aslantas

where i , r1 , r2 , and r3 are mutually different indexes selected from the current
generation f1; 2; : : :; NP g. As a result, the population size must be greater than 3.
F 2 Œ0; 2 is a user-defined real constant called step size which control the perturba-
tion and improve convergence by scaling the difference vector. Smaller values for F
result in faster convergence in the generated population and larger values in higher
diversity.
Crossover: This operator combines successful solutions of the previous genera-
tion in order to maintain diversity in the population. By considering the elements
of the mutated vector, vi;G C 1 , and the elements of the target vector, xi;G , the trial
vector ui;G C 1 is produced as:

vj;i;GC1 if randj;i  CR or j D Irand

uj;i;GC1 D (13.3)
xj;i;G if randj;i > CR and j ¤ Irand

where j D 1; 2; : : :; D; randj;i 2 Œ0I 1 is the random number; CR 2 Œ0I 1 is user-

defined crossover constant and Irand 2 Œ1; 2; : : :; D is a randomly chosen index.
Irand ensures that the trial vector gets at least one parameter from the mutant vector,
i.e., vi;GC1 ¤ xi;G .
Selection: The selection operator chooses the vectors that are going to compose
the population in the next generation by employing the so called “greedy” criterion.
If the fitness value of the trial vector is better than or equal to fitness value of the
target vector, the latter is replaced by the former otherwise the latter is preserved in
the population of the next generation.

ui;GC1 if f .ui;GC1 /  f .xi;G /

xi;GC1 D i D 1; 2; : : : ; NP (13.4)
xi;G otherwise

13.3.2 Optimisation of Scaling Factors

In order to accomplish the optimal performance of digital image watermarking

techniques, the proposed methods employ the DE algorithm to search for opti-
mal parameters that are the scaling factors. By applying DE algorithm, the SFs
(k) are obtained for optimal watermarking depending on both the robustness and
the transparency requirements under certain attacks at each generation of optimi-
sation process. Figure 13.1 illustrates the flowchart of the developed DE based
watermarking techniques in this study [12–14].
In the application of DE, each member vector in the population represents a pos-
sible solution to the problem and hence consists of a set of SFs. At the beginning
of the optimization process, DE utilizes randomly generated initial solutions pro-
duced by random number generator. By employing each vector (the values of SFs),
the watermarked images of the current generation are computed according to the
watermark embedding schemes described in Tables 13.1 and 13.2 in Section 13.2.
13 SVD and DWT-SVD Domain Robust Watermarking 153

I
W Correlation
computation IW
mutation
W*
DE Algorithm

crossover Watermark
extraction

Fitness evaluation
Attacks
selection IW

Watermarking
No embedding
Final generation
k W
Yes
Watermark Host image
Watermarked
image k

Fig. 13.1 Block diagram of DE-based watermark embedding scheme

In practice, different watermarking applications may require robustness to differ-

ent attacks. Hence, the level of robustness of the watermark differs from application
to application. The robustness of the developed watermarking schemes is evaluated
with respect to the commonly employed attacks in literature. They include average
filtering (AV), rotation (RT), sharpening (SH), and resizing (RS). As can be seen
from Fig. 13.1, the developed system is very flexible. Therefore, the other attacking
methods can easily be included to it or replaced with those used in this work [4,13].
By employing the extraction procedures given in Section 13.2, the watermarks
are extracted from the attacked watermarked images. The correlation values are
computed between the host and watermarked images (corrI / and between the origi-
nal watermark and the extracted watermarks (corrW ). These values are then used to
compute the fitness of a solution in the population of DE at each generation. Having
assigned the fitness values for all vectors in the population, the vectors producing the
better fitness values are selected for the next generation. By employing the operators
of DE (mutation, crossover, and selection), the population of the next generation is
then generated from these vectors. The above process on the population is repeated
until the predefined criterion is fulfilled.
The fitness function designed for SVD based method is formulated as:
" , ! #1
1X
t
 
fi D 1 corrw W; Wi corrI .I; IW / (13.5)
t
i D1

where fi and t are the fitness value of the i th solution and the number of attack-
ing methods, respectively. corrI () and corrW () are related to transparency and
154 V. Aslantas

Table 13.3 Outlines of the developed DE based watermarking algorithms

1. Define the fitness function, numbers of variables (D) and the values for population
size (NP), control parameters (F and CR), and number of generations (or any other
terminating criteria)
2. Generate an initial population of potential solutions at random
3. Calculate watermarked images using the solutions in the population by means of
embedding process (Tables 13.1 and 13.2)
4. Apply the attack functions upon the watermarked images one by one
5. Extract out the watermarks from the attacked images using the extraction procedure
(Tables 13.1 and 13.2)
6. Obtain the NC values between the host and each watermarked images
7. Compute the NC values between the watermark and the extracted ones
8. Evaluate the fitness value for each corresponding solution
9. Apply mutation and crossover operators
10. Apply the selection process
Repeat Steps 3–11 until a predefined termination criteria are satisfied

Fig. 13.2 (a) Host image, (b) watermark, (c) SVD domain watermarked image, (d) DWT-SVD
domain watermarked image

robustness measure, respectively. In DWT-SVD domain method, corrW () in Eq.

(13.5) is replaced with max.corrW / which represents the maximum correlation
value produced by the best quality watermark among the watermarks extracted from
the four subbands. Table 13.3 illustrates the outlines of the developed DE based
watermarking algorithms: The steps given in the table are the same for both of the
techniques provided that the watermark embedding and extracting stages must be
done according to the procedures of each technique explained in Section 13.2.

13.4 Results

To evaluate the performances of the proposed schemes, several experiments were

conducted using the 256  256 Lena (host) and 32  32 watermark images that are
illustrated in Fig. 13.2a and b. The attacks employed in the fitness evaluation process
were: AV (3  3), SH .3  3/, RS (bicubic: 256 ! 128 ! 256), and RT (30ı/. The
control parameters of DE employed for SVD and DWT-SVD domain methods are
listed in Table 13.4.
13 SVD and DWT-SVD Domain Robust Watermarking 155

Table 13.4 Control parameters used for SVD and DWT-SVD domain methods
Control parameters SVD domain DWT-SVD domain

NP 150 16
F 0.6 0.5
CR 0.8 0.6
Number of parameters (D D SFs) 32 4
Maximum generation number 400 200
Mutation method DE/best/1/exp DE/best/1/exp

In the representation scheme of solutions, each string consisted of 32 parame-

ters for SVD domain and 4 parameters for DWT-SVD domain techniques. Every
parameter represented a possible SF. Table 13.5 shows the optimal SVD domain
SFs obtained. In DWT-SVD domain, the optimized SFs for LL, HL, LH and HH
subbands are 0.873, 0.083, 0.092 and 0.012, respectively. By employing the SFs
obtained, the watermarked images of both techniques are computed and are shown
in Fig. 13.2c and d. As can be observed from those figures, the watermarked images
are not distinguishable from the original one. In addition to the attacks used in the
optimisation process, the effectiveness of the algorithms were also tested to cope
with the Gaussian noise (GN) (mean D 0, variance D 0:001), 30  30 pixel transla-
tion (TR), cropping (CR) on left half and JPEG compression (50:1) attacks [12–14].
Distorted images after attacks and corresponding watermarks extracted by both
techniques are illustrated in Fig. 13.3. The extracted watermarks by DWT-SVD
given in the figure show the best quality watermarks obtained from the 4 subbands
which are also indicated in the figure in parenthesis.
The results obtained are given in Tables 13.6 and 13.7 by comparing the corre-
lation values. The same experiments were also carried out with single SF as shown
in the tables [15, 16]. As can be seen from the tables, the larger the SF, the more the
distortion of the quality of the host image (transparency) and the stronger the robust-
ness. In contrast, the smaller the SF, the better the image quality and the weaker the
robustness. On the other hand, results of the multiple SFs obtained by DE show both
the significant improvement in transparency and the robustness under attacks.

13.5 Conclusions

In this chapter, two optimal watermarking schemes based on SVD and DWT-SVD
domain are presented. As different images (or subbands) may have different sin-
gular values, the proposed techniques provide a general solution for determining
the optimum values of the scaling factors for each image/subband. Differential
 156

Table 13.5 SVD domain SFs obtained by DE (NV = number of variables)

NV 1 2 3 4 5 6 7 8 9 10 11
SFs 51:05 242:21 1:65 0:71 0:76 1:11 0:97 2:30 0:96 1:08 0:85
NV 12 13 14 15 16 17 18 19 20 21 22
SFs 0:95 0:77 0:63 7:90 13:61 1:13 0:59 1:22 0:77 0:99 1:10
NV 23 24 25 26 27 28 29 30 31 32
SFs 0:85 1:22 0:62 1:07 1:15 1:24 20:31 19:48 27:54 7:75
V. Aslantas
13 SVD and DWT-SVD Domain Robust Watermarking 157

RT RS GN TR

SVD Domain

DWT-SVD Domain

(LL) (LL) (LL) (HL)

AV SH CR JPEG

SVD Domain

DWT-SVD Domain

(LL) (HH) (HL) (LL)

Fig. 13.3 Distorted images and corresponding extracted watermarks

evolution algorithm is adopted in the techniques in order to achieve the highest

possible robustness without degrading image quality. Experimental results show the
feasibility of the developed techniques and their superiority over the pure SVD and
DWT-SVD domain techniques.
158 V. Aslantas

Table 13.6 SVD domain correlation values . corrI .I; I  / and corrW .W; W  //
DE Single SF
SFs 0.1 0.3 0.5 0.7 0.9

corrI 0.9995 1.0000 0.9996 0.9989 0.9986 0.9981

RT 0.9965 0.9790 0.9799 0.9783 0.9703 0.9741
RS 0.9947 0.9827 0.9820 0.9820 0.9838 0.9840
AV 0.9963 0.9829 0.9829 0.9829 0.9846 0.9847
SH 0.9843 0.9673 0.9728 0.9713 0.9710 0.9733
corrW GN 0.9944 0.9678 0.9744 0.9749 0.9774 0.9837
TR 0.9809 0.9596 0.9708 0.9773 0.9776 0.9804
CR 0.9815 0.9625 0.9708 0.9710 0.9751 0.9784
JPEG 0.9996 0.9763 0.9922 0.9973 0.9975 0.9988

Table 13.7 DWT-SVD domain correlation values . corrI .I; I  / and corrW .W; W  //
DE Single SF
SFs 0.005 0.01 0.05 0.1 0.5
corrI 0.9990 0.9999 0.9999 0.9996 0.9985 0.9750
RT 0.9910 0.2512 0.2693 0.3551 0.4121 0.6851
RS 0.9815 0.2916 0.3040 0.4574 0.6227 0.9619
AV 0.9868 0.5281 0.5154 0.4353 0.6176 0.9780
SH 0.9840 0.8213 0.8332 0.8859 0.9227 0.9761
corrW GN 0.9982 0.7527 0.7929 0.9859 0.9950 0.9933
TR 0.9896 0.5733 0.9298 0.9978 0.9994 0.9996
CR 0.8339 0.6942 0.6999 0.7297 0.7998 0.9755
JPEG 0.9974 0.6320 0.7105 0.9315 0.9634 0.9971

References

1. Hartung F, Kutter M. Multimedia watermarking techniques. Proceedings of IEEE 1999;87:

1079–1107.
2. Potdar VM, Han S, Chang E. A survey of digital image watermarking techniques. IEEE third
International Conference on Industrial Informatics, INDIN’05, 2005, pp. 709–716.
3. Zhang F, Zhang H. Applications of a neural network to watermarking capacity of digital
image. Neurocomputing 2005;67:345–349.
4. Shieh CS, Huang HC, Wang FH, Pan JS. Genetic watermarking based on transform-domain
techniques. Pattern Recognition 2004;37:555–565.
5. Shih FY, Wu YT. Enhancement of image watermark retrieval based on genetic algorithm.
Journal of Visual Communication and Image Representation 2005;16:115–133.
6. Chang YL, Sun KT, Chen YH. ART2-based genetic watermarking. 19th International
Conference on Advanced Information Networking and Applications, 2006, pp. 729–734.
7. Aslantas V, Ozer S, Ozturk S. A novel clonal selection algorithm based fragile watermarking
method. LNCS 2007; 4628:358–369.
8. Khan A, Mirza A. Genetic perceptual shaping: utilizing cover image and conceivable attack
information using genetic programming. Information Fusion, 2007;8:354–365.
9. Aslantas V, Ozer S, Ozturk S. A novel fragile watermarking based on particle swarm
optimization. IEEE International Conference on Multimedia and Expo, 2008, pp. 269–272.
13 SVD and DWT-SVD Domain Robust Watermarking 159

10. Kumsawat P, Attakitmongcol K, Srikaew A. A new approach for optimisation in image

watermarking by using genetic algorithms. IEEE Transactions on Signal Process, 2005;53:
4707–4719.
11. Lee D, Kim T, Lee S, Paik J. Genetic algorithm-based watermarking in discrete wavelet
transform domain. LNCS 2006; 4113:709–716.
12. Aslantas V; Dogan A, Ozturk S. DWT-SVD based image watermarking using Particle Swarm
Optimizer. IEEE International Conference on Multimedia and Expo, 2008, pp. 241–244.
13. Aslantas V. A singular-value decomposition-based image watermarking using genetic algo-
rithm, International Journal of Electronics and Communications (AEU), 2008;62:386–394.
14. Aslantas V. Optimal SVD based robust watermarking using differential evolution algorithm.
World Congress of Engineering, London, UK, 2008, pp. 629–631.
15. Liu R, Tan T. An SVD-based watermarking scheme for protecting rightful ownership. IEEE
Transactions on Multimedia, 2002;4:121–128.
16. Ganic E, Eskicioglu AM. Robust DWT-SVD domain image watermarking: embedding data
in all frequencies. ACM multimedia and security workshop, Magdeburg, Germany, 2004,
pp. 166–174.
17. Huang F, Guan ZH. A hybrid SVD-DCT watermarking method based on LPSNR. Pattern
Recognition Letter, 2004;25:1769–1775.
18. Chang CC, Tsai P, Lin CC. SVD-based digital image watermarking scheme. Pattern Recogni-
tion Letter, 2005;26:1577–1586.
19. Bao P, Ma X. Image adaptive watermarking using wavelet domain singular value decomposi-
tion. IEEE Transactions on Circuits Systems Video Technology, 2005;15:96–102.
20. Shieh JM, Lou DC, Chang MC. A semi-blind digital watermarking scheme based on singular
value decomposition. Computer Standards Interfaces 2006;28:428–440.
21. Chung, KL, Yang, WN, Huang, YH, Wu, ST, Hsu, YC. On SVD-based watermarking
algorithm. Applied Mathematics and Computation 2007;188:54–57.
22. Chang CC, Hu YS, Lin CC. A digital watermarking scheme based on singular value
decomposition. ESCAPE 2007, pp. 82–93.
23. Mohan BC, Kumar SS. A robust image watermarking scheme using singular value decompo-
sition. Journal of Multimedia 2008;3:7–15.
24. Storn R, Price K. Differential evolution – A simple and efficient heuristic for global
optimization over continuous spaces. Journal of Global Optimization, 1997;11:341–359.
Chapter 14
Design and Performance Evaluation
of a Prototype Large Ring PET Scanner

M. Monjur Ahasan and David J. Parker

Abstract Positron emission tomography (PET) is a powerful radiotracer imaging

technique. Over the last 2 decades, PET has been widely applied as a non-invasive
tool for in-vivo diagnosis of human diseases. The macroPET system described here
was constructed as a prototype to demonstrate the feasibility of performing PET
scans on a large scale. It was created by reconfiguring components from an orig-
inal ECAT 951 system, which had 32 detector modules (“buckets”) mounted in
two rings with an inner diameter of 100 and a 10.8 cm axial field of view. In the
macroPET system the 32 buckets are mounted in a single ring with an inner diam-
eter of 2.34 m. This paper presents the macroPET design and initial measurements
of its characteristics.

Keywords Positron emission tomography  radiotracer imaging  macroPET design

 detector module  PET scanner

14.1 Introduction

Positron emission tomography (PET) is a powerful radiotracer imaging technique, in

which the distribution of radiotracer is measured by detecting pairs of back-to-back
gamma rays produced in positron-electron annihilation. Over the last 2 decades,
PET has been widely applied as a non-invasive tool for in-vivo diagnosis of human
diseases [1–3]. Considerable work has also been done on developing small diameter
PET scanners for small animal imaging [4–6]. At Birmingham, the technique is also
used for studying engineering processes [7].
We believe that PET also has considerable potential in the area of large ani-
mal veterinary medicine. However, a PET scanner designed to accommodate human

M.M. Ahasan (B)

School of Physics and Astronomy, The University of Birmingham, Birmingham B15 2TT, UK,
E-mail: mma532@bham.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 161
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
162 M.M. Ahasan, D.J. Parker

subjects will not be suitable for large animals such as horses. The macroPET sys-
tem described here was constructed as a prototype to demonstrate the feasibility of
performing PET scans on a large scale. It was created by reconfiguring components
from an original ECAT 951 system [8], which had 32 detector modules (“buckets”)
mounted in two rings with an inner diameter of 100 and a 10.8 cm axial field of
view. In the macroPET system the 32 buckets are mounted in a single ring with
an inner diameter of 2.34 m. This paper presents the macroPET design and initial
measurements of its characteristics. To the best of our knowledge, this is the first
attempt to construct and test a PET scanner with such a large diameter.

14.2 The Macropet Design

The ECAT 951 scanner, manufactured by CTI Inc, is based on bismuth germanate
(BGO) block detectors. Each block consists of a BGO crystal 50  60  30 mm3 ,
backed by four photomultiplier tubes (PMTs). The front of the crystal is cut into an
8  8 array of crystal elements, each 6:25  6:75 mm2 . By comparing the light inten-
sity detected in the four PMTs, a  -ray interaction can be unambiguously assigned
to an individual crystal element. The blocks are grouped into buckets each com-
prising four detector blocks together with their associated electronics (preamplifiers
and discriminators) under the control of a microprocessor. Data from all the buckets
are fed to a set of coincidence processors which identify events in which a pair of
511 keV  -rays have been detected in two buckets within the resolving time of the
system (6 ns).
In the original ECAT 951 the 32 buckets were mounted in two adjacent rings,
each with an inner diameter of 100 cm. The 8,192 individual detector elements
were thus arranged in 16 rings, each of 512 elements. For each bucket, coincidences
could be detected with any of the seven opposing buckets in the same ring or in
the other ring. The resulting field of view was a cylinder approximately 60 cm in
diameter.
For macroPET (Fig. 14.1), the detector blocks have been remounted in a single
horizontal ring of 128 blocks. Because the blocks are tapered to fit tightly together
in the original 64 block rings, in the new arrangement there are gaps of approxi-
mately 7.5 mm between the front faces of adjacent blocks. The inner diameter of the
single ring is 234 cm. For convenience, the blocks are mounted on eight separate alu-
minium base plates, each corresponding to a 45ı segment of the ring. The blocks are
connected in fours to the original sets of bucket electronics, with bucket controllers
from the two original rings alternating around the new single ring. Overlapping of
controllers is made possible by displacing alternate controllers vertically.
By alternating the buckets from the original rings, the original coincidence com-
binations of buckets are appropriate in the new geometry, and enable imaging to
be performed over a field of view approximately 125 cm in diameter. The eight
rings of detector elements span an axial (vertical) field of view of 5.4 cm, which
14 Design and Performance Evaluation of a Prototype Large Ring PET Scanner 163

Fig. 14.1 A photography of the macroPET scanner

is divided into 15 image planes (eight direct planes and seven cross planes) with a
plane separation of 3.375 mm.
The initial results reported here were obtained by accepting data with a maxi-
mum ring difference of 3, as is generally the case in acquisition of “2D” PET data.
Purpose-built septa were not constructed for macroPET, but in an attempt to inves-
tigate whether such septa would be valuable some studies were performed with the
original ECAT 951 septa mounted inside the macroPET ring. This limits the FOV
to 82 cm diameter. Apart from these septa, macroPET has no shielding against out
of field activity.
Events within the energy window from 250 to 850 keV were accepted from the
system. A delayed timing window was used to record the distribution of random
coincidences. Data were initially recorded in histograms (prompt minus delayed
events) appropriate to the image planes of the original geometry and were then
rebinned into the correct sinograms for the new geometry. In this rebinning the gaps
between blocks were allowed for by treating each large ring as if it consisted of
1,152 (128  9) detector elements with every ninth detector absent, and interpo-
lation was used to complete the missing sinogram elements. Arc correction of the
projection data was achieved by performing a further rebinning with linear weight-
ing. To compensate for the difference in efficiency of different detector elements, a
normalisation was applied based on the response to a central uniform phantom.
Simple 2D filtered backprojection has been used to reconstruct all the images pre-
sented here, using a Hamming filter of cutoff frequency 0.4 and a 256 by 256 matrix
with zoom factor from 1 to 4. Results measured with septa in place are reported in
brackets.
164 M.M. Ahasan, D.J. Parker

14.3 Performance Evaluation

14.3.1 Spatial Resolution

The transaxial spatial resolution of the system, with or without septa, was measured
using data from one or more 68-Ge line sources mounted parallel to the scanner axis.
In each case, a Gaussian fit was performed on the profile across the central plane of
the reconstructed images (Fig. 14.2),
p and the full-width at half-maximum (FWHM)
was calculated as FWHM D 2 2 ln 2, where  is the standard deviation of the
fitted Gaussian function. The pixel size in the reconstructed images was determined
experimentally from the image obtained using two 68-Ge line sources separated by
a known distance (40 cm apart).
On the central axis, the transaxial resolution (FWHM) was determined as
11.2 mm (11.7 mm). The resolution degrades slightly as the source is moved off
axis. At 40 cm off-centre, values of 12.0 mm (11.4 mm) were measured, and at 50 cm
off-centre the no-septa resolution was measured as 12.1 mm.
The axial resolution of the system was measured using data from an 18-F point
source positioned at the scanner centre (no-septa mode), and fitting a Gaussian func-
tion to the profile across the 15 image planes (Fig. 14.3). This central point source
gave a transaxial resolution of 10.8 mm and an axial resolution of 9.3 mm.

40000
Unnormalized counts

30000

20000

10000

0
50 75 100 125 150 175 200
-10000
Distance (pixels)

Fig. 14.2 Reconstructed image for two 68-Ge line sources and its profile

1500
Activity (a.u.)

1000

500

Fig. 14.3 The axial image 0

profile for a central 18-F point 1 3 5 7 9
source Image plane
14 Design and Performance Evaluation of a Prototype Large Ring PET Scanner 165

These resolution values are all significantly larger than the spatial resolution
quoted for the original ECAT 951 system (around 6 mm). Much of the difference
can be attributed to the effect of acollinearity of the two photons emitted in elec-
tron/positron annihilation. The mean angular deviation from collinearity is around
0:4ı [9] and the resultant effect on resolution is 0.0022D, where D is the ring diam-
eter [10]. This is expected to contribute around 5.1 mm to the FWHM of images
measured using a detector ring of diameter 2.34 m. It is also possible that errors in
positioning individual blocks in the large ring contribute to the poor spatial resolu-
tion. Nevertheless, the scanner has been shown to give acceptable spatial resolution
over a FOV of at least 1 m diameter.

14.3.2 Sensitivity

The sensitivity of the scanner was determined from the count rate measured using
bare sources of known activity in both acquisition modes. To avoid problems of
dead-time etc., these measurements were performed using sources of relatively low
activity. A 14 cm long 68-Ge line source of total activity 200 kBq was mounted
along the scanner axis, so that the activity in each 3.375 mm slice was approxi-
mately 4.8 kBq, and the count rate in each image plane was determined. In another
measurement, an 18-F point source with an activity of 15 MBq was mounted at the
centre of the field of view, and the total count rate (all planes) was found. The sen-
sitivity calculations were corrected for the positron branching ratios of 18-F (0.967)
and 68-Ge (0.89).
Figure 14.4 shows the axial sensitivity profiles determined using a 68-Ge line
source on the scanner axis. The profiles are approximately as expected based on
the number of detector rings which can contribute to each image plane, using a
maximum ring difference of 3 (four combinations in each cross plane from 4 to 12,
three combinations in each direct plane from 3 to 13, and fewer combinations in the
end planes). The septa reduce the sensitivity by approximately 55% in each plane.
The absolute values of sensitivity without septa (approximately 1.87 and
1.39 cps/kBq for cross and direct plane respectively) are significantly lower (by

(no-septa mode) (Septa mode)

2.5 1
2 0.8
Sensitivity
(cps /kBq)
Sensitivity
(cps /kBq)

1.5 0.6
1 0.4 at the centre
at the centre
0.5 0.2 at 25cm off-axis

0 0
1 3 5 7 9 11 13 15 1 3 5 7 9 11 13 15
Plane /sinogram Plane /sinogram

Fig. 14.4 Axial sensitivity profile for a line source positioned along the scanner axis
166 M.M. Ahasan, D.J. Parker

approximately 35%) than expected from a simple theoretical calculation, assuming

a detector efficiency of 0.5 and neglecting the gaps between blocks. In a detector
ring of diameter 234 cm the 128 gaps, each 7.5 mm wide, represent 6.5% of the
solid angle. Overall, we might expect these gaps to reduce the sensitivity for detect-
ing a pair of  -rays by approximately 13%. This is insufficient to account for the
observed discrepancy.
The sensitivity measured using a central point source was 2.16 (0.80) cps/kBq,
which is closer to the expected value. This may suggest an error in determining the
activity of the line source.
The sensitivity of a PET scanner is often quoted in terms of the count rate
achieved using a cylindrical phantom uniformly filled with activity. Results have
been reported using cylinders of various sizes [11,12]. For macroPET using a 20 cm
diameter cylinder filled with 18-F solution, sensitivity values of approximately 900
(400) cps kBq1 ml1 (averaged over all planes) were measured. Naturally, this is
significantly lower than the equivalent sensitivity values quoted for clinical PET
scanners [6, 13].
The total sensitivity for the original ECAT 951, measured using a 20 cm diameter
phantom in 2D mode, was quoted as 110 cps Ci1 ml1 .4; 070 cps kBq1 ml1 /.
In general the sensitivity of a scanner is proportional to Z2 =D (where Z is the axial
length and D the ring diameter), since the activity within each slice is proportional
to Z and the solid angle subtended by the detector ring is proportional to Z/D. Com-
pared with the ECAT 951, macroPET has Z reduced by a factor 1=2 and D increased
by a factor 2.34, so we expect the sensitivity to be reduced overall by approximately
a factor 0.107, giving a figure of 435 cps kBq1 ml1 . The observed sensitivity of
400 cps kBq1 ml1 (with septa) is entirely consistent with this. Actually we would
expect the sensitivity to be somewhat reduced because of the gaps between blocks.

14.3.3 Scatter Fraction

Detection of Compton scattered events is a major source of background in PET,

especially when imaging large objects without any collimating septa. To find the
effect of septa on scattered photons, the scatter fraction was measured in no-septa
and septa acquisition modes. As is conventional, the scatter fraction was measured
for a 68-Ge line source mounted at the centre of a 20 cm diameter cylinder filled
with inactive water. Figure 14.5 shows the profile across the resulting sinograms.
The scatter fraction was then determined from each profile by estimating the number
of counts within the peak, and using the equation SF D S= .T C S/. For the 20 cm
diameter cylinder, the scatter fraction was measured as 26% (17%). To investigate
the effect of scatter within a large animal, measurements were also performed in
no-septa mode using three 68-Ge line sources mounted inside a large rectangular
water tank .62  38 cm2 /. The three sources were mounted vertically at intervals
of 25 cm along the long axis of the tank, with the middle source at the centre of the
14 Design and Performance Evaluation of a Prototype Large Ring PET Scanner 167

(20cm D Phantom; (20cm D Phantom;

no-septa) with septa)
10000 100000

10000
1000
Counts

Counts
1000
100
100

10 10
0 100 200 300 400 0 100 200 300 400
Projection bin Projection bin

Fig. 14.5 Profile across the sinogram measured for a 68-Ge line source at the centre of a 20 cm
diameter cylinder of water without septa (left) and with septa (right)

Fig. 14.6 Central plane (plane 8) sinogram measured from three 68-Ge line sources inside a large
(62  38 cm2 / water tank (no-septa mode)

tank, and coincidence events were acquired for 66 h. Figure 14.6 shows the resulting
sinogram.
Values for scatter fraction were obtained by analysing the profiles at angles A
(looking directly along the long axis of the water tank, with the three sources in line)
and B (looking along the short axis). SF values of approximately 63% and 38% were
obtained along A (62 cm long) and B (38 cm wide) respectively. This measurement
was repeated with septa in plane but the results were inconclusive because the count
rate was very low and there was significant influence from out-of-field activity.
168 M.M. Ahasan, D.J. Parker

14.3.4 Count Rate Performance

Count rate performance is evaluated by calculating the noise equivalent count rate
(NEC) which takes into account the statistical noise due to scatter and random
events. The general formula [14], to calculate the NEC rate is NEC D T2 =.T C
S C kR/, where T, S and R are the true, scatter and random rates respectively. The
parameter k is the randoms correction factor, with a value of 1 or 2 depending on
the randoms correction method. A value of 2 is used when randoms are measured
using a delayed coincidence window as in the present work.
NEC rates are usually quoted using a 20 cm diameter cylindrical phantom with
an axial length of 20 cm. For macroPET, a similar phantom was used but it was only
filled to a depth of around 5 cm to avoid the effect of out-of-field of activity due to
the lack of any shielding on the current scanner. NEC rates were measured in both
no-septa and septa acquisition mode. The phantom was filled with 1,500 ml water
containing an initial activity of 425 kBq/ml (1.1 MBq/ml) of 18-F, and was then
placed in the centre of the scanner. A set of 10 min scans was acquired every half
hour for almost 12 h. The contribution of scattered events was estimated assuming
a scatter fraction of 26% (17%). The noise equivalent count rates were calculated
from the NEC formula.
Figure 14.7 shows the various components of the count rate in both acquisition
modes, as a function of source activity, and also the resulting NEC rate. Although
the scanner delivers coincidence rates of up to 100 kcps, the NEC rate peaks at
around 30 kcps in the no-septa case due to the significant contributions of random
and scattered events. The peak NEC rate was also measured as 28 kcps in septa con-
figuration. The NEC performance would be even poorer in both acquisition modes
if out of field activity were present.

150 Trues 150

125 Scatter 125
Count rate (kcps)

Randoms
100 100
Multiples
75 NEC 75

50 50

25 25

0 0
0 100 200 300 400 0 200 400 600 800
Activity concentration (kBq/ml) Activity concentration (kBq/ml)

Fig. 14.7 Count rates for a 20 cm diameter phantom: without septa (left), with septa (right)
14 Design and Performance Evaluation of a Prototype Large Ring PET Scanner 169

14.3.5 Imaging Studies

14.3.5.1 Jaszczak Phantom

The Jaszczak SPECT phantom is often used for assessing the image quality in
nuclear medicine images. It consists of a cylinder with perspex inserts which cre-
ate cold lesions when the phantom is filled with radioactive solution. In the present
work a Jaszczak phantom with internal diameter 18.6 cm was used, containing six
different sets of cylindrical rods with diameters 4.8, 6.4, 7.9, 9.5, 11.1, 12.7 mm.
The images reported here were obtained in no-septa mode by filling only the
lower half of the phantom with approximately 2.5 l of water, and adding approxi-
mately 200 MBq of 18-F. Data were acquired for 600 s per image, with the phantom
mounted at the centre of the field of view and at different radial offsets. Each image
was reconstructed using filtered backprojection and incorporating the calculated
attenuation correction for a water-filled cylinder.
Due to the curved nature of the PET geometry, lines of response (LORs) near the
periphery of the field of view are more closely spaced than those near the centre. To
avoid distortion due to this effect, the sinogram data were rebinned before recon-
struction (a process referred to as “arc correction”). Figure 14.8 shows the centre
plane image from the phantom mounted with its axis displaced by 45 cm from the
centre of the field of view, (a) before arc correction and attenuation correction, and
(b) after applying these corrections. It can be seen that the distortion due to chang-
ing LOR separation is removed by the arc correction process, and a reliable image is
obtained out to a radius of at least 55 cm. In this image, only the two largest sets of
cold rods (12.7 and 11.1 mm diameter) are clearly visible. In other images measured
with the phantom at the centre of the FOV the 9.5 mm rods can also be seen.

14.3.5.2 Uniform Phantom

Images of a central uniform 20 cm diameter cylindrical phantom filled with 18-F

solution to the depth of the FOV were used to assess the uniformity and noise level

Fig. 14.8 Forty-five centimeters off-axis Jaszczak phantom images: before and after attenuation
and arc correction
170 M.M. Ahasan, D.J. Parker

35
30
25

Arbitrary unit
20
15
10
5
0
-5

Fig. 14.9 Uniform phantom image (slice 3) and the corresponding horizontal profile across the
centre of the image

Fig. 14.10 Contrast phantom image (central slice): no-septa (left) and septa (right)

in no-septa mode. Figure 14.9 shows the image (slice 3) for a measurement in which
150 M counts were obtained in 600 s. Also shown is the horizontal profile across this
image. From the ratio of standard deviation () to the mean pixel counts () within
a circular region of interest (ROI) covering the central part of the phantom in the
reconstructed image, the coefficient of variation (CV) was calculated as 5.9%.

14.3.5.3 Contrast Phantom

A laboratory-made 35.7 cm square phantom with warm background embedded by

three hot and two cold lesions (five small cylinders) was scanned in no-septa and
septa mode. The cylinders containing high activity (hot lesions) were 4, 2, and
1 cm diameter and all have approximately the same uptake value in each mode.
The diameters of cylinders containing inactive water (cold lesions) were 3 and 1
cm. The phantom was only filled with F-18 to a depth of around 5 cm to avoid the
effect of out-of-field of activity due to the lack of any side shielding on the system.
Figure 14.10 shows the image slice for the square phantom generated without
and with septa. For both configurations, three hot and two cold lesions are visible
14 Design and Performance Evaluation of a Prototype Large Ring PET Scanner 171

in each case. Contrast recovery coefficient (CRC) [15] values close to unity can
be achieved for the 2 and 4 cm hot lesions and for the 3 cm cold lesion. For both
the 1 cm hot and cold lesions the CRC is around 0.3.The 1 cm hot lesion is more
prominent than the 1 cm cold lesion.

14.4 Conclusions

After design, setup and calibration of the macroPET system, initial results of its
performance characteristics were obtained. The spatial resolution is around 11 mm,
significantly poorer than for a standard scanner, probably because of the significant
contribution from acollinearity. Without septa, the sensitivity measured for a 20 cm
diameter phantom is 900 cps/kBq/ml and the peak NEC rate is 30 kcps. The septa
used (which were not designed specifically for this scanner) reduce the scatter
fraction by around 35% at the cost of reducing the sensitivity by 55%.
The main weakness of the prototype scanner is the lack of side shielding, so
that it is sensitive to out-of-field activity. Most of the measurements reported here
were obtained with the activity confined axially to the field of view. Out of field
activity will contribute significantly to the randoms rate. Also, the scatter fraction
grows dramatically as larger objects are imaged. The septa used have only a limited
effect in blocking out-of-field activity (and can actually make the situation worse
as they reduce the true rate more than the background rate). To date all the images
in this work have been reconstructed using simple 2D FBP. In the future it may be
appropriate to use 3D image reconstruction and/or iterative techniques.
With these improvements, the results so far indicate that a geometry like
macroPET offers a practical solution for large animal imaging.

References

1. M. E. Phelps. Molecular imaging with positron emission tomography, Annu. Rev. Nucl. Part.
Sci. 52, 303–338 (2002).
2. C. Das, R. Kumar, Balakrishnan, B. Vijay, M. Chawla, and A. Malhotra. Disseminated tuber-
culosis masquerading as metastatic breast carcinoma on PET-CT, Clin. Nucl. Med. 33(5),
359–361 (2008).
3. R. Myers, S. Hume, P. Bloomfield, and T. Jones. Radio-imaging in small animals, Psychophar-
macol 13(4), 352–357 (1999).
4. A. P. Jeavons, R. A. Chandler, C. A. R. Dettmar. A 3D HIDAC-PET camera with sub-
millimetre resolutionnfor imaging small animals, IEEE Trans. Nucl. Sci. 46, 468–473
(1999).
5. J. Missimer, Z. Madi, M. Honer, C. Keller, A. Schubiger, and S. M. Ametamey. Performance
evaluation of the 16-module quad-HIDAC small animal PET camera, Phy. Med. Biol. 49(10),
2069–2081 (2004).
6. W. W. Moses, P. R. G. Virador, S. E. Derenzo, R. H. Huesman, and T. F. Budinger. Design
of a high-resolution, high-sensitivity PET camera for human brains and small animals, IEEE
Trans. Nucl. Sci. 44(4), 1487–1491 (1997).
172 M.M. Ahasan, D.J. Parker

7. A. Sadarmomtaz, D. J. Parker, and L. G. Byars. Modification of a medical PET scanner for

PEPT studies, Nucl. Instr. Meth. A 573, 91–94 (2007).
8. R. D. Badawi, M. P. Miller, D. L. Bailey, and P. K Marsden. Randoms variance reduction in
3D PET, Phys. Med. Biol. 44, 941–954 (1999).
9. S. DeBenedetti, C. E. Cowan, W. R. Konneker, and H. Primakoff. On the angular distribution
of two-photon annihilation radiation, Phys. Rev. 77(2), 205–212 (1950).
10. C. S. Levin and E. J. Hoffman. Calculation of positron range and its effect on the fundamental
limit of positron emission tomography system spatial resolution, Phys. Med. Bio. 44, 781–799
(1999).
11. V. Bettinardi, M. Danna, A. Savi, M. Lecchi, I. Castiglioni, M. C. Gilardi, H. Bammer,
G. Lucignani, and F. Fazio. Performance evaluation of the new whole-body PET/CT scanner:
Discovery ST, E. J. Nucl. Med. Mol. Imag. 31(6), 867–881 (2004).
12. T. DeGrado, T. Turkington, J. Williams, C. Stearns, and J. Hoffman. Performance characteris-
tics of a whole-body PET scanner, J. Nucl. Med. 35, 1398–1406 (1994).
13. T. K Lewellen, S. G. Kohlmyer, R. S. Miyaoka, M. S. Kaplan, C. W. Stearns, and
S. F. Schubert. Investigation of the Performance of the General Electric ADVANCE Positron
Emission Tomograph in 3D Mode, IEEE Trans. Nucl. Sci. 43(4), 2199–2206 (1996).
14. S. C. Strother, M. E. Casey, and E. J. Hoffmann. Measuring PET scanner sensitivity: relating
count rates to image signal-to-noise ratios using noise equivalent counts, IEEE Trans. Nucl.
Sci. 37, 783–788 (1990).
15. M. E. Daube-Witherspoon, J. S. Karp, M. E. Casey, F. P. DiFilippo, H. Hines, G. Muehllehner,
V. Simcic, C. W. Stearns, L. E. Adam, S. Kohlmyer, and V. Sossi. PET performance
measurements using the NEMA NU 2-2001 standard, J. Nucl. Med. 43(10), 1398–1409
(2002).
Chapter 15
Robust Wavelet-Based Video Watermarking
Using Edge Detection

Tamás Polyák and Gábor Fehér

Abstract In this paper a robust video watermarking method is presented, which

embeds data in the wavelet domain using edge detection. The algorithm uses the
luminance values around the edges where changes are less noticeable for the human
visual system. Watermark is embedded in an additive way using spread spectrum
noise. The complexity of the method is low. The algorithms used for wavelet
transformation and edge detection are fast and simple. Results show that the water-
mark signal is imperceptible, and the method is robust against low bitrate lossy
compressions, like H.264.

Keywords video watermarking, wavelet, edge detection, low bitrate lossy compres-
sion

15.1 Introduction

Digital video streaming is a more and more popular service among content providers.
These streams can be displayed on various types of devices: computer, PDA etc. The
content can be easily distributed through the internet, but digital contents carry a big
security risk, i.e. the copying and reproduction of the content is very easy and effort-
less. Even users without special knowledge can easily share the downloaded content
with other people.
To avoid illegitimate access, digital content is often encrypted, and travels in
an encrypted form to the consumer. Although encryption secures the content on
the way to the consumer, during playback it must be decrypted, and this is the
point where it exposed to illegal copying. In these cases watermarks can be used to

T. Polyák (B)
Budapest University of Technology and Economics, Department of Telecommunications and
Media Informatics, Magyar tudósok krt. 2., 1117 Budapest, Hungary, BME TMIT,
E-mail: tpolyak@tmit.bme.hu

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 173
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
174 T. Polyák, G. Fehér

give some extra protection to the content, since watermarking can embed additional
information in the digital content.
Watermarks are embedded in the digital content in an imperceptible way. Water-
marks can carry some information about the content: owner of the content, metadata,
QoS parameters etc. This protection does not eliminate the need for encryption, but
is a supplemental technique to secure the content, or store additional data. These
watermarks are called robust watermarks because they must survive transformations
of the underlying content e.g. lossy compression.
Digital watermarking algorithms are generally divided into two groups: algo-
rithms that hide data in the spatial domain. It means that information is embedded
by modifying the pixel values directly [1–3], and algorithms that use a transform
domain for data hiding. Discrete cosine transform (DCT) and discrete wavelet trans-
form (DWT) are often used at transform domain watermarking [4–7]. These water-
marking techniques modify coefficients of the given transform domain. Wavelet
transform is commonly used because it has many advantages over DCT transform. It
is closer to the human visual system (HVS), instead of processing 8  8 pixel blocks
it processes the whole frame. It divides the frame into four parts with different
frequencies (LL, LH, HL and HH) and directions.
Some watermarking techniques embed data in object borders, and other percep-
tually important areas. It has several advantages: the HVS is less sensitive to changes
made in these components, and modern lossy compression algorithms leave them
relatively untouched to maintain high video quality. These properties make these
regions ideal for watermarking. The detection of suitable regions can be realized in
the uncompressed domain. Pröfrock, Schlauweg and Müller use the Normed Centre
of Gravity (NCG) algorithm to find the blocks that contain object borders or other
significant changes in the video [8]. Ellinas presents an algorithm that embeds data
in images using four level DWT and edge detection. This algorithm also modifies
the surroundings of the edges; it is accomplished using a morphological dilatation
with a structuring element of 9  9.
The proposed algorithm is designed for watermarking low resolution (CIF and
QCIF) video streams. During the design of a video watermarking algorithm com-
plexity has to be taken into count. A trade-off has to be made between complexity,
quality loss, and robustness. The algorithm uses the wavelet domain for data hiding.
It modifies the wavelet coefficients that belong to object borders. Visible artifacts
will not appear on the source video. The suitable coefficients are selected by an edge
detection algorithm. Watermark is inserted in an additive way that a spread spec-
trum pseudorandom noise is added to the luminance pane of the middle frequency
components.

15.2 The Watermark Embedding Process

Figure 15.1 shows the process of watermark embedding. First the input video signal
(Xi;j ) is transformed using a fast n-level (n D 1; 2 or 3) DWT transform, the Haar
wavelet [9].
15 Robust Wavelet-Based Video Watermarking Using Edge Detection 175

N+D R Te

Xi,j Xu,v Yu,v

X'u,v
Sobel edge Insert
DWT IDWT
detector watermark

Fig. 15.1 Block diagram of the watermark embedding process

Fig. 15.2 Video frame after two-level wavelet transformation

The transformation produces a low frequency, low resolution approximation sub-

band, and 3 n detail subbands. Watermark is inserted into the middle frequency
components (HL and LH) of each level. Figure 15.2 shows the video frame after a
two-level wavelet decomposition
The transformed video then gets through an edge detecting filter, where edge
detection is applied on the middle frequency components. Edge detection is per-
formed using the Sobel edge detector [10].
The Sobel operator performs a 2-D spatial gradient measurement on an image.
It is used to find the approximate absolute gradient magnitude at each point in an
input grayscale image. The Sobel edge detector uses a pair of 3  3 convolution
masks, one estimating the gradient in the x-direction (Gx ) and the other estimating
the gradient in the y-direction (Gy ). A convolution mask is usually much smaller
than the actual image. As a result, the mask is slid over the image, manipulating a
square of pixels at a time. The two kernels are shown in (15.1).
2 3 2 3
1 0 1 1 2 1
Gx D 4 2 0 2 5 ; Gy D 4 0 0 0 5 ; (15.1)
1 0 1 1 2 1

The magnitude of the gradient (G) is then calculated using (15.2)

q
jGj D Gx2 C Gy2 (15.2)
176 T. Polyák, G. Fehér

Fig. 15.3 Video frame with Sobel edge detection in the middle frequency DWT coefficients

To improve performance an approximate magnitude is used (15.3)

ˇ ˇ
jGj D jGx j C ˇGy ˇ (15.3)

Figure 15.3 shows the frame after edge detection in the middle frequencies.
During watermark insertion the edges with value greater than a given threshold
(Te ), and the pixels around them in a given radius (R) are selected. The value of
the radius is different at the different levels of the transformed image. Radius of
two pixels is used at the first level, and radius of one pixel is used at the second
and third level considering that the higher level coefficients contain lower frequency
components, which affect the quality of the video.
Data is embedded into the 2 n selected middle frequency areas by adding a
pseudo random spread spectrum noise.
First, data is extended with a chip rate cr to improve robustness.

di D Dx ; where i D x  cr; :::; .x C 1/  cr  1 (15.4)

Insertion is done in an additive way using (15.5).

0
Yu;v D Xu;v C ˛  N i  di (15.5)
0
where Yu;v are the modified wavelet coefficients, Xu;v are the selected wavelet coef-
ficients for watermarking, ˛ is the strength of the watermark (the value which the
luminance values are modified by), Ni is the pseudorandom Gaussian noise con-
sisting of 1 and 1, di is the data to embed. The data also consists of 1 and 1
according to 0 and 1.
15 Robust Wavelet-Based Video Watermarking Using Edge Detection 177

N R Te

Yi,j Yu,v Y'u,v dj D

Sobel edge Detect
DWT watermark Decision
detector

Fig. 15.4 Block diagram of the watermark detection process

The pseudorandom noise Ni is calculated from a seed, and from the u and v coor-
dinates of the selected coefficients. This seed is used as the key for the watermarking
method.
The same data are embedded into consequent frames to improve robustness.

15.3 The Watermark Detection Process

The process of the watermark detection is also made in the wavelet domain. After
the n level DWT the middle frequency blocks get through an edge detection.
Figure 15.4 shows the block scheme of the watermark detecting process.
The embedded data are extracted by correlating the pseudorandom noise Ni with
the area containing the detected edges and their surroundings. The detection is blind.
Assuming that the value of Yu;v coefficients is almost equal,

1 X
dj D Yu;v  Ni ; (15.6)
K u;v

can be used, where K is the number of suitable coefficients and dj is the extracted
value. The embedded bit is calculated the following way:

ı D 0; if d < Tw
; (15.7)
ı D 1; if d > Tw

where Tw is a threshold used to decrease the number of false results.

Its value is:
˛
Tw D (15.8)
2
Because the embedded information changes only after nine frames, simple voting is
applied on the detected bits.

15.4 Experimental Results

The watermarking algorithm was tested on four video streams: “Mobile”, “Bus” and
“Flower”, which are highly detailed streams, and “Foreman”, which contains more
smooth regions. The size of all video streams is 352  288 pixels. Video quality was
evaluated using two metrics: PSNR and SSIM [11, 12].
178 T. Polyák, G. Fehér

15.4.1 Quality Results

Figure 15.5 shows the original and the watermarked version of the same video
frame. For better visibility it also contains cropped, 20  20 pixels blocks of the
two frames, which contain significant edges (situated at the top of the bus on the
right side of the frame). It can be seen that the noise caused by watermark signal is
imperceptible.
Figure 15.6 shows the PSNR and SSIM error maps between the original and the
watermarked frame. The error maps are converted to be negative for better visibility.
It can be seen that watermark is inserted around the edges. The PSNR error map
shows that numerous pixels have been altered (PSNR = 40.31 dB), but the SSIM
error map shows a better quality (SSIM = 0.9952).
Table 15.1 presents the objective quality values of the tested videos using two
level DWT.
It can be seen that the video quality depends on the content. If the content con-
tains more edges or textured components, then more noise is added to it. Although
it improves robustness of the watermark, the objective quality of the video becomes
worse.
Table 15.2 shows the objective quality values of the test sequence “Bus” after
embedding watermark into different number of DWT levels. At the first level R D 2,
at the second and the third levels R D 1 was used as radius to select the pixels to
modify.

Fig. 15.5 Original (a) and watermarked (b) video frame and blocks of 20  20 pixels from the
original (c) and watermarked (d) video frame using two level DWT
15 Robust Wavelet-Based Video Watermarking Using Edge Detection 179

Fig. 15.6 Error maps between original and watermarked frame using two-level DWT (converted
to negative for better visibility): (a) PSNR; (b) SSIM

Table 15.1 Average PSNR and SSIM values of watermarked test sequences
Videos Mobile Foreman Bus Flower
PSNR [dB] 35.29 45.17 37.37 35.27
SSIM 0.9842 0.9969 0.9905 0.9905

Table 15.2 Average PSNR and SSIM values after embedding data using different levels of DWT
DWT level 1 2 3
PSNR [dB] 40.90 37.37 35.88
SSIM 0.9967 0.9905 0.9818

As expected, the more DWT levels used the more pixels are modified, this results
worse video quality.

15.4.2 Robustness of the Algorithm

Robustness of the algorithm has been tested against lossy compression. The H.264/
AVC with Baseline profile was used, this codec is widely used for low bitrate video
coding.
First, test data of 128 bits were embedded in the video streams; 8 bits were
embedded at every frame using a watermarking strength (˛) of 5. Then the water-
marked videos have been compressed using the H264/AVC codec with differ-
ent bitrates. Finally, the accordance was measured between the original and the
extracted data produced by the detector.
Accordance of 100% means that every embedded bit can be correctly identified,
while 50% means that watermark signal can not been retrieved, the embedded and
the extracted data are uncorrelated.
The algorithm was tested using one-, two- and three-level DWT.
180 T. Polyák, G. Fehér

100%
95%
90%
85%
80% Erroneous bits
75% Unrecognized bits
70% Correct bits
65%
60%
55%
50%
Original 1008 kbps 720 kbps 384 kbps
Compression bitrate

Fig. 15.7 Robustness results against H.264/AVC lossy compression, 1-level DWT

15.4.2.1 Robustness Results Using One-Level DWT

Figure 15.7 shows the test results of the watermarking algorithm with one-level
DWT after H.264/AVC compression.
“Correct bits” means the number of bits that can be detected correctly.
“Erroneous bits” is the number of false bits. The embedded bit was 0 and the
detected was 1 and vice versa.
“Unrecognized bits” is the number of bits that the detecting algorithm can not
decide whether the embedded bit was 0 or 1. Results show that the watermark can
be extracted even from videos compressed using low bitrate H.264 compression.

15.4.2.2 Robustness Results Using Two-Level DWT

Figure 15.8 shows the test results of the watermarking algorithm with two-level
DWT after H.264/AVC compression
“Correct bits”, “Erroneous bits” and “Unrecognized bits” mean the same.
Using two levels produces better results: even at bitrate of 384 kbps the 94% of
the embedded data can be extracted.
The number of false bits also has reduced.

15.4.2.3 Robustness Results Using Three-Level DWT

Figure 15.9 shows the test results of the watermarking algorithm with three-level
DWT after H.264/AVC compression
15 Robust Wavelet-Based Video Watermarking Using Edge Detection 181

100%
95%
90%
85%
80% Erroneous bits
75% Unrecongnized bits
70% Correct bits
65%
60%
55%
50%
Original 1008 kbps 720 kbps 384 kbps
Compression bitrate

Fig. 15.8 Robustness results against H.264/AVC lossy compression, 2-level DWT

100%
95%
90%
85%
80% Erroneous bits
75% Unrecongnized bits
70% Correct bits
65%
60%
55%
50%
Original 1008 kbps 720 kbps 384 kbps
Compression bitrate

Fig. 15.9 Robustness results against H.264/AVC lossy compression, 3-level DWT

“Correct bits”, “Erroneous bits” and “Unrecognized bits” mean the same. It can
be seen that using the third level of DWT does not improve the robustness of the
algorithm.

15.5 Conclusion

In this paper a novel video watermarking algorithm is presented. Data embedding

is made in the wavelet domain. It uses the Sobel edge detector for finding the
appropriate areas for watermarking – these areas are the significant edges of the
182 T. Polyák, G. Fehér

frames and their surroundings. The HVS is less sensitive to modifications on middle
and high frequencies. Compression algorithms make only minor changes to these
areas. The watermark itself is a pseudo random noise, which is calculated from the
input data and a seed value. Watermark is detected by correlating the pixel values of
the selected area with a pseudo random noise.
Test results show that the embedded watermark can be imperceptible to the HVS,
and the method is resistant to the H.264/AVC lossy compression algorithms.

References

1. R. Schyndel, A. Tirkel, and C. Osborne, “A digital watermark,” IEEE Proceedings of the

International Conference Image Processing, vol. 2, pp. 86–90, 1994.
2. W. Bender, D. Gruhl, N. Morimoto, and A. Lu, “Techniques for data hiding,” IBM Systems
Journal, vol. 35, no. 3–4, pp. 313–336, 1996.
3. R. B. Wolfgang and E. J. Delp, “A watermark for digital images,” in IEEE Proceedings of the
International Conference Image Processing, vol. 3, pp. 219–222, 1996.
4. I. J. Cox, J. Kilian, T. Leighton, and T. Shamoon, “Secure spread spectrum watermarking for
multimedia,” IEEE Transactions on Image Processing, vol. 6, no. 12, pp. 1673–1687, 1997.
5. M. D. Swanson, B. Zhu, and A. H. Tewfik, “Transparent robust image watermarking,” in IEEE
Proceedings of the International Conference Image Processing, vol. 3, pp. 211–214, 1996.
6. M. Barni, F. Bartolini, and A. Piva, “Improved wavelet-based watermarking through pixel-
wise masking,” IEEE Trans. Image Processing, vol. 10, no. 5, pp. 783–791, 2001.
7. J. N. Ellinas, “A robust wavelet-based watermarking algorithm using edge detection,”
International Journal of Computer Science, vol. 2, no. 3, pp. 206–211, 2007.
8. D. Pröfrock, M. Schlauweg, and E. Müller, “A new uncompressed-domain video watermark-
ing approach robust to H.264/AVC compression,” Proceedings of the IASTED International
Conference on Signal Processing, Pattern Recognition, and Applications, pp. 99–104, 2006.
9. P. Schrder, W. Sweldens, M. Cohen, T. DeRose, and D. Salesin, “Wavelets in Computer
Graphics,” SIGGRAPH 96 Course Notes.
10. M. Nixon and A. S. Aguado, “Feature extraction & image processing,” Academic, 2007.
11. S. Winkler, “Digital Video Quality: Vision Models and Metrics,” Wiley, 2005.
12. Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From
error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4,
pp. 600–612, 2004.
Chapter 16
High Speed Soft Computing Based Circuit
for Edges Detection in Images

Nashaat M. Hussein and Angel Barriga

Abstract In this chapter a technique for detecting edges in images is presented.

The technique is based on applying soft computing techniques such as fuzzy logic
and Lukasiewicz algebra operator. The utility of this technique is related to the sim-
plicity of the operations for edge calculation that makes it very suitable for low cost
hardware implementation and high processing speed.

Keywords Edge detection  image  soft computing  fuzzy logic  Lukasiewicz alge-
bra operator

16.1 Introduction

Many image processing techniques requires the detection of edges. An edge is

defined as a sharp change in the luminosity intensity between a pixel and another
adjacent one. Most of the edge detection techniques can be grouped in two cat-
egories: gradient based techniques and Laplacian based methods. The techniques
based on gradient use the first derivative of the image and look for the maximum
and the minimum of this derivative. Examples of this type of strategies are: the
Canny method [1], Sobel method, Roberts method [2], Prewitt method [3], etc. On
the other hand the techniques based on Laplacian look for the cross by zero of
the second derivative of the image. An example of this type of techniques is the
zero-crossing method [4].
Normally the edge extraction mechanisms are implemented executing the cor-
responding software realisation on a processor. Nevertheless in applications that
demand constrained response times (real time applications) the specific hardware
implementation is required. The main drawback of implementing the edge detec-

N. M. Hussein and A. Barriga

Instituto de Microelectrónica de Sevilla (CNM-CSIC)/University of Seville, Spain
E-mails: nashaat@imse.cnm.es,barriga@imse.cnm.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 183
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
184 N.M. Hussein, A. Barriga

tion techniques in hardware is the high complexity of the existing algorithms. In

this chapter a technique for detecting edges in images is presented. The technique is
based on applying soft computing techniques such as fuzzy logic and Lukasiewicz
algebra operator. The utility of this technique is related to the simplicity of the
operations for edge calculation [5]that makes it very suitable for low cost hardware
implementation and high processing speed.

16.2 Edge Detection Algorithm

The process of edge detection in an image consists of a sequence of stages. The first
stage receives the input image and applies a filter in order to eliminate noise. Then
a binary image is obtained applying a threshold in order to classify the pixels of the
image under two categories, black and white. Finally, in the last stage the detection
of edges is performed.

16.2.1 The Filter Stage

The filter stage allows to eliminate noise patterns. The target of the filter step
consists in eliminating all those points that do not provide any type of informa-
tion of interest. The noise corresponds to nonwished information that appears in
the image. It comes principally from the capture sensor (quantisation noise) and
from the transmission of the image (fault on transmitting the information bits).
Basically we consider two types of noise: Gaussian and impulsive (salt & pep-
pers). The Gaussian noise has its origin in differences of gains of the sensor, noise
in the digitalization, etc. The impulsive noise is characterized by some arbitrary
values of pixels that are detectable because they are very different from its neigh-
bours. A way to eliminate these types of noise is by means of a low pass filter.
This filter makes smoothed of the image replacing high and low values by average
values.
The filter used in the proposed edge detection system is based on the bounded-
sum Lukasiewicz’s operator. This operator comes from multi-valued Lukasiewicz
algebra and is defined as:

BoundedSum.x; y/ D min.1; x C y/ (16.1)

where x; y 2 Œ0; 1. The main advantage of applying this operator comes from the
simplicity of the hardware realisation as it is seen in Section 16.3.
The Lukasiewicz’s bounded sum filter performs the smoothing of the image and
is suitable for salt & peppers noise as well as Gaussian. Figure 16.1 shows the effect
of applying this type of filter.
16 High Speed Soft Computing Based Circuit for Edges Detection in Images 185

Fig. 16.1 (a) Input image with salt & peppers noise, (b) Lukasiewicz’s bounded sum filter output

The application of the filter has been made using a mask based on a 3  3 array.
For the pixel xij the weighted mask is applied to obtain the new value yij as is
shown in the following expression:

1 X X
1 1
yij D min.1; xi Ck;j Cl / (16.2)
8
kD1 lD1

16.2.2 The Threshold Stage

The techniques based in thresholding an image allow classifying the pixels in two
categories (black and white). This transformation is made to establish a distinc-
tion between the objects of the image and the background. The way of generating
this binary image is making the comparison of the values of the pixels with a
threshold T .
0 if xi;j < T
yi;j D (16.3)
255 if xi;j > T
where xi;j is a pixel of the original image and yi;j is the pixel corresponding to
the binary image. In the case of a monochrome image in that the pixels are codified
with 8 bits the range of values that the pixels take corresponds to the range between
0 and 255. It is usual to express the above mentioned range with normalized values
between 0 and 1.
A basic technique for threshold calculation is based on the frequency of grey
level. In this case, for an image with n pixels and ni pixels with the grey level i; the
threshold T is calculated by means of the following expression:

P
255 P
255
T D pi i with pi D ni =n and pi D 1 (16.4)
i D1 i D1
186 N.M. Hussein, A. Barriga

where pi represents the grey level frequency (also known as the probability of the
grey level).
In 1965 Zadeh [6] proposed the fuzzy logic as a reasoning mechanism that uses
linguistic terms. The fuzzy logic is based on the fuzzy set theory in which an element
can belong to several sets with different membership degree. This is opposed with
the classic set theory in which an element belongs or not to a certain set. Thus a
fuzzy set A is defined as

A D f.x; .x// jx 2 X g (16.5)

where x is an object of the set of objects X and .x/ is the membership degree of
the element x to the set A. In the classic set theory .x/ takes values 0 or 1 whereas
in the fuzzy set theory .x/ belongs to the range of values between 0 and 1.
An advantageous aspect of applying fuzzy logic in the calculation of the thresh-
old T is that the calculation mechanism improves the processing time since it only
requires processing the image once and allows calculating in a direct way the value
of the threshold.
The fuzzy system has an input that receives the pixel that is going to be evalu-
ated and an output that corresponds to the result of the fuzzy inference. Once read
the image the output shows the value of threshold T . Basically the operation that
makes the fuzzy system corresponds to the calculation of the centre of gravity of the
histogram of the image with the following expression:
,M R
XM X R XX
T D ˛ij cij ˛ij (16.6)
i D1 j D1 i D1 j D1

where T is the threshold, M is the number of pixels of the image, R is the number
of rules of the fuzzy system, c is the consequent of each rule and ˛ is the activation
degree of the rule.
The knowledge base of the fuzzy system contains both the membership functions
and the rule base. It is made a partition of the universe of discourse of the histogram
in a set of N equally distributed membership functions. Figure 16.2a shows a par-
tition example for N D 9. Triangular membership functions have been used since
they are easier for hardware implementation. These functions have an overlapping
degree of 2 what allows to limit the number of active rules. The membership func-
tions of the consequent are singletons equally distributed in the universe of discourse
of the histogram. The use of singleton-type membership functions for the conse-
quent allows applying simplified defuzzification methods such as the Fuzzy Mean.
This defuzzification method can be interpreted as one in which each rule proposes
a conclusion with a “strength” defined by its grade of activation. The overall action
of several rules is obtained by calculating the average of the different conclusions
weighted by their grades of activation. These characteristics of processing based on
active rules and simplified defuzzification method allows a low cost and high speed
hardware implementation.
16 High Speed Soft Computing Based Circuit for Edges Detection in Images 187

L1 C1
L2 L3 L4 L5 L6 L7 L8 L9 C2 C3 C4 C5 C6 C7 C8 C9

0 a) 255 0 b) 255

if x is L1 then c is C1;
if x is L6 then c is C6;
if x is L2 then c is C2;
if x is L7 then c is C7;
if x is L3 then c is C3;
if x is L8 then c is C8;
if x is L4 then c is C4;
if x is L9 then c is C9;
if x is L5 then c is C5;
c)

Fig. 16.2 Membership functions for N = 9, (a) antecedent, (b) consequent, (c) rulebase

The rule base of the system of Fig. 16.2c used the membership functions pre-
viously defined. The knowledge base (membership functions and the rule base) is
common for any images, for that reason the values can store in a ROM memory.
It is possible to optimize the expression shown in Eq. (16.6) if the system is
normalized. In this case the sum extending to the rule base of the grades of activation
P
of the consequent takes value 1 . R j D1 ˛ij D 1/. Then Eq. (16.6) transforms in:

1 XX
M R
T D ˛ij cij (16.7)
M
i D1 j D1

For each pixel the system makes the inference in agreement with the rule base of
Fig. 16.2c. The output of the system accumulates the result corresponding to the
numerator of Eq. (16.7). The final output is generated with the last pixel of the
image.

16.2.3 The Edge Detection Stage

The input image for the edge detection is a binary image in which pixels take value
black (0) or white (255). In this case the edges appear when a change between black
and white takes place between two consecutive pixels. The edge generation consists
of determining if each pixel has neighbours with different values. Since the image is
binary every pixel is codified with a bit (black = 0 and white = 1). This operation of
edge detection is obtained calculating the xor logic operation between neighbouring
pixels using a 3  3 mask. Using the 3  3 mask it is possible to refine the edge
generation detecting the orientation of the edges. Figure 16.3c shows an example
of applying the xor operator on the binary image. Figure 16.4 shows the obtained
results when making the edge detection on a set of test images.
188 N.M. Hussein, A. Barriga

Fig. 16.3 (a) Lena‘s image, (b) binary image, (c) edge detection

Fig. 16.4 Test images and edge detection results

16.3 Hardware Implementation

The circuit that realizes the edge detection has been implemented on a low cost
FPGA device of the Xilinx Spartan3 family. This circuit receives the image from a
vision sensor based on a CMOS chip that gives a CIF resolution .352  288/. The
image is stored in a double port RAM memory. The data memory width is 32 bits.
This allows to read two words simultaneously.
Figure 16.5 shows the block diagram of the system [5]. The image is captured by
the camera. The camera provides an 8 bits pixel in each clock cycle. Then the pixel is
stored in the RAM memory under the control of the memory control circuit. During
the reading of the image the threshold is calculated since the threshold generation
circuit receives the pixel coming from the camera. As soon as the image is stored
in the memory the threshold generation circuit produce the threshold value of the
image.
Next, the edge detection circuit initiates its operation reading from the memory
eight pixels in each clock cycle (2 words of 32 bits). This way the edge detection
circuit is able to provide four output data in parallel that are stored in the external
memory. Each data corresponds to a pixel of the edge image. This image is binary
16 High Speed Soft Computing Based Circuit for Edges Detection in Images 189

Double port
Camera
RAM

8 32 32

Memory control
8
4 32

Threshold 8 Edge detection

generation
circuit
circuit

Fig. 16.5 Block diagram of the system

reason why only a bit is needed to represent the value of the pixel (0 if edge or 1 if
background). The new image of the edges is stored in the above mentioned memory.

16.3.1 The Threshold Generation Circuit

In order to implement the fuzzy system circuit that generates the threshold the
design methodology is based on the hardware description language VHDL as a
way to describe and model the system at high level [7]. To achieve a behavioural
modelling of the fuzzy inference module a VHDL description style will be used. In
this description the system structure (fuzzy sets, rule base) and the operator descrip-
tion (connectives, fuzzy operations) are defined separately. This makes it possible
to describe independently both the fuzzy system structure and the processing algo-
rithm. The fuzzy system description must be synthesizable in order to generate the
hardware realizations.
Figure 16.6 shows the VHDL architecture of the fuzzy system. The rule base is
described in the architecture body. It contains a rule base structure with the nine
rules of Fig. 16.2. Each rule can be divided into two components: the antecedent of
the rule and the consequent. The antecedent is an expression of the input variables
related to their linguistic values. The consequent sets the linguistic value of the rule
output.
The processing mechanisms of the fuzzy operation is (=) and the inference then
(rule(,)) are not defined in the VHDL description. Only the structure of the rulebase
is defined. Such a description is a high level description because it does not assume
any specific implementation criteria. It only describes the knowledge base in terms
of a behavioral rule base.
Linguistic labels represent a range of values within the universe of discourse of
input and output variables. These labels can be described by functions in order to
compute the membership degree of a certain input value. Membership functions
190 N.M. Hussein, A. Barriga

architecture knowledge_base of threshold is

signal R: consec;
-- MF for x
constant L1: triangle:=((0, 0, 31),(0, 32));
constant L2: triangle:=((0, 31, 63),(32,32));
. . .
constant L9: triangle:=((223,255,255),(32,
0));
--MF for z
constant C1: integer := 0; constant C2: integer := 31;
. . .
constant C9: integer := 255;
begin
R(1)<=rule((x=L1),C1); R(2)<=rule((x= L2),C2);
R(3)<=rule((x=L3),C3); R(4)<=rule((x= L4),C4);
R(5)<=rule((x=L5),C5); R(6)<=rule((x= L6),C6);
R(7)<=rule((x=L7),C7); R(8)<=rule((x= L8),C8);
R(9)<=rule((x=L9),C9);
Zout<=defuzz(R);
end knowledge_base;

Fig. 16.6 VHDL architecture of the knowledge base

associated to a linguistic label can be triangular or trapezoidal. The declarative part

of the architecture of Fig. 16.8 shows the definition of such membership functions.
The data type triangle is defined in a VHDL package. This data type contains
the definitions of the points that define a triangle as well as the slopes. On the other
hand the rule base expresses the knowledge of Fig. 16.2. The function rule is also
defined in the VHDL package. This function makes the inference of the rule. The
set of rules is evaluated concurrently since the signal assignments in the architecture
body are concurrent. The operator “ = ” has been redefined taking advantage of the
functions overload properties in VHDL.
The functions used in the description of the fuzzy system have been described in
agreement with the restrictions of VHDL for synthesis. This has allowed to generate
the circuit that implements the fuzzy system using conventional circuit synthesis
tools. The fuzzy inference module is a combinational circuit that makes the fuzzy
inference.
The circuit can operate at a frequency of 137 MHz which would allow that the
processing of an CIF image (352  288 pixels) will carry out in 0.7 ms. In case of an
VGA image (1; 024  768) require a time of 5.7 ms to calculate the threshold.

16.3.2 Design of the Edge Detection System

The edge detection algorithm basically is constituted by three stages. In the first
stage the Lukasiewicz bounded-sum is performed. After the filter stage a threshold-
ing step is applied. This gives rise to a black and white monochrome image. In the
third stage the edges of the image are obtained. For it the final value of each pixel is
16 High Speed Soft Computing Based Circuit for Edges Detection in Images 191

R1 R2 R3

R4 R5 R6

UF1

pixel 1 pixel 2 pixel 3 R7 R8 R9

UF2

pixel 4 pixel 5 pixel 6 R10 R11 R12

UF3
Unidad
pixel 7 pixel 8 pixel 9 Funcional R13 R14 R15
(UF)
UF4

R16 R17 R18

UF5

R19 R20 R21

UF6

R22 R23 R24

Fig. 16.7 Block diagram of the (a) 3  3 architecture and (b) 8  3 architecture

evaluated. Only those pixels that are around the target pixel are of interest (a 3  3
mask). By this reason, if in the surroundings of a pixel the value is the same (all
white or all black) then it means no edge and the output value associates the above
mentioned pixel with the background of the image. When it is detected that some
value of the surroundings of the pixel changes indicates that the pixel at issue is in
an edge, reason why the black value is assigned to it.
Figure 16.7a shows the basic processing scheme to calculate the value of one
pixel. Pixels 1–9 correspond to the 3  3 mask register that moves through the
image. The Functional Unit (FU) makes the processing of the data stored in the
mask registers.
In order to improve the image processing time the mask was spread to an 8  3
matrix as is shown in Fig. 16.7b. Each Functional Unit (FU) operates on a 3  3
mask in agreement with the scheme of Fig. 16.7a. The data are stored in the input
registers .R3; R6; R9; : : :/ and they move in each clock cycle to their interconnected
neighbours registers. In the third clock cycle the mask registers contain the data of
the corresponding image pixels. Then the functional units operate with the mask
data and generate the outputs. In each clock cycle the mask advances a column in
the image. Pixels enter by the right and shift from one stage to another outgoing at
the left side. It is a systolic architecture that follows a linear topology and allows
processing several pixels in parallel.
The system receives two input data of 32 bits (these mean eight pixels). These
data come from a double port memory that stores the image. The circuit also receives
as input data the threshold (T) that has been calculated previously. The circuit
192 N.M. Hussein, A. Barriga

DinR1
DinR2
DinR3
DinR4
DinR5
DinR6
DinR7
DinR8
DinR9
1

BW4
BW5
BW6
BW7
BW8
BW9
1
8

8 Dout
Lukasiewicz 0
“0000”DinR1(7:3)
9
Filter ...
“0000”DinR9(7:3)
+ carry

Thresholding b)
Logic
BW3
1 1
Dout
Pipeline
Register 1 8
BW2 Din 0 0

Pipeline
8 <
Th
Register 2
BW1
c)

Ege
Detection BW1-BW9 Dout

Dout

a) d)

Fig. 16.8 (a) Functional Unit (FU) circuit schematic, (b) Lukasiewicz filter, (c) thresholding logic
circuit, (d) edge detection circuit

generates as output the 4 bits corresponding to the output values of the processed
pixels stored in R5, R8, R11 and R14.
The functional unit operates on the 3  3 mask and generates the output value
corresponding to the evaluated element of the mask. A block diagram of a functional
unit is shown in Fig. 16.8a. The circuit consists of two pipeline stages so that the data
has a latency of two clock cycles. The first stage is the image filter. Then threshold
T is applied. The edge detector, in the output stage, operates on the binary mask
(black and white image).
Figure 16.8 shows the circuits corresponding to the different blocks of the func-
tional unit (FU). As we can observe in Fig. 16.8b the filter based on Lukasiewicz’s
bounded sum receives the data stored in registers R1 to R9. These data are scaled
by the factor 0.125 entailing division by 8, which signifies a displacement of three
places to the left. The sum of the data is compared (using the carry as control signal)
with value 1. The threshold logic (Fig.16.8c) compares the pixel with the threshold
value. The output is a binary image (black and white) and only therefore requires
one bit. Finally, the edge detection circuit receives a 3  3 binary image. It carries
out the xor operation of the bits. If all the bits of the mask are equal the output pixel
is in the background, whereas if some bit is different the output is an edge pixel.
Figure 16.9 shows the chronogram of the circuit. It can be observed that the
operation of the system begins with the falling edge of signal CS. Whenever a row
begins two clock cycles are needed to initialize the mask registers. In the third clock
16 High Speed Soft Computing Based Circuit for Edges Detection in Images 193

Fig. 16.9 Chronogram of the circuit

cycle Dvalid signal take value 1, indicating a valid output. Input data are provided
in each clock cycle. Once Dvalid has been activated the output data in the following
cycles is also valid since Dvalid = 1 (data of the following columns being processed
in successive cycles).
The circuit occupies an area of 293 slices on an FPGA of the Spartan3 Xilinx
family that supposes a cost of 5,361 equivalent gates. It can operate at a 77 MHz
frequency, although the development board used has a 50 MHz clock. This supposes
that the required time to process an image is 0.335 ms what allows to process 2,985
images per second.

16.4 Conclusion

The hardware implementation of an edge extraction system has been described. The
edge detection algorithm is based on soft computing technique. Thus the calculation
of the threshold that allows to obtain a binary image is realized by means of a fuzzy
logic inference mechanism. The conditioning of the image for its later processing is
based on a filter that uses Lukasiewicz’s bounded sum. The main advantage of this
edge detection technique is that it allows a very efficient hardware implementation
in terms of cost and speed. This makes it specially indicated in applications that
require a real time processing.

References

1. J. F. Canny, A computational approach to edge detection, IEEE Transactions on Pattern

Analysis and Machine Intelligence, 679–698 (1986).
2. L. G. Roberts, Machine perception of three-dimensional solids, in J. T. Tippet et al., editor,
Optical and Electro-Optical Information Processing (MIT, Cambridge, MA, 1965), pp. 159–
197.
194 N.M. Hussein, A. Barriga

3. J. M. S. Prewitt, Object enhancement and extraction, in A. Rosenfeld and B. S. Lipkin, editors,

Picture Processing and Psychophysics (Academic, New York, 1970), pp. 75–149.
4. D. Marr and E. Hildreth, Theory of Edge Detection, Proceedings of the Royal Society London,
207, pp. 187–217 (1980).
5. A. Barriga and N. M. Hussein, Design and implementation of an edge detection circuit based
on soft computing, Communications of SIWN, vol. 3, pp. 135–139 (2008).
6. L. A. Zadeh, Fuzzy Sets, Information and Control, vol. 8, pp. 338–353 (1965).
7. A. Barriga, S. Sánchez Solano, P. Brox, A. Cabrera, and I. Baturone, Modelling and Implemen-
tation of Fuzzy Systems based on VHDL, International Journal of Approximate Reasoning,
vol. 41, issue 2, pp. 164–278 (2006).
Chapter 17
A Comparison Between 3D OSEM and FBP
Image Reconstruction Algorithms in SPECT

Khalid S. Alzimami, Salem A. Sassi, and Nicholas M. Spyrou

Abstract In this chapter, we will only concern our selves with the two most widely
used algorithms in SPECT reconstruction, namely the FBP and Ordered-Subsets
Expectation Maximization (OSEM) algorithms. This chapter describes the basic
principles of these two algorithms, and summarises SPECT image attenuation and
scatter correction methods. Finally, it presents a study evaluating the volumetric iter-
ative reconstruction algorithm OSEM 3-D, and compares its performance with the
FBP algorithm.

Keywords Image Reconstruction  Single Photon Emission Computed Tomogra-

phy  Filtered Back Projection  ordered-subsets expectation maximization  image
attenuation

17.1 Introduction

Single Photon Emission Computed Tomography (SPECT) is a nuclear medicine

imaging technique used to visualize the distribution of radioactive tracer uptake in
the patient [1]. In SPECT, the radio-pharmaceutical emits single gamma () ray
photons isotropically. A gamma camera system is used to acquire a 2-D image of
the distribution of the radiotracer uptake within the object of interest. These images
often called views or projections and are acquired at discrete angles around the
patient [2].
Mathematical algorithms are then used to reconstruct 3-D images of selected
planes within the object from the 2-D projection data. This process is known as
Emission Computed Tomography (ECT) [3]. Images produced by SPECT result

K.S. Alzimami (B)

Department of Physics, University of Surrey, Guildford, Surrey GU2 7XH, UK,
E-mail: K.Alzimami@surrey.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 195
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
196 K.S. Alzimami et al.

in a significant improvement of object contrast, and can allow more accurate

quantification of radiotracer uptake at specific locations within the body [1].
Even though, the first tomographic imaging systems used iterative reconstruc-
tion methods, the analytical Filtered Back Projection (FBP) method is still the
most widely used method in the reconstruction of SPECT images. However, with
the rapid increase in the processing speed of modern computers, interest in the
development and implementation of iterative reconstruction methods in emission
tomography modalities has grown rapidly and a plethora of iterative reconstruction
algorithms have been developed. These include algebraic methods like the algebraic
reconstruction technique (ART) [4], simultaneous iterative reconstruction technique
(SIRT) and iterative least-squares technique (ILST), as well as Statistical recon-
struction methods like maximum likelihood expectation maximization (MLEM)
[5], ordered-subsets expectation maximization (OSEM) [6], maximum likelihood
algorithm (RAMLA), maximum a posteriori (MAP) algorithms and gradient and
conjugate gradient (CG) algorithms [7].
Many of these methods have been implemented in SPECT imaging systems. In
this chapter, however, we will only concern our selves with the two most widely
used algorithms in SPECT reconstruction, namely the FBP and Ordered-Subsets
Expectation Maximization (OSEM) algorithms. This chapter describes the basic
principles of these two algorithms, and summarises SPECT image attenuation and
scatter correction methods. Finally, it presents a study evaluating the volumetric iter-
ative reconstruction algorithm OSEM 3-D, and compares its performance with the
FBP algorithm for the reconstruction of 99 Tcm SPECT imaging.

17.2 Background

The purpose of tomographic reconstruction is to obtain axial cross-sections of an

object from projections of that object. In 1917 Radon showed the reconstruction of
a 2-D object using an infinite number of its 1-D projections [8]. Modified realiza-
tions of this method have been adopted to reconstruct 3-D SPECT images from the
acquired 2-D projections. These methods have been used widely in SPECT image
reconstruction due to their computational speed and simple implementation.
Historically, all image reconstruction methods have aimed to reduce the problem
of 3-D image reconstruction into a 2-D problem by dividing the 2-D projection data
into 1-D profiles to produce cross-sectional images of the object. These algorithms
are often referred to as 2-D reconstruction algorithms. Recently, due to the massive
increase in processing speeds of modern computers and the availability of cheap
high capacity memory chips, a new generation of image reconstruction algorithms
has evolved. These algorithms have permitted the full 3-D reconstruction of the
projection data.
In order to correct the SPECT images accurately for photon crosstalk between
trans-axial slices, one needs fully 3D reconstruction. Unlike 2-D reconstruction,
slice by slice, full 3-D reconstruction uses a large matrix making it possible for the
17 A Comparison Between 3D OSEM and FBP Image Reconstruction 197

photons that are detected in the out-of-slice projection pixels to be accounted for [9].
Although the computational burden of full 3-D iterative reconstruction is somewhat
high, the accuracy is significantly improved and there are noteworthy improvements
of signal-to-noise ratio (SNR) in comparison to 2D reconstruction [10]. A detailed
description of the 3-D reconstruction problem is given in [9, 11, 12].
In general, image reconstruction in SPECT can be divided into two main
approaches. The first approach is based on direct analytical methods, while the
second is based on algebraic and statistical criteria and iterative algorithms.

17.2.1 Analytical Reconstruction Algorithms: Filtered

Backprojection

Filtered back projection (FBP) is the most commonly used reconstruction algorithm
in SPECT imaging. This is principally due to the simple concept of the algorithm
and relatively quick processing time. FBP can be viewed as a two-step process:
filtering of the data and back projection of the filtered projection data. There is
extensive literature on these reconstruction algorithms and reviews of FBP and its
applications to SPECT [1, 4, 13, 14].

17.2.1.1 Filtering

3-D SPECT images are created by back projecting planar 2-D views of the object
into image space along the line of response [15]. This process will result in blurred
images and reconstruction star like artefacts. To overcome this problem, a ramp
filter is applied to the projection data, which in turn increases the magnitude of
the high frequency components of the image including statistical noise. Further fil-
tration is therefore required to temper this effect. This can be done by applying
an image smoothing kernel, but because convolution is a computationally inten-
sive task, the data is first transferred to the frequency domain and projection data
filtration is performed in the frequency (Fourier) domain [1, 2].
There are basically three types of reconstruction filters, these can be categorised
as low pass, high pass and band pass filters. While high pass filters allow high fre-
quency information including image noise to pass, these usually referred to as image
enhancement filters. Band-pass filter allows only data with frequencies within the
pass band to be retained, while suppressing or eliminating all other frequencies
[15, 16].
Reconstruction filters can be described by two important parameters: the “cut-
off frequency” and “order or power” of the filter function. The cut-off frequency
determines the filter rolling off to an infinitely low gain whereas the shape of the
curve is determined by the order of the filter function. The location of the cut-off
frequency determines how the filter will affect both image noise and resolution.
198 K.S. Alzimami et al.

17.2.1.2 Back Projection

The main reconstruction step involves the backprojection of acquired data into the
image domain. For simplicity, the position of photon detection within the gamma
camera head is assumed to be perpendicular the location photon emission. So by
smearing back the number of counts at each point in projection profiles into the
image space, a 3-D image of radioactive tracer distribution can be built. Regions
with higher concentration of back projected lines (ray sums) and greater number of
counts form a reconstructed image of radiotracer concentration within the object.
This process is done in spatial (image) domain unlike image filtration, which is
normally carried out in the frequency domain [1, 13, 14].

17.2.2 Iterative Reconstruction Algorithms

There is increasing recognition that iterative reconstruction plays a key role in

improving the quality of reconstructed images and may improve the accuracy of
SPECT image quantification, particularly where attenuation is non homogeneous
or where a more exact model of the emission and detection processes is required
[17,18]. Iterative reconstruction algorithms, such as OSEM have become a clinically
practical alternative to FBP [19].
The iterative reconstruction algorithm begins with a preliminary estimate of the
object source distribution. This is done by assuming a homogenous object distri-
bution or created using a single measured projection. A set of projection data is
estimated from the initial estimate using a mathematical projector that models the
imaging process. Differences between estimated projection data and measured pro-
jection data are calculated for every projection angle. Using an algorithm derived
from specific statistical criteria, these differences are used to update the initial image
estimate. The updated image estimate is then used to recalculate a new set of esti-
mated projection data that are again compared with the measured projection data.
The number of iterations is either limited by the user or is controlled by pre-specified
conversion criteria [9].
The iterative reconstruction techniques can be classified into three groups, defined
mainly by the underlying assumptions regarding the nature of the data, whether
statistical or non statistical. These are:
Algebraic reconstruction algorithms
Statistical algorithms assuming Gaussian noise and
Statistical algorithms assuming Poisson noise

17.2.2.1 Ordered Subsets Expectation Maximization (OSEM)

The ML-EM algorithm has proven to be effective, but also to be very slow for daily
use. Splitting up the measured datasets into different subsets and using only one
17 A Comparison Between 3D OSEM and FBP Image Reconstruction 199

subset for each iteration speeds up the algorithm by a factor equal to the number
of subsets. The method is known as Ordered Subsets Expectation Maximization
(OSEM) [6].
In SPECT, the sequential processing of ordered subsets is natural, as projection
data is collected separately for each projection angle. In other words, counts on
single projections can form successive subsets. With OSEM, the careful selection
of subset ordering can greatly accelerate convergence. However, the lack of con-
vergence of OSEM algorithms is of theoretical importance since in practice, the
algorithm is terminated after only a few iterations [20].
In recent years many attempts have been made to improve the quality of the
3D OSEM algorithms by incorporating corrections for the major image-degrading
factors in the projection and backprojection operations of the iterative steps. Specif-
ically, several commercial algorithms (such as Flash 3-D by Siemens [21], Astonish
3-D by Philips [22], HOSEM by HERMES [23] and Wide-Beam Reconstruction
by UltraSPECT [24]) have been developed to improve the SNR of the image by
modelling Poisson noise that results from low counts as well as implementing reso-
lution recovery algorithms to restore resolution degradation due to collimator spread
function and source detector distance in the case of WBR.

17.2.3 Physical Factors Affecting Quantitative and Qualitative

Accuracy

Several physical factors, the most important being photon attenuation, Comp-
ton scatter and spatially varying collimator response, degrade the qualitative and
quantitative accuracy of SPECT images. Theoretically, the number of counts in
reconstructed SPECT images should be directly proportional to the absolute con-
centration and distribution of the radiotracer within the imaged object. It has been
reported that the presence of scatter and attenuation in the images limits the accu-
racy of quantification of activity. The uncertainty could be as high as 50–100% if
the scatter and attenuation are not corrected [25].

17.2.3.1 Attenuation Correction

Attenuation causes inconsistent projection data that can increase or decrease counts
in the image, especially near hot regions. It also causes shadows or regions of dimin-
ished activity, within the patient. Thus, attenuation may potentially affect tumour
detectablility or artificially enhance a noise blob [26].
AC methods can be classified as: (a) constant attenuation coefficient (), known
as the Chang method [27]; or (b) variable , or transmission source method. The
most widely used AC method is the Chang method, but it can only be used in the
brain or abdomen regions as they can be considered essentially uniform. Obviously,
the Chang method will only work well where  is, in fact, approximately constant.
200 K.S. Alzimami et al.

However, this is not the case when the attenuating medium is non-uniform (e.g.
cardiac studies). Therefore, AC requires a variable attenuation coefficient dependent
on the spatial location of the pixel in the object (i.e. patient or phantom) [28].

17.2.3.2 Scatter Correction

Scattered photons limit resolution and act as a modifier to the effective attenuation
coefficients in a depth dependent way. Moreover, the presence of scattered photons
in SPECT projection data leads to reduced contrast and loss of the accuracy of
quantification in reconstructed images [25].
In order to compensate for the effects of scatter, several scatter compensation
techniques have been developed. These techniques can be categorised into two
main groups; subtraction based and reconstruction based scatter correction. In the
subtraction-based approach the scatter component is estimated and subtracted from
the projection data prior to reconstruction. Typically the scatter estimate is produced
from multiple energy window acquisition. Kadrmas et al. have illustrated the advan-
tages of using multiple energy windows when imaging isotopes that have multiple
emission peaks [29].
In the reconstruction-based approach, the scatter is incorporated in the algorithm
model. The main advantage of this approach is that the noise increase associated
with the scatter subtraction method is avoided. This is because there is no explicit
subtraction of scatter counts. A number of studies have shown that iterative recon-
struction with accurate modelling of scatter is superior to pre-reconstruction scatter
subtraction [25]. However, the major setback to model based reconstruction is that
it is not obvious how the scatter originating from radio-nuclides in object regions
out of the field of view (FOV) sources is compensated [10].

17.2.3.3 Distance-Dependent Resolution Compensation

The distance-dependent acceptance of photons through the holes of the collima-

tor is the main contributor to the degradation of the spatial resolution in SPECT.
3D modelling of the collimator line spread function (LSF) utilizing the collimator
parameters and assumed distance between collimator and source is used to compen-
sate for collimator and distance dependent SPECT image blurring. This resolution
recovery is achieved by applying stationary or non-stationary deconvolution filters
to the acquired projection data prior to image reconstruction or to the reconstructed
data. Deconvolution methods using Wiener, Metz, or power spectrum equalisation
filtering have been used in 2-D by ignoring the interslice blurring, but the 3-D filter-
ing gives better results. These methods have been used mainly in conjunction with
the FBP reconstruction technique [30].
In an Alternative approach, a point spread function (PSF) of the camera system is
modelled at each point of the reconstructed area for each camera projection, and is
included into the projection/back-projection operator in the iterative reconstruction
17 A Comparison Between 3D OSEM and FBP Image Reconstruction 201

process. This modelling may be made in two dimensions, by ignoring the interslice
blur. However, for best results calculations should be done in 3-D [30, 31].
Both deconvolution and modelling methods improve contrast of the recon-
structed images. However, the main drawback of deconvolution methods is that
they may also substantially increase high frequency noise. To compensate for this
consequence one needs to apply a low pass filter, which, in turn, degrades image
resolution. The modelling method on the other hand has a regularising effect on the
reconstructed image. Compared to reconstruction without blurring compensation,
high frequency noise is reduced, while contrast and resolution are improved [30,31].

17.3 A Comparison of 3D OSEM with FBP Image

Reconstruction Algorithms

17.3.1 Methods

The performance of 3-D OSEM, with and without AC, was characterized with
respect to subset and iteration number for a constant pixel size and the same amount
of post reconstruction filtering. Due to space constraints, more details about methods
and data analysis are given in [32].

17.3.2 Results and Discussions

As shown in Fig. 17.1, for the same number of subsets, there is a linear relation-
ship between noise termed as variation coefficient and the number of iterations. The
graph also shows a linear relationship between noise and number of subsets for the
same number of iterations.
This linear relationship (R2 1) leads to a predictable and accurate characteri-
zation of noise. Also, Fig. 17.1 shows that the effect of subset and iteration number
is additive over noise, in accordance with OSEM theory [33].
As shown in Fig. 17.2, the number of iterations needed to reach a constant
FWHM was approximately 17 for the periphery and 22 for the centre of the FOV.
These results agree with those reported by Yokoi et al. and Kohli et al. [34, 35].
Kohli et al. concluded that the reason for the slow convergence in the central region
was the effect of attenuation [35]. However this is not right as simulations by Yokoi
et al. assumed no attenuating medium for a point source [34]. According to Yokoi
et al. the reason of the slow convergence is that the SPECT resolution is originally
better at the periphery than at the centre [34]. Another reasonable explanation of the
slow convergence could be due to the fact that many more ray sums contribute to
the formation of objects located at the centre of the image than in the periphery [1].
202 K.S. Alzimami et al.

24
23 1 4 8 16 32
22 Subsets Subsets Subsets Subsets Subsets
21 R2 = 0.9146
20
Noise

19
18
17
16
15

24

12

24
8

12

1
1

24

NO. of Iteration

Fig. 17.1 Relationship between noise and number of iteration at different number of subsets

S=4 (P) S=16 (P)

S=4 (C) S=16 (C)
31

26
FWHM(mm)

21

16

11

6
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
No of Iterations

Fig. 17.2 Variation of the spatial resolution (FVHM) with the number of iterations for a range of
subsets at the centre (C) and the periphery (P)

Figure 17.3 demonstrates a statistically significant (P < 0:01) improvement in

spatial resolution using 3-D OSEM compared to FBP. This could be attributed to
the resolution recovery properties of Flash 3D [21].
Figure 17.4 illustrates that 3D OSEM is the best choice for low count statis-
tics study as 3D OSEM is significantly (P < 0:01) better than FBP in terms of
noise. This could be due to the fact that IA takes the Poisson nature of the data
into account, and unlike conventional 1-D or 2-D reconstruction methods like FBP
and 2-D OSEM respectively, OSEM 3-D allows the reconstruction of all slices
simultaneously, thus improving count statists [9].
As shown in Fig. 17.5, there was no significant difference (P > 0:01) in image
contrast between FBP and 3D OSEM without applying AC. However, with AC 3-D
OSEM image contrast has improved significantly in comparison to FBP with AC.
This means that 3D OSEM with AC may extend the range of detectable lesion
sizes and contrast levels, and may provide clarity in situations in which FBP is
17 A Comparison Between 3D OSEM and FBP Image Reconstruction 203

3D OSEM, 8 subsets FBP, Wiener, Order=20

15
14
FWHM (mm)

13
12
11
10
9
8
0.5/4 0.6/8 0.7/12 0.8/25 0.9/30
Cut-off Frequency/Iteration NO

Fig. 17.3 Spatial resolution versus number of iteration in 3D OSEM with eight subsets and cut-off
frequency in FBP with Wiener filter with order of 20

FBP (Metz, 10sec, Order = 15) OSEM(10sec, 8 subsets)

FBP (Metz, 30sec, Order = 15) OSEM(30sec, 8 subsets)
30
28
26
24
22
Noise

20
18
16
14
12
10
0 0.4/1 0.5/4 0.6/8 0.7/12 0.8/24
Cutoff Frequency/Iteration NO.

Fig. 17.4 Estimated noise versus number of iteration in 3D OSEM and cut-off frequency in FBP
with Metz filter with order of 15 for 10 s per projection angle

FBP, Butterworth (Order = 15) OSEM (Subsets = 8)

FBP, Butterworth (Order = 15) with AC OSEM (Subsets = 8) with AC
16
14
12
Contrast

10
8
6
4
2
0
0.4 / 1 0.5 / 4 0.6 / 8 0.7 / 12 0.8 / 25
Cut-off Frequency / Iteration NO.

Fig. 17.5 Measured contrast with and without AC versus number of iteration in 3D OSEM and
cut-off frequency in FBP with Butterworth filter with order of 15
204 K.S. Alzimami et al.

Fig. 17.6 Reconstructed images of bone scan (above): (a) using FBP – Wiener filter (Cut-off:
0.7/Order: 15); and (b) using 3D OSEM (Subsets: 8/Iterations: 16) and reconstructed images of
adrenal scan (bottom); (c) using FBP – Metz filter (Cut-off: 0.7/Order: 15); and (d) using 3D
OSEM (Subsets: 8/Iterations: 12)

ambiguous. It must be acknowledged that Chang AC is only approximate; there-

fore, the benefit gained by the AC may not be as great as could be gained with a
more accurate correction method such as one based on a CT scan.
To demonstrate the performance of 3-D OSEM vs. FBP in clinical studies,
Fig. 17.6 shows SPECT bone and adrenal clinical scans reconstructed with 3D
OSEM and FBP with an optimised set of parameters. Visual assessment of these
images suggests that the superiority of 3D OSEM in terms of spatial resolution and
contrast.

17.4 Summary

We have described the basic principles of the two most commonly used reconstruc-
tion algorithms in SPECT imaging: FBP and OSEM, and presented a brief review
of current compensation methods for the major image-degrading effects. Iterative
image reconstruction methods allow the incorporation of more accurate imaging
models rather than the simplistic Radon model assumed in the FBP algorithm.
These include scatter and attenuation corrections as well as collimator and distance
response and more realistic statistical noise models. In our study we have demon-
strated the superior performance of 3-D OSEM compared to FBP particularly for
low count statistics studies, including improved image contrast and spatial resolu-
tion. Using 3-D OSEM with suitable AC may improve lesion detectability due to
the significant improvement of image contrast. Indeed, 3-D iterative reconstruction
algorithms are likely to replace the FBP technique for most SPECT and PET clinical
17 A Comparison Between 3D OSEM and FBP Image Reconstruction 205

applications. However, for the full potential of these methods to be fully realised,
more exact image compensation methods need to be developed and optimal image
reconstruction parameters need to be used. The full impact of these methods on
quantitative SPECT imaging is yet to be assessed. Finally, with development of new
faster and more accurate 3-D and possibly 4D iterative algorithms, the future of
SPECT image reconstruction is certain to be based on iterative techniques rather
than any analytical methods.

References

1. S. R. Cherry, J. A. Sorenson and M. E. Phelps, Physics in Nuclear Medicine (Saunders,

Philadelphia, PA, 2003).
2. H. H. Barrett, Perspectives on SPECT. SPIE 671, 178–183 (1986).
3. K. Kouris, N. M. Spyrou and D. F. Jackson, Imaging with Ionizing Radiation (Surrey
University Press, London, 1982).
4. G. T. Herman, Image Reconstruction from Projections: the Fundamental of Computerised
Tomography (Academic, New York, 1980).
5. L. A. Shepp and Y. Vardi, Maximum likelihood reconstruction for emission tomography, IEEE
Transactions on Medical Imaging 1(2), 113–122 (1982).
6. H. M. Hudson and R. S. Larkin, Accelerated image reconstruction using ordered subsets of
projection data, IEEE Transactions on Medical Imaging 13(4), 601–609 (1994).
7. B. F. Hutton, J. Nuyts and H. Zaidi, Iterative Reconstruction Methods, edited by H. Zaidi
(Springer, New York, 2005) pp. 107–140.
8. J. Radon, Uber die bestimmung von funktionen durch ihre integral-werte langs gewisser
mannigfaltigkeiten, Ber Verh Sachs Akad Wiss 67, 226 (1917).
9. P. Gilbert, Iterative methods for the three-dimensional reconstruction of an object from
projections. Journal of Theoretical Biology 36(1), 105–117 (1972).
10. F. J. Beekman, W. A. M. de Jong, and S. van Geloven, Efficient Fully 3-D Iterative SPECT
reconstruction with Monte Carlo-based scatter compensation, IEEE Transactions on Medical
Imaging 21(8), 867–877 (2002).
11. M. Defrise, D. W. Townsend, R. Clack, Three-dimensional image reconstruction from
complete projections, Physics in Medicine and Biology 34(5), 573–587 (1987).
12. D. R. Gilland, R. J. Jaszcak, T. Riauka, R. E. Coleman., Approximate 3D iterative reconstruc-
tion for SPECT, Medical Physics 24(9), 1421–1429 (1997).
13. M. W. Tsui and E. C. Frey, Analytical Image Reconstruction Methods in Emisssion Computed
Tomography, edited by H. Zaidi (Springer, New York, 2005) pp. 82–106.
14. G. L. Zeng, Image reconstruction: a tutorial, Computerized Medical Imaging and Graphics
25(2), 97–103 (2001).
15. W. R. Hendee and E. R. Ritenour, Medical Imaging Physics (Wiley, New York, 2002).
16. D. R. Gilland, B. M. W. Tsui, W. H. McCartney, J. R. Perry, and J. Berg, Determination of
the optimum filter fuction for SPECT imaging, Journal of Nuclear Medicine 29(5), 643–650
(1988).
17. S. Vandenberghe, Y. D’Asseler, R. Van de Walle, T. Kauppinen, M. Koole, L. Bouwens, K. Van
Laere, I. Lemahieu and R. A. Dierckx, Iterative reconstruction algorithms in nuclear medicine,
Computerized Medical Imaging and Graphics 25, 105–111 (2001).
18. P. P. Bruyant, Analytic and iterative reconstruction algorithms in SPECT, Journal of Nuclear
Medicine 43(10), 1343–1358 (2002).
19. W. Koch, C. Hamann, J. Welsch, G. Pöpperl, P. E. Radau, and K. Tatsch, Is iterative recon-
struction an alternative to filtered backprojection in routine processing of dopamine transporter
SPECT studies?, Journal of Nuclear Medicine 46(11), 1804–1811 (2005).
206 K.S. Alzimami et al.

20. R. L. Byrne, Editorial recent developments in iterative image reconstruction for PET and
SPECT, IEEE Transactions on Medical Imaging 19(4), 257–260 (2000).
21. Flash 3D and e-soft, 2008, Siemens Medical Solutions (September 4, 2008); http://www.
medical.siemens.com/siemens
22. Astonish 3-D, 2008, Philips (September 4, 2008); http://www.medical.philips.com
23. HOSEM, 2008, HERMES (September 4, 2008); http://www.hermesmedical.com
24. WBR, 2008, UltraSPECT (September 4, 2008); http://www.ultraspect.com
25. H. Zaidi, Scatter modeling and compensation in emission totmography. European Journal of
Nuclear Medicine and Molecular Imaging 31(5), 761–782 (2004).
26. T. Kauppinen, M. O. Koskinen, S. Alenius, E. Vanninen and J. T. Kuikka, Improvement of
brain perfusion SPET using iterative reconstruction with scatter and non-uniform attenuation
correction, European Journal of Nuclear Medicine 27(9), 1380–1386 (2000).
27. L. T. Chang, A method for attenuation correction in radionuclide computed tomography, IEEE
Transactions on Nuclear Science NS-25, 638–643 (1978).
28. M. G. Erwin, SPECT in the year 2000: basic principles, Journal of Nuclear Medicine 28(4),
233–244 (2000).
29. D. J. Kadrmas, E. C. Frey, and B. M. W. Tsui, Application of reconstruction-based scatter
compensation to thallium-201 SPECT: implementations for reduced reconstructed image
noise. IEEE Transactions on Medical Imaging 17(3), 325–333 (1998).
30. B. F. Hutton and Y. H. Lau, Application of distance-dependent resolution compensation and
post-reconstruction filtering for myocardial SPECT, Physics in Medicine and Biology 43,
1679–1693 (1998).
31. P. Gantet, P. Payoux, P. Celler, P. Majorel, D. Gourion, D. Noll and J. P. Esquerré, Iterative
three-dimensional expectation maximization restoration of single photon emission computed
tomography images: application in striatal imaging. Medical Physics 33(1), 52–59 (2006).
32. 32. K. Alzimami, S. Sassi, and N. M. Spyrou, Optimization and comparison of 3D-OSEM with
FBP SPECT imaging, The 2008 International Conference of Signal and Image Engineering
Proceeding I, London July 2008, 632–636.
33. M. Brambilla, B. Cannillo, M. Dominietto, L. Leva, C. Secco and E. Inglese, Characterization
of ordered-subsets expectation maximization with 3D post-reconstruction Gauss filtering and
comparison with filtered backprojection in 99m Tc SPECT, Annals of Nuclear Medicine 19(2),
75–82 (2005).
34. T. Yokoi, H. Shinohara and H. Onishi, Performance evaluation of OSEM reconstruction algo-
rithm incorporating three-dimensional distance-dependent resolution compensation for brain
SPECT: A simulation study, Annals of Nuclear Medicine 16(1), 11–18 (2002).
35. V. Kohli, M. King, S. J. Glick, T. S. Pan, Comparison of frequency-distance relationship and
Gaussian-diffusionbased method of compensation for distance-dependent spatial resolution in
SPECT imaging, Physics in Medicine and Biology 43, 1025–1037 (1998).
Chapter 18
Performance Improvement of Wireless MAC
Using Non-Cooperative Games

Debarshi Kumar Sanyal, Matangini Chattopadhyay, and

Samiran Chattopadhyay

Abstract This chapter illustrates the use of non-cooperative games to optimize the
performance of protocols for contention-prone shared medium access in wireless
networks. Specifically, it proposes and analyses a set of new utility functions con-
forming to a recently proposed generic game-theoretic model of contention control
in wireless networks [5]. The functions ensure that the game has a unique non-trivial
Nash equilibrium. Simulations carried out for IEEE 802.11 style networks indicate
that the aggregate throughput of the network at the Medium Access Control (MAC)
layer has improved while the collision overhead is reduced.

Keywords Wireless networks  medium access control  game theory  nash equilib-
rium  throughput

18.1 Introduction

In most wireless networks, including ad-hoc and sensor networks, access to the
shared wireless medium is random-access based. In absence of a central regulat-
ing authority, it is only natural that multiple wireless nodes will attempt to access
the medium simultaneously, resulting in packet collisions. IEEE 802.11 (here-
after briefly called 802.11) is the most commonly followed standard in wireless
networks. It defines a distributed mechanism called the distributed coordination
function (DCF) to resolve contention among the contending nodes. It involves chan-
nel sensing to assess the state of the shared medium and adjusting the channel access
probability accordingly to minimize chances of collision. In 802.11, the channel

D.K. Sanyal (B)

Senior Member of Technical Staff, Interra Systems (India) Pvt. Ltd., Salt Lake Electronics Com-
plex, DN-53, Kolkata-700 091, India (West Bengal)
E-mail: debarshi@cal.interrasystems.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 207
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
208 D.K. Sanyal et al.

access probability is determined by a contention window (CW) maintained by a

node. DCF uses a binary feedback signal (collision or no collision) and sharply
updates CW using binary exponential backoff to adapt to the contention.
However, various studies [1–4] report that this contention control scheme results
in drastic reduction in throughput with increasing number of contending nodes.
Every new transmission begins with the same channel access probability and thus
the history of contention level in the network is not used. Moreover, doubling
the contention window for every failed transmission only reduces the chances of
transmission by the nodes that were already unsuccessful.
These inferences indicate that instead of using drastic update strategies, it is bet-
ter to steadily guide the system to a state that is optimal with respect to the current
contention level in the network. Since this state is optimal, it is sustainable and has
high throughput, greater fairness and sparing collisions. This requires an analyti-
cal framework where these desirable features can be mathematically characterized.
Towards this end, we follow [5–8] and model the nodes as selfish players in a non-
cooperative game [9]. Unlike common studies [10], [11], we do not reverse-engineer
DCF as a game but use game theory as a tool for optimization. We define util-
ity functions reflecting the gain from channel access and the loss from collision.
The strategy set of each node is the set of allowable channel access probabilities
of the node. We next use gentle update strategies that drive the network to its Nash
equilibrium from which no player has an incentive to unilaterally deviate. This char-
acterizes the desired stable operating point or the optimal state of the network. A
continuous feedback signal is used to assess the contention level.
The main contribution of this work is to propose new utility functions that allow
a large range of channel access probabilities (thus ensuring good performance in
a wide spectrum of network sizes) and result in a unique non-trivial Nash equilib-
rium. Non-triviality and uniqueness ensure that the equilibrium is efficient and leads
to high short-term fairness. As we show through extensive simulations, the resulting
design is able to provide far better results than DCF used in 802.11. Throughput
gets higher, and collision overhead and time slots wasted in collisions are drasti-
cally reduced in the new design. The nodes do not need to know the total number of
nodes in the network nor is any message passing necessary. A completely distributed
mechanism that allows each node to behave selfishly to maximize its own payoff
is used to ensure a socially beneficial state. The chapter is organized as follows:
Section 18.2 provides a background of this study, journeying through the recent
past of game-theoretic models of medium access control in wireless networks. We
also point out the differentiation and novelty of our work as regards related stud-
ies in the literature. Section 18.3 presents the proposed game model and its many
properties. Section 18.4 investigates distributed mechanisms to achieve the equi-
librium of the game in real networks. The model is evaluated via simulations in
Section 18.5.
After a brief discussion in Section 18.6 conclusions follow in Section 18.7.
18 Performance Improvement of Wireless MAC Using Non-Cooperative Games 209

18.2 Background

Game theory [9] provides a tool to study situations in which there is a set of rational
decision makers who take specific actions that have mutual, possibly conflicting,
consequences. A game models an interaction among parties (called the players)
who are rational decision makers and have possibly conflicting objectives. Each
player has a set of actions called its strategies and with each strategy is associated a
payoff function. Rationality demands each player maximize its own payoff function.
Non-cooperative games are commonly used to model problems in medium access
control in telecommunication networks. In such a game, the solution concept is a
notion called a stable set or Nash equilibrium that identifies a set of strategies for
all the participating players, from which no player has an incentive to unilaterally
deviate as any unilateral deviation will not result in an increase in payoff of the
deviating player.
In [12] the authors model the nodes in an Aloha network as selfish players in
a non-cooperative Aloha game. They assume that the number of backlogged users
is always globally known. This limitation is addressed in [13] where only the total
number of users is known. More recent research includes modeling the DCF in
802.11 in a game-theoretic framework [11], and reverse engineering exponential
backoff as a strategy [10] of non-cooperative games.
In contradistinction to these works, we do not explicitly reverse-engineer 802.11
or DCF into a game model but use ideas from non-cooperative game theory to
optimize the performance of 802.11. We consider each node in a wireless net-
work as having a set of strategies which are its channel access probabilities. Unlike
common practices in optimization of 802.11 like [4], we do not assume that each
node knows the number of users in the network. This is more reflective of prac-
tical situations where such knowledge is difficult to acquire. Non game-theoretic
approaches that do not depend on the nodes’ knowing the network size include the
ones presented in [2] and [3]. We use selfishness to achieve optimized system-wide
performance. In [5, 6, 8] the authors use game theory to study and optimize 802.11.
Although they remain our motivation, their illustrative utility functions are differ-
ent from ours. Their utility functions are well-behaved in a much more restricted
domain than ours. Moreover, as we show, our utility functions give far better results
for large network sizes. Supermodular games for contention control are presented
in [7].

18.3 Game-Theoretic Model of Medium Access Control

The system we consider is a network of N nodes that are all able to hear one another.
To facilitate analysis we will adopt a description of our access mechanism in terms
of channel access probabilities. It has been shown in [1] that in a saturated regime
the constant channel access probability p relates to the corresponding contention
210 D.K. Sanyal et al.

window cw according to
2
pD (18.1)
cw C 1
Now we turn to the game model. We define the game G as a 3-tuple G D< N ,
fSi gi 2N ; fui gi 2N >, where N is a set of players (wireless nodes), player i 2 N ,
each player having a strategy set Si D fpi jpi 2 Œvi ; wi g with 0 < vi < wi <
1 and payoff ui . The strategy set of each player is the set of its channel access
probabilities. Note that it is a generalization to a continuous space of the simpler
strategy set fwait, transmitg denoting the two deterministic actions that a node can
perform. Denote the strategy profile of all nodes by p D .p1 ; p2 ; : : : ; pjN j /. The
payoff is naturally of the form: ui D Ui .pi /  pi qi .p/. Here Ui .pi / is the utility
function denoting the gain from channel access and qi .p/ is the conditional channel
probabilty given by Y
qi .p/ D 1  .1  pj / (18.2)
j 2N fi g

Thus pi qi .p/ is the cost of channel access.

We propose a novel utility function:
pi .ln wi  ln pi C 1/
Ui .pi / D (18.3)
ln ri
where pi = channel access probability .0 < vi  pi  wi < 1/, ri D wi =vi > 1.
In non-cooperative games, the final outcome is described by investigating its Nash
equilibrium (NE). The strategy profile of all nodes but i is denoted by pi D
.p1 ; p2 ; : : : ; pi 1 ; pi C1 ; : : : ; pjN j /. Then the NE is defined as the strategy profile
p D .p1 ; p2 ; : : : ; pjN 
j / with the property that for every player i and every strat-
egy pi of player i , ui .pi ; pi / ui .pi ; pi /. So no player gains by unilaterally
changing its strategy.
Note in passing that our utility function does not correspond to a physical quan-
tity like throughput, delay, or the like but is a mathematical expression formulated
to satisfy the desirable property of unique non-trivial NE of the game. The rest of
the paper proves these characteristics and shows that it leads to performance far
superior to DCF in contention control.
Theorem 18.1. The game G has an NE.
Proof. The strategy set Si D Œvi ; wi  .0 < vi < wi < 1/ of each player is a non-
empty, convex and compact subset of the Euclidean space. Further ui is continuous.
Since
@2 ui 1
2
D <0
@pi p i ln ri
and pi 2 Si , ui is quasi-concave in pi . Hence the game has an NE.
An NE is non-trivial if the following condition holds for all nodes i 2 N :
0
Ui .pi / D qi .p / (18.4)
and trivial otherwise.
18 Performance Improvement of Wireless MAC Using Non-Cooperative Games 211

Theorem 18.2. The game G has a non-trivial NE.

Proof. Let us define fi .p/ as:

0
fi .p/ D .Ui /1 .qi .p// (18.5)

Since
0 ln.wi =pi /
Ui .pi / D (18.6)
ln ri
is one-to-one in Si , the inverse function
0 qi
.Ui /1 .qi / D wi ri
0
exists, which is continuous and decreasing in qi . Now observe, .Ui /1 .0/ D wi and
0
.Ui /1 .1/ D vi . Thus, fi .p/ maps any p 2 S1  S2      SjN j into a point pi 2
Œvi ; wi . Hence the vector function f .p/ D .f1 .p/; f2 .p/; : : : ; fjN j .p// maps the set
S1  S2      SjN j into itself. Hence, by Brouwer’s fixed point theorem, there is a
fixed point of f .p/ in i 2N Si , that is, fi .p / D pi for each i 2 N . It immediately
follows from the definition of fi .p/ in (18.5) above that at this fixed point, the
condition (18.4) for non-trivial NE holds. In other words, G has a non-trivial NE.
The existence of non-trivial NE signifies that the channel access probability of
any node in NE is not at the boundary of strategy space. This reduces unfair sharing
of the wireless channel among the contending nodes.
wi
Theorem 18.3. The NE is unique for all pi in Œvi ; wi  for all i 2 N if vi > .
e 1=wi 1

Proof. First consider the function:

0
'i .pi / D .1  pi /.1  Ui .pi //

Putting the value of Ui .pi / and differentiating with respect to pi ,

0 1 1 pi
'i .pi / D .  1  ln. //
ln ri pi vi
pi 0
If vi > 1 1 , then, ln vi C . p1i  1/ > ln pi , which means 'i .pi / > 0. Since
e pi
pi wi 0
is strictly increasing in pi , if vi > , we have 'i .pi / > 0. For the
1
pi 1 e 1=wi 1
e
rest of the proof, assume that the above relation between vi and wi holds. Recollect
from condition (18.4) that at non-trivial NE,
0
Y
Ui .pi / D qi .p / D 1  .1  pj /
j 2N fi g

making 'i .pi / identical for all nodes i . Now assume for the sake of contradiction,
that there exist at least two distinct non-trivial NE x and y of the game G. Let
212 D.K. Sanyal et al.

'i .xi / D 1 and 'i .yi / D 2 for some node i . Since 'i is strictly increasing,
1 ¤ 2 . Assume without loss in generality, 1 > 2 . Then xi > yi for all
0
i 2 N . From definition of qi .p/, since qi .p/ is increasing in p, Ui .xi / D qi .x / >
0 00 0
qi .y / D Ui .yi /. But Ui .pi / D 1=.pi ln ri / showing that Ui .pi / is decreasing,
 
thus a contradiction. Hence x D y proving the uniqueness of the NE.

Theorem 18.4. Suppose all the players in G have identical strategy sets. Then at
the unique non-trivial NE of G, all the players choose the same strategies.

Proof. Since all players have the same strategy sets, for any two players i and j ,
wi D wj and vi D vj . Now suppose that the unique non-trivial NE is p . Let
there exist pi and pj in p corresponding to nodes i and j such that pi ¤ pj .
Denote the set of strategies in some order of all nodes except i and j as pij so
that the NE can be described as an ordered set .pij ; pi ; pj /. By symmetry, every
permutation of the NE must be an NE. Interchange the strategies of nodes i and
j to get a new NE .pij ; pj ; pi /. But this contradicts the uniqueness of the NE,
violating Theorem 18.3. Hence, pi D pj . i and j being arbitrary, it follows that
pi D pj for all i; j 2 N .
This has a very important and useful consequence: the wireless channel is
accessed by the nodes with identical access probabilities at the equilibrium point
thus resulting in fair sharing of the resource among the contenders, and consequently
short-term fairness.

18.4 Distributed Mechanisms to Achieve Nash Equilibrium

The non-cooperative game G is played repeatedly by the nodes, that is, it is a multi-
stage game. A stage may be a single transmission or a sequence of K transmissions
for a fixed K > 1.
Suppose that each node after playing a round observes the cumulative effect (in
the sense of conditional collisional probability) of the actions of all players in the
previous round. This knowledge may now be used by the node to select its strategy
for the next round. Based on how this knowledge is used, there are many common
techniques to push the system to its Nash equilibrium. Two more popular ones are:

Best response: This is the most obvious mechanism where at each stage every
player chooses the strategy that maximizes its payoff given the actions of all
other players in the previous round:

pi .t C 1/ D arg max .Ui .pi .t//  pi .t/qi .p.t///

pi .t /2Œvi ;wi 

for each node i 2 N .

Gradient play: In this case, each player adjusts its persistent probability gradu-
ally in a gradient direction suggested by the observations of the effect of other
18 Performance Improvement of Wireless MAC Using Non-Cooperative Games 213

players’ actions. Mathematically,

0
pi .t C 1/ D pi .t/ C fi .pi .t//.Ui .pi .t/  qi .p.t///

for each node i 2 N where step size fi .:/ > 0 can be a function of the strategy
of player i . It has an intuitively appealing interpretation: if at any stage of the
0
game, the marginal utility Ui .pi .t// exceeds the “price of contention” qi .p.t//,
the persistence probability is increased, but if the price is greater, the persistence
probability is reduced.
To eliminate the need to know the current strategies of all other players, a
node needs to estimate qi .p/ in an indirect way. A simple method is to track the
number of packet transmissions (nt) and the number of packet losses (nl) over
a period of time and calculate qi D nl=nt. A more sophisticated approach [3]
involves observing the number of idle slots which follows a geometric distri-
Q I D p =.1  p id le / where p id le D probability of a slot
id le
bution with mean
being idle = i 2N .1  pi /. Thus by observing I , qi can be estimated as qi D
1  .I =.I C 1//=.1  pi / allowing a completely distributed update mechanism.
See [5] for a thorough discussion of the convergence of these procedures.

18.5 Performance Evaluation

In this section, we report results from numerical simulations performed to analyze

the model with our utility functions. The simulation results are obtained in a satu-
ration regime. The performance is compared with the standard DCF in basic access
802.11b. We choose to measure the channel access probabilities, aggregate through-
put of all nodes, collision overhead incurred and the average number of time slots
wasted in collisions per successful transmission.
The saturation throughput Si of a node i in an 802.11 network is given by the
following expression in Bianchi’s model [1]:

pisuc p t rans }
Si D
p id le  C p t rans pisuc  C p col  col

where } = packet payloadQ (assumed constant), p t rans = probability of a transmis-

sion by any node = 1  i 2N .1  pi /, pisuc = conditional success probability
of node i = probabilityQof a successful transmission by the node given that there
p g .1pj /
is a transmission = i j 2Npfi , p id le = probability of a slot being idle =
Q tr
P
i 2N .1  pi /, and finally p D 1  p id le  i 2N pisuc .
col

 is the slot time,  suc is the average duration of a slot in which there is a successful
transmission occurs and  col is the average duration of a slot in which a collision
occurs.
In our experiments, the slot time is 20 μs, SIFS = 10 μs, DIFS = 20 μs, basic
rate = 1 Mbps, data rate = 1 Mbps, data rate = 11 Mbps, propagation delay = 1 μs,
214 D.K. Sanyal et al.

Fig. 18.1 Network size vs. channel access probabilities

PHY header = 24 bytes, MAC header = 34 bytes, ACK = 14 bytes and the packet
payload is fixed at 1,500 bytes. We assume all nodes use the same parameters for
the utility function. To compare with DCF, we set bounds of the contention window
to powers of 2 (specifically 32 and 1,024) and derive the bounds of the strategy sets
using (18.1). This gives wi D 2=33 D 0:06061 and vi D 2=1025 D 0:00195. Note
wi
that in this range, the condition of Theorem 18.3 is satisfied since vi > 1=w i 1
e
1  108 . In the following sub-sections, we call our game-theoretic distributed
coordination function GDCF.

18.5.1 Channel Access Probabilities

Figure 18.1 shows the variation of the probability with which a node accesses the
wireless channel in the saturation regime in GDCF and DCF as the network size
increases. In both protocols, the domain of the probability is the set [0.06061,
0.00195] since contention window varies from 32 to 1,024. Recollect from Sec-
tion 18.3 that the probability with which any given node in GDCF accesses the
channel at the NE is identical for all nodes. We observe that the channel access
probabilities are always observed to be higher in DCF than in GDCF. This has an
important consequence as will be evident in the ensuing discussion.

18.5.2 Aggregate Throughput

The aggregate throughput (see Fig. 18.2) which measures the total throughput of all
nodes under saturation conditions is higher in DCF when the network size is small
but it gets far lower than GDCF as the network size grows. This is because the chan-
nel access probabilities are higher in DCF. In a small network, collisions are few and
18 Performance Improvement of Wireless MAC Using Non-Cooperative Games 215

Fig. 18.2 Variation of aggregate throughput with network size

so higher channel access probabilities translate to higher throughput. But as network

size grows, collisions increase if all nodes aggressively access the channel, trigger-
ing a drastic reduction in throughput. Since channel access probabilities are lower
in GDCF, it supersedes DCF as network size grows. Consequently, large beds of
wireless nodes achieve much greater throughput if they run GDCF instead of DCF.

18.5.3 Collision Overhead

We observe from Fig. 18.3 that the conditional collision probability is higher in case
of DCF than in GDCF. This is expected as the channel access probabilities are lower
in case of GDCF resulting in reduced medium contention even when the number of
nodes increases. Indeed GDCF is designed with the aim of keeping the price of
contention low.

18.5.4 Slots Wasted in Collisions

In 802.11, a large number of slots are wasted due to collisions. This seriously affects
the performance and lifetime of the power-constrained devices in ad-hoc and sensor
networks. Figure 18.4 makes clear that the number of slots wasted in collisions for
every successful transmission is much higher in DCF than in GDCF, making the
later attractive in these networks.
216 D.K. Sanyal et al.

Fig. 18.3 Variation of conditional collision probability with network size

Fig. 18.4 Number of slots wasted in collisions per successful transmission

18 Performance Improvement of Wireless MAC Using Non-Cooperative Games 217

Table 18.1 Brief comparison of utility functions

Number of nodes Throughput Conditional collision probability
GDCF U Chen GDCF U Chen
80 5.26 5.13 0.48 0.56
90 5.15 4.93 0.55 0.59
100 5.06 4.73 0.57 0.63

18.6 Discussion

The motivation for this work is [5]. So we make a brief comparison with it in this
section. Our utility function produces unique non-trivial NE for a wide range of vi
and wi while the one in [5] has more strict constraints on vi once wi is specified.
This makes our utility functions admit a much larger game space and hence is supe-
rior for widely varying network sizes since the channel access probability should be
very small for very large network sizes and high for small network sizes. Table 18.1
presents a performance comparison as regards throughput and collision for large
networks. Here, as in our experiments, we vary contention windows from 32 to
1,024 as is the default for DSSS PHY in 802.11 standard while in [5] (call it U Chen)
the window ceiling is limited to 256 if the base is 32 and only powers of 2 are cho-
sen. As expected, with these settings, our utility function exhibits particularly better
performance for large network sizes.

18.7 Conclusion

This chapter describes an elegant game-theoretic model aimed at optimizing the

aggregate throughput and distributing it fairly among the contending nodes in
an 802.11 network, thus inducing a desirable state of the entire network. It also
drastically brings down the collision overhead, reducing energy loss due to failed
transmissions. We believe it contends as a serious and superior alternative to the
traditional distributed coordination function in 802.11 MAC. The game-theoretic
model uses selfish behavior of the participating players to steer the network to
desirable system-wide characteristics. Our utility functions assure the existence
of unique, non-trivial Nash equilibrium where all nodes possess identical channel
access probabilities. This results in fairly distributed high throughput with low col-
lision overhead. Rigorous analysis and simulations are used to prove the ideas. In
future we intend to simulate the algorithm in settings with bit errors to gather more
performance measures. Further, we intend to investigate other utility functions with
similar characteristics and explore if this design can be used in real networks. We
also intend to analyse the more generic features of the model.
218 D.K. Sanyal et al.

References

1. Bianchi, G. (2000). “Performance analysis of the IEEE 802.11 distributed coordination

function”, IEEE Journal on Selected Areas in Communication, 18(3):535–547.
2. Cali, F., M. Conti, and E. Gregori. (2000). “Dynamic IEEE 802.11: design, modeling and
performance evaluation”, IEEE Journal on Selected Areas in Communication, 18(9): 1774–
1786.
3. Heusse, M., F. Rousseau, R. Guillier, and A. Dula. (Aug 2005). “Idle sense: an optimal access
method for high throughput and fairness in rate-diverse wireless LANs”, Proceedings of ACM
Sigcomm.
4. Toledo, A. L., T. Vercauteren, X. Wang. (Sep, 2006). “Adaptive optimization of IEEE 802.11
DCF based on Bayesian estimation of the number of competing terminals”, IEEE Transactions
on Mobile Computing, 5(9):1283–1296.
5. Chen, L., S. H. Low, and J. C. Doyle. (2006). “Random access game and medium access
control design”, Caltech CDS, Technical Report.
6. Chen, L., S. H. Low, and J. C. Doyle. (2007). “Contention Control: A game-theoretic
approach”, Proceedings of IEEE Conference on Decision and Control.
7. Chen, L., T. Cui, S. H. Low, and J. C. Doyle. (2007). “A game-theoretic model for medium
access control”, Proceedings of WICON.
8. Cui, T., L. Chen, S. H. Low, and J. C. Doyle. (2007). “Game-theoretic framework for medium
access control”, Caltech CDS, Technical Report.
9. Fudenburg, D., and J. Tirole. (1991). “Game Theory”, MIT, Cambridge, MA.
10. Tang, A., J. W. Lee, J. Huang, M. Chiang, and A. R. Calderbank. (2006). “Reverse engineering
MAC”, Proceedings of WiOpt.
11. Xiao, Y., X. Shan, and Y. Ren. (Mar 2005). “Game theory models for IEEE 802.11 DCF in
wireless ad-hoc networks”, IEEE Communications Magazine, 43(3):22–26.
12. MacKenzie, A. B., and S. B. Wicker. (Oct, 2001). “Selfish users in aloha: a game-theoretic
approach”, Proceedings of IEEE Vehicular Technology Conference.
13. Altman, E., R. E. Azouzi, and T. Jimenez. (2002). “Slotted Aloha as a stochastic game with
partial information”, Proceedings of WiOpt.
Chapter 19
Performance Evaluation of Mobile Ad Hoc
Networking Protocols

Nadia Qasim, Fatin Said, and Hamid Aghvami

Abstract In this paper performance evaluations of mobile ad hoc network’s pro-

tocols are made with its quality of service factors. It is seen that mobile ad hoc
networks will be an integral part of next generation networks because of its flex-
ibility, infrastructure less nature, ease of maintenance, auto configuration, self
administration capabilities, and cost effectiveness. This research paper shows com-
parative evaluation within mobile ad hoc networks’ routing protocols from reactive,
proactive and hybrid categories. We have comprehensively analyzed the results of
simulation for mobile ad hoc routing protocols over the performance metrics of
packet delivery ratio, end to end delay, media access delay and throughput for opti-
mized link state routing, temporary ordered routing algorithm and ad hoc on demand
distance vector protocol. In mobile ad hoc networks, mobile nodes must collaborate
with each other in order to interconnect, organize the dynamic topology as mobility
cause route change and establish communication over wireless links. The simulation
results showed the lead of proactive over reactive and hybrid protocols in routing
traffic for dynamic changing topology. Proactive protocol, optimized link state rout-
ing, a protocol for building link tables for ad-hoc networks, can transmit traffic more
rapidly though involve less processing speed in packet forwarding.

Keywords mobile ad hoc network  Performance Evaluation  routing protocol 

Proactive protocol  packet forwarding

19.1 Introduction

Mobile ad hoc networks (MANETs) are rapidly evolving as an important area of

mobile communication. MANETs are infrastructure less and wireless. It has several

N. Qasim (B)
The Centre for Telecommunications Research, King’s College London, Strand,
WC2R 2LS, London, UK,
E-mail: nadia.qasim@kcl.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 219
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
220 N. Qasim et al.

Fig. 19.1 Mobile ad hoc networks

routers, which are free to move arbitrarily and perform management. MANETs as
shown in Fig. 19.1. It has characteristics that network topology changes very rapidly
and unpredictably where many mobile nodes (MNs) can move to and from a wire-
less network without any fixed access point, consequently routers and hosts move,
hence topology is dynamic.
MANETs has to support multi hop paths for mobile nodes to communicate with
each other and can have multiple hops over wireless links; therefore its’ connection
point to the internet may also change. If mobile nodes are within the communi-
cation range of each other than source node can send message to the destination
node otherwise it can send through intermediate node. Now-a-days MANETs have
robust and efficient operation in mobile wireless networks as it can include routing
functionality into MNs which is more than just mobile hosts and reduces the rout-
ing overhead and saves energy for other nodes. Hence, MANETs are very useful
when infrastructure is not available [1], unfeasible, or expensive because it can be
rapidly deployable, without prior planning. Mostly mobile ad hoc networks are used
in military communication by soldiers, automated battlefields, emergency manage-
ment teams to rescue [1], search by police or fire fighters, replacement of fixed
infrastructure in case of earthquake, floods, fire etc., quicker access to patient’s
data from hospital database about record, status, diagnosis during emergency sit-
uations, remote sensors for weather, sports stadiums, mobile offices, electronic
payments from anywhere, voting systems [2], vehicular computing, education sys-
tems with set-up of virtual classrooms, conference meetings, peer to peer file sharing
systems [2].
Major challenges in mobile ad hoc networks are routing of packets with fre-
quently MNs movement, there are resource issues like power and storage and there
are also wireless communication issues. As mobile ad hoc network consists of wire-
less hosts that may move often. Movement of hosts results in a change in routes.
In this paper we have used routing protocols from reactive, proactive and hybrid
categories to make evaluation.
19 Performance Evaluation of Mobile Ad Hoc Networking Protocols 221

19.2 Mobile Ad Hoc Network’s Routing Protocols

Mobile ad hoc network’s routing protocols are characteristically subdivided into

three main categories. These are proactive routing protocols, reactive on-demand
routing protocols and hybrid routing protocols. Each category has many protocols
and some of these protocols are shown in Fig. 19.2.
Proactive routing protocols maintain regular and up to date routing information
about each node in the network by propagating route updation at fixed time intervals
throughout the network, when there is a change in network topology. As the routing
information is usually maintained in tables, so these protocols are also called table-
driven protocols i.e. ad hoc on demand distance vector protocol (AODV), dynamic
source routing (DSR), admission control enabled on-demand routing (ACOR) and
associativity based routing (ABR). Reactive routing protocols establish the route to
a destination only when there is a demand for it, so these protocols are also called
on demand protocols i.e., destination sequenced distance vector (DSDV), optimized
link state routing (OLSR), wireless routing protocol (WRP) and cluster head gate-
way switch routing (CGSR). When a source wants to send to a destination, it uses
the route discovery mechanisms to find the path to the destinations by initiating
route request. When a route has been established, then route remains valid till the
destination is reachable or when the route is expired. Hybrid routing protocols is the
combination of both proactive and reactive routing protocols i.e. temporary ordered
routing algorithm (TORA), zone routing protocol (ZRP), hazy sighted link state
(HSLS) and orderone routing protocol (OOPR). Proactive and reactive algorithms
are used to route packets. The route is established with proactive routes and uses
reactive flooding for new MNs. In this paper, we have compared MANETs routing
protocols from reactive, proactive and hybrid categories, as we have used randomly
one protocol from each categories as from reactive AODV, proactive OLSR, hybrid
TORA.
Ad hoc on demand distance vector protocol is reactive protocol and construct
route on demand and aims to reduce routing load [3]. It uses a table driven rout-
ing framework and destination sequence numbers for routing packets to destination
mobile node (MN) and has location independent algorithm. It sends messages only

MANETs Routing Protocol

Reactive Proactive Hybrid

Protocols Protocols Protocols

AODV DSR ACOR ABR DSDV OLSR WRP CGSR TORA ZRP ARPAM OORP

Fig. 19.2 Mobile ad hoc network’s routing protocols

222 N. Qasim et al.

when it was demanded and has bi-directional route from the source to destination.
When it has packets to send from source to destinations MN then it floods the net-
work with route request (RREQ) packets. All MNs that receive the RREQ checks
its routing table to find out that if it is the destination node or if it has fresh route
to the destination then it unicast route reply (RREP) which is routed back on a tem-
porary reverse route generated by RREQ from source node, or else it re-broadcast
RREQ.
Optimized link state routing is a proactive routing protocol [4]. In which each
node periodically broadcasts its routing table allowing each node to build a global
view of the network topology. The periodic nature of the protocol creates a large
amount of overhead. In order to reduce overhead it limits the number of MN
that can forward network wide traffic and for this purpose it uses multi point
relays (MPRs) which is responsible for forwarding routing messages and opti-
mization for controlled flooding and operations. Each node independently elects
a group of MPRs from its one hop neighbors. All MNs periodically broadcast a
list of its MPR selectors instead of the whole list of neighbors. MPRs are also
used to form a route from MN to destination node and perform route calcula-
tion. All MNs maintain the routing table that contains routes to all reachable
destination nodes. OLSR does not notify the source immediately after detecting a
broken link.
Temporary ordered routing algorithm is hybrid protocol, which is distributed and
routers only maintain information about adjacent routers [5]. During reactive oper-
ation, sources initiate the establishment of routes to a given destination on demand.
Where in dynamic networks, it is efficient with relatively sparse traffic patterns; as
it does not have to maintain routes at all the time. It does not continuously execute
a shortest path computation and the metric used to establish the routing structure
does not represent a distance. TORA maintains multiple routes to the destination
when topology changes frequently. It consists of link reversal of the Directed Acyclic
Graph (ACG). It uses internet MANET encapsulation protocol (IMEP) for link sta-
tus and neighbor connectivity sensing. IMEP provide reliable, in-order delivery of
all routing control messages from a node to all of its neighbors, and notification to
the routing protocol whenever a link neighbors is created or broken. As TORA is
for multihop networks which is considered to minimize the communication over-
head associated with adapting to network topological changes by localization of
algorithmic reaction. Moreover, it is bandwidth efficient and highly adaptive and
quick in route repair during link failure and providing multiple routes to destination
node in wireless networks.

19.3 Simulation Setup

We have conducted extensive simulation study to evaluate the performance of dif-

ferent mobile ad hoc networks routing protocols reactive AODV, proactive OLSR,
hybrid TORA. We have used OPNET 14.0 simulator to carry out simulation study
19 Performance Evaluation of Mobile Ad Hoc Networking Protocols 223

[6], which is used for network modeling and simulation results as it has fastest
event simulation engine. In our Mobility Model, MNs in the simulation area move
according to random waypoint model [1]. Radio Network Interface used are the
physical radio characteristics of each mobile node’s network interface, such as the
antenna gain, transmit power, and receiver sensitivity, were chosen to approximate
the direct sequence spread spectrum radio [2]. Media Access Control is the distribu-
tion coordination function (DCF) of IEEE 802.1 1b, which was used for underlying
MAC layer [2]. Default values are used for MAC layer parameters. For Network
Traffic, in order to compare simulation results for performance of each routing pro-
tocol, communication model used for network traffic sources is FTP. For Traffic
Configuration, all experiments have one data flow between a source node to a sink
node consisting of TCP file transfer session and TCP transmits with the highest
achievable rate. TCP is used to study the effect of congestion control and reliable
delivery [7].

19.4 Simulation Environment

It consists of 50 wireless nodes which were placed uniformly and forming mobile
ad hoc network, moving about over a 1; 000  1; 000 m area for 900 s of simulation
time [8].
The AODV simulation parameters used are the same as in [8] except the active
route timeout which was set to 30 s, the TORA parameters we used are simi-
lar to those in [9]; moreover, the OLSR parameters we used are similar to those
in [8]. In OLSR’s Hello interval and TC interval were set to 2 and 5 s respectively,
its neighbor hold time was 6 s, and entries in topology table expire after 15 s. In
OLSR scenario the MNs in the network are grouped in clusters and each clus-
ter has MPR. The transmission range is 300 m. The MPRs are selected in each
cluster, which are the MNs which have high willingness parameter. We have five
clusters each having its own five MPRs that move towards stable state. MPR of
MNs in the network sends topology control messages periodically. The numbers of
MPR in network are directly proportional to the number of TC traffic sent. Each MN
sends periodically Hello messages in the network that consists of list of neighbors
and node movement’s changes [6] (Tables 19.1–19.3).

Table 19.1 Constants used in the AODV simulation

Route discovery parameter Gratuitous reply

Active route timeout 30s

Hello interval Uniform (10,10.1)
Allowed hello loss 10
TTL parameter 2s
224 N. Qasim et al.

Table 19.2 Constant of IMEP used in the TORA simulation

Beacon periods 3s
Max beacon timer 9s
Max tries Three attempts

Table 19.3 Constants used in the OLSR simulation

Hello interval 2s
TC interval 5s
Neighbor hold time 6s
Topology table entries expire time 15s

In our simulation study, performance evaluations are made using following

parameters:
Throughput is the total number of packets received by the destination.
End to End Delay is the average end to end delay of data packets from senders to
receivers.
Media Access Delay is the media transfer delay for multimedia and real time
traffics’ data packets from senders to receivers.
Packet delivery ratio (PDR) is ratio between the number of packets received by
the TCP sink at the final destination and number of packets generated by the
traffic sources. Moreover, it is the ratio of the number of data packets received by
the destination node to the number of data packets sent by the source MN [10].
There have been previous papers [2, 7–9, 11] to provide a comparative analysis of
routing protocols for ad hoc networks, although those simulations used substan-
tially different input parameters than ours. Yet, there has been no comprehensive
evaluation study done to compare the performance based on categories of routing
protocols, which are reactive, proactive and hybrid routing protocols. Specifically,
the total simulation time was 900 s over which the performance statistics are col-
lected. Another important difference between our study and previous studies was
that we aim to evaluate the varying state behavior of routing protocols from different
categories.

19.5 Simulation Results

When the mobile ad hoc network simulations were run than result shows that all
MNs were in range of each other, no data packets experience collisions in presence
of ftp traffic load and all MNs were capable of sending packets. Hence, it shows that
carrier sense and back off mechanisms of the 802.11b were working precisely. All
results were obtained by averaging over ten random mobility scenarios of mobile ad
hoc networks.
19 Performance Evaluation of Mobile Ad Hoc Networking Protocols 225

The most of events were simulated by OLSR which are 143,571,00. Conse-
quently, average number of events simulated by TORA and AODV are 199,354,5
and 229,537 respectively. On the other hand, high simulation speed for most of
events simulated per seconds was observed in TORA routing protocol simulation
runs that was 544,829 events per second, than it was in AODV and OLSR for
about 398,557 and 232,943 events per seconds. These statistics shows that proactive
protocol can simulate millions of more event than reactive and hybrid protocols.

19.5.1 Throughput

Throughput which is the number of routing packets received successfully by each

routing protocol was shown in Fig. 19.3a. When comparing the routing through-
put packets received by each of the protocols, OLSR has the high throughput.
Throughput is a measure of effectiveness of a routing protocol. OLSR receives about
1,950,000 routing packets at start of simulation time, then fluctuates for 60 s and
gradually becomes stable around 1,600,200 data packets received. In Fig. 19.3b
AODV and TORA are plotted on the different scales to best show the effects
of varying throughput. TORA’s throughput increases IMEP’s neighbor discovery
mechanism, which requires each node to transmit at least one Hello packet per BEA-
CON period (3 s). For 900 s of simulations with 50 MNs, this results in a maximum
throughput of 1, 4,500 packets. In reactive protocol AODV as the number of sources
nodes increases than the number of routing packets receive increases to 8,000 pack-
ets as it maintains cache of routes in routing table to destination and unicast reply by
reversing route generated by source node or re broadcast route request. Delivery of
broadcast packets are not reliable at receiver as there cannot be reservation for the
wireless medium at the receivers before transmitting a broadcast packet by exchange
of request to send or clear to send (RTS/CTS) packets. The source nodes generate
packets and broadcast packets which are received by many MNs, so number of pack-
ets received is much higher than the number of packets sent. This difference does
not exist in wired networks and shows fundamental limitation of wireless networks.
Overall, proactive routing protocol has highest throughput in MANETs.

19.5.2 End to End Delays

Figure 19.3c shows that OLSR has lowest steady end to end delays which are about
0.0004 s. Further on, the end to end delay start to rise and fall abruptly in AODV and
TORA therefore ends up in less end to end delays in AODV as compare to TORA
that is around on average 0.0015 s and 0.0032 respectively. TORA had higher delays
because of network congestion. As created loop where the number of routing pack-
ets sent caused MAC layer collisions, and data, Hello and ACK packets were lost
that resulted in assuming that links to neighbors was broken by IMEP. Therefore,
TORA reacted to these link failures by sending more UPDATEs, in turn that created
226 N. Qasim et al.

Fig. 19.3 Comparison between the MANETs routing protocol’s simulation results: (a) Through-
put, (b) Throughput for TORA and AODV, (c) End to End Delay, (d) Media Access Delay, (e)
Packet Delivery Ratio, (f) Packet Delivery Ratio for AODV and TORA
19 Performance Evaluation of Mobile Ad Hoc Networking Protocols 227

Table 19.4 Simulation results over simulation time of 900s

Protocols Average number of events simulated Average speed
AODV 229,576 398,557 events/s
TORA 199,354,5 544,829 events/s
OLSR 143,571,00 232,943 events/s

more congestion as failure to receive an ACK from retransmitted UPDATEs was

considered as link failure indication.

19.5.3 Media Access Delay

In Fig. 19.3d we plotted media access delay which is very important for multime-
dia and real time traffic; furthermore it is vital for any application where data is
processed online. Media access delay was low for OLSR that is around 0.0001 s.
However, the media access delay for AODV and TORA fluctuates more frequently
but AODV fluctuates more frequently above and below its mean while TORA
mainly around its mean, thus in both case fluctuation is higher and more frequent as
compared to OLSR that remains steady over 900s of simulation time (Table 19.4).

19.5.4 Packet Delivery Ratio

The fraction of the originated application data packets by each protocol was able
to deliver at varying time as shown in Fig. 19.3e. As packet delivery ratio shows
both the completeness and correctness of the routing protocol and also the measure
of efficiency. We used packet delivery rate as the ratio of number of data packets
received at the sink node to the number of data packets transmitted by source nodes
having a route in its routing table after a successful route discovery. For all protocols
packet delivery ratio is independent of offered traffic load, where routing protocols
OLSR, AODV, TORA delivering about 81%, 53.6% and 53.1% of the packets in all
cases. OLSR provides better packet delivery rate than all other routing protocols,
on the other hand AODV has higher delivery ratio as compared to TORA. AODV
and TORA takes time for computing route to send data packets as these protocols
constructs route on demand whereas OLSR is an proactive routing protocol and uses
routing table to send data packets at once. As packet delivery ratio indicates the loss
rate that can be seen by the transport protocols that effects the maximum throughput
that the network can handle. OLSR have MRPs for each cluster which maintains
routes for the group of destination, packets that the MAC layer is unable to deliver
are dropped since there are no alternate routes. In Fig. 19.3f, we have used different
scales of axes to show results of packet delivery ratio visibly for reactive and hybrid
protocol. In AODV and TORA graph starts after 100 s because AODV and TORA
228 N. Qasim et al.

takes time for computing route to receive data packets on destination as these proto-
cols constructs route on demand whereas OLSR is a proactive routing protocol and
uses routing table to send data packets at once. TORA in 50 MN wireless networks
delivered around 53% of data packets over simulation time, TORA fall short to con-
verge because of increased congestion. In TORA mainly data packets were dropped
because of short lived routing loops, which are part of its link reversal process.
When the packet of next hop and previous hop are same then more data packets
were dropped because the packets were looped until time to live expires or when
loop exited; moreover, data packets which were in loops interfered by broadcast
UPDATE packet from neighbor MNs which in turn can resolve routing loop. It was
observed that packet delivery ratio was less in TORA than AODV. Moreover, routing
protocols have differed in how much protocols can deliver packets to destination
MN.

19.6 Conclusion

The mobile nodes’ mobility management is key area since mobility causes route
change and frequent changes in network topology, therefore effective routing has to
be performed immediately. This paper makes contributions in two areas. Firstly, this
paper compared the performance of reactive ad hoc on demand distance vector pro-
tocol; proactive optimized link state routing protocol and hybrid temporary ordered
routing algorithm protocol in mobile ad hoc networks under ftp traffic. Secondly, we
have presented the comprehensive results of packet delivery ratio, throughput, media
access delay, and end to end delay over mobile ad hoc networks of 50 MNs moving
about and communicating with each other. The simulation results were presented
for a range of node mobility at varying time. OLSR performs quite predictably,
delivering virtually most data packets at node mobility. In [8] also shows that OLSR
shows the best performance in terms of data delivery ratio and end-to-end delay.
TORA, although did not perform adequate in our simulation runs in terms of rout-
ing packet delivery ratio, delivered over fifty three percentage of the packets. Since
the network was unable to handle all of the traffic generated by the routing protocol
and a significant fraction of data packets were dropped. As well as in [9] shows that
the relative performance of TORA was decisively dependent on the network size,
and average rate of topological changes; TORA can perform well in small network
size but TORA’s performance decreases when network size increases to 50 nodes.
On the other hand, AODV performed better than TORA in most performance met-
rics with response to frequent topology changes. Finally, the overall performance of
OLSR was very good when MNs movement was changing over varying time. We
have analyzed that all routing protocol successfully delivers data when subjected to
different network stresses and topology changes. Moreover, mathematical analysis
and simulation results both show that optimized link state routing protocol, from
proactive protocol category, is a very effective, efficient route discovery protocol for
MANETs.
19 Performance Evaluation of Mobile Ad Hoc Networking Protocols 229

References

1. H. Pucha, S. M. Das, Y. C. Hu, “The Performance Impact of Traffic Patterns on Routing

Protocols in Mobile Ad Hoc Networks”, Journal (COMNET), vol. 51(12), pp. 3595–3616,
August 2007.
2. J. Broch, D. A. Maltz, D. B. Johnson, Y. C. Hu, J. Jetcheva, “A Performance Comparison of
Multi-Hop Wireless Ad-Hoc Network Routing Protocols,” Proceedings of the 4th ACM/IEEE
International Conference on Mobile Computing and Networking MOBICOM’98, pp. 85–97,
Texas, October 1998.
3. C. E. Perkins, E. M. Royer, I. D. Chakeres, “Ad hoc On-Demand Distance Vector (AODV)
Routing Protocol”, draft-perkins-manet-aodvbis-00.txt, October 2003.
4. T. Clausen, C. Dearlove, P. Jacquet, “The Optimized Link State Routing Protocol version
2”, MANET Working Group, Available: http://www.ietf.org/internet-drafts/draft-ietf-manet-
olsrv2-05.txt, February 2008.
5. V. Park, S. Corson, “Temporally-Ordered Routing Algorithm (TORA) Version 1”, Internet
draft, IETF MANET working group, Available: http://tools.ietf.org/id/draft-ietf-manet-tora-
spec-04.txt, July 2001.
6. Opnet.com (2008), The OPNET Modeler 14.0, Available: http://www.opnet.com.
7. E. Nordstrom, P. Gunningberg, C. Rohner, O. Wibling, “A Comprehensive Comparison
of MANET Routing Protocols in Simulation, Emulation and the Real World ”, Uppsala
University, pp. 1–12, 2006.
8. C. Mbarushimana, A. Shahrabi, “Comparative Study of Reactive and Proactive Routing Pro-
tocols Performance in Mobile Ad Hoc Networks”, 21st International Conference on Advanced
Information Networking and Applications Workshops (AINAW’07), IEEE Computer Society,
March 2007.
9. S. R. Das, R. Castaneda, J. Yan, R. Sengupta, “Comparative Performance Evaluation of Rout-
ing Protocols for Mobile Ad hoc Networks”, Proceedings of the International Conference on
Computer Communications and Networks, pp. 153–161, 1998.
10. G. Ivascu, S. Pierre, A. Quintero, “QoS Support based on a Mobile Routing Backbone for
Ad Hoc Wireless Networks”, Proceedings of the International Conference on Wireless Com-
munications and Mobile Computing IWCMC’06, pp. 121–126, Vancouver, Canada, July
2006.
11. V. D. Park, M. S. Corson, “A Performance Comparison of TORA and Ideal Link State
Routing”, Proceedings of IEEE Symposium on Computers and Communication, June 1998.
Chapter 20
IEEE 802.11E Block Acknowledgement Policies

O. Cabral, A. Segarra, and F. J. Velez

Abstract Optimization of IEEE 802.11e MAC protocol performance is addressed

by modifying several parameters left open in the standard, like block size and
acknowledgement policies in order to improve the channel efficiency. The use of
small block sizes leads to a high overhead caused by the negotiation on the other
hand, the use of large block sizes causes long delays, which can affect negatively
real-time applications (or delay sensitive applications). An event driven simulator
was developed, and results with a single service and several services running simul-
taneously were extracted. By using the Block Acknowledgement (BA) procedure,
for video and background traffics in a single service situation, the capacity was
improved in the case when the number of stations is equal or higher than 16 and
12, respectively. However, for lower values of the number of stations, the use of
BA leads to a slightly worst system performance. In a scenario with mixture of ser-
vices the most advised block size is 12 (less delay in a highly loaded scenario). The
number of supported user (total) increases from 30 to 35.

Keywords IEEE 802.11e  simulator  block acknowledgement  quality of service

20.1 Introduction

Recent years have seen an immense growth in the popularity of wireless applications
that require high throughput. To support such growth, standardization bodies such
as the IEEE 802.11 have formed task groups to investigate and standardize features
providing increased quality of service (QoS) and higher throughputs. One of these
extensions is the acknowledgement (ACK) policy feature included in the ratified
IEEE 802.11e amendment for QoS support [3], which is the focus of this work. In

O. Cabral (B)
Instituto de Telecomunicacoes, DEM-UBI Calcada Fonte do Lameiro 6201-001 Covilha Portugal,
E-mail: orlando@ubi.pt

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 231
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
232 O. Cabral et al.

particular, we investigate the policy regarding the block size for a video application.
The Block Acknowledgement (BA) procedure improves system throughput results
by reducing the amount of overhead required by a station to acknowledge a burst of
received traffic [1,2]. It acknowledges a block of packets by a single ACK, instead of
using several ACKs, one for each packet, saving the Arbitration Inter-frame Spacing
(AIFS) period, the backoff counter time, and the acknowledgement time. The num-
ber of frames that can be transmitted within a block is called block size. It is limited
and is not specified in the standard. In this chapter, to find the most suitable block
size we have tried several block sizes and several loaded scenarios with and without
mixture of services. This chapter is organised as follows. In Section 20.2, a brief
introduction to the IEEE 802.11e standard is presented along with the main direc-
tive lines of the BA procedure. Section 20.3, gives the description of the state of the
art. Section 20.4 defines the problem and the scenario, including details on traffic
parameters. Section 20.5 gives the simulation results obtained for several scenarios
with and without the use of the BA procedure, with and without mixtures of traffic.
Conclusions are given in Section 20.6, as well as suggestions for further work.

20.2 IEEE 802.11e and Block Acknowledgement Description

20.2.1 IEEE 802.11e User Priorities and Access Categories

The so-called enhanced distributed channel access (EDCA) provides differentiated,

distributed access to the medium for Quality Stations (terminals that support IEEE
802.11e) using four access categories (ACs) voice (VO), video (VI), best effort
(BE), and background (BK). This differentiation is achieved by mapping the traffic
to four queues that correspond to the four AC. The traffic prioritisation is performed
by varying the amount of time a station queue senses the channel to be idle before
backoff or transmission, or the length of the contention window to be used for the
backoff, or the duration a station queue may transmit after it acquires the channel.
Each AC contends to access the medium with a single CSMA instance, [3] [1].
Each queue has an Arbitration Inter-frame Spacing (AIFS) period preceding the
next backoff contention window (see Fig. 20.1).

20.2.2 Block Acknowledgement

The BA mechanism (see Fig. 20.2) improves channel efficiency by aggregating

several acknowledgements into one frame [4,5]. There are two types of BA mech-
anisms: immediate and delayed. Immediate BA is suitable for high-bandwidth,
low-latency traffic while the delayed BA is suitable for applications that tolerate
moderate latency. The QSTA with data to send using the BA mechanism is referred
 AIFS Contention Window from
Immediate access
when medium is idle [1,Cwmin[AC]+1]
>= AIFS[AC]+SlotTime PIFS

AIFS [AC] Busy SIFS Backoff

Medium Next frame
+SlotTime Window

Slot Time
20 IEEE 802.11E Block Acknowledgement Policies

Defer Access
Select Slot and decrement
backoff as long as the
medium stays idle

Fig. 20.1 Timing relationship in EDCA

233
234 O. Cabral et al.

ADDBA Request

ACK (a) Setup

ADDBA Response

ACK

Data

Multiple (b) Data & BlockACK

frames
Block ACKReq

BlockACK

DELBA Request
(c) Tear Down
ACK

Originator Recipient

Fig. 20.2 Block ACK sequence

to as the originator, and the receiver of that data as the recipient. The BA mech-
anism is initialized by an exchange of Add Block Acknowledgment (ADDBA)
Request/Response frames. After initialization, blocks of QoS data frames can be
transmitted from the originator to the recipient. A block may be started within a
polled TXOP or by winning EDCA contention. The number of frames in the block
is limited, and the amount of state that is to be kept by the recipient is bounded. The
MAC Packet Data Units (MPDUs) within the block of frames are acknowledged
by a BA control frame, which is requested by a BlockACKReq control frame. The
response to the BlockACKReq will acknowledge all the correctly received frames
and request the badly received to be transmitted again.

20.3 State of the Art

Analytical frameworks to model BA have been published [6-8] but the results are
not based on a realistic approach to the problem nor account for the achievable
QoS, because the use of several service classes with different priorities (the base
20 IEEE 802.11E Block Acknowledgement Policies 235

for EDCA) is not considered at all. The existing theoretical approaches [6-8] do
not consider the hidden terminal problem, assume that the buffer is always full,
nor assume a multi-rate scenario. In [6,7], an analytical framework was presented
to model an ad-hoc network under the standard IEEE 802.11e with the BA proce-
dure for a completely simplified scenario. The hidden/exposed terminal problem is
one fundamental issue in WLANs but most of the existing analytical models either
assume it does not exist, or do not consider the EDCA features of IEEE 802.11e
(or do not account for the delay or any other QoS metric) [6,7]. Results presented
in [6,7] express the block size as a function of the goodput in saturation conditions.
Results show that the block size should be as high as possible, which can be mis-
leading when we consider QoS. In [8] an analytical approach to model the BA in
IEEE 802.11e was proposed without accounting for the hidden terminal problem.
The multi-rate feature in the same environment was also not considered. Further,
the packet loss due to errors in the channel was not taken in consideration. Finally,
the use of EDCA procedures, like several virtual queues, for the several classes of
service are also not considered, i.e., this work does not consider the IEEE 802.11e
standard at all. From the simulation approaches presented in the literature the one
that is most similar to the work proposed here was proposed in [9]. In [9] several
combinations for the block size are presented where a scheduler based on the delay
and amount of data in the buffer is proposed. The work presented here is an improve-
ment of this approach and provides a more extensive study on the block size while
considering use-perceived QoS.

20.4 System, Scenario and Assumptions

A cellular WiFi system was considered where each cell has a set of N+1 IEEE
802.11e stations communicating through the same wireless channel. While station
0 is the QoS Access Points (QAP), the other N are QoS stations (QSTA). Each
station has four buffers whose size depends on the kind of service being dealt in
order to guarantee a given value for the goodput (payload of the packet at MAC
level). These buffers will be filled with a MAC Service Data Units (MSDU) gener-
ated that characterises the service being dealt in the given buffer. If the MSDU is
bigger than a fragmentation threshold, it will be fragmented. In order to cope with
service quality the packet transmission follows the Enhanced Distributed Channel
Access (EDCA) IEEE 802.11e MAC procedure. Due to collisions or interference
a packet may not be correctly received. The interference issues are addressed by
using a radio propagation model. Each packet exits the buffer only after the recep-
tion of an acknowledgement, or if it has suffered more than a given collision
threshold. The users are assumed to be static, and are distributed uniformly in a
square area of 2,500 square meter. The physical layer specification assumed in this
work is the IEEE 802.11a standard 802, 1999, [10], that defines an Orthogonal Fre-
quency Division Multiplexing (OFDM) based PHY layer that operates in the 5 GHz
frequency bands, being able to achieve bit-rates as high as 54 Mbps. The physical
236 O. Cabral et al.

Table 20.1 MAC and PHY parameters

Slot time 0.009 ms
ACK size 112 bit
SIFS 0.016 ms
DIFS 0.034 ms
RTS threshold 4,000 bit
RTS size 160 bit
CTS size 112 bit
CWmin 31 slots
CWmax 1,280 slots
Collisions threshold 7
Fragmentation threshold 8,192
Simulation time 100 s

Table 20.2 Traffic parameters

AC Voice (VO) Video (VI) Background (BK)
Packet size 1,280 bit 10,240 bit 18,430 bit
Packet interval 20 ms 10 ms 12.5 ms
Usage 50% 30% 20%
Symmetry Symmetric Asymmetric (downlink) Asymmetric (downlink)

and MAC parameters are presented in Table 20.1. The use of the Request/Clear-to-
send (RTS/CTS) procedure is implemented only if the packet size is larger than a
given threshold, RTS threshold in Table 20.1.
In the algorithm proposed in [10] the sender chooses the bit-rate that achieves the
highest throughput taking into account the SINR estimate. More details on physi-
cal layer implementation used in the simulator can be found in [11]. Three types
of traffic sources were chosen, namely high priority voice (VO), medium prior-
ity video (VI), and low priority FTP data as background traffic (BK). The traffic
sources parameters are presented in Table 20.2 as well as the Access Categories
(AC) of each type of traffic.
We implemented the BA procedure with minor modification to the existing fea-
tures like TXOP and RTS/CTS as already explained and without disturbing the
overall TXOP and RTS/CTS procedure. When a TXOP is gained, the transmission
starts and the origin will know if a BA procedure is implemented with this desti-
nation. If this is true, it will not wait for an acknowledgement, but just a SIFS and
start the next transmission for the same destination or not depending on which is
the next packet in the buffer. Figure 20.3 presents this procedure for destination 1
that has the BA procedure implemented and for destination 2 without the BA. At the
beginning of a transmission for a user with active BA procedure, if the block size
threshold is reached, a block ACK request packet is added to the buffer to the top-1
place in the queue to be transmitted as the next in line packet.
20 IEEE 802.11E Block Acknowledgement Policies 237

Fig. 20.3 Illustration of the

MPDU Dest1

MPDU Dest1

MPDU Dest2

MPDU Dest2
BA procedure within a TXOP

ACK

ACK
TXOP Limit

20.5 Results

20.5.1 Block Acknowledgement with Standalone Services

The objective of the simulations was to investigate the gain of the BA procedure for
a single service and for a mixture of services. In terms of grade of service, the voice
application supports delays up to 30 ms, the video application supports delays up
to 300 ms, while the delay for background applications can be up to 500 ms [13].
As previously mentioned, the BA mechanism improves the channel efficiency by
aggregating several acknowledgments into one frame. To investigate the proposed
BA procedure for the stand-alone service a BA procedure with a block size of 16
fragments was adopted. This procedure is simulated only for BK and VI traffic. For
VO application it is certain that BA is not a solution because of the large delays
that occur when the buffer is filled with 16 packets (i.e., the delay of the first one
would be 320 ms, application). Figure 20.4 presents the delay for BK and VI traffic.
It starts to increase around 12 stations for BK and at 16 stations for VI and increases
more for a higher number of stations. As expected, the delay is lower when the BA
procedure is used. The improvement (reduction) for BK traffic is 300 ms on the
average after 12 stations, while for VI traffic it is 420 ms on the average after 16
stations. The improvement in the goodput is of 2 and 2.2 Mb/s in average, after 12
stations, for BK traffic and after 16 stations for VI, respectively [12].

20.5.2 Block Acknowledgement with Mixtures of Applications

This part of the work explores which should be the policy regarding the block size
with mixtures of applications. For the purpose we have investigated the BA policy
for a video service, an application that is delay-sensitive. Additionally, we have
investigated the BA block size for background service that is not delay sensitive.
Simulations were performed for 100 s and around 40 times for each number of
stations, 15, 20, 25, 30, 35, 40, 45 and each block size, 4, 8, 12, 16. The users
started the negotiation to initiate BA procedure in the beginning of the simulation
238 O. Cabral et al.

Fig. 20.4 Delay for BK and VI with and without Block ACK

1 2 3 4 2 5 6

Delay = actual_delay +
Bad transmission clock-time_tx_packet (3) +clock-
time tx packet (4)

Fig. 20.5 Procedure to count delay

so all the packets of video and background traffic were transmitted within the BA
procedure. The use of the BA procedure in a scenario with a mixture of applications
running at the same time was further studied. On the one hand, the overhead caused
by the negotiation to establish the BA procedure, and to maintain the BA causes bad
performance when a small block size is used. On the other hand, the packet losses
caused by the voice users with higher priority in the system, influences the overall
QoS on the system and mainly on the video service since we considered in this
simulations, that if a packet is not correctly received at a given time the following
packets sent within this block will have the delay of the badly sent packet after
transmitted added to their delay, Fig. 20.5.
The impact of packet retransmissions in the delay therefore increased. Figure 20.6
presents the delay for each access category. For BK traffic the delay starts to increase
considerably with more than 20 stations (12 in the standard procedure). This service
class will transmit very rarely since the voice and video traffic will always be sched-
uled first. For video traffic the delay starts to increase in the 35 stations. In contrast
with BK and VI, VO applications present a negligible delay.
The delay impacts the most the video application. Figure 20.7 shows the results
for various block sizes for transmitted video application. Regardless on the number
of stations, the BA procedure reduces the delay. Certain sensitivity is observed for
a block size of 16, which is the size that gives the highest delay for a total of 1–25
stations. For a larger number of station the best block size is 16 as smaller block
sizes are not that efficient in decreasing the delays.
When 30 stations are being served in the system the results for the delay with
BA is near 5 ms for all block sizes, while without BA the values of the delay
20 IEEE 802.11E Block Acknowledgement Policies 239

Delay for all traffic classes

x 104
2
block size 4(BK)
block size 8(BK)
block size 12(BK)
1.5 block size 16(BK)
standard(BK)
block size 4(VI)
Delay [ms]

block size 8(VI)

1 block size 12(VI)
block size 16(VI)
standard(VI)
0.5 block size 4(VO)
block size 8(VO)
block size 12(VO)
block size 16(VO)
0
15 20 25 30 35 40 45
Number of stations

Fig. 20.6 Delay for all services with Block ACK implemented for VI and BK traffic

Video traffic delay

4
x 10 15
2
Delay [ms]

10

1.5 5
Delay [ms]

0
15 20 25 30 35
1 block size 4 Number of stations
block size 8
block size 12
0.5 block size 16
standard ACK

0
15 20 25 30 35 40 45
Number of stations

Fig. 20.7 Delay for video traffic

goes up to 80 ms. For 35 stations the difference is even higher. A minimum near
5 ms is obtained for block size 12 while for delay without BA the delay is near
2.3 s. Results for 40 and 45 advise the use of block size 16, although the network
is already overloaded (and the delay will just keep growing up to infinity). When
less then 40 stations are present the confidence intervals are very small for all the
buffers. When there are 40 or more stations the confidence intervals increase. One
can therefore extrapolate that there are stations that manage to transmit and have a
small delay while others transmit from time to time, causing very high delays. The
behaviour observed in Fig. 20.7 occurs due to the overhead caused by the blockACK
request, and the delays caused by bad receptions affected mostly the block size 16,
but providing lower delays for higher loaded scenarios. The solution is neither to
use a large block size (as large as 16) nor a small block size (as low as 4). The
240 O. Cabral et al.

Fig. 20.8 Total goodput with BA implemented for video and background traffic

Fig. 20.9 Video goodput with Block ACK implemented for video traffic

former increases the delay causing problems to the application using these packets
(for more than 35 stations), while the latter causes some unnecessary overhead by
requesting very often block ACK requests (for less than 35 stations). The advised
block size shall be 12 since it is the one that provides lower delay in a scenario
where the load is still manageable, at least for the case where BA is used.
Figure 20.8 presents the results of the goodput in the system in the downlink.
Without the BA the maximum goodput achieved is near 11 Mbit/s, while with BA is
near 14 Mbit/s for block sizes 12 and 16. The decreasing behaviour after 25 stations
occurs due to the high number of collision, and as the background traffic has low
priority ends up not being transmitted giving its turn to voice and video users. As
voice traffic provides higher overhead the resulting goodput is lower.
20 IEEE 802.11E Block Acknowledgement Policies 241

Figure 20.9 presents the goodput only for the video service class. Only after 30
stations the achieved throughput is different when using and not using BA. The
highest goodput is found for block size 16, and is more than 10% higher relatively
to the case of standard ACK. The increasing behaviour of the goodput is verified up
to 40 stations.

20.6 Conclusions

This work investigated the BA size as a policy to decrease delays and ensure a stable
system. The investigation was based on tuning several parameters and investigated
the effect for a procedure with and without BA. Several policies were investigated.
Future work will test policies that provide access to the medium, that ensure some
degree of service, based on the channel SINR, delays, bit error rate, can be tested.
Although, for lower values of the number of stations, the use of BA leads to a slightly
worst system performance, the BA procedure provides an improvement in highly
loaded scenarios. The improvement is of 2 and 2.2 Mb/s in average, for BK traffic
and VI traffic, respectively. In a scenario with mixture of services the most advised
block size is 12 since it is the one that provides lower delays in a highly loaded
scenario while the users are still within the capacity of the AP. The number of sup-
ported user increases from 30 to 35. Note that 35 stations is the total number of VO
plus VI and BK user.

Acknowledgements This work was partially funded by CROSSNET (Portuguese Foundation

for Science and Technology, FCT, POSC project with FEDER funding), by IST-UNITE, by the
FCT PhD grant SFRH/BD/28517/2006, by Fundação Calouste Gulbenkian, and by “Projecto
de Re-equipamento Cientı́fico” REEQ/1201/EEI/2005 (a Portuguese Foundation for Science and
Technology project).

References

1. A. R. Prasad, and N. R. Prasad, (2005). 802.11 WLANs and IP Networking. Artech House,
Boston, MA.
2. J. L. Mukerjee, R. ă Dillinger, M. ăMohyeldin and Schulz, E.ă (2003). Investigation of
radio resource scheduling in wlans coupled with 3g cellular network. IEEE Communications
Magazine, 41(6):108–115.
3. T. Li, Q. Ni, T. Turletti and Y. Xiao, (2005). Performance analysis of the ieee 802.11e block ack
scheme in a noisy channel. In IEEE BroadNets 2005 – The Second International Conference
on Broadband Networks, Boston, MA.
4. T. Li, Q. Ni and Y. Xiao, (2006). Investigation of the block ack scheme in wireless ad-hoc
networks. Wiley jornal of Wireless Communications and Mobile Computing, 6(6):877–888.
5. I. Tinnirello, and S. Choi, (2005). Efficiency analysis of burst transmissions with block
ack in contention-based 802.11e wlans. In ICC 2005 – IEEE International Conference on
Communications, Seoul, Korea.
242 O. Cabral et al.

6. V. Scarpa, G. Convertino, S. Oliva and C. Parata, (2005). Advanced scheduling and link adap-
tation techniques for block acknowledgement. In 7th IFIP International Conference on Mobile
and Wireless Communications Networks (MWCN 2005), Marrakech, Morocco.
7. Ni, Quiang (2005). Performance analysis and enhancements for ieee 802.11e wireless
networks. IEEE Networks, 19(4):21–27.
8. A. Grilo, (2004). Quality of Service in IP-based WLANs. Ph.D. thesis, Instituto Superior
Tcnico, Technical University of Lisbon, Lisbon, Portugal.
9. O. Cabral, A. Segarra and F. J. Velez, (2008). Event-driven simulation for ieee 802.11e
optimization. IAENG International Journal of Computer Science, 35(1):161–173.
Chapter 21
Routing in a Custom-Made IEEE 802.11E
Simulator

João Miguel Ferro and Fernando J. Velez

Abstract The custom-made IEEE 802.11E simulator in this chapter is an evolution

of the previous one. To provide support for multi-hop environment it was neces-
sary to implement the routing algorithm above the already existing Medium Access
Control (MAC) plus Physical (PHY) layers. After the random placement of all the
stations in the field, the simulator determines which can communicate directly. With
this data, the selected routing algorithm determines the shortest path from a station
to all the others. For these initial tests we chose the well-known Dijkstra’s algo-
rithm. In this work, however, we present no cross-layer at all, and the routing table
is not modified during the simulations. The engine starts collecting events at the
beginning of the simulation. For the end user, this simulator allows for simulating a
network with an unrestricted number of nodes, in any configuration type. By using
the chosen routing protocol, it supports connections to every reachable station from
its neighbours, allowing for a message to reach any destination.

Keywords IEEE 802.11E Simulator  Custom-Made  Routing  multi-hop environ-

ment  network simulation

21.1 Introduction

Wireless networks are gaining more and more importance in our world. Cellular
phones with GPRS/UMTS, Wi-Fi, and WiMAX networks are very common these
days, and they share a common feature: they require some sort of backbone infras-
tructure in order to allow for packets from different communication peers to reach
each other. For example, if someone makes a phone call, the conversation will
always pass from the cell phone to the operators’ infrastructure, and then to the

J. M. Ferro (B)
Instituto de Telecomunicações, DEM-UBI, Portugal,
E-mail: ferro@lx.it.pt

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 243
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
244 J.M. Ferro, F.J. Velez

receivers phone, even if they are both in the same building. Resources would be
certainly saved if somehow the cell phones could connect directly to each other.
In an ad-hoc network, all the participants (also called nodes) can communicate
directly with their neighbours. Two nodes are considered neighbours if their com-
munication devices can reach each other. Nodes wanting to communicate to others
that are not neighbours will simply send a message to another node which is located
nearer the destination. As so, a centralised infrastructure to establish the connec-
tivity is not required, since each node will determine by itself to where it should
forward its data. The specification IEEE 802.11 refers to this type of network as
Independent Basic Service Set (IBSS).
This chapter is organized as follows. Section 21.2 presents some background
information about the previous simulator and about IEEE 802.11e. Section 21.3
describes the features of the new simulator highlighting the modifications we per-
formed. Section 21.4 presents the results of the initial simulations, while
Section 21.5 presents conclusions and suggestions for future work.

21.2 Previous Work

This simulator is based on the one developed at the Instituto de Telecomunicações,

Laboratório da Covilhã, as part of the IST-UNITE project. The purpose of that
previous version was to create a trustworthy simulator to be used in the context
of IST-UNITE, allowing the study of the interoperability between Wi-Fi and High
Speed Downlink Packet Access (HSDPA), [1]. This involved the creation of an
HSDPA time-driven simulator and a Wi-Fi event-driven simulator that may be
able to run separately or together for the coexistence scenario. In the latter case,
there is communication between the two simulators which will run separately in a
synchronous way.
The Wi-Fi simulator was initially built only for the infrastructure mode, i.e.,
an architecture with one access point (AP) and several wireless stations (STAs)
attached to it. It can simulate traffic from one STA to the AP and vice-versa, and
calculates the throughput (total in the simulation and per unit of time), packet delay
(total in the simulation and average), lost packets, retransmissions, and collisions.
This computation is oriented to the support of the Quality of Service (QoS), i.e., it
implements IEEE 802.11e with four access categories. As a consequence, these met-
rics are calculated for each traffic type: voice (VO), video (VI), background (BK)
and best-effort (BE).
The IEEE 802.11e standard was conceived to implement Quality of Service
(QoS) in IEEE 802.11a/b/g. QoS refers to the way resources are reserved for certain
applications, guaranteeing that the most urgent data flows will have higher priority
levels. The IEEE 802.11 Medium Access Control (MAC) coordination functions
are the Distributed Coordination Function (DCF) and the Point Coordination Func-
tion (PCF). They establish the timing and sequence for each station to access the
shared medium. The latter is only used in the infrastructure mode, with the AP
21 Routing in a Custom-Made IEEE 802.11E Simulator 245

acting as the coordinator. It provides some basic QoS support, but since it is defined
as optional and is more complex and costly to implement, only a few APs support it.
The former does not present QoS guarantees at all. To ensure the QoS performance,
IEEE 802.11e introduced the Hybrid Coordination Function (HCF), which defines
the HCF Controlled Channel Access (HCCA) and Enhanced Distributed Channel
Access (EDCA). In EDCA, a Transmission Opportunity (TXOP) is won according
to the traffic type, scheduling firstly higher priority traffic, like voice and video. The
HCCA is more complex, but allows greater precision in terms of QoS. All these
enhancements are fully explained by Prasad [2].
In the EDCA simulator each station (and AP) has four buffers, one for each traffic
type. By using an event-driven approach, for each new packet arrived at a machine
a new event is generated. The simulator engine uses its internal clock to pick up the
next event and pass the corresponding packet to the developed MAC + PHY layers.
A more comprehensive description of the lower layers of the simulator can be found
in [3].

21.3 Overview

The current simulator is an evolution of the previous one. To provide support for
multi-hop environment it was necessary to implement the routing algorithm above
the already existing Medium Access Control (MAC) plus Physical (PHY) layers, as
it is presented in Fig. 21.1.
After the random placement of all the stations in the field, the simulator deter-
mines which can communicate directly. With this data, the selected routing algo-
rithm determines the shortest path from a station to all the others. For these initial
tests we chose the well-known Dijkstra’s algorithm, [4], which calculates the least
cost path from one node to all the remaining ones. This characteristic allows the
routing calculation to take place before the simulation starts, making the routing
table available to all nodes since time instant t D 0 s. Besides this initial calcula-
tion, the routing table can be accessed and modified at any time, allowing the use
of cross-layer design (i.e., gather information from physical and MAC layers and
use it in the routing algorithm) for optimisation purposes. In this work, however,
we present no cross-layer at all, and the routing table is not modified during the
simulations.
The engine starts collecting events at the beginning of the simulation. When a
packet arrives at a wireless machine, a new simulation module verifies if the packet
has another destination by checking the two additional fields now included in the
packet header: besides the “payload”, “origin station”, and “destination station”,
among others, each packet now includes the “initial origin” and the “final destina-
tion” fields. If it is found that the destination is another node, the routing table is
accessed to determine the next hop, and the packet is added to the machine buffer
with the new destination stamped.
246 J.M. Ferro, F.J. Velez

Fig. 21.1 Multi-hop simulator flowchart, modules added to the previous version are displayed in
grey

The simulation will run until the time limit is reached, and will be repeated during
a pre-determined number of times. The final output of the simulator is obtained as
an average of the results for these simulations.
For the end user, this simulator allows for simulating a network with an unre-
stricted number of nodes, in any configuration type. By using the chosen routing
protocol, it supports connections to every reachable station from its neighbours,
allowing for a message to reach any destination. Stations are randomly placed in a
field and even if a station is isolated the simulator will still run (unless this station is
the source/final destination of a packet, which will terminate the simulation). Sev-
eral parameters can be tuned by the user to its needs, some of them are presented in
Table 21.1.

21.4 Results

In order to verify if the modified simulator was able to simulate a multi-hop envi-
ronment, we ran a few tests. In the first one we tried to verify if the simulation time
was going to affect the end-to-end delay. For this purpose, we considered a fixed
21 Routing in a Custom-Made IEEE 802.11E Simulator 247

Table 21.1 Some values that can be modified by the user

Parameter Default value Description

Side 60 m Side of the square where the sta-

tions are placed
Range 35.365 m Transmission range of each station
Simulation Time 10,000 ms Period of time to be simulated
Payload XX – bits Length of each packet of the
type XX
Inter Arrival XX – ms Interval of time to generate a new
packet of the type XX

Fig. 21.2 Video traffic being transmitted from station 1 to station 3

configuration of 14 stations, with a video stream being generated from station 1 to

station 3, Fig. 21.2.
The following parameters were used:
Video traffic starts at: t D 1 ms
New packet generated each: 10 ms
Simulation ends at: t = 0.1/0.5/1/5/10/15 s
Video payload size: 10,240 bits (each packet will be sent in three fragments)
Figure 21.3 presents the average delay for each simulation time. As one can observe,
the delay is constant with the simulation time, despite the slight differences for the
shortest simulations.
248 J.M. Ferro, F.J. Velez
Delay [ms]

Simulation time [s]

Fig. 21.3 Delay versus simulation time

Fig. 21.4 Station deployment for the tests

After this initial test, 10 s was assumed as a reasonable simulation time, and the
subsequent tests were made with this value for the duration.
Besides this, further tests were made to evaluate the influence of multiple streams
in the delay and packet loss. For these tests we used the same deployment of stations
as in the previous test, Fig. 21.4.
The inner circle has a radius of 15 m, the next one 30 m, and the outer one 45 m.
The square has 120  120 m2. In the next set of tests we added sequentially video
streams between the following stations:
A. From station 1–3
B. From station 4 to 2
21 Routing in a Custom-Made IEEE 802.11E Simulator 249

C. From station 5 to 6
D. From station 8 to 9
For each of these streams the routing algorithm established the following paths:
A. 1 ! 9 ! 13 ! 14 ! 10 ! 3
B. 4 ! 11 ! 14 ! 7 ! 2
C. 5!6
D. 8 ! 14 ! 13 ! 9
Figure 21.5 presents the results obtained for the delay of video stream from station
1 to station 3. By increasing the number of stations that generate packets, the end-
to-end delay (latency) increases.
Another interesting metric is the number of lost packets, which is presented as the
darker areas in Fig. 21.6. When the number of packets to be transmitted increases,
there are more collisions and packet losses.
One interesting characteristic of this simulator is the possibility of simulating
four different types of traffic and their mixtures. By using this feature we produced
a new set of tests, keeping the same topology but modifying the type of packets in
each stream:
A. Video stream from station 1–3
B. Background traffic from station 4–2
C. Voice traffic from station 5–6
Delay [ms]

Number of video streams

Fig. 21.5 End-to-end delay for video stream 1

250 J.M. Ferro, F.J. Velez

Number of packets

Number of video streams

Fig. 21.6 Total of packets received/lost

Table 21.2 Configuration of each traffic type

Type Packet size (bits] Period (ms]
Video 10,240 10
Voice 1,280 20 (bi-directional)
Background 18,430 12.5

The details for the configuration of each traffic type are presented in Table 21.2 and
were taken from Qiang [5], while the results for the number of packets successfully
delivered (displayed in light grey) or lost (displayed darker) are presented in
Figs. 21.7 and 21.8.
While Fig. 21.7 presents the results for each packet, Fig. 21.8 addresses each
stream, i.e., the latter only counts a successful packet when this one arrives at the
final destination, while the former presents results for the accounting of packets that
arrive at any station. For this reason, and looking at the results for the video stream
alone (VI), in the former case, 5,000 packets were successfully delivered, while for
the latter only 1,000 packets arrived at the final destination. Remember that this
stream is transmitted in a 5-hop path, and that the most relevant result is the latter.
With only one stream, the system behaves perfectly, and all of the packets are
correctly delivered. This happens because there are no collisions, since the period
is longer than the end-to-end delay for every string. However, when more than one
type of traffic is being generated simultaneously, the packets start to collide and
some of them are considered lost (current policy establishes that a station will drop
21 Routing in a Custom-Made IEEE 802.11E Simulator 251

Number of packets

Traffic generated

Fig. 21.7 Total of packets received/lost

Number of packets

Traffic generated

Fig. 21.8 Total of packets lost/received at destination

a packet after it experiences two collisions). The extension of this problem is larger
for background traffic, which can be explained by the largest size of each packet,
which will require the fragmentation into four smaller packets. For voice traffic, the
delay is very low and it seems that it does not affect the remaining traffic at all. This
is due to the small packets generated by this traffic type (no fragmentation required),
and because these stations are in line-of-sight, so no multi-hop is needed.
For the next tests, a new deployment was made changing the position of station 6,
while keeping all the others in the same position, Fig. 21.9.
252 J.M. Ferro, F.J. Velez

Fig. 21.9 Station deployment for the next set of tests

The traffic is the same generated before, and as so the routing algorithm com-
puted the following paths:
A. 1 ! 9 ! 13 ! 14 ! 10 ! 3
B. 4 ! 11 ! 14 ! 7 ! 2
C. 5 ! 13 ! 14 ! 6
D. 8 ! 14 ! 13 ! 9
To have faster results the simulation time was cut to 1 s (and not 10 s anymore). The
results for the new simulation presented in Fig. 21.10 also differ from the previous
because a new policy establishes that a packet will be discarded after experiences
eight collisions (previously it was discarded after just two collisions). By increasing
the collision limit, one should expect that the number of packets lost will be reduced
(for each packet it will be given eight chances to be transmitted, four times more
than previously). Comparing the new results with the previous (Fig. 21.8), one can
see that in fact the number of packets lost was reduced, while the number of packets
successfully delivered increased.

21.5 Conclusions

In this chapter we presented the potentialities of a custom-made IEEE 802.11e sim-

ulator, with multi-hop capabilities. Based on a previous work that developed a MAC
plus PHY layers simulator for this standard, we added above these layers a routing
21 Routing in a Custom-Made IEEE 802.11E Simulator 253

Traffic generated

Traffic generated

Fig. 21.10 Total of packets lost (grey)/received (light grey) at destination

algorithm in order to allow the simulation of traffic among stations that can not
communicate directly. We performed some initial tests to verify if the simulator
was performing as expected, and the results were encouraging. A few improve-
ments for this simulator are being considered, as well as the release of its source
code to the scientific community for research purposes. Another possibility cur-
rently under study, is to adapt the simulator to allow simulations in the field of
Wireless Sensor Networks (WSN) – another research the authors of this chapter are
interested in. The simulator was built using the MAC and PHY layers of the standard
IEEE 802.11e, and it is intended to replace it by using a MAC layer specification
developed specifically for WSN.

References

1. H. Holma and A. Toskala, in: WCDMA for UMTS: HSPA Evolution and LTE, edited by Harri
Holma and Antti Toskala (Wiley, Chichester, 2007).
2. A. R Prasad and N. R. Prasad, in 802.11 WLANs and IP Networking, edited by Anand R. Prasad
and Neeli R. Prasad (Artech House, London, 2005).
3. O. Cabral and F. J. Velez, Event-Driven Simulation for IEEE 802.11e Optimization, IAENG
International Journal of Computer Science 35(1), 161–173 (2008).
4. E. W. Dijkstra, A note on two problems in connexion with graphs, Numerische Mathematik 1,
269–271 (1959).
5. N. Qiang, Performance analysis and enhancements for IEEE 802.11e wireless networks, IEEE
Network 19(4), 21–27 (2005).
Chapter 22
Web-Based Management of Distributed Services

George Oikonomou and Theodore Apostolopoulos

Abstract This paper presents WebDMF, a Web-based Framework for the Manage-
ment of Distributed services. It is based on the Web-based Enterprise Management
(WBEM) family of standards and introduces a middleware layer of entities called
“Representatives”. WebDMF can be integrated with existing WBEM infrastructures
and is not limited to monitoring. On the contrary, it is capable of actively mod-
ifying the run-time parameters of a managed service. Due to its abstract design,
it is suitable for the management of a variety of distributed services, such as
grids and content delivery networks. The paper includes a discussion on Web-
DMF’s design, implementation and advantages. We also present experiments on
an emulated network topology as an indication of the framework’s viability.

Keywords Common information model  distributed services management  web-

based enterprise management  WebDMF

22.1 Introduction

Legacy management approaches such as the Simple Network Management Protocol

(SNMP) [1], target single nodes and are mainly used for the management of devices.
The current paradigm of highly decentralized, distributed systems poses new chal-
lenges and increased complexity in the area of network and service management.
There is need for solutions that are better-suited for software and services. Such
solutions should also take into account the managed system’s distributed nature.

G. Oikonomou (B)
Department of Informatics, Athens University of Economics and Business, Athens, Greece
E-mail: geo@aueb.gr

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 255
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
256 G. Oikonomou, T. Apostolopoulos

We present the design, implementation and evaluation of WebDMF, a Web-based

Management Framework for Distributed services. It uses standard web technologies:
Its core is based on the Web-based Enterprise Management (WBEM) family of spec-
ifications [2–4]. It is not limited to monitoring but is also capable of modifying the
run-time parameters of a managed service. Due to its abstract design it has wide
scope and is suitable for the management of a variety of services. We have partic-
ularly studied its application for the management of grids [5] and content delivery
networks [6]. This paper is a revision of our previous work, originally published
in [7].
In this context, the paper’s contribution is as follows:
It proposes a framework for the management of distributed services, called Web-
DMF. We demonstrate how WebDMF can be integrated with existing WBEM
infrastructures and, in doing so, it can achieve its goal without need for modifi-
cations to the managed service.
It provides indications for the viability of the approach through a preliminary
performance evaluation.
Section 22.2 briefly outlines existing efforts in the field of distributed service
management. In Section 22.3 we describe WebDMF’s architectural design. Imple-
mentation details and preliminary evaluation results are presented in Section 22.4.
Finally, in Section 22.5 we discuss our conclusions.

22.2 Related Work

Existing research and development efforts investigate techniques for the manage-
ment of distributed applications and services. The Open Grid Forum’s GMA [8]
and the Relational GMA (R-GMA) [9] focus on monitoring grids. MonALISA [10]
and the CODE toolkit [11] have wider scope but still only perform monitoring.
There are some proposals that can go beyond monitoring. The Unified Grid
Management and Data Architecture (UGanDA) [12] contains an infrastructure man-
ager called MAGI. MAGI has many features but is limited to the management of
UGanDA deployments. MRF is a Multi-layer resource Reconfiguration Framework
for grid computing [13]. It has been implemented on a grid-enabled Distributed
Shared Memory (DSM) system called Teamster-G [14].
Ganglia is another noteworthy proposal [15]. It performs monitoring of dis-
tributed environments (mainly computer clusters) by adopting a hierarchical
approach and by breaking down a cluster into “federations”. Astrolabe [16] parti-
tions a distributed system into non overlapping zones. Each zone has a “representa-
tive” that is chosen automatically among the zone’s nodes through a gossip protocol.
Users can monitor a zone’s operational parameters by specifying aggregation func-
tions. Lastly, UPGRADE-CDN [17] introduces “AMonitor”, a monitoring frame-
work for content delivery networks. AMonitor uses state-of-the-art technologies but
is limited to monitoring.
22 Web-Based Management of Distributed Services 257

Thorough research reveals that, existing management systems with many fea-
tures tend to target specific distributed applications. On the other hand, generic
approaches with wider scope offer less features and are often limited to moni-
toring. Furthermore, many current research efforts are based on proprietary tech-
nologies and are limited in terms of interoperability and ease of integration with
existing infrastructures. Lastly, there are questions about the scalability of some
approaches.
Compared to those necessary efforts, the framework discussed in this paper has
wide scope while maintaining a rich set of features, including the ability to perform
modifications on the managed service. It is based on open standards and is easy to
integrate with existing WBEM infrastructures.

22.3 WebDMF: A Web-Based Management Framework

for Distributed Services

WebDMF stands for Web-based Distributed Management Framework. It treats a

distributed system as a number of nodes that are interconnected over a network
and share resources to provide services to the end user. The proposed framework’s
aim is to provide management facilities for the nodes. Through those, we achieve
management of the entire deployment.
WebDMF’s architecture is based on the “Web-based Enterprise Management”
(WBEM) family of specifications, published and maintained by the Distributed
Management Task Force (DMTF). WBEM adopts the client-server paradigm and
specifies a method for the modeling of management data, called the Core Infor-
mation Model (CIM) [2]. A WBEM server is also called a CIM Object Manager
(CIMOM) and interacts with clients through a set of well-defined request-response
HTTP packets, as specified in “CIM Operations over HTTP” [4]. The payload of
those packets is CIM data encoded in XML [3].
WebDMF nodes function as WBEM entities; clients, servers or both, depending
on their role in the deployment. The framework’s design introduces a middle-
ware layer of entities that we call “Management Representatives”. They act as
peers and form a management overlay network, which can be integrated with an
existing WBEM-based management infrastructure. Representatives act as interme-
diaries between existing WBEM clients and CIM Object Managers. In our work,
we use the terms “Management” and “Service” node when referring to those enti-
ties. Figure 22.1 displays the three management entities mentioned above, forming
a very simple topology.
This resembles the “Manager of Managers” (MoM) approach. However, in MoM
there is no direct communication between domain managers. In WebDMF, represen-
tatives are aware of the existence of their peers. Therefore, it adopts the “Distributed
Management” approach. By distributing management over several nodes throughout
the network, we can increase reliability, robustness and performance, while network
communication and computation costs decrease [18].
258 G. Oikonomou, T. Apostolopoulos

Fig. 22.1 WebDMF entities

22.3.1 Management and Service Nodes

A “Management Node” is a typical WBEM client. It is used to monitor and config-

ure various operational parameters of the distributed service. Any existing WBEM
client software can be used without modifications.
A “Service Node” is the term used when referring to any node – member of the
distributed service. For instance, in the case of a computational grid, the term can
describe an execution host. Similarly, in a content delivery network a service node
is the term used to refer to an intermediate relay node or a node hosting content.
The role of a node in a particular distributed deployment does not affect the design
of our framework.
Typically, a Service Node executes an instance of the (distributed) managed ser-
vice. As displayed in Fig. 22.2a, a WebDMF request is received by the CIMOM on
the Service Node. A WBEM provider, specifically written for the service, handles
the execution of the management operation. The existence of such a provider is a
requirement. In other words, the distributed service must be manageable through
WBEM. Alternatively, a service may be manageable through SNMP, as shown in
Fig. 22.2b. In this scenario, the node may still participate in WebDMF deployments
but some functional restrictions will apply.

22.3.2 Management Representative

The framework introduces an entity called the “Management Representative”. This

entity initially receives a request from a WBEM client (management node) and
performs management operations on the relevant service nodes. After a series of
message exchanges, it will respond to the initial request. A representative is more
than a simple ‘proxy’ that receives and forwards requests. It performs a number of
other operations including the following:
Exchanges messages with other representatives regarding the state of the system
as a whole
22 Web-Based Management of Distributed Services 259

Representative

CIM-XML SNMP

CIMOM SNMP Agent

Managed Service

Service Node
Provider Managed Service
MIB

Managed Service
CIM Schema

(a) Managed Service (b)

Fig. 22.2 Two methods of communication between the WebDMF representative and service nodes

Initial Request
CIMOM
Other
Management Representative
Node WebDMF Representative
Provider Service Node
WBEM Client CIMOM

SNMP Adapter Service Node

SNMP Agent
Representative Node

Fig. 22.3 WebDMF representative and its building blocks

Keeps a record of Service Nodes that participate in the deployment

Redirects requests to other representatives
An initial request does not state explicitly which service nodes are involved in the
management task. The functionality implemented in the representative is partly
defined by its ability to determine the destination of intermediate messages. All
intermediate message exchange is transparent to the management node as well as to
the end user.
In order to achieve the above functionality, a representative is further split into
building blocks, as shown in Fig. 22.3. It can act as a WBEM server as well as
a client. Initial requests are received by the CIMOM on the representative. They
260 G. Oikonomou, T. Apostolopoulos

are delegated to the WebDMF provider module for further processing. The module
performs the following functions:
Determines whether the request can be served locally.
If the node can not directly serve the request then it selects the appropriate
representative and forwards it.
If the request can be served locally, the representative creates a list of service
nodes that should be contacted and issues intermediate requests.
It processes intermediate responses and generates the final response.
It maintains information about the distributed system’s topology.
In some situations, a service node does not support WBEM but is only manageable
through SNMP. In this case, the representative attempts to perform the operation
using SNMP methods. This is based on a set of WBEM to SNMP mapping rules.
There are limitations since it is not possible to map all methods. However, even
under limitations, the legacy service node can still participate in the deployment.

22.3.3 Domains

In a WebDMF deployment, a representative is responsible for the management of

a group of service nodes, either on its own or in cooperation with other peers. We
use the term “Domain” when referring to such groups. The relationship between
domains and representatives is many-to-many. Thus, a representative may be respon-
sible for the management of more than one domain.
Domains are organized in a hierarchical structure. The hierarchy’s top level (root
node of the tree) corresponds to the entire deployment. The exact rationale behind
the domain hierarchy of each individual deployment can be based on a variety of
criteria. For example, a system may be separated into domains based on the geo-
graphical location of nodes or on the structure of an organization. In any case, the
hierarchy can affect the performance, ease of management and scalability of the
system. The domain hierarchical structure and the relationships between domains
and nodes are depicted in the class diagram in Fig. 22.4.

22.3.4 CIM Schemas and Operations

WebDMF defines two categories of management operations: (i) Horizontal (Cate-

gory A) and (ii) Vertical (Category B).
Horizontal Operations enable management of the WebDMF overlay network
itself. Those functions can, for example, be used to perform topology changes. The
message exchange that takes place does not involve Service Nodes. Therefore, the
managed service is not affected in any way. When a service node joins the network,
it registers itself with a representative. This can be performed by a script on the
22 Web-Based Management of Distributed Services 261

WebDMF_DomainHierarchy

0..1

WebDMF_Domain
* *
1

WebDMF_DomainRepresentativeNodes

WebDMF_DomainServiceNodes

WebDMF_Node

*
WebDMF_NodeService
*
WebDMF_NodeRepresentative

Fig. 22.4 Domains and nodes – the WebDMF core schema

service node itself or manually by a user. However, there is no automated discovery

mechanism for new service nodes.
The registration operation is very lightweight. However, its “semi-manual”
nature makes WebDMF more suitable for deployments with relatively infrequent
topology changes. It is less suitable for systems with very frequent and abrupt
topology changes (disconnecting nodes, roaming nodes), such as wireless, ad-hoc
networks.
Vertical operations read and modify the CIM schema on Service Nodes, thus
achieving management of the target application. Typical examples include:
Setting new values on CIM objects of many service nodes
Reading operational parameters from service nodes and reporting an aggregate
(e.g. sum or average)
In line with the above, we have designed two CIM Schemas for WebDMF, the
core schema (“WebDMF Core”) and the request factory. They both reside on
representative repositories.
The former schema models the deployment’s logical topology, as discussed ear-
lier. The class diagram presented in Fig. 22.4 is the actual diagram representing
WebDMF Core. Horizontal functions correspond to WBEM operations on instances
of classes declared in this schema.
The latter schema corresponds to vertical functions. Users can call WBEM meth-
ods on instances of this schema. In doing so, they can define management operations
262 G. Oikonomou, T. Apostolopoulos

that they wish to perform on the target application. Each request towards the dis-
tributed deployment is treated as a managed resource itself. For example, users can
create a new request. They can execute it periodically and read the results. They can
modify it, re-execute it and finally delete it.
In order to complete a vertical operation, the following message exchange takes
place:
The management node sends a CreateInstance() WBEM message to any
representative. This requests the creation of a new instance for class WebDMF
RequestWBEM. This instance defines the management operation that needs to
be performed on service nodes.
The representative determines whether the request can be served locally. If not, it
chooses the appropriate representative and issues a new CreateInstance()
request.
If the representative can serve the request, it generates a list of service nodes that
must be contacted, based on the values of properties of the recently generated
instance.
The representative sends the appropriate requests to service nodes. The type
of CIM operation used for those requests is also based on values of proper-
ties in the instance. This operation is usually a WBEM GetInstance() or
a ModifyInstance().
After all service nodes have been contacted and responses have been sent,
the instance on the first representative contains results. It remains available to
the user for potential modification and/or re-execution. All other intermediate
instances are deleted.
Request factory classes are generic. They are not related in any way with the CIM
schema of the managed application. This makes WebDMF appropriate for the man-
agement of a wide variety of services. Furthermore, no re-configuration is needed
with each new release of the target service.

22.4 Implementation and Performance Evaluation

22.4.1 Implementation Details

The WebDMF representative is implemented as a single shared object library file

(.so). It is comprised of a set of WBEM providers. Each one of them implements
CIM management operations for a class of the WebDMF schemas.
The interface between the CIMOM and providers complies with the Common
Manageability Programming Interface (CMPI). Providers themselves are written in
C++. This does not break CIMOM independence, as described in [19].
The representative was developed on Linux 2.6.20 machines. We used gcc 4.1.2
and version 2.17.50 of binutils. Testing took place using version 2.7.0 of the Open
Pegasus CIMOM.
22 Web-Based Management of Distributed Services 263

22.4.2 Performance Evaluation

In order to evaluate WebDMF, we installed a testbed environment using ModelNet

[20]. Results presented here have been obtained from actual code execution on an
emulated network topology and are used as an indication of the solution’s viability.
They also bring out one of WebDMF’s key features, the ability to perform changes
to the managed service.
The topology emulated by ModelNet represents a wide-area network. It con-
sists of 250 virtual nodes situated in three LANs with each LAN having its own
gateway to the WAN. The three gateways are interconnected via a backbone net-
work, with high bandwidth, low delay links. We have also installed two WebDMF
representatives (nodes R1 and R2).
In this scenario, the distributed managed service is a content delivery net-
work implemented with the OpenCDN (oCDN) open source software [21]. Service
nodes correspond to oCDN Origin nodes. Each of those registers with an oCDN
centralized control entity called “Request Routing and Distribution Management”
(RRDM). We have also designed a CIM schema and the relevant WBEM providers
that enable management of OpenCDN nodes. Figure 22.5 portrays the emulated
topology and test scenario. For clarity reasons, we omit service nodes residing in
other domains.
In our experiment, we wish to perform a change in the 200 service nodes residing
in domain “ccslab.aueb.gr”. To be more specific, we wish to set them to register with
a different RRDM. This involves changing properties of the oCDN Origin instance
on the service node’s CIMOM.
The client forms a WBEM CreateInstance() request for class WebDMF
RequestWBEM of the request factory. It is initially sent to the WebDMF Repre-
sentative (R1). The request is forwarded to R2. R2 sends ModifyInstance()
requests to the 200 service nodes. R2 reports to R1. R1 sends the final response to
the client.

R1 R2

WBEM
client

200 service nodes

ccslab.aueb.gr

Fig. 22.5 Emulated topology and test scenario

264 G. Oikonomou, T. Apostolopoulos

Table 22.1 Evaluation results

Metrics Values
Repetitions N 200
Central tendency Arithmetic mean 3.851944
Median 3.869485
Dispersion Variance 0.005281
Standard deviation 0.072670
Quartiles Q0 (min) 3.710042
Q1 3.797917
Q3 3.910481
Q4 (max) 3.997615
95% Confidence interval for the mean
From 3.841873
To 3.862016

The above experiment was repeated 200 times. Table 22.1 summarizes the results
with times measured in seconds. Consider the fact that each repetition involves 204
request-response exchanges among various nodes. Furthermore, consider that pack-
ets crossing the network are of a small size (a few bytes). The total execution time
includes the following:
Communication delays during request–response exchanges. This includes TCP
connection setup for all WBEM message exchanges.
Processing overheads on R1 and R2. This is imposed by WebDMF’s functio-
nality.
Processing at the service nodes to calculate the requested value and generate a
response.
The absolute value of the average completion time may seem rather high. However,
in general terms, processing times are minimal compared to TCP connection setup
and message exchange. With that in mind, we can see that each of the 204 request-
responses completes in less than 20 ms on average. This is normal.
After 200 repetitions we observe low statistical dispersion (variance and standard
deviation). This indicates that the measured values are not widely spread around the
mean. We draw the same conclusion by estimating a 95% confidence interval for
the mean. This indicates that the same experiment will complete in the same time
under similar network load conditions.

22.5 Conclusions

Existing monitoring solutions are generic. However, ones that offer more features
and the ability to “set” tend to have narrow scope. We wanted to design a frame-
work that would be generic enough and suitable for a wide variety of services. At
the same time, it should not be limited to monitoring. WebDMF achieves that by
22 Web-Based Management of Distributed Services 265

detaching the details of the managed service from the representative logic. Manage-
ment functions for a specific service are realised by WBEM providers on the service
nodes. Representatives unify those on a deployment scale. WebDMF has some other
noteworthy features:
It is based on WBEM. This is a family of open standards. WBEM has been
considered adequate for the management of applications, as opposed to other
approaches (e.g. SNMP) that focus on the management of devices.
It provides interoperability with existing WBEM-based management infrastruc-
tures without need for modifications.

References

1. W. Stallings, SNMP, SNMPv2, SNMPv3, RMON 1 and 2. Addison Wesley, Redwood City, CA,
1999.
2. CIM Infrastructure Specification, DMTF Standard DSP0004, 2005.
3. Representation of CIM in XML, DMTF Standard DSP0201, 2007.
4. CIM Operations over HTTP, DMTF Standard DSP0200, 2007.
5. G. Oikonomou, and T. Apostolopoulos, Using a web-based framework to manage grid
deployments, in Proceedings of The 2008 International Conference on Grid Computing and
Applications (part of WORLDCOMP 08), Las Vegas, 2008, pp. 10–16.
6. G. Oikonomou, and T. Apostolopoulos, Web-based management of content delivery networks,
in Proceedings of 19th IFIP/IEEE International Workshop on Distributed Systems: Opera-
tions and Management (DSOM). Managing Large Scale Service Deployment (MANWEEK
08), Samos, Greece, 2008, pp. 42–54.
7. G. Oikonomou, and T. Apostolopoulos, WebDMF: A web-based management framework for
distributed services, in Proceedings of The 2008 International Conference of Parallel and
Distributed Computing (part of WCE 08), Vol. I, London, 2008, pp. 593–598.
8. A Grid Monitoring Architecture, Open grid Forum GFD.7, 2002.
9. W. Cooke, et al., The Relational Grid Monitoring Architecture: Mediating Information about
the Grid, Journal of Grid Computing, 2(4), 2004, 323–339.
10. I. C. Legrand, H. B. Newman, R. Voicu, C. Cirstoiu, C. Grigoras, M. Toarta, and C. Dobre,
MonALISA: An Agent based, Dynamic Service System to Monitor, Control and Optimize
Grid based Applications, in Proceedings of Computing in High Energy and Nuclear Physics
(CHEP), Interlaken, Switzerland, 2004.
11. W. Smith, A System for Monitoring and Management of Computational Grids, in Proceedings
of International Conference on Parallel Processing (ICPP ’02), 2002, p. 55.
12. K. Gor, D. Ra, S. Ali, L. Alves, N. Arurkar, I. Gupta, A. Chakrabarti, A. Sharma, and
S. Sengupta, Scalable enterprise level workflow and infrastructure management in a grid
computing environment, in Proceedings of Fifth IEEE International Symposium on Cluster
Computing and the Grid (CCGrid ’05), Cardiff, UK, 2005, pp. 661–667.
13. P.-C. Chen, J.-B. Chang, T.-Y. Liang, C.-K. Shieh, and Y.-C. Zhuang, A multi-layer resource
reconfiguration framework for grid computing, in Proceedings of 4th International Workshop
on Middleware for Grid Computing (MGC ’06), Melbourne, Australia, 2006, p. 13.
14. T.-Y. Liang, C.-Y. Wu, J.-B. Chang, and C.-K. Shieh, Teamster-G: A grid-enabled software
DSM system, in Proceedings of Fifth IEEE International Symposium on Cluster Computing
and the Grid (CCGrid ’05), Cardiff, UK, 2005, pp. 905–912.
15. M. L. Massie, B. N. Chun, and D. E. Culler, The ganglia distributed monitoring system:
Design, implementation, and experience, Parallel Computing, 30, 817–840, 2004.
16. K. P. Birman, R. V. Renesse, and W. Vogels, Navigating in the storm: Using astrolabe to
adaptively configure web services and their clients, Cluster Computing, 9, 127–139, 2006.
266 G. Oikonomou, T. Apostolopoulos

17. G. Fortino, and W. Russo, Using p2p, grid and agent technologies for the development of
content distribution networks, Future Generation Computer Systems, 24, 180–190, 2008.
18. M. Kahani, and P. H. W. Beadle, Decentralised approaches for network management, ACM
SIGCOMM Computer Communication Review, 27(3), 36–47, 1997.
19. Common Manageability Programming Interface, The Open Group, C061, 2006.
20. A. Vahdat, K. Yocum, K. Walsh, P. Mahadevan, D. Kostic, J. Chase, and D. Becker, Scalabil-
ity and Accuracy in a Large-Scale Network Emulator, in Proceedings of 5th Symposium on
Operating Systems Design and Implementation (OSDI), December 2002.
21. OpenCDN Project [Online]. Available: http://labtel.ing.uniroma1.it/opencdn/
Chapter 23
Image Index Based Digital Watermarking
Technique for Ownership Claim and Buyer
Fingerprinting

Sarabjeet S. Bedi, Rabia Bano, and Shekhar Verma

Abstract Digital Image Watermarking is a technique for inserting information into

an image that can be later extracted for a variety of purpose including identifica-
tion, ownership authentication, and copyright protections. The goal of this paper
is to design a secure watermarking scheme in spatial domain to identify the true
buyer and to protect the ownership copyright of digital still images. The scheme
inserts a binary watermark in an original image that serves as a fingerprint for a
specific buyer. Utilizing the one-way property of the hash function, it generates
the unique buyer fingerprint. An image index concept has been used to split the
image randomly into disjoint blocks. A threshold is defined for manipulation of
pixel intensities of these blocks. The amount of manipulation in pixel intensities
absolutely depends upon the buyer fingerprint. The recovery of the watermark not
only identifies the buyer but also protects the owner’s copyright. The extracted buyer
fingerprint identifies the buyer of the original image. The scheme has the ability to
trace the buyer involved in forgery. The experimental results show that the water-
mark can be retrieved from the attacked watermarked image even if the opponent
has the complete information of watermarking scheme.

Keywords Buyer fingerprint  cryptographic hash function  Digital Watermarking 

Image key

23.1 Introduction

Information security aspects come into role when it is necessary or desirable to

protect information as it is being shared during transmission or storage from an
immediate future opponent who may present a threat to confidentiality, authenticity,

S. S. Bedi (B)
MJP Rohilkhand University, Bareilly (UP), India-243006
E-mail: dearbedi@gmail.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 267
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
268 S.S. Bedi et al.

integrity, access control, and availability. The need for information security has been
termed as security attack, mechanism, and services [1]. The various data hiding tech-
niques like cryptography, stenography, digital signatures, finger printing, have been
developed to address the information security issues but fail to provide the com-
plete solutions to protect the intellectual property rights of digital multimedia data.
The existing basket of technologies like cryptography secures the multimedia data
only during storage or transmission and not while it is being consumed [2]. Digital
Watermarking provides an answer to this limitation as the watermark continues to
be in the data throughout its usage.
Digital Watermarking (DWM) is the process of embedding information into dig-
ital multimedia contents such that the embedded information (watermark) can be
extracted later [3]. The extracted information can be used for the protection of
intellectual property rights i.e. for establishing ownership right, ensuring authorized
access and content authentication. The watermarking system can be implemented
using either of two general approaches. One approach is to transform the origi-
nal image into its frequency domain representation and embed the watermark data
therein. The second is to directly treat the spatial domain data of the host image to
embed the watermark.
According to Hartung [4] most proposed watermark method utilize the spatial
domain, this may be due to simplicity and efficiency. The spatial domain method
is about embedding watermark information directly into image pixels proposed
by [5]. These techniques embed the watermark in the LSB plane for perceptual
transparency which is relatively easy to implement but their significant disadvan-
tages includes the ease of bypassing the security they provide [5,6] and the inability
to lossy compression the image without damaging the watermark.
The methods [6, 7] extended the work to improve robustness and localization in
their technique, in which watermark is embedded by adding a bipolar M-Sequence
in the spatial domain and detection is via a modified correlation detector. But these
schemes were not very much capable to protect the watermark and also not resist
with lossy compression.
Regarding security and content authentication a new method [8] introduce the
concept of hash function, in which author insert the binary watermark into the LSB
of original image using one-way hash function. The technique is too sensitive since
the watermark is embedded into the LSB plane of the image and algorithm also does
not very resist with lossy compression. Thus the limitations of spatial domain meth-
ods are that, in general they are not robust to common geometric distortion and have
a low resistance to JPEG lossy compression. Therefore a scheme is required to fulfill
the existing gap in the use of watermarking and cryptographic techniques together.
In this paper, the robust and secure digital invisible watermarking scheme in
spatial domain is proposed. The proposed scheme combines the advantages of cryp-
tographic concept and imperceptibility feature of digital image watermarking. The
security and perceptual transparency are achieved by using cryptographic one-way
hash function and computation of threshold value respectively. For the robustness
the proposed technique does not depend upon perceptually significant regions;
rather it utilizes the concept of image key and buyer fingerprint generator. The
23 Image Index Based Digital Watermarking Technique 269

unique binary sequence serves as the buyer authenticator of a particular image. The
recovery of watermark also protects the copyrights.
The rest of the paper is organized as follows. Section 23.2 describes the concept
of generation of image key and buyer fingerprint. Section 23.3 explains the proposed
watermarking scheme with watermark insertion and extraction process. Experimen-
tal results with discussion and conclusion are given in Section 23.4 and Section 23.5
respectively.

23.2 Image Key and Buyer Fingerprint

The image key and generation of buyer fingerprint used in proposed scheme are
described as below.

23.2.1 Image Key

An Image key, I for any grayscale image, Im of size X  Y pixels (let, X D 2x

and Y D 2y / is of the same size as of image. The original image is spatially divided
into z D 2n .D 2x D 2y / disjoint blocks. Each block is denoted by an index k,
for 1  k  2n and for every block B.k/, Bi .k/ \ Bj .k/ D ∅ for 1  k 
2n ; i ¤ j . The length of blocks which contains different number of locations may
vary from each other. Each location for two-dimensional image key, I is represented
as .i; j /, where i corresponds to row for 1  i  2n and j corresponds to column
for 1  j  2n . The image key, I store the index numbers of image pixels. As
the indexes are random, so it is not possible for attacker to guess the image key.
The pixel value of blocks in the image is modified corresponding to the indexes of
the image key. Now even if the attacker knows the factor by which manipulation is
done, he/she will not be able to locate the pixels whose values are modified. The
image key is stored with the owner of the image and is used during the extraction of
the watermark.

23.2.2 Buyer Fingerprint

The Buyer fingerprint, F is a binary sequence of length 2n and will be equal to

the number of blocks. Each location of buyer fingerprint is denoted by index, k for
1  k  2n . The unique buyer fingerprint, F is generated using cryptographic hash
function. A cryptographic hash function H.S / D ff1 ; f2 ; : : :; fp g where S is string
of data of arbitrary length, fi is binary output bits of the hash function, and p is a
size of the output bit string, has the two important properties [1]. First property it
is computationally infeasible to find any input which maps to pre-specified output.
270 S.S. Bedi et al.

Second is computationally infeasible to find any two distinct inputs that map to
same output referred as collision-resistant. The generation of Buyer fingerprint is
discussed in Section 23.3.

23.3 Proposed Technique

23.3.1 Basic Idea

In this watermarking scheme, the original image, Im is divided into blocks based
on the image key. The intensity value of pixels in each block is modified depend-
ing upon the bit of watermark to get the watermarked image. Only those pixels
are modulated whose value is greater than the threshold, T .k/ for 1  k  2n .
The motivation for selecting the threshold is to increase the perceptual transparency
as it filters out the pixels having intensity values less than threshold. The thresh-
old is calculated for each block. The arithmetic mean of pixels for each group is
calculated separately which is taken as the threshold. In the extraction phase water-
marked blocks are compared with the original image block and the buyer fingerprint
is generated.

23.3.2 Generation of Watermark

The watermark is generated through the creation of unique Buyer fingerprint, F of

an original image using one-way hash function. The Buyer fingerprint is created
as F D H.X; Y; I; M / where X is an image height, Y is an image width, I is an
image key and M is a MSB array of block. The two parameters image key and MSB
array of block makes the Buyer fingerprint unique. The MSB of pixels at index, k
are summed together to form the kt h element of array M . The image key, I store
the index numbers of block of image pixels. As the indexes are generated randomly,
so it is not possible for the attacker to guess the image key. The generated Buyer fin-
gerprint is of length 2n and shall be equal to the number of blocks of original image.

23.3.3 Insertion of Watermark

The original image, Im is spatially divided into z D 2n .D2x D 2y / blocks. The

arithmetic mean of each block is calculated which is taken as threshold for that
block. As the length of the generated Buyer fingerprint is equal to the number of
blocks of original image. Therefore for each bit of watermark the intensities of the
pixels of correspond indexed blocks shall be modified. This modification is based
on threshold value of the specific block. If the watermark bit is ‘1’, then the pixels
having intensity value greater than threshold are increased by a factor, ˛. Where as
23 Image Index Based Digital Watermarking Technique 271

for the watermark bit ‘0’, no modification is recommended. The detailed algorithm
of watermark insertion is as given in the following steps:

1. For 1  k  2n
(a) Let T .k/ be the threshold of the block having index k.
(b) Suppose F .k/ is the kt h bit of watermark.
2. For 1  i  2x , 1  j  2y
(a) Let Im .i; j / be pixel intensity of original image at location .i; j /.
(b) Assume that .i; j / belongs to the block B.k/.
(c) If F .k/ D 0, then Wm .i; j / D Im .i; j /.
(d) If F .k/ D 1, then
(e) If Im .i; j / > T .k/, then Wm .i; j / D Im .i; j / C ˛ and d.k/ D d.k/ C ˛.

The factor ˛ is taken as positive throughout the insertion. The value of ˛ is chosen
so as to maintain the fidelity. The larger values will degrade the quality of the image.
From step 2e of algorithm it is clear that the factor d.k/ records the increase in the
value of intensities for each block.
The watermark insertion process is described in the block diagram as shown
in Fig. 23.1. The buyer signature, threshold and image key are the required input
attributes for watermark insertion.

23.3.4 Extraction of Watermark

This section illustrates the extraction of watermark form watermarked image, Wm .

The extraction procedure requires original image, and image key. The algorithm for
the extraction of watermark is as given in the following steps:
1. For 1  k  2n

Image Threshold
Key

Image Width (X)

Watermark Watermarked
Image Height (Y) Buyer
H(X,Y,I,M) Insertion Image
Image Key (I) Signature
Process
MSB Array (M)

Watermark Watermarked
Generation Image
Generation

Fig. 23.1 Watermark insertion process

272 S.S. Bedi et al.

a. Set values s.k/ D 0.

b. Set bit values b.k/ D 0.
c. Let T 0 .k/ be the new threshold of the group having index k.
2. For 1  i  2x ; 1  j  2y
If jWm .i; j / – Im .i; j /j > j˛j, then
Wm .i; j / D Im .i; j / C ˛.
3. If .i; j / belongs to the block B.k/ and Wm .i; j / > T 0 , then
s.k/ D s.k/ C .Wm .i; j / – Im .i; j //:
4. If ˇs.k/  d.k/ and s.k/ ¤ 0, then b.k/ D 1
Else b.k/ D 0
5. The retrieved watermark is b.k/.
The value of ˛ is used to reduce the watermarked pixel intensities (may be attacked)
to optimum value. As the value of ˛ is known, the limits of pixel values can be deter-
mined after watermarking. This fact can be used to rectify the values of attacked
pixels. The value of ˛ is utilized for the calculation of the new value of threshold in
step 2 of algorithm 2. For the smaller value ˛ there will smaller change in the value
of threshold. The difference of watermarked and original pixel is found for every
block in step 3 of algorithm 2. The damping factor ˇ.<1/ decreases the value of
s.k/ in step 4 of algorithm 2 so as to satisfy the inequality. The value of s.k/ D 0
for unaltered watermarked image. The extracted watermark, b is obtained which is
equal to inserted watermark; F as there is exact bit-wise match.
The Fig. 23.2 describes the process of extraction of the watermark using the
original image, the watermarked image, the threshold and the image key. The buyer
signature output of this phase is used to verify the ownership claim.

23.4 Results and Discussion

The proposed scheme is applied on different grayscale original images of size

128  128 pixels. The unique watermark of 128 bits and the unique image key of
size 128  128 pixels are generated. The threshold is computed for every block and
the watermark is inserted in the spatial domain. The different watermarked images
of size 128  128 pixels are obtained for different buyer fingerprint. The com-
mon image processing operations like modification, low pass filter, medium pass
filter, cropping, combination of rotation and scaling and compression are imposed
on watermarked image.
The Normalized Cross Correlation (NCC) values between original image and the
attacked watermarked image is computed to demonstrate the robustness and fidelity
of the proposed scheme.
23 Image Index Based Digital Watermarking Technique 273

Image
Key Threshold

Watermark
Watermarked Buyer
Extraction
Image Signature
Process

Owner
Claimed

Original
Image

Fig. 23.2 Extraction of the watermark

Table 23.1 NCC values of images for some common image processing operations
Attacks Scaling Scaling C Compression MPF LPF Modify Cropping
image rotation

Lena 0.9999 0.9878 0:9943 0:98461 0:96294 0:68474 0:66389

Cameraman 0.9999 0.9878 0:99523 0:9502 0:93504 0:71265 0:88474

The simulation results have been produced on various sets of images. The orig-
inal gray-scale image of “Lena” of size 128  128 pixels is taken as shown in
Fig. 23.3a. Unique image key of size 128  128, and MSB array of 128 bits are
taken as input to MD5 hash function. The 128-bit string buyer fingerprint is gener-
ated and inserted in original image which produced the watermarked image of size
128  128 as shown in Fig. 23.3b. The result shows that the watermark is invisible.
The NCC value between the original image and the watermarked image is 0.99998.
However the watermark is extracted from watermarked image. The exact bit-wise
match between extracted watermark and the inserted buyer fingerprint identifies the
true buyer of the original image.
The effect of some attacks on the watermarked image is also shown in Table 23.1
and Fig. 23.1. Table 23.2 shows the bit-wise match of inserted buyer fingerprint and
extracted buyer fingerprint. The graph of the resultant values have been plotted as
shown in Fig. 23.4.
In the low pass filter attack, a mask of 3  3 is used. The NCC value of 0.96294
is obtained between the original and the modified watermarked image whereas it is
0.98461 for median-pass filter attack. The modified watermarked image is shown in
Fig. 23.3c. An exact match of 128 bits is obtained for both the filtering operations
as illustrated in Table 23.2. The watermarked image is scaled to twice of its size as
shown in Fig. 23.3d and the measured value of NCC is 0.9999. For the combined
274 S.S. Bedi et al.

Fig. 23.3 (a) Original Lena

image. (b) Watermarked Lena
image. (c)–(h) Results of
watermarked images with
some attacks: (c) low pass
filtered; (d) scaled; (e) rotated
(17ı ); (f ) cropped from top
and left border; (g) random
modification; (h) JPEG lossy
compression (quality factor
D 90)

attack of rotation of 17ı followed by resizing to the size of 128128 pixels, the NCC
value between original and modified watermarked image is 0.9878 (Fig. 23.3e).
The 126–127 bits are recovered for scaling and combined attack. The cropped and
randomly modified images are shown in Fig. 23.3f–g. In case of modification, the
watermarked image has been tampered at specific locations by changing the pixel
values. In case of severe manipulation to pixel intensities, a bit-wise match of 120–
23 Image Index Based Digital Watermarking Technique 275

Table 23.2 BIT-wise match between inserted and extracted buyer fingerprint (LENA IMAGE)
Attacks ˇ Scaling Scaling C rotation Compression MPF LPF Modify Cropping

1 127 126 128 128 128 120 123

0.9 128 128 128 128 128 128 128

Fig. 23.4 Wise match between the inserted and extracted buyer signature for the image, “Lena”

126 bits is obtained. With the use of damping factor of 0.9, exact 128 bits is obtained
for the buyer fingerprint and the value of NCC is 0.68474. In case of cropping the
NCC value becomes 0.66389. In a rigorously cropped image 3–4 bits of inserted
Buyer fingerprint are lost which can be recovered by using a damping factor of 0.9.
The robustness against lossy JPEG compression with quality factor 90 is demon-
strated in Fig. 23.3h and the NCC value 0.9843 is obtained with all the 128 bits of
Buyer fingerprint are recovered.
The NCC values in Table 23.1 shows that scaling attack and the combined attack
of scaling and rotation as the pixel values are not changed intensely at the center
of the image, so there is not much distortion in the NCC values that are obtained.
In the Low Pass Filter (LPF) attack only the pixels having lower intensities are
retrievable, so depending on the image if the LPF’s threshold for a block of the
image is higher than the threshold assumed in the watermark insertion algorithm
then the NCC value is higher than that obtained if the threshold of the LPF is lower
than the one assumed. In the Medium Pass Filter (MPF) attack only the pixels hav-
ing medium intensities are retrievable, so depending on the image if the threshold
assumed in the watermark insertion algorithm lies towards the lower range of the
MPF’s thresholds for a block of the image then the NCC value is higher than that
276 S.S. Bedi et al.

obtained if the assumed threshold lies towards the higher threshold of the MPF. In
the modification and cropping attacks as some of the pixels are tampered badly the
original pixel values cannot be identified. Therefore results demonstrates that pro-
posed scheme is more robust to geometric attacks and compression, whereas robust
to modification and cropping.

23.5 Conclusion

The proposed watermarking technique is for copyright protection and buyer finger-
printing. The image key and the unique watermark are the key features for sustaining
the security of algorithm. The watermark generated uses the property of the hash
function which is an important feature in making the watermark cryptographically
secure. When the positive value of ˛ is used the watermark increased the intensity
of an image block. An attacker who is known to the watermarking process might
intentionally utilize this fact and try to remove the watermark. For this purpose the
attacker must know the image key. The attacker cannot guess the secret image key as
the image key has been kept secret and the indexes have been generated randomly.
The watermark extraction algorithm rectifies the pixel values in case the attacker
increases or decreases the pixel values. An invalid image key would be unsuccessful
in forging the watermark. The technique survives with common image transforma-
tions as well as intentional attacks, maintaining the objective of buyer fingerprinting
and ownership claim.

References

1. W. Stallings, “Cryptography and Network Security: Principles and Practices,” Pearson Educa-
tion, Inc., NJ, 3rd Ed., 2005, ISBN 81–7808–902–5.
2. S. S. Bedi and S. Verma, “Digital Watermarking Technology: A Demiurgic Wisecrack Towards
Information Security Issues,” Invertis Journal of Science and Technology, vol. 1, no. 1,
pp. 32–42, 2007.
3. A. Kejariwal, “Watermarking,” Magazine of IEEE Potentials, October/November, 2003,
pp. 37–40.
4. F. Hartung and M. Kutter, “Multimedia Watermarking Techniques,” Proceedings of IEEE,
vol. 87, no. 7, pp. 1079–1106, July 1999.
5. M. Yeung and F. Mintzer, “Invisible watermarking for image verification,” Journal of Electric
Imaging, vol. 7, no. 3, pp. 578–591, July 1998.
6. R. Wolfgang and E. Delp, “Fragile watermarking using the VW2D watermark,” Proceedings of
the IS & T/SPIE Conference on Security and Watermarking of Multimedia Contents, pp. 204–
213, San Jose, CA, January 1999.
7. J. Fridrich, “Image watermarking for temper detection,” Proceedings of the IEEE International
Conference on Image Processing, vol. 2, pp. 404–408, Chicago, IL, October 1998.
8. P. W. Wong and N. Memon, “Secret and Public key Image Watermarking Schemes for Image
Authentication and Ownership Verification,” IEEE Transaction on Image Processing, vol. 10,
no. 10, October 2001.
Chapter 24
Reverse Engineering: EDOWA Worm Analysis
and Classification

Madihah Mohd Saudi, Emran Mohd Tamil, Andrea J Cullen,

Mike E Woodward, and Mohd Yamani Idna Idris

Abstract Worms have become a real threat for computer users for the past few years.
Worm is more prevalent today than ever before, and both home users and system
administrators need to be on the alert to protect their network or company against
attacks. It is coming out so fast these days that even the most accurate scanners
cannot track all of the new ones. Indeed until now there is no specific way to classify
the worm. To understand the threats posed by the worms, this research had been
carried out. In this paper the researchers proposed a new way to classify the worms
which later is used as the basis to build up a system which is called as the EDOWA
system to detect worms attack. Details on how the new worm of classification which
is called as EDOWA worm classification is produced are explained in this paper.
Hopefully this new worm classification can be used as the basis model to produce a
system either to detect or defend organization from worms attack.

Keywords Classification  worm analysis  payload  worm classification

24.1 Introduction

Computer worm can caused millions dollars of damage by infecting hundreds and
thousands of host in a very short period of time. A computer worm is a computer
program or a small piece of software that has the ability to copy itself from machine
to machine. It uses computer networks and security holes to replicate itself. Com-
puter worm is classified as a highly threat to the information technology world.
McCarthy [1] defines computer worm as a standalone malicious code program that
copies itself across networks. Meanwhile Nachenberg [2] stated that the computer
worm is a program that is designed to copy itself from one computer to another,
dominate some network medium such as through email. The computer worm would

M. M. Saudi (B)
Faculty Science & Technology, Universiti Sains Islam Malaysia (USIM), Nilai, Malaysia
E-mail: madihah@usim.edu.my

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 277
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
278 M. M. Saudi et al.

infect as many machines as possible on the network. The prototypical computer

worm infects a target system only once; after the initial infection, the worm attempts
to spread to other machines on the network.
While there are thousands of variations of computer worms, the classification
of computer worm can be done via several ways Nazario [3] proposed a function
structure framework that consist six components. The components are reconnais-
sance capabilities, specific attack capabilities, a command interface, communication
capabilities, intelligence capabilities and unused attack capabilities. The framework
mainly predicts the future research on network worms. Another form of classifica-
tion of computer worm Weaver et al. [4], they classified computer worm into five
major classifications: target discovery, carrier, activation, payload and attackers. As
for Kienzle et al. [5], they classified computer worm into three basic classes by prop-
agation strategies. The computer worms are classified into e-mail worm, windows
file-sharing worm and traditional worm.

24.2 Method of Testing

In order to produce a new worm classification, the researchers’ had conducted

few testing and researches. A worm code analysis laboratory is build to test and
analyze the worm. A controlled laboratory environment is built to conduct the
testing. This laboratory is not connected to the real network. Three machines
were used and connected in LAN using a hub. Figure 24.1 below illustrates the
laboratory.

Fig. 24.1 Worm code analysis lab architecture

24 Reverse Engineering: EDOWA Worm Analysis and Classification 279

24.2.1 Worm Analysis Process

For a new computer worm, the main purpose of analyzing it is to know the intention
of the code. As for computer worm that has been released in real time, the worm
analysis process is for verification of what the worm source code intention and to
verify as what has been published in anti virus website or CERT website or it might
be a new worm feature. The analysis techniques can be divided into two techniques,
either the static analysis or dynamic analysis. Before loading the computer worm
specimen into the machine, researchers make sure all the preparation and the veri-
fication has been done. While conducting the analysis, all the process is document
in writing. A written record of analytic techniques and the computer worm action is
useful in understanding how the computer worm works, tracing through its function
in a repeatable fashion and improving the worm analyst skills.

24.2.2 Loading Specimen

When all the preparation and the verification have been done, the lab is disconnected
from any production network. The USB thumb drive is used to transfer the computer
worm specimen onto the lab system.
Once the specimen already placed in the lab system, the analysis can be carried
out. To determine the purpose and the capabilities of this piece of code, researchers
can used either the static analysis or dynamic analysis.

24.2.2.1 Static Analysis

Static analysis is also known as white box analysis. It involves analyzing and under-
standing source code. If only binary code available, the binary code has to be
compiled to get the source code. White box analysis is very effective in finding
programming errors and implementation errors in software and to know the flow
of the program. The static analysis looks at the files associated with the computer
worm on the hard drive without running the program. With static analysis, the gen-
eral idea of the characteristics and purpose of the computer worm can be analyzed.
The static analysis phase involves antivirus checking with research, analyzing the
strings, looking for scripts, conducting binary analysis, disassembling and reverses
compiling.

Antivirus Checking with Research

When computer worm specimen already copied to the testing machine, researchers
will run the antivirus to check if the antivirus installed detects anything. If the
antivirus detects the computer worm, check the name of the computer worm and
280 M. M. Saudi et al.

search it in any antivirus website for further information. If the computer worm is
in compressed or archived form, researchers will open the archive to get its con-
tent. The researchers need to verify if the information available from the antivirus
website is correct.

String Analysis

The strings that extracted from the computer worm could help the researchers to
know more about the computer worm characteristics. A few tools such as TDS3 and
Strings.exe (from Sysinternal) were used to extract the strings. The information that
could be retrieved from the extracted strings are consist of the worm specimen’s
name, user dialog, password for backdoors, URLs associated with the malware,
email address of the attacker, help or command-line options, libraries, function calls
and other executables used by the malware.

Looking for Script

The language written for the computer worm can be identified based on strings
extracted from it. The following Table 24.1 gives some clues:

Disassemble Code

Disassemble and debugger is used to convert a raw binary executable into assembly
language for further analysis. Researchers use the tools that have been listed in
Appendix A to disassemble and debug the computer worm.

24.2.2.2 Dynamic Analysis

Dynamic analysis involves executing the computer worm and watching its actions.
The computer worm is activated on a controlled laboratory system.

Table 24.1 Scripting language

Scripting language Identifying characteristics inside File’s common suffix
the file

Bourne shell scripting Starts with line !#/bin/sh .sh

language
Perl Start with line !#usr/bin/perl .pl, .perl
JavaScript Includes the word javascript or .js, .html, .htm
JavaScript, especially in the form
<Script language = “JavaScript”>
Visual basic script Includes the word VBScript, or .vbs, .html, .htm
(VBScript) characters vb scattered throughout
the file
24 Reverse Engineering: EDOWA Worm Analysis and Classification 281

Monitoring File Activities

Most computer worms read from or write to the file system. It might attempts to
write files, altering existed programs, adding new files or append itself to the file
system. By using tool such as Filemon all actions associated with opening, reading,
writing, closing and deleting files can be monitored.

Monitoring Process

The monitoring tool such as Prcview v3.7.3.1 or Process Explorer, displays each
running program on a machine, showing the details of what each process is doing.
With this kind of tool the files, registry keys and all of the DLLs that each process
has loaded can be monitored. For each running process, the tool displays its owner,
its individual privileges, its priority and its environment variables.

Monitoring Network Activities

From a remote machine which will be in the same LAN with the infected testing
machine, the port scanner, Nmap program and a sniffer will be installed. The port
scanner and Nmap program are used to monitor the listening port. A sniffer will
be installed to sniff the worm traffic. All of the related tools like Ethereal, NeWT
and TDS-3 use the sniffer. Using the sniffer, details of individual packets and all
packets transmitted across the LAN can be monitored. As for the local network
monitoring tool (TDIMon), it will monitor and records all request to use the net-
work interface and show how the worm grabbed the network resources and used
them.
The computer worm might have placed the network interface in promiscu-
ous (broadcast) mode which allowed it to sniff all packets from LAN. To deter-
mine if the infected machine in promiscuous mode state of the interface, run the
Promiscdetect.exe tool.

Monitoring Registry Access

The registry need to be monitored as the registry is the hierarchical database con-
taining the configuration of the operating system and most programs installed on the
machine. The monitoring registry access can be done by using the Regmon.

24.3 EDOWA Classification

After the analysis done in the laboratory, it leads researchers to produce a new
classification for the EDOWA system. A proposal of the classification of worm is
made. This classification is based on several factors: infection, activation, payload,
282 M. M. Saudi et al.

Fig. 24.2 EDOWA worm classification

operating algorithms and propagation. Figure 24.2 is an overview of the EDOWA

classification.
Infection is the first step before a worm infects a computer. The activation is
a mechanism that will activate a worm. Payload is a code that carries a destructive
mechanism of a worm. Operating algorithms is a techniques used to avoid detection.
Finally, the propagation mechanisms are how the worm spread to reproduce.
Sections below will elaborate more details on the EDOWA classification.

24.3.1 Infection

Infection is the phase on how a computer gets infected by a worm. From the research
of eight classifications that is available, we found out that only one research made
infection as the first phase. Albanese et al. [6] says that infection refers to how a
worm gains initial control of a system. Worms rely on two general methods to infect
a host. Either they exploit an error in software running on a system, or they are the
result of some action taken by a user. Here we proposed two techniques:
24 Reverse Engineering: EDOWA Worm Analysis and Classification 283

24.3.1.1 Host

Host is a mechanisms needed by the worm copy itself to a new systems that are not
yet been infected. It cannot propagate autonomously across the network. Host com-
puter worms where the original terminates itself after launching a copy on another
host so there is only one copy of the worm running somewhere on the network at
any given moment. It requires human help to move from one machine to another.
CD, Floppy Disks, USB (thumb-drive and external hard disk) and File are the most
common host available now.

24.3.1.2 Network

Network is the fastest way in moving worm. It consists of multiple parts, each run-
ning on different machines and possibly performing different actions also using
the network for several communication purposes. Propagating from one machine
to another is only one of those purposes. It can infect a computer without human
interaction. Most simply copy themselves to every computer with which the host
computer can share data. Most Windows networks allow machines within defined
subgroups to exchange data freely, making it easier for a worm to propagate itself.

24.3.2 Activation

For this classification, activation is defined as a trigger mechanism of a worm. This

phase is where the worm enters the host once it found a machine. According to
Nazario [3] in his book “Defense and Detection Strategies against Internet Worms”,
stated these are used to launch an attack against an identified target system.

24.3.2.1 No Activation

Worm with no activation will just stay in the computer doing nothing. It just used
up some hard disk space.

24.3.2.2 Human Trigger

Human trigger is the slowest activation mechanisms. Usually this approach use
worm that propagates using emails. The social engineering technique is used to
attract user to click on the file to activate the worm [7]. According to Christoffersen
et al. [8], some worms are activated when the user performs some activity, like reset-
ting the machine, logging onto the system and thereby running the login scripts or
executing a remotely infected file. Evidently, such worms do not spread very rapidly.
284 M. M. Saudi et al.

24.3.2.3 Schedule Process

Based on Weaver et al. [4], the second fastest worms activate is by using scheduled
system processes. Schedule process is an activation that is based on specific time
and date. Many desktop operating systems and applications include auto-updater
programs that periodically download, install and run software updates.

24.3.2.4 Self Activation

The worms that are fastest activated are able to initiate their own execution by
exploiting vulnerabilities in services that are always on and available (e.g., Code
Red [9] exploiting IIS Web servers) or in the libraries that the services use (e.g.,
XDR [10]). Such worms either attach themselves to running services or execute
other commands using the permissions associated with the attacked service.

24.3.2.5 Hybrid Launch

Hybrid Launch uses the combination of two or more activation mechanism to launch
a worm. ExploreZip [2] is an example of a hybrid-launch worm. ExploreZip send
an e-mail that required a user to launch the infected attachment to gain control of
the system. Once running on the computer system, ExploreZip would automati-
cally spread itself to other computers over the peer-to-peer network. These targeted
machines would then become infected on the next reboot without any known user
intervention.

24.3.3 Payload

For this classification, payload is defined as a destructive mechanism of a worm. A

payload is code designed to do more than spread the worm. Many worms have been
created which are only designed to spread, and don’t attempt to alter the systems
they pass through.

24.3.3.1 No Payload

Worm with no payload does not do any harm to the computer system. This kind
of worm will just propagate without infecting any destructive mechanisms to the
computer.
24 Reverse Engineering: EDOWA Worm Analysis and Classification 285

24.3.3.2 Installing Backdoor

Backdoor is a term used to describe a secret or undocumented means of getting

into a computer system. Many programs have backdoors placed by the programmer
to allow them to gain access to troubleshoot or change the program. Some back-
doors are placed by hackers once they gain access to allow themselves an easier
way in next time or in case their original entrance is discovered. Example of back-
door attack is the worm called Blaster [11] that used the backdoor mechanism to
transfer the worm payload to newly infected systems.

24.3.3.3 Denial of Services

A denial of service (DoS) attack floods a network with an overwhelming amount of

traffic, slowing its response time for legitimate traffic or grinding it to a halt com-
pletely. The more common attacks use built-in “features” of the TCP/IP protocol to
create exponential amounts of network traffic. Example of DoS attack is the well-
known worm called Code Red [9] was programmed to unleash a denial-of-service
attack on the Whitehouse.gov that targeting the actual Whitehouse.gov IP address.

24.3.3.4 Destructive

This will do harm to the machine or the host. According to Shannon et al. [12],
Witty worm deletes a randomly chosen section of the hard drive, which, over time,
renders the machine unusable.

24.3.3.5 Phishing

Phishing [13] is a criminal activity using social engineering techniques. Phishers

attempt to fraudulently acquire sensitive information, such as usernames, passwords
and credit card details, by masquerading as a trustworthy entity in an electronic com-
munication. eBay and PayPal are two of the most targeted companies, and online
banks are also common targets. Phishing is typically carried out by email or instant
messaging, and often directs users to give details at a website, although phone con-
tact has been used as well. Attempts to deal with the growing number of reported
phishing incidents include legislation, user training, and technical measures.

24.3.4 Operating Algorithms

Operating algorithms is defined as a detecting techniques used by worms to avoid

detection. Among the eight classifications that are available, we found out that only
286 M. M. Saudi et al.

one research have operating algorithms in their classification. Albanese et al. [6]
classified it as survival. Operating algorithms are the mathematical and logical ways
that a worm attempts to avoid detection. It can be categorized as:

24.3.4.1 Polymorphic

A polymorphic worm is a worm that changes all part of their code each time they
replicate, this can avoid scanning software. Kruegel et al. [14] paper defined poly-
morphic worms as a worm that is able to change their binary representation as part
of the spreading process. It can be achieved by using self-encryption mechanisms or
semantics-preserving code manipulation techniques. As a consequence, copies of a
polymorphic worm might no longer share a common invariant substring of sufficient
length and the existing systems will not recognize the network streams containing
the worm copies as the manifestation of a worm outbreak.

24.3.4.2 Stealth

Stealth worm use a concealment mechanisms. It spread slow, evokes no irregular

communication pattern and spread in an approach that makes detection hard. Chee-
tancheri [15] stated in his thesis that the goal of stealth worm is to spread to as
many hosts as possible without being detected. However, once such a worm has
been detected, manual means of mitigation are possible.

24.3.4.3 Terminate and Stay Resident

Terminate and stay resident (TSR) worm exploit a variety of techniques to remain
resident in memory once their code has been executed and their host program has
terminated. These worms are resident or indirect worm, known as such because they
stay resident in memory, and indirectly find files to infect as they are referenced by
the user.

24.3.4.4 Anti Anti-virus

Anti anti-virus will corrupt the anti-virus software by trying to delete or change
the anti-virus programs and data files so the anti-virus does not function properly.
According to Nachenberg [16], anti anti-virus or are usually called retroviruses, are
computer viruses that attack anti-virus software to prevent themselves from being
detected. Retroviruses delete anti-virus definition files, disable memory resident
anti-virus protection and attempt to disable anti-virus software in any number of
ways.
24 Reverse Engineering: EDOWA Worm Analysis and Classification 287

24.3.5 Propagation

Propagation is defined as a worm that spread itself to another host or network. After
researching the propagation issue, we strongly believe that there are two ways for a
worm to reproduce itself: scanning and passive.

24.3.5.1 Scanning

Scanning is a method used by worms to find their victims. We strongly agree with
the method proposed by Weaver et al. [4]. There are two possible scanning method
that is random scanning and sequential scanning.

Random Scanning

It is the most popular method where the worm simply picks a random IP address
somewhere in the Internet Address space and then tries to connect to it and infect it.
Example of a random scanning worm is the Blaster [11] that picks a random num-
ber to determine whether to use the local address it just generated or a completely
random one.

Sequential Scanning (Hitlist)

The worm releaser scans the network in advance and develops a complete hitlist
of all vulnerable systems on the network. According to Staniford [17], the worm
carries this address list with it, and spreads out through the list.

Passive

Worms using a passive monitoring technique are not actively searching for new
victims. Instead, they are waiting for new targets to contact them or rely on the user
to discover new targets. Christoffersen et al. [8] says passive worms tend to have a
slow propagation rate, they are often difficult to detect because they generate modest
anomalous reconnaissance traffic.

24.4 Conclusion

This new classification is produced based on the research and testing that have been
done in the laboratory. The classifications are divided into five main categories:
Infections, Activation, Payload, Operating Algorithms and Propagation. Efficient
288 M. M. Saudi et al.

Detection of Worm Attack (EDOWA) system is produced based on this classifica-

tion. This EDOWA system is not discussed in this paper. Hopely this paper can be
used as the basis model for worm classification and can be used for other upcoming
research.

References

1. L. McCarthy, “Own Your Space: Keep Yourself and Your Stuff Safe Online (Book),” Addison-
Wesley, Boston, MA, 2006.
2. C. Nachenberg, “Computer Parasitology,” Proceedings of the Ninth International Virus
Bulletin Conference, September/October 1999, pp. 1–25.
3. J. Nazario, “Defense and Detection Strategies against Internet Worms” (BOOK), Artech House
Inc., 2003. Or a paper entitles “The Future of Internet Worm” by Nazario, J., Anderson, J.,
Wash, R., and Connelly, C. Crimelabs Research, Norwood, USA, 2001.
4. N. Weaver, V. Paxson, S. Staniford and R. Cunningham, “A Taxonomy of Computer
Worms,” Proceedings of the ACM CCS Workshop on Rapid Malcode (WORM), pp. 11–18,
2003.
5. D.M. Kienzle and M.C. Elder, “Recent Worms: A Survey and Trends,” Proceedings of the
ACM CCS Workshop on Rapid Malcode (WORM), pp. 1–10, 2003.
6. D.J. Albanese, M.J. Wiacek, C.M. Salter and J.A. Six, “The Case for Using Layered Defenses
to Stop Worms (Report style),” UNCLASSIFIED-NSA Report, pp. 10–22, 2004.
7. C.C. Zou, D. Towsley and W. Gong, “Email worm modeling and defense,” Computer
Communications and Networks, ICCCN 2004, pp. 409–414, 2004.
8. D. Christoffersen and B.J. Mauland, “Worm Detection Using Honeypots (Thesis or Disser-
tation style),” Master dissertation, Norwegian University of Science and Technology, June
2006.
9. H. Berghel, “The Code Red Worm: Malicious software knows no bounds,” Communication of
the ACM, vol. 44, no. 12, pp. 15–19, 2001.
10. CERT. CERT Advisory CA-2002–25 Integer Overflow in XDR Library, http://www.
cert.org/advisories/ca-2002–25.html
11. M. Bailey, E. Cooke, F. Jahanian, D. Watson and J. Nazario, “The Blaster Worm: Then and
Now,” IEEE Security & Privacy, vol. 3, no. 4, pp. 26–31, 2005.
12. C. Shannon and D. Moore, “The Spread of the Witty Worm,” IEEE Security & Privacy, vol. 2,
no. 4, pp. 36–50, 2004.
13. A. Tsow, “Phishing With Consumer Electronics: Malicious Home Routers,” 15th International
World Wide Web Conference (WWW2006), Edinburgh, Scotland, May 2006.
14. C. Kruegel, E. Kirda, D. Mutz, W. Robertson and G. Vigna, “Polymorphic Worm detection
using structural information of executables,” 8th International Symposium on Recent Advances
in Intrusion Detection (RAID), 2005.
15. S.G. Cheetancheri, “Modeling a computer worm defense system (Thesis or Dissertation
style),” Master dissertation, University of California, 1998.
16. C. Nachenberg, “The Evolving Virus Threat,” 23rd NISSC Proceedings, Baltimore, Maryland,
2000.
17. S. Staniford, President of Silicon Defense. “The Worm FAQ: Frequently Asked Questions on
Worms and Worm Containment,” The Worm Information Center, 2003
Chapter 25
Reconfigurable Hardware Implementation
of a GPS-Based Vehicle Tracking System

Adnan Yaqzan, Issam Damaj, and Rached Zantout

Abstract In this chapter, we build on a recently produced VTS (The Aram Locator)
offering a SOC replacement of the microcontroller-based implementation. Although
the microcontroller-based system has acceptable performance and cost, an FPGA-
based system can promise less cost and a more cohesive architecture that would
save processing time and speeds up system interaction. The performance of the
proposed implementations is evaluated on different FPGAs. The suggested designs
show enhancement in terms of speed, functionality, and cost.

Keywords Hardware Implementation  Vehicle Tracking System  GPS-Based 

Reconfigurable  FPGA

25.1 Introduction

After a great evolution, reconfigurable systems fill the flexibility, performance and
power dissipation gap between the application specific systems implemented with
hardwired Application Specific Integrated Circuits (ASICs) and systems based on
standard programmable microprocessors. Reconfigurable systems enable extensive
exploitation of computing resources. The reconfiguration of resources in different
parallel topologies allows for a good matching with the inherent intrinsic parallelism
of an algorithm or a specific operation. The reconfigurable systems are thus very
well-suited in the implementation of various data-stream, data-parallel, and other
applications.
The introduction of a new paradigm in hardware design called Reconfigurable
Computing (RC ) offers to solve any problem by changing the hardware configu-
rations to offer the performance of dedicated circuits. Reconfigurable computing

I. Damaj (B)
Electrical and Computer Eng’g Dept, Dhofar University, P.O. Box 2509, 211 Salalah, Oman
E-mail: i damaj@du.edu.om

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 289
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
290 A. Yaqzan et al.

enables mapping software into hardware with the ability to reconfigure its con-
nections to reflect the software being run. The ability to completely reprogram the
computer’s hardware implies that this new architecture provides immense scope for
emulating different computer architectures [1, 2].
The progression of field programmable gate arrays (FPGAs) RCs has evolved to
a point where SOC designs can be built on a single device. The number of gates and
features has increased dramatically to compete with capabilities that have tradition-
ally been offered through ASIC devices only. FPGA devices have made a significant
move in terms of resources and performance. The contemporary FPGAs have come
to provide platform solutions that are easily customizable for system connectivity,
digital signal processing (DSP), and data processing applications. Due to the impor-
tance of platform solutions, leading FPGA vendors are coming up with easy to-use
design development tools [3, 4].
As the complexity of FPGA-based designs grows, a need for a more efficient
and flexible design methodology is required. Nowadays, hardware implementations
have become more challenging, and the device densities have increased at a pace that
such flows have become cumbersome and outdated. The need for a more innovative
and higher-level design flow that directly incorporates model simulation with hard-
ware implementation is needed. One of the modern tools (also used in the proposed
research) is Quartus II, started with Altera. Quartus II is a compiler, simulator,
analyzer and synthesizer with a great capability of verification and is chosen to be
used for this implementation. It can build the verification file from the input/output
specification done by the user. Quartus II design software provides a complete, mul-
tiplatform design environment that easily adapts to your specific design needs. It is
a comprehensive environment for system-on-a-programmable-chip (SOPC) [5, 6].
An application that needs real-time, fast, and reliable data processing is GPS-
based vehicle tracking. In this chapter, we build on a recently produced VTS (The
Aram Locator) offering a SOC replacement of the microcontroller-based implemen-
tation. Although the microcontroller-based system has acceptable performance and
cost, an FPGA-based system can promise less cost and a more cohesive architecture
that would save processing time and speeds up system interaction.
This chapter is organized so that Section 25.2 presents the currently avail-
able existing microprocessor-based VTS and its proposed update. Section 25.3
presents different designs and implementations with different levels of integration.
In Section 25.4, performance analysis and evaluation of results are presented.
Section 25.5 concludes the chapter and sheds light by summarizing the achieve-
ments described in the chapter and suggesting future research.

25.2 Upgrading the Aram Locator GPS System

One recently implemented VTS is the Aram Locator [5, 7]. It consists of two main
parts, the Base Station (BS ) and the Mobile Unit (MU ). The BS consists of a PIC
Microcontroller based hardware connected to the serial port of a computer. The MU
25 Reconfigurable Hardware Implementation of a GPS-Based Vehicle Tracking System 291

Fig. 25.1 Modules of the FPGA system

is a self-contained PIC Microcontroller based hardware and a GPS module. The

latter would keep track of all the positions traversed by the vehicle and records
them in its memory. The system has a great storage capacity, and could perform a
significant recording with a considerable sampling rate. The mobile unit (MU ) of
the addressed Aram Locator consists of two communicating microcontrollers inter-
faced with memory. There is also a GPS unit and RF transceiver (simply sketched
in Fig. 25.1) [7–9].
The processor of the Aram follows a sequential procedure of different parallel
components. For instance, although two processes P1 and P 2 hardware both exist
parallel, these processes run sequentially. An interrupt, sent by the I2C controller, is
used to activate one of the processes (P1 or P 2/. But the inner part of the processes
contain several parallel operations like bit assignments and selectors (corresponding
to if-statements). In Fig. 25.2, the state diagram describes the behavior of a system,
some part of a system, or an individual object (See Fig. 25.3).

25.3 The FPGA-Based Aram System

The microcontrollers make use of the same memory using a specific protocol.
The system is performing properly and has a big demand in the market. However,
FPGAs promise a better design, with a more cohesive architecture that would save
processing time and speeds up system interaction.
The two microcontrollers along with memory would be incorporated into or bet-
ter supported with a high-density PLD. This will transform the hard slow interface
between them into a faster and reliable programmable interconnects, and therefore
makes future updates simpler. This design estimated to save a considerable per-
centage of the overall cost of one working unit. For a large number of demands,
there would be a significant impact on production and profit. Hence, PLDs such as
FPGAs are the way, for a better design in terms of upgradeability and speed, and it
292 A. Yaqzan et al.

Fig. 25.2 The state diagram of the FPGA system

Fig. 25.3 Block diagram of the FPGA system

is a promising advancement for the production cost and revenue. The block diagram
of the FPGA system is depicted in Fig. 25.3.
Hiding the detailed architecture of the underlying FPGA, The proposed system
is of two communicating processes, P1 and P 2, along with a shared memory. In
addition to the FPGA-based system, the GPS antenna and the mobile unit play sig-
nificant roles. The memory block of the microcontroller-based design is replaced by
hardware entity controlled by the I 2C .
25 Reconfigurable Hardware Implementation of a GPS-Based Vehicle Tracking System 293

Table 25.1 The parameters sent by the GPS

NMEA Description

HPGGA Global positioning system fixed data

GPGLL Geographic position-latitude/longitude
GPGSA GNSS DOP and active satellites
GPGSV GNSS satellites in view
GPRMC Recommended minimum specific GNSS data
GPVTG Course over ground and ground speed
GPMSS Radio-beacon signal-to-noise ratio, signal strength, frequency, etc.
GPZDA PPS timing message (synchronized to PPS)

25.3.1 The Intersystem Process 1

This process has to deal with the message received from the GPS. The default com-
munication parameters for NMEA (the used protocol) output are 9,600 bps baud rate,
eight data bits, stop bit, and no parity. The message includes information messages
as shown in Table 25.1.
$GPGGA,161229.487,3723.2475,N,12158.3416,W,1,07,1.0,9.0,M,,,,0000*18
$GPGLL,. . . $GPGSA,. . . $GPGSV,. . . $GPGSV,. . .
$GPRMC,161229.487,A,3723.2475,N,12158.3416,W,0.13,309.62,120598,*10,
$GPVTG,. . . $GPMSS,. . . $GPZDA,. . .
From these GPS commands, only necessary information is selected (i.e. longi-
tude, latitude, speed, date, and time). The data needed is found within the commands
RMC and GGA; others are of minor importance to the FPGA. The position of the
needed information is located as follows:

$GPRMC: < time > , < validity > , < latitude > , latitude
hemisphere, < longitude > , longitude hemisphere, < speed > , < course over
ground > , < date > , magnetic variation, check sum [7]

$GPGGA, < date > , latitude, latitude hemisphere, longitude, longitude hemi-
sphere, < GPS quality > , < # of satellites > , horizontal dilution, < altitude > ,
Geoidal height, DGPS data age, Differential reference, station Identity (ID), and
check sum. This information is stored in memory for every position traversed.
Finally and when the VTU reaches its base station (BS), a large number of positions
is downloaded to indicate the route covered by the vehicle during a time period and
with a certain download speed.
Initially, a flag C is cleared to indicate that there’s no yet correct reception of
data. The first state is “Wait for GPS Parameters”, as mentioned in the flow chart,
there’s a continuous reception until consecutive appearance of the ASCII codes of
“R, M, C” or “GGA” comes in the sequence. For a correct reception of data, C is
set (i.e. C D “1”), indicating a correct reception of data, and consequently make
the corresponding selection of parameters and saves them in memory. When data
294 A. Yaqzan et al.

storing ends, there is a wait state for the I 2C interrupt to stop P1 and start P 2,
P 2 download the saved data to the base station (BS). It is noted that a large num-
ber of vehicles might be in the area of coverage, and all could ask for reserving the
channel with the base station; however, there are some predefined priorities that are
distributed among the vehicles and therefore assures an organized way of commu-
nication. This is simply achieved by adjusting the time after which the unit sends its
ID when it just receives the word “free”.

25.3.2 The Intersystem Process 2

The Base station is continuously sending the word “free”, and all units within the
range are waiting to receive it and acquire communication with the transceiver. If the
unit received the word “free”, it sends its ID number, otherwise it resumes waiting. It
waits for acknowledge, if Acknowledge is not received, the unit sends its ID number
and waits for feedback. If still no acknowledgement, the communication process
terminates, going back to the first step. If acknowledge is received, process 2 sends
Interrupt to process 1, the latter responds and stops writing to memory.
Process 2 is then capable of downloading information to the base station. When
data is transmitted, the unit sends the number of points transmitted, to be compared
with those received by the base station. If they didn’t match, the unit repeats down-
loading its information all over again. Otherwise, if the unit receives successful
download, it terminates the process and turns off.
Initially, the circuit shown in Fig. 25.4 is off. After car ignition, current passes
through D1, and continues its way towards the transistor. This causes the relay to
switch and supports an output voltage of 12 V. The circuit (C  ) is now powered and
could start its functionality. Using the two regulators, it becomes feasible to provide
an adequate voltage to the FPGA, which in turn navigates the whole switching tech-
nique of the system. In other words, the FPGA adapts itself so that it can either put
a zero or 5 V at the side connecting D2. For the 5 V, the circuit is all on, and the
vehicle is in normal functionality. When data download ends, the FPGA perceives
that, and changes the whole circuit into an idle one, and waits for another car igni-
tion. So, it is well known now that the FPGA will be the controller of the behavior
of the VTS system.

25.3.3 The Memory Block

The suggested memory blocks are addressed by a 12-bit address bus and stores 8-bit
data elements. This means that the memory can store up to 4 KB of data. The mem-
ory controller navigates the proper memory addressing. Multiplexers are distributed
along with the controller to make the selection of the addressed memory location
and do the corresponding operation (See Fig. 25.5).
25 Reconfigurable Hardware Implementation of a GPS-Based Vehicle Tracking System 295

Fig. 25.4 The electric components driving the FPGA

Fig. 25.5 The memory block of the FPGA system

296 A. Yaqzan et al.

25.3.4 Communication Protocols: I2C and UART

The I 2C bus is a serial, two-wire interface, popularly used in many systems because
of its low overhead. It is used as the interface of process 1 and process 2 with the
shared memory. It makes sure that only one process is active at a time, with good
reliability in communication. Therefore, it writes data read from the GPS during
process 1, and reads from memory to output the traversed positions into the base
station. The Universal Asynchronous Receiver Transmitter (UART) is the most
widely used serial data communication circuit ever. UART allows full duplex
communication over serial communication links as RS232. The UART is used to
interface Process 1 and the GPS module from one side, and Process 2 and the Base
Station (BS) from the other side.

25.4 Performance Analysis and Evaluation

Three different systems are to be tested for the FPGA implementation. The sug-
gested systems gradually add more parts to the designed FPGA implementation till
we reach a complete stand alone system. The three suggested integrations are as
follows:

First Integration: Process 1, process 2, and I2C

Second Integration: Process 1, process 2, I2C, and memory
Third integration: Process 1, process 2, I2C, UART, and memory (standalone
FPGA).

According to the local market cost, around 9.6% could be saved per unit if the
FPGA-based all-in-one system is adopted. For the kind of memory implemented in
the system, the vehicle cannot store many locations, so the Vehicle Tracking System
(VTS) is to be used within a small city. If the sampling rate is to be set one reading
every 2 min, one could get a general but not very specific overview of the tracks
traversed.
The vehicle tracking needs 344 bits of data to store the five important parameters
(longitude, latitude, speed, date, and time). Consequently, this would need 43 mem-
ory locations. In other words, the system needs 43 locations for one reading lapsed
2 min.
With 4,096 memory locations, the system makes 95.25 readings, meaning that
the vehicle should come back to its base station every 3 h and 10 min to download
its information. This would be 4 h and 45 min if the rate is one reading every 3 min.
This is not very satisfactory but is tolerated as long as the intended area has a small
range. However, this is one problem subject to upgradeability. One solution is to use
an FPGA with a large memory.
Table 25.2 shows the results of simulation done on integration of parts (mod-
ules) forming the FPGA-based system. Each integrated system is tested on the
component, integration, and system levels.
25 Reconfigurable Hardware Implementation of a GPS-Based Vehicle Tracking System 297

Table 25.2 Results summary taken from the syntheses of integrations on STRATIX
EP1S10F484C5. The execution time in some cases varies according to “RMC”, “GGA”, or “free”
Integration % Area in Prop. delay Execution time Max. op. freq.
logic elem. (ns) (ns) (MHz)
First 18 118:77 Varies 8:149
Second 50 116:00 Varies 8:62
Third 50 94:47 Varies 10:58

Table 25.3 The readings obtained from the integrations compiled on STRATIX EP1S10F484C5
Integration Size (bits) Number of Prop. delay Speed of
clock cycles (ns) processing (μs)
First 1,448 181 118:773 21.497
Second 1,448 181 116:003 20.996
Third 1,448 181 94:474 17.099

The first design used 1,910 logic elements out of 10,570 (STRATIX EP1S10
F484C5, 175.47 MHz), and a maximum operating frequency of 8.149 MHz, leav-
ing 82% free space of the FPGA capacity. However, after adding memory to the
integration, the number of logic elements increased to 5,303, with 50% usage of the
capacity. The propagation delay decreased slightly inside the FPGA, The decrease in
propagation delay means that the optimizer found a better way to reduce its critical
path.
Similar results are shown when the UART part is added (standalone FPGA), with
an improvement in propagation delay. Although the number of logic elements has
increased, it contributed to better interaction among the parts and raised the operat-
ing frequency to 10.58 MHz. Therefore, integration of parts has enhanced the delay
with an expected increase in number of logic elements positively affects the process-
ing speed when propagation finds its paths among combinations of gates. Suppose
that the GPS message “M” received by the UART has come in the following
sequence:

$GPGGA,161229.487,3723.2475,N,12158.3416,W,1,07,1.0,9.0,M,,,,0000*18
$GPGLL,3723.2475,N,12158.3416,W,161229.487,A*2C
$GPRMC,161229.487,A,3723.2475,N,12158.3416,W,0.13,309.62,120598,*10

“M” is to be tested on the three obtained integrations (Table 25.3), taking into
account that the GPS parameters needed for the application come as represented
in Section 25.3. The system takes selective parameters according to their position
in the sequence, and checks if they correspond to the desired information. Every
character is represented by its 8-bit ASCII representation. Table 25.3 shows an exact
interpretation of data when received from the GPS and processed via several mod-
ules of the integration. After integrations have been inspected, the proposed system
synthesized on different FPGAs, and the results appear in Table 25.4 [7, 8].
298 A. Yaqzan et al.

Table 25.4 Syntheses of the VTS on different FPGAs. The execution time in some cases varies
according to “RMC”, “GGA”, or “free”
FPGA Logic area in Prop. delay Exec. time Max. operating
logic elements (ns) (ns) frequency (MHz)
STRATIX EP1S10F484C5 50% 94:474 Varies 10:58
STRATIX-II EP2S15F484C3 36% 151:296 Varies 6:609
MAX3000A EPM3032ALC44–4 Doesn’t fit NA NA NA
Cyclone EP1C6F256C6 90% 90:371 Varies 11:06
FLEX6000 EPF6016TI144–3 Doesn’t fit NA NA NA
APEXII EP2A15B724C7 30% 181:418 Varies 5:512

From the readings of Table 25.4, the following could be concluded testing the
all-in-one system:
STRATIX EP1S10F484C5 (175.47 MHz) has enough number of logic elements,
and the capacity taken by the project is one half its total capacity. The propagation
delay is 94.474 ns, thus, the system runs with a frequency of 10.58 MHz.
STRATIX-II EP2S15F484C3 (420 MHz) has a larger number of logic elements,
so the project took lesser capacity (36%), and caused more propagation delay.
The project fits 90% in Cyclone EP1C6F256C6 (405 MHz), but with minimum
propagation delay of 90.371 ns and thus 11.065 MHz operating frequency.
APEXII EP2A15B724C7 (150 MHz) has the largest capacity among the listed
devices allocating 30% only for the ARAM project with a largest propagation
delay (181.418 ns) and minimum frequency (5.512 MHz).

25.5 Conclusion

In this chapter, we have presented an alternative design of an existing modern GPS-

based VTS using FPGAs. The performance of the proposed implementations is
evaluated on different FPGAs. The suggested designs show enhancement in terms
of speed, functionality, and cost. Future work includes refining the proposed designs
in order to eliminate the sequential alternation of the two main internal processes,
and investigating larger buffering by providing more memory elements and using
state-of-art FPGAs.

References

1. C. Ashraf, Y. Arjumand, An Approach to Task Allocation for Dynamic Scheduling in

Reconfigurable Computing Systems, International Multitopic Conference, 2005, pp. 1–6.
2. V. Subramanian, J. Tront, S. Midkiff, C. Bostian, A Configurable Architecture for High
Speed Communication Systems, Military and Aerospace Applications of Programmable Logic
Devices (MAPLD) International Conference, Vol. 3, pp. E11 1–9, Sept. 2002.
3. I. Damaj, “Parallel Algorithms Development for Programmable Logic Devices”, Advances in
Engineering Software, Elsevier Science, 2006. Issue 9, Vol. 37, pp. 561–582.
25 Reconfigurable Hardware Implementation of a GPS-Based Vehicle Tracking System 299

4. V. Subramanian, J. Tront, C. Bostian, S. Midkiff, Design and Implementation of a Config-

urable Platform for Embedded Communication Systems, IPDPS, 2003, p. 189.
5. M. Shanblatt, B. Foulds, A Simulink-to-FPGA Implementation Tool for Enhanced Design
Flow, International Conference on Microelectronic Systems Education, 2005, pp. 89–90.
6. F. Vahid, T. Givargis, Embedded System Design, A Unified Hardware/Software Introduction
(Wiley, New York, 2001).
7. Z. Osman, M. Jrab, S. Midani, R. Zantout, Implementation of a System for Offline Tracking
Using GPS, Mediterranean Microwave Symposium, 2003.
8. NMEA Reference Manual SiRF Technology, Inc. 148 East Brokaw Road San Jose, CA 95112,
USA; http://www.nmea.org
9. R. Kühne, R. Schäfer, J. Mikat, K. Thiessenhusen, U. Böttger, S. Lorkowski, New Approaches
for Traffic Management in Metropolitan Areas, Proceedings of the IEEE Conference on ITS,
2003.
Chapter 26
Unknown Malicious Identification

Ying-xu. Lai and Zeng-hui. Liu

Abstract The detection of unknown malicious executables is beyond the capability

of many existing detection approaches. Machine learning or data mining methods
can identify new or unknown malicious executables with some degree of success.
Feature set is a key to apply data mining or machine learning to successfully
detect malicious executables. In this paper, we present an approach that conducts
an exhaustive feature search on a set of malicious executables and strives to obviate
over-fitting. To improve the performance of Bayesian classifier, we present a novel
algorithm called Half Increment Naı̈ve Bayes (HIB), which selects the features by
carrying an evolutional search. We also evaluate the predictive power of a classifier,
and we show that our classifier yields high detection rates and learning speed.

Keywords Unknown malicious detection  Half Increment Naı̈ve Bayes  classifi-

cation

26.1 Introduction

As network-based computer systems play increasingly vital roles in modern society,

a serious security risk is the propagation of malicious executables. Malicious exe-
cutables include viruses, Trojan horses, worms, back doors, spyware, Java attack
applets, dangerous ActiveX and attack scripts. Identifying malicious executables
quickly is an important goal, as they can cause significant damage in a short time.
Consequently detecting the presence of malicious executables on a given host is a
crucial component of any defense mechanism.
Traditional malicious executables detection solutions use signature-based meth-
ods, in that they use case-specific features extracted from malicious executables in

Y.-x. Lai (B)

College of Computer Science, Beijing University of Technology, Beijing 100124, China
E-mail: laiyingxu@bjut.edu.cn

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 301
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
302 Y.-x. Lai, Z.-h. Liu

order to detect those same instances in the future [1]. Security products such as
virus scanners are examples of such application. While this method yields excel-
lent detection rates for existing and previously encountered malicious executables,
it lacks the capacity to efficiently detect new unseen instances or variants. Due
to detect malicious accurately is a NP problem [2, 3], heuristic scanners attempt
to compensate for this lacuna by using more general features from viral code,
such as structural or behavioral patterns [4]. Although proved to be highly effec-
tive in detecting unknown malicious executables, this process still requires human
intervention.
Recently, attempts to use machine learning and data mining for the purpose of
identifying new or unknown malicious executables have emerged. Schultz et al.
examined how data mining methods can be applied to malicious executables detec-
tion [5] and built a binary filter that can be integrated with email servers. Kolter et al.
used data mining methods, such as Naı̈ve Bayes, J48 and SVM to detect malicious
executables [6]. Their results have improved the performance of these methods.
Bayes or improved Bayes algorithm has the capability of unknown malicious detec-
tion, but it spends more time to study. A new improved algorithm (half-increment
Bayes algorithm) is proposed in this paper.
In this paper, we are interested in applying data mining methods to malicious
executables detection, and in particular to the problem of feature selection. Two
main contributions will be made through this paper. We will show how to choose
features which are most representative properties. Furthermore, we propose a new
improved algorithm and will show that our method achieve high learning speed
and high detection rates, even on completely new, previously unseen malicious
executables.
The rest of this paper is organized as follows: Section 26.2 is a brief discussion
of related works. Section 26.3 gives a brief description of Bayesian algorithm. Sec-
tion 26.4 presents details of our methods to obtain high learning speed. Section 26.5
shows the experiment results. Lastly, we state our conclusions in Section 26.6.

26.2 Related Work

At IBM, Kephart et al. [7] proposed the use of Neural Networks to detect boot
sector malicious binaries. Using a Neural Network classifier with all bytes from
the boot sector malicious code as input, it had shown that 80–85% of unknown
boot sector malicious programs can be successfully identified with low false pos-
itive rate (<1%). The approach for detecting boot-sector virus had incorporated
into IBM’s Anti-Virus software. Later, Arnold et al. [8, 9] applied the same tech-
niques to win32 binaries. Motivated by the success of data mining techniques in
network intrusion system [10, 11], Schultz et al. [5] proposed several data min-
ing techniques to detect different types of malicious programs, such as RIPPER,
Naı̈ve Bayes and Multi-Naı̈ve Bayes. The authors collected 4,301 programs for the
Windows operating system and used MacAfee Virus Scan to label each as either
26 Unknown Malicious Identification 303

malicious or benign. The authors concluded that the voting naı̈ve-Bayesian classifier
outperformed all other methods. In a companion paper [12] the authors developed
an Unix mail filter that detect malicious Windows executables based on the above
work. Kolter et al. [6] also used data mining methods, such as Naı̈ve Bayes, J48 and
SVM to detect malicious codes. The authors gathered 1,971 benign and 1,651 mali-
cious codes and encoded each as a training example using n-grams of byte codes
as features, boosted decision trees outperformed other methods with an area under
the ROC curve of 0.996 (applied detectors to 291 malicious executables discovered,
and boosted decision trees achieved a TP rate of 0.98 for a desired FP rate of 0.05).
Zhang et al. [13] used SVM and BP neural network to virus detection, the D-S the-
ory of evidence was used to combine the contribution of each individual classifier
to give the final decision. It showed that the combination approach improves the
performance of the individual classifier significantly. Zhang et al. [14, 15] estab-
lished methods based on fuzzy pattern and K-nearest neighbor recognition applying
to detect malicious executables for the first time.

26.3 Naı̈ve Bayesian and Application

The goal of our work was to improve a standard data mining technique to compute
accurate detectors for new binaries. We gathered a large set of programs from public
sources and separated the problem into two classes: malicious and benign executa-
bles. We split the dataset into two subsets: the training set and the test set. The data
mining algorithms used the training set while generating the rule sets. We used a
test set to check the accuracy of the classifiers over unseen examples.
In a data mining framework, features are properties extracted from each example
in the data set – such as strings or byte sequences. A data mining classifier trained
with features can use to distinguish between benign and malicious programs. We
used strings that are extracted from the malicious and benign executables in the data
set as features.
We propose an exhaustive search for strings. Typically there are many “disorder-
words” when files are read as ASCII code, like “autoupdate.exe ?soft-
ware*abbbbbb download”, etc. The string selection program extracts consec-
utive printable characters from files. To avoid yielding a large number of unwanted
features, characters are filtered before they are recorded. Our feature selection
involves an extraction step followed by an elimination step.

26.3.1 String Extraction and Elimination

In string extraction step, we scan files and read character one by one, record con-
secutive printable characters (like English letter, number, symbolic and etc.), and
construct lists of all the strings. The length of the string is specified as a number
304 Y.-x. Lai, Z.-h. Liu

of characters. The shorter the length, the more likely the feature is to have general
relevance in the dataset. But a short length will yield a larger number of features.
Extracted strings from an executable are not very robust as features because some
strings are unmeaning, like “abbbbbb”, so we select a subset of strings as feature set
by an eliminate step.
Many have noted that the need for a feature filter is to make use of conventional
learning methods [13,14], to improve generalization performance, and to avoid over-
fitting. Following the recommendation of those, the glossary filter criterion is used
in the paper to select a subset of strings.
The glossary is a computer glossary which includes 7,336 items:
Computer words, such as function name, API name
Abbreviations on computer networks, like “QQ”, “msn”, etc
Postfixs, like “.dll”, presenting dynamic link libraries

26.3.2 Naı̈ve Bayes

The naive Bayes classifier computes the likelihood that a program is malicious given
the features that are contained in the program. We treat each executable’s features as
a text document and classified based on that. Specifically, we want to compute the
class of a program given that the program contains a set of features F . We define
C to be a random variable over the set of classes: benign, and malicious executables.
That is, we want to compute P .C jF /, the probability that a program is in a certain
class given the program contains the set of features F . We apply Bayes rule and
express the probability as:

P . F jC /  P .C /
P . C jF / D (26.1)
P .F /
To use the naı̈ve Bayes rule we assume that the features occur independently from
one another. If the features of a program F include the features F1 ; F2 ; F3 : : :; Fn ,
then Eq. (26.1) becomes:
Qn
P . Fi jC /  P .C /
P . C jF / D i D1Qn   (26.2)
j D1 P Fj

Each P .Fi jC / is the frequency that string Fi occurs in a program of class C . P .C /

is the proportion of the class C in the entire set of programs.
The output of the classifier is the highest probability class for a given set of
strings. Since the denominator of Eq. (26.1) is the same for all classes we take
the maximum class over all classes C of the probability of each class computed in
Eq. (26.2) to get:
!
Yn
Most Likely Class D max P .C / P .Fi jC / (26.3)
C
i D1
26 Unknown Malicious Identification 305

In Eq. (26.3), we use maxC to denote the function that returns the class with the
highest probability. Most Likely Class is the class in C with the highest probability
and hence the most likely classification of the example with features F:

26.4 Increment Naı̈ve Bayes

26.4.1 Naı̈ve Bayes

To train the classifier, we record how many programs in each class contained each
unique feature. We use this information to classify a new program into an appropri-
ate class. We first use feature extraction to determine the features contained in the
program. Then we apply Eq. (26.3) to compute the most likely class for the program.
The Naı̈ve Bayes algorithm requires a table of all features to compute its prob-
abilities. This method requires a machine with one gigabyte of RAM, because the
size of the binary data was too large to fit into memory.
To update the classifier, when new programs are added to the training set, we
update feature set at first, and apply Eq. (26.3) to compute the most likely class for
the program again. So it is time-consuming to update the classifier by NB algorithm.

26.4.2 Multi-naı̈ve Bayes

To correct NB algorithm problem the training set is divided to smaller pieces that
would fit in memory. For each set we train a Naı̈ve Bayes classifier. Each classifier
gives a probability of a class C given a set of strings F which the Multi-Naı̈ve Bayes
uses to generate a probability for class C given F over all the classifiers.
For each classifier, the probabilities in the rules for the different classifiers may
be different because the underlying data the each classifier is trained on is different.
The prediction of the Multi-Naı̈ve Bayes algorithm is the product of the predictions
of the underlying Naı̈ve Bayes classifier.

Y
n
P .C jF / D Pk .C jF / (26.4)
kD1

When new programs are added to the training set, these new programs as a subset
and train a Naı̈ve Bayes classifier over the subproblem. Based on the Eq. (26.4), we
update the probability. The Multi-Naı̈ve Bayes algorithm does not need to compute
its probabilities over all the training set, but the accuracy of the classifier will be
worsen.
306 Y.-x. Lai, Z.-h. Liu

26.4.3 Half Increment Bayes (HIB)

The above NB and MNB algorithms obtain feature set over all training set or subset
at first. Once the final feature set was obtained, we represent our positive and zero
data in the feature space by using, for each feature, “1” or “0” to indicate whether
or not the feature is present a given executable file. The probability for a string
occurring in a class is the total number of times it occurred in that class’s training
set divided by the total number of times that the string occurred over the entire
training set.
We can derive a method purely from the NB algorithm for increment update. In
our method, feature set is increased with studying of classifier. That is, composed
there are k1 string features extracted from the first sample, so k1 strings are elements
of set F . If there are S2 strings extracted from the second sample, k2 elements are
not found in F , these elements should be added in feature set F . Set F will includes
k1 C k2 elements. Classifier is trained based on the evolutional feature set.
 ˇ 
Claim 1 We can obtain the class-conditional probability P F .nC1/ ˇC over
n C 1 samples by that h 
of former
ˇ n samples andi the .n C 1/th sample, that is
 ˇ .n/ ˇ
P Fi ˇCj  n C P . x .nC1/ jCj /  ˇ 
.nC1/ ˇ
P Fi ˇCj D nC1
. Where, P F .nC1/ ˇ C is the
 ˇ 
class-conditional probability over n Cˇ 1 samples, P F .n/ ˇC is class-conditional
 
probability over n samples, P x .nC1/ ˇC is class-conditional probability over the
.n C 1/th sample.
Proof. Composed there are a features obtained from n samples, that is F .n/ D
fF1 ; F2 : : :; Fa g, then
 ˇ Y
a
 
ˇ
P F .n/ ˇC D P Fi jCj .j D 1; 2/ (26.5)
i D1
  P .Fi Cj / count.F D Fi ^ C D Cj /
P Fi jCj D D (26.6)
P .Cj / count.C D Cj /

Where, count.F D Fi ^ C D Cj / is sample number for F D Fi and class C D Cj ,

count.C D Cj / is sample number for class C D Cj .

ˇ  with C D Cj in training set, then the class-conditional

If there are n samples
probability P F .n/ ˇC for n samples is:

 ˇ count..F D Fi.n/ / ^ .C D Cj //
ˇ
P Fi.n/ ˇCj D (26.7)
n
If the (n C 1)th sample for class C D Cj was added, there are two cases.
Case 1. the strings in the (n C 1)th sample are all found in F , then
 ˇ count..F D Fi.n/ / ^ .C D Cj //
ˇ
P Fi.nC1/ ˇCj D (26.8)
nC1
26 Unknown Malicious Identification 307

Case 2. there are b strings in the (n C 1)th sample are not found in F , then these
ˇ D fF1 ; F2 : : :; Fa ; FaC1 ;   ; FaCb g. For those
.nC1/
strings are added in F , that is F

ˇ
new b probabilities, due to P Fi.n/ ˇCj D 0 a < i  a C b, then

 ˇ
ˇ 1
P Fi.nC1/ ˇCj D a <i aCb (26.9)
nC1
We rewrite Eqs. (26.8) and (26.9) as Eq. (26.10):
h  ˇ  ˇ i
ˇ
 ˇ P Fi.n/ ˇCj  n C P x .nC1/ ˇCj
ˇ
P Fi.nC1/ ˇ Cj D (26.10)
nC1
Therefore, information of n C 1 samples can obtained from those of former n
samples and the (n C 1)th sample.

Claim 2 For NB and HIB, based on same training set, same feature sets can be
obtained by same string extraction and elimination methods, that is FN D FH .
Where FH is feature set obtained by half increment algorithm, FN is feature set
obtained by naı̈ve bayes algorithm.

Proof. Composed there are a features obtained from n samples, that is FN D

fF1 ; F2 : : :; Fa g, Set FN is made of strings that extracted from n samples and
eliminated based on computer dictionary.
FH is increased as training samples. When all of training samples are all extracted,
Set FH is made of strings that extracted from n samples and eliminated based on
computer dictionary. So FN D FH .

Based on Claim 1 and 2, we can obtain Theory 1.

Theory 1 Most Likely Classes computed by half increment classification and Naı̈ve
Bayes classification are same, that is C1 D C2 . Where C1 is the most likely class
obtained by naı̈ve Bayes, and C2 is the most likely class obtained by half-increment
classification.

Proof. Composed a features were obtained from n samples, that is FN D fF1 ; F2

: : :; Fa g, Based on Eq. (26.3),
!
Y
a
C1 D max P .C / P .Fi jC / (26.11)
i D1

When a new sample is added, FN D fF1 ; F2 : : :; Fa ; FaC1 ;   ; FaCb g the most

likely class computed by naı̈ve Bayes is

Y
aCb
C1 D arg max P .C /  P .Fi jC / (26.12)
i D1
308 Y.-x. Lai, Z.-h. Liu

As FN D FH , based on Eq. (26.10)

Y
aCb
C2 D arg max P .C /  P .Fi jC / (26.13)
i D1

Therefore, the C1 D C2 .

26.4.4 Complexity

Based on former algorithm, time-consuming of NB and HIB are made of two parts:
1. Extract unrepeated strings from samples
2. Fix on feature set and training set, build up classifier
For step (1), time-consuming of two algorithms are same. But for step (2), they are
different.
For HIB, feature set is increased with studying of classifier. That is, composed
there are S1 strings in the first sample, and k1 strings are unrepeated, so k1 strings
are elements of set F . If there are S2 strings in the second sample, the time-
consuming of computing whether k1 elements are found in S2 or not is TH 2 D
O.S2  k1 /. If k2 elements are not found in F , these elements should be added
in feature set F . Set F will includes k1 C k2 elements. For n samples, the all
time-consuming of fix on set F is

TH D O.S2  k1 C S3  .k1 C k2 / C : : : C Sn  .k1 C k2 C : : : C kr // (26.14)

For NB, the set F are obtained before classifier study. Composed there are K ele-
ments in F , the time consuming to indicate whether or not the feature is present a
given executable file is

TN D O.K  .S1 C S2 C : : : C Sn // (26.15)

Based on Claim 2, F N DF H , K D k1 C k2 C : : : C kr .
Based on Eq. (26.14),

TH D O.K .S2 CS3 C: : :CSn /S2 .k2 C: : :Ckr /: : :Sn1 kr / (26.16)

If k2 D k3 D : : : D kr D 0, or K D k1 , TH D TN , the time-consuming of
two algorithm are same. But malicious executables include viruses, Trojan horses,
worms, back doors, spyware, Java attack applets, dangerous ActiveX and attack
scripts, a sample cannot include all feature elements.
26 Unknown Malicious Identification 309

26.5 Experimental Results

26.5.1 Experimental Design

Our experiment are carried out on a dataset of 2,995 examples consisting of 995
previously labeled malicious executables and 2,000 benign executables, collected
from desktop computers using various versions of the Windows operating system.
The malicious executables are taken from an anti-Virus software company.
For each run, we extract strings from the executables in the training and testing
sets. We select the most relevant features from the training data, apply elimination
method, and use the resulting classifier to rate the examples in the test set.

26.5.2 Performance Analysis

26.5.2.1 Time-Consuming

In our experiments, we used VC CC implementation of the NB, MNB and HIB

classifiers. In this section, we evaluate the performance of NB, MNB and HIB three
algorithms in comparison.
For HIB algorithm, increasing rule of feature elements is based on order of
studied samples. Figure 26.1 is the curve between number of studied samples and
number of feature elements. In Fig. 26.1, the slope of the curve is steep at some
points. When classifier begins to study some type sample, the slope of the curve is
steep, that is, the number of feature elements increase quickly. After classifier has
studied some samples, the slope of the curve is smooth.

4000

3500
Number of feature elements

3000

2500

2000

1500

1000

500

0
0 200 400 600 800 1000 1200 1400 1600 1800
Number of samples

Fig. 26.1 Curve between number of feature elements and number of samples
310 Y.-x. Lai, Z.-h. Liu

14 NB

12 MNB

Time-consuming (h)
HIB
10

0
0 500 1000 1500 2000
Sample

Fig. 26.2 NB, MNB and HIB curves of time-consuming

1
0.9
0.8
True Positive Rate

0.7
0.6
0.5
NB
0.4
0.3 MNB
0.2
HIB
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False Positive Rate

Fig. 26.3 NB, MNB and HIB ROC

Figure 26.2 is the curve of three algorithm’s time-consuming. The slope of NB

algorithm is initially much steeper than the HIB and MB algorithms. HIB algorithm
has better efficiency.
The ROC curves in Fig. 26.3 show a more quickly growth than the NB and MNB
until the false positive rate climbed above 4%. Then the three algorithms converged
for false positive rates greater then 6% with a detection rate greater then 95%.

26.5.2.2 Comparison Against Benchmark Model

To evaluate our system we are interested in several quantities:

1. True Positives (TP), the number of malicious executable examples classified as
malicious code
2. True Negatives (TN), the number of benign programs classified as benign
3. False Positives (FP), the number of benign programs classified as malicious code
4. False Negatives (FN), the number of malicious codes classified as benign
TP CT N
The Overall Accuracy is defined as TP CT N CFP CF N .
26 Unknown Malicious Identification 311

Table 26.1 Experimental results using our method and traditional methods
Method Feature Classifier Accuracy (%)

Our method Strings NB 96:8

Our method Strings MNB 93:3
Our method Strings HIB 97:9
Schultz String NB 97:11
Schultz Bytes MNB 96:88
Kolter 4-gram SVM 93
Henchiri 8-gram J48 93:65

We compare our method with a model used in previous research [7, 15]. The
results, displayed in Table 26.1, indicate that a virus classifier can be made more
accurate by using features representative of general viral properties, as generated by
our feature search method. With up to 97.9% overall accuracy, our system outper-
forms NB and MNB algorithms and achieves better results than some of the leading
research in the field, which performs at 97.11%.

26.6 Conclusion

The naı̈ve Bayes classifier is widely used in many classification tasks because its
performance is competitive with state-of-the-art classifiers, it is simple to imple-
ment, and it possesses fast execution speed. In this paper, we discussed the problem
of how to classify a set of query vectors from the same unknown class with the
naı̈ve Bayes classifier. Then, we propose the method HIB algorithm and compare it
with naı̈ve Bayes and multi-naı̈ve Bayes. The experimental results show that HIB
algorithm can take advantage of the prior information, can work well on this task.
Finally, HIB algorithm was compared with a model used in previous research [7,15].
Experimental results reveal that, HIB can reach a higher level of accuracies as
97.9%. HIB’s execution speed is much faster than MNB and NB, and HIB has
low implementation cost. Hence, we suggest that HIB is useful in the domain of
unknown malicious recognition and may be applied to other application.

References

1. G. McGraw, G. Morrosett, Attacking malicious code: a report to the infosec research council,
IEEE Transactions on Software, vol. 2, Aug. 1987, pp. 740–741.
2. F. Cohen, Computer Viruses Theory and Experiments, Computers & Security, vol. 6, Jan.
1987, pp. 22–35.
3. D. Spinellis, Reliable Identification of Bounded-Length Viruses Is NP Complete, IEEE
Transactions on Information Theory, vol. 49, Jan. 2003, pp. 280–284.
4. MacAfee. Homepage-MacAfee.com. Online Publication, 2000. http://www.macafee.com
312 Y.-x. Lai, Z.-h. Liu

5. M.G. Schcltz, E. Eskin, E. Zadok, S.J. Stolfo, Data mining methods for detection of new
malicious executables, In: Proceedings of IEEE Symposium in Security and Privacy, 2001.
6. J.Z. Kolter, M.A. Mloof, Learning to detect malicious executables in the wild, In: Proceed-
ings of the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, ACM Press, New York, 2004, pp. 470–478.
7. J.O. Kepahart, G.B. Sorkin, W.C. Arnold, D.M. Chess, G.J. Teauro, S.R. White, Biologically
inspired defenses against computer viruses. In: Proceedings of IJCAI’95, Montreal, 1995.
pp. 985–996.
8. W. Arnold, G. Tesauro, Automatically generated Win32 heuristic virus detection, In: Proceed-
ings of the 2000 International Virus Bulletin Conference, 2000.
9. G. Tesauro, J.O. Kephart and G.B. Sorkin, Neural networks for computer virus recognition,
IEEE Expert, vol. 11, Apr. 1996, pp. 5–6.
10. W. Lee, S. Stolfo, K. Mok, A Data Mining Framework for Building Intrusion Detection
Models, IEEE Symposium on Security and Privacy, 1999.
11. W. Lee, S.J. Stolfo, P.K. Chan, Learning patterns from UNIX processes execution traces
for intrusion detection, AAAI Workshop on AI Approaches to Fraud Detection and Risk
Management, AAAI Press, 1997, pp. 50–56.
12. M.G. Schcltz, E. Eskin, E. Zadok, M. Bhattacharyya, S.J. Stolfo, MEF: Malicious Email filter,
A Unix mail filter that detects malicious windows executables, In: Proceeding of USENIX
Annual Technical Conference, 2001.
13. B. Zhang, J. Yin, J. Hao, Unknown computer virus detection based on multi-naive Bayes
algorithm. Computer Engineering, vol. 32, Oct. 2006, pp. 18–21.
14. B. Zhang, J. Yin, J. Hao, Using fuzzy pattern recognition to detect unknown malicious
executables code, Lecture Notes in Computer Science, vol. 36, Mar. 2005, pp. 629–634.
15. B. Zhang, J. Yin, D. Zhang, J. Hao, Unknown computer virus detection based on K-nearest
neighbor algorithm, Computer Engineering and Applications, vol. 6, 2005, pp. 7–10.
Chapter 27
Understanding Programming Language
Semantics for the Real World

Trong Wu

Abstract Computer is used virtually everywhere everyday in the world. Before

1990s computer systems are generally used for mathematics, engineering, and busi-
ness computations. In this period, mainly use FOTRAN, COBOL, and PL/1 for
computation on mainframe systems. In the last two decades scientists found that
the natural world is complicated that has overwhelmed with data that required
more sophisticated computation facilities and better languages for computation and
that created new computing disciplines. This paper addresses the understanding of
programming language semantics that will help user in selection of programming
language features for various applications needs and that can help programmers in
designing reliable, accurate, efficient, and user-friendly software systems for the real
world.

Keywords User-defined types  exception handling  multitasking  communication

between program units  and user-defined precision

27.1 Introduction

It is a chaotic world, it is a systematic world; it is a confused age, it is illuminate

age; it is disordered society, it is organized society. Everyday we are struggling in a
complex, intricate, and difficult environment for our life. Computer science can give
one abilities to solve a category of complicated, obscure, and oppressive problems.
We can take the advantage of computer speed and large volume of storage spaces to
add our ability for solving complex problems in the world. This is a very challenge
task and interesting assignment to youngsters. It attracts many of them to study
computer science.

T. Wu
Department of Computer Science at Southern Illinois University Edwardsville, IL 62026, USA
e-mail: twu@siue.edu

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 313
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
314 T. Wu

In October 1962, Purdue University established the first department of computer

science in the United States [1]. Since then Computer Science education has become
an integral discipline in colleges and universities across the country. Initially, stu-
dents in physics, engineering, and mathematics were advised to take one course
in the FORTRAN language while students in the school of business particularly in
Management Information Systems ware required to take a course in the COBOL
language. Students who majored in Computer Science were required to enroll in
FORTRAN, COBOL, and PL /1 courses. It was thought that this would prepare them
potentially to work in industrial firms or business companies after completion of
their degrees.
Since then, computer software and its use have spread to nearly every aspect
of peoples’ lives from work to play. Today almost every one uses computer and
software everywhere. Most of the equipment and devices, which use software are
designed and manufactured by engineers. For these applications, users need not only
robust hardware, but also reliable software.
Today, the most serious problems with software products are expense, ease of
use, and reliability [2]. Computer scientists not only need to study computer mech-
anisms and how to increase the productivity and efficiencies of a computer system,
but also need to design computer systems, write programs, and execute them. The
former is the state of the practice of software engineering and the latter is the state
of the art. Unfortunately, there exists a gap between the former and the latter. The
task for computer scientists is to eliminate or narrow this gap. To implement this,
one should thoroughly analyze system requirements, carefully design the system
and programs, and then perform various tests including unit tests and system tests.
Finally computer scientists need to deliver systems and make sure that they meet
their requirements and fit in their environments [3].
However, computer scientists need a good language tool to implement these soft-
ware engineering principles. Human languages such as English, Chinese, Spanish,
Russian, etc. are all ambiguous languages. None of them can be used to communi-
cate with a computer system. Therefore, mathematicians, computer scientists, and
linguists must work hard to develop semantically unambiguous languages, starting
by designing grammar rules and then developing the sentential structures of the
language
Nevertheless, most engineers have had only one computing course in FORTRAN
or in the C language or both while they were in college and perhaps they believe
that FORTRAN or the C language is enough for their engineering application needs.
In fact, the FORTRAN language lacks user-defined types that limit its application
domain. Moreover, both the FORTRAN and C languages do not have predefined
accuracy features, friendly concurrency facilities, and effective exception handling
methods. Today, neither FORTRAN language nor the C language remains adequate
to program in this complex and chaotic world for designing and implementing
more sophisticated and reliable systems for biological, physical, and other natural
related systems. Therefore, we need understand programming Language semantics
thoroughly so that we can deal with natural world computation.
27 Understanding Programming Language Semantics for the Real World 315

This paper describes some necessary elements in the understanding program-

ming language semantics that is important for the computation of the natural world.
It requires computation more accurate, more reliable, more precise, more efficient,
and friendlier. Therefore, we should consider the extends of application domain, the
application of accuracy and efficient facilities, the use of modularization and com-
munication, the utilization of parallelism or concurrency, and applying exception
handling and for reliability critical projects in this real world.

27.2 Extension of the Applications to a Real World

A department store usually consists of tens’ of thousands merchandises; it is very

difficult to manage them properly, efficiently, and cost-effectively. One of the most
effective ways to manage such department stores is to partitioning all the merchan-
dises into departments such as man’s ware, lady’s ware, house ware, juniors, sports,
electronics, hardware, shores, bath and bed, pharmacy, etc. Any merchandise in a
department store must belong to one and only one department, and department-
to-department are discriminates to each other. No merchandises can belong to two
distinct departments at the same time. Each department consists of a set of similar
or the same kind merchandises and a set of operation rules and policies for manage
these merchandises. For example, hardware and electronic departments have dif-
ferent return or refund policies; some electronic items are not allowed customers to
test or to tryout. Pharmacy department has its operational procedures or rules; med-
ications are classified into prescription and non-prescription. For any prescription
medication is strictly required a medical doctor’s signed prescription [4].
In a special season of the year, like Christmas time, most department stores want
to make some additional business for this splendor season; they create a new depart-
ment called “Santa;” it is for children to take pictures with the Santa. Likewise,
in the springtime they create a new department called “Plants” by selling saplings,
flowers, seedling, soils, seeds, fertilizer, rocks, and other goods for the garden [4].
Today real world projects can be very large and complicated for applications
in newly developed domains. These projects can include hardware projects, soft-
ware projects, and firmware projects. Many engineering projects need to take years
or tens’ of years to complete them such as the Yangtze River Three Gorges Dam
project that was launched in 1993 and the water level in the reservoir will reach to
175 m in 2009, when the project is finally completed [5]; the Airbus A380 project
that was started in 2000 for worldwide market campaign and made its maiden
flight in 2005 [6]; and the Boeing 777 program was launched in October 1990
and the first 777–300 was delivered to Cathay Pacific Airways in June 1998 [7].
These projects involved thousands designers, hundreds subcontractors, and tens of
thousands manufactures.
A computer system commonly manages a large amount of data objects. In struc-
ture it is similar to a department store, with its thousands of items of merchandise,
or a large engineering project, with hundreds of subcontractors. Therefore, we need
316 T. Wu

to group data objects into types. Types are discriminates to each other, unless a
type conversion function is applied that can converts from one type to another. Data
objects in a computer system are analogous to merchandise in a department store or
devices, parts, and equipment in an engineering project. The association rules used
to manage merchandise, parts, devices, and equipments are similar to respect oper-
ations of types in a computer system. To create a new department in a department
store is to expand its business; much like creating a new type in a computer system
enlarges its application domain [4] to multidisciplinary areas and to the control of
more complex physical systems. These new types are called user-defined types.
Programming language likes FORTRAN IV have only basic types and do not
allow users to define their own types. Therefore, its application domain is so limited.
However, the Ada and the C CC programming languages do provide user-defined
type capability; and their application domain is not limited [4]. Today, these two
languages are considered general purpose programming languages. Hence, user-
defined types are vital for this complex and chaotic world in computing.

27.3 Modularization and Communication in the Chaotic World

Modularity partitions a complicated problem into sub-problems, and we implement

them as sub-programs. A sub-program is a unit of a program declared out of line
and invoked via calls. The purposes of a subprogram are many folds:
1. Result of modular design of program
2. Factoring of common logic that occurs several places in a program
3. Parameterized calls allow operating on different objects at different times, and
4. Simplifying and easing the complexity of the problem
In general, there two distinct subprogram types exist in commonly used program-
ming languages.
1. Procedures. It represents computational segments, and its results may be passed
back via parameters or side effects. These are used wherever statements are
permitted.
2. Functions. It returns values of a designated type. These are used anywhere an
expression is accepted.
Communication between the main program and subprograms or from one subpro-
gram to another occurs via parameter passing. Each programming language defines
its own parameter passing mechanisms. There are five parameter-passing mecha-
nisms in current commonly used programming languages such as parameter passing
by value, by reference, by name, by value-result, and by copy rules that including
copy- in, copy-out, and copy-in out rules [8, 9].
Among these five parameter-passing mechanisms, parameter passing by value,
parameter passing by reference, and parameter passing by name can be defined
mathematically below: For a parameter pass, in evaluating a variable name to get
27 Understanding Programming Language Semantics for the Real World 317

a value by finding the location associated with the name and extracting the value
from the location. Let names, locations, and values be three sets, we define two
mappings:

W names ! locations; and

 W locations ! values

Differences in the parameter passing mechanism are defined by when and  are
applied.
1. Parameter passing by value
and  both applied at point of call, argument completely evaluated at point of
call.
2. Parameter passing by reference (location)
Location is determined at the point of call and location is bound as the value of
parameter. is applied at point of call and  is applied with every reference to
the parameter.
3. Parameter passing by name
At time of call, neither nor the  is applied. and  are applied with every
reference to the parameter.
These three parameter-passing mechanisms formed hierarchical structures; the dia-
gram is given in Fig. 27.1. For parameter passing by value, both and  are applied
at calling time; for parameter passing by reference, is applied at calling time and 
is applied at reference time; for parameter passing by name both and  are applied
with every reference to the parameter.
The ALGOL, C , C CC, and Pascal languages provide parameter passing by
value mechanism; and it is a convenient and effective method for enforcing write
protection. The ALGOL, C , C CC, and Pascal languages also implement param-
eter passing by reference. This eliminates duplication of memory. But, there are
disadvantages in the parameter passing by reference. First, it will likely be slower
because one additional level memory addressing is needed compared to parameter
passing by value. Second, if only one-way communication to the called subprogram
is required, unexpected and erroneous changes may occur in the actual parameter.

Name
Reference

Value

Fig. 27.1 Hierarchical structures of parameter passing

318 T. Wu

Finally, parameter passing by reference can create aliases, which are harmful to
readability and reliability. They also make program verification difficult.
The ALGOL programming language by default provides parameter passing by
name. When implementing parameter passing by name, the system will create
a run-time subprogram to evaluate the expression in the calling unit of the pro-
gram and return the result to the called unit of the program. Therefore, it requires
some additional overhead to implement such a run-time subprogram. In the ALGOL
programming language, if one wants parameter passing by value, he or she must
declare with the word “value” for that variable in the actual parameter. ALGOL
treats parameter passing by reference, as a special case of parameter passing by
name; therefore, the programmer does not need to specify anything. We can define
a hierarchical structure for parameter passing by value, reference, and name in the
ALGOL language.
In the programming language PL/1, for actual parameters with a single variable,
PL/1 uses parameter passing by value. However, for a constant or an expression as
an argument in the calling statement, PL/1 refers to a dummy formal parameter in
the called subprogram and it implements a default value or parameter passing by
value.
In most FORTRAN implementations before FORTRAN 77, parameters were
passed by reference. In later implementations parameter passing by value-result
has been used commonly.
For the Ada language, parameter passing has three modes: mode in, mode out,
and mode in out. These are different from parameter passing by value and by
reference [10–13].
Mode in
This is the default mode (i.e., in may be omitted).
The actual parameter is copied into a local variable. The actual parameter must
have a defined value at the point of call.
The actual parameter may be an expression of compatible type.
Mode out
The result is copied into the actual parameter upon exit.
The actual parameter may be a variable of compatible type.
The actual parameter need not have a value upon entry.
Mode in out
The value of the actual parameter is copied into a local variable upon entry.
The value of local parameter is copied into the actual parameter upon exist.
The actual parameter must be a variable with a defined value upon entry.
These parameter-passing mechanisms are serving communication facilities between
main program and its subprograms or between one subprogram and another. The
purpose of engineering computing is to solve problems in the complicated world by
means of granulation [14,15], organization, and causation. Granulation subdivides
the problem into a set of more manageable sub-problems. Granulation is an effective
27 Understanding Programming Language Semantics for the Real World 319

tool to modularize the original problem and to write it into subprograms. From a pro-
grammer viewpoint, communication between the main program and subprograms
and the communication from one subprogram to another is the causation within
the program. This paper emphasizes using features of programming languages for
engineering computation. It is worth noting that the Ada programming language
is designed for embedded systems, safety-critical software, and large projects that
require high portability, reliability, and maintainability. For example, over 99% of
the aviation software in the Boeing 777 airplane uses the Ada language [16]. It is
not surprisingly; the Ada language was the first object-oriented design programming
language to be accepted as an International Standard.
Today, we design software systems to meet complex application requirements.
Almost all the activities in human society, the biological world, physical systems,
and engineering projects are concurrent or parallel; and purely sequential activities
are special cases. Therefore, concurrency reflects the nature of designing software
projects. In the next section, we will address the multitasking features in the Ada
programming language [17, 18].

27.4 Parallel or Concurrency Is the Nature of the World

Among all commonly used programming languages, the Ada language has the most
complete and best features for multitasking. Multitasking permits a programmer
to partition a big job into many parallel tasks [1,13]. Other programming languages
like the C and C CC programming languages can only apply some predefined
functions. Thus, they are lack of flexibility and limit their applicability.
A task is a unit of computation that can be scheduled independently and in par-
allel with other such units. An ordinary Ada program can be thought of as a single
task; in fact, it would be called the main task. Other tasks must be declared in the
main task (as subtasks) or be defined in a package [7, 10, 13, 19]. Several indepen-
dent tasks are often to be executed simultaneously in an application. Ada tasks can be
executed in true parallelism or with apparent concurrency simulated by interleaved
execution. Ada tasks can be assigned relative priorities and the underlying operat-
ing system can schedule them accordingly. A task is terminated when its execution
ends. A task can be declared in packages, subprograms, blocks, or other tasks. All
tasks or sub-tasks must terminate before the declaring subprogram, block, or task
can be terminated.
A task may want to communicate with other tasks. Because the execution speed
of tasks cannot be guaranteed, a method for synchronization is needed. To do
this, the Ada language requires the user to declare entry and accept statements in
two respective tasks engaged in communication. This mechanism provides for task
interaction and is called a rendezvous in the Ada language [19].
The Ada language also gives an optional scheduling called a priority that is asso-
ciated with a given task. A priority expresses relative urgency of the task execution.
An expression of a priority is an integer in a given defined range. A numerically
320 T. Wu

smaller value for priority indicates lower level of urgency. The priority of a task, if
defined, must be static. If two tasks with no priorities or two tasks of equal priority
exist, they will be scheduled in an arbitrary order. If two tasks of different priorities
are both eligible for execution, they could sensibly be executed on the same proces-
sor. A lower priority task cannot execute while a higher priority task waits. The Ada
language forbids time-sliced execution scheduling for tasks with explicitly speci-
fied priorities. If two tasks of prescribed priorities are engaged in a rendezvous, the
rendezvous is executed with the higher of the two priorities. If only one task has a
defined priority, the rendezvous is executed at least at that priority.
A task may delay its own execution or put itself to sleep and not use process-
ing resources while waiting for an event to occur by a delay statement. The delay
statement is employed for this purpose. Zero or negative values have no effect. The
smallest delay time is 20 ms or 0.020 s. The maximum delay duration is up to 86,400
s or 24 h. The duration only specifies minimum delay; the task may be executed any
time thereafter, if the processor is available at that time.
The Ada language also provides a select statement. There are three forms of
select statements. Selective wait, conditional entry call, and timed entry call. A
selective wait may include (1) a terminate alternative, (2) one or more delay alter-
natives, or (3) an else part, but only one of these possibilities is legal. A task may
designate a family (an array) of entries by a single name. They can be declared as:
Entry request (0..10) (reqcode: integer);
Entry alarm (level); – where type level
– must be discrete
An accept statement may name an index entry
Accept request (reqcode: integer) (0) do . . .
Accept request (1) (reqcode: integer) do . . .
Accept alarm (level);

An entry family allows an accepting task to select entry calls to the same function
deterministically.
Among all commonly used programming languages, the Ada language is the
unique one that provides multitasking features at the programming level, and it
is very important useful feature for modeling and simulating of real-time and
concurrent events in programming.
The goals of computing are reliability, efficiency, accuracy, and ease of use.
From the programming point of view, to provide reliable computation is to prevent
or eliminate overflow, underflow, and other unexpected conditions so that a program
can be executed safely, completely, and efficiently. Efficient computation requires
an effective computational algorithm for the given problem using proper program-
ming language features for that computation. For accurate computation, one should
consider problem solving capability, accuracy features, and parallel computation
abilities in a given programming language. For ease of use, the software engineer
should put himself in the situation of the user. A software engineer should remember
that users have a job to be done, and they want the computer system to do the job
27 Understanding Programming Language Semantics for the Real World 321

with a minimum of effort. In the next section, we will discuss issues of accuracy
and efficiency of the Ada language in numerical computation capability [8].

27.5 Accuracy and Efficiency Are Required by This Sophicificate

World

The area of numerical computation is the backbone of computer science and all
engineering disciplines. Numerical computation is critical to real world engineering
applications. For example, on November 10, 1999, the U.S. National Aeronautics
and Space Administration (NASA) reported that the Mars Climate Orbiter Team
found:
The ‘root cause’ of the loss of the spacecraft was the failed translation of English units into
metric units in a segment of ground-based, navigation-related mission software as NASA
previously announced [20].

This example indicates that numerical computation and software design are crucial
tasks in an engineering project. The goal of numerical computation is to reach to
a sufficient level of accuracy for a particular application. Designing software for
efficient computation is another challenge.
For engineering applications, we need to deal with many of numerical compu-
tations. Among most commonly used programming languages, the Ada language
has the best numerical computation capability. From the precision aspect, the Ada
language allows a user to define his own accuracy requirement. This section will
address the Ada language numerical computation capability [8]. To deal this, we
should consider the following four criteria: problem solving capability, accuracy of
computation, execution time for solving problems, and the capability of parallelism
1. Problem solving capability: The Ada language provides user-defined types and
separate compilation. The former supports programmers solving a wide range
of engineering problems and the latter permits development of large software
systems. The Ada language provides data abstraction and exception handling
that support information hiding and encapsulation for writing a reliable program.
In the real world, many engineering projects consist of concurrent or parallel
activities in their physical entities. Ada multitasking meets this requirement [21,
22]. In fact, multiprocessor computer systems are now available, thus simulating
a truly parallel system becomes possible.
2. Precision and accuracy: The Ada language’s real number types are subdivided
into float-point types and fixed-point types. Float-point type have values are
numbers
with the format, ˙:dd::d  10˙dd . Fixed-point types have values with the formats
˙dd.ddd; ˙dddd:0 or ˙0:00ddd [1, 11, 13].
For the float-point number types, model numbers other than zero, the numbers
that can be represented exactly by a given computer, are of the form:

sign  mantissa  .radix   exponent/

322 T. Wu

In this form, sign is either C1 or 1; mantissa is expressed in a number base given
by radix and exponent is an integer. The Ada language allows the user to specify the
number of significant decimal digits needed. A floating-point type declaration with
or without the optional range constraint is shown:

type T is digit DŒrange L::RI

In addition, most Ada compilers provide the types long float and long long float
(used in package standard) and f float, d float, g float, and h float (used in package
system) [22]. The size and the precision of each of the Ada floating-point types are
given as follows:

Type Size(bits) Precision (digits)

f float 32 6
d float 64 9
g float 64 15
h float 128 33

The goal of computation is accuracy. Higher accuracy will provide more reliability
in the real-time environment. Sometimes, a single precision or a double precision
of floating point numbers in FORTRAN 77 [21] is not enough for solving some
critical problems. In the Ada language one may use the floating point number type:
long long float (h float) by declaring digit 33 to use 128 bits for floating point num-
bers provided by Vax Ada [23] to provide a precision of 33 decimal digits accuracy,
and the range of exponent is about from 10134 to 10C134 or -448 to +448 of base
2 [1,11,13,24] for the range. The author has employed this special accuracy feature
in the computation of hypergeometric distribution function [25, 26].
For the fixed-point types, the model numbers are in this form:

sign  mantissa  small

The sign is either C1 or 1; mantissa is a positive integer; small is a certain positive
real number. Model numbers are defined by a fixed-point constraint, the number
small is chosen as the largest power of two that is not greater than the delta of a
fixed accuracy definition. The Ada language permits the user to determine a possible
range and an error bound which is called delta for computational needs. Examples
are the follows:

Overhead has a delta of 0:01I

Overhead has a range  10E5–1:0E5I

These indicate small is 0.0078125 which is 27 and model numbers are
12; 800; 000small toC12; 800; 000small. The predetermined range provides a
reliable programming environment. The user assigned error bound delta guarantees
an accurate computation. These floating-point number and fixed-point number
27 Understanding Programming Language Semantics for the Real World 323

types not only provide good features for real-time critical computations, but also
give extra reliability and accuracy for general numerical computations.

3. The Ada for parallel computation: The author has used exception handlings
and tasks [17, 19] for computation of a division of a product of factorials and
another product of factorials in the computation of the hypergeometric distri-
bution function [26, 27]. Exception handling is used to prevent an overflow or
underflow of multiplications and divisions, respectively. The tasks are used to
compute the numerator and denominator concurrently. In addition, tasks and
exception handling working together can minimize the number of divisions
and maximize the number of integer multiplications in both of the numerator
and denominator, reduce round off errors, and obtain the maximum accuracy.
When both products in the numerator and denominator have reached a maxi-
mum before an overflow occurs, both task one and task two stop temporarily and
invoke task three to perform a division of the products that have been obtained
in the numerator and denominator before an overflow occurs. After task three
completes its job, task one and task two resume their computation and repeat this
procedure until the final result is obtained. These tasks work together and guar-
antee that the result of this computation will be the most accurate and the time
for the computation is reasonable. The author has performed these computations
on a single processor machine, so the parallelism is a logical parallelism. If one
has a multiprocessor machine, he can perform an actual parallelism, tasking and
exception handling can easily be employed in the computation of the hypergeo-
metric distribution function and some computation results and required time for
this problem are given in [19], along with those for the computation of the multi-
nomial distribution function, multivariate hypergeometric distribution function,
and other comparable functions. We conclude here that it is not possible to carry
out the computation without using these Ada special features.
4. Execution time: In the 1990s, compiler technology was inadequate to support
many Ada features. In a real-time system, the response time for multitasking fea-
tures was seemed not fast enough. Therefore, speed was an important criterion for
choosing an Ada compiler for real-time applications. However, the second gen-
eration of Ada compilers has doubled the speed of the first generation compilers.
Today, compilers are fast enough to support all Ada features. Currently, Ada com-
pilers are available for supercomputers, mainframe computers, minicomputers,
and personal computers at reasonable prices.
In running an application, a program crash is a disaster. If it is not possible to prevent
a crash from occurring, the programmer should provide some mechanisms to handle
it, to eliminate it, or to minimize its damage. In the next section, we will address
exception-handling features for these purposes.
324 T. Wu

27.6 Exception Handling Is the Safe Guard in the Dangerous

World

An exception is an out-of-the-ordinary condition, usually representing a fault state

that can cause a program crash. The Ada language provides for detection and han-
dling of such conditions [1,11,25]. It is raised implicitly from an operation of integer
overflow during the evaluation of an expression, or assigning a negative value to a
type with positive data item. It is also raised explicitly as a result of checking when
a determinant is found to be zero during matrix inversion or a stack is found to
be empty during “pop” operation. It is not always possible or practical to avoid
all exceptions. Allowing exceptions to occur without providing a means to handle
the condition could lead to an unreliable program or a crash, however. Therefore,
exception handling is very important in designing a programming language [12,25].
The Ada language provides five predefined exceptions. They are:
Constraint error

It possibly the most frequently used exception in Ada programs.

Constraint error is raised when a range constraint is violated during assignment
or a discriminant value is altered for a constrained type.

Numeric error
When an arithmetic operation cannot deliver the correct result
Overflow or underflow condition occurs
Storage error
It is raised if out of memory during the elaboration of a data object
During a subprogram execution, creation of a new-access type object

Tasking error
It is raised during inter-tasking communication

Program error
A program attempts to enter an unelaborated procedure, e.g. a forward is not
declared
The Ada language encourages its user to define and raise his exceptions in order
to meet the specific purposes needed. An exception handler is a segment of
subprogram or block that is entered when the exception is raised. It is declared
at the end of a block or subprogram body. If an exception is raised and there is
no handler for it in the unit, then the execution of the unit is abandoned and the
exception is propagated dynamically, as follows:

If the unit is a block, the exception is raised at the end of the block in the
containing unit.
27 Understanding Programming Language Semantics for the Real World 325

If the unit is a subprogram, the exception is raised at the point of call rather
than the statically surrounding units; hence the term dynamic is applied to the
propagation rule.
If the unit is the main program, program execution is terminated with an appro-
priate (and nasty) diagnostic message from the run time support environment.
The Ada language strongly encourages its user to define and use his own excep-
tions even if they manifest in the form of predefined exceptions at the outset, rather
than depend on predefined ones. The following is an example that shows the excep-
tion, handler, and propagation working together for the computation of factorial
function.
Function Power (n, k: natural) return natural is
Function Power (n, k: natural) return natural is
begin – inner
if k < 1 then
return 1;
else
return n * Power(n, k1);
end if;
end Power; – inner
begin – outer
return Power(n, k);
Exception
when numeric error = >
return natural’last;
end Power; – outer
The advantage of this segment of code is that when an overflow condition occurs,
the inner Power function will exit. Once the outer function returns the
natural’last, it is not possible to get back to the exact point of the exception.
Function Power (n, k: natural) return natural is
begin
if k < 1 then
return 1;
else
return n * Power (n, k1);
end if;
Exception
when numeric error = >
return natural’last;
end Power;

If the function had been written as above, each execution of the function would
encounter an exception when an overflow occurs. An undesired example is given as
follows for comparison. This function will continuously raise exceptions when the
first overflow is detected and at the end of all the recursive calls.
326 T. Wu

Designing a programming language for application to all human activities and

their needs is not an easy task. However, the language we want for engineering
computation must be a reliable one. The Ada language is a language that provides
complete exception-handling facilities. For a program to be considered reliable, it
must be operate sensibly even when presented with improper input data as well as
a software or hardware malfunction. A truly reliable program must monitor itself
during execution and take some appropriate actions when a computation is entering
an unexpected state. This is the purpose of an exception handler:

27.7 Conclusion

In this paper we have discussed the goal of engineering computing as accuracy,

efficiency, reliability, and ease of use. To accomplish the goals of engineering
computation is not easy; it involves human intelligence, knowledge of variety of
programming languages, and various programming facilities. In this paper, we
have examined many programming language features that are critically important
for engineering applications. Some of these features such as subprograms, user-
defined types, and basic numerical computation facilities are provided by most
commonly used languages like FORTRAN 90, the C , the C CC, Pascal, and Ada
languages. This paper has highlighted the use of the Ada programming language’s
strong typed feature, predefined exception handling, user defined exception han-
dling, and user-defined types for developing reliable programs. All of these good
features in the Ada programming languages are unique among all commonly used
programming languages. In particular, the use of the Ada language’s delta, dig-
its, and model numbers for designing engineering projects, which require accurate
critical numerical computation, are very important. For the Ada language, most of
their compilers allow users to have a 128-bit floating-point number or 33 significant
digits for numerical computation. Programmers may specify required accuracies
through digits and delta clauses for floating-point numbers and fixed-point num-
bers, respectively. In addition, Ada’s exception handling can prevent overflow or
underflow during the execution of programs and multitasking can perform paral-
lel computations. Therefore, from the software reliability point of view, the Ada
language is better than FORTRAN, the C , the C CC and languages in numerical
computation for engineering applications.
Today, most engineers have only FORTRAN or the C =C CC programming lan-
guages for computing. Perhaps some of them have used Mathematica, Mathlab,
or Maple for computation. All these mathematical packages have one or more of
following drawbacks. Each mathematical software package has its own input and
output formats; these formats might not compatible with an engineering project.
Each of these software packages has its predefined accuracy, which might not meet
the needs of engineering projects. All of these mathematical packages are designed
for classroom teaching or laboratory research computations; efficiency is not crit-
ically important for these purposes. Moreover, these mathematical packages are
27 Understanding Programming Language Semantics for the Real World 327

not for real-time or embedded systems applications; they lack exception-handling

capabilities. To do this, each engineering student needs to take at least two courses
in computing, one is to learn the basic computation capabilities, a FORTRAN or
the C =C CC course will serve for this purpose, the other is to study how to build
efficient, accurate, reliable, and ease of use software systems to satisfy all engineer-
ing domain needs. However, the instructor must have knowledge about engineering
experiences in real world, and backgrounds in inter-disciplinary applications in
order to qualify for this purpose and lead students to design, implement, and handle
engineering projects for the challenges of the real world.

References

1. Demillo, R. and Rice, J. eds. Studies in Computer Science: In Honor of Samuel D. Conte,
Plenum, New York, 1994.
2. Myers, G. J. Software Reliability, Wiley, New York, 1976.
3. Pfleeger, S. L. Software Engineering, the Production of Quality Software, 2nd edition,
Macmillan, New York, 1991.
4. Wu, T. Some tips in computer science, a talk given at University of Illinois at Springfield, Dec.
2, 2004.
5. China Daily, CHINAdaily.com.cn, October 28, 2006.
6. The Aviation Book, A visual encyclopedia of the history of aircraft, www. Chronicale-
Books.com
7. Boeing 777, globalsecurity.org/military/systems/ aircraft/b777.html
8. Pratt, T. W. and Zelkowitz, M. V. Programming Languages, Design and Implementation, 3rd
edition, Prentice-Hall, Englewood Cliffs, NJ, 1996.
9. Sebesta, R. W. Concepts of Programming Languages, 6th edition, Addison-Wesley, Reading,
MA, 2003.
10. Barnes, J. G. P. Programming in Ada, 3rd edition, Addison-Wesley, Reading, MA, 1989.
11. Barnes, J. Programming in Ada 95, Addison-Wesley, Reading, MA, 1995.
12. Smedema, C. H., et al. The Programming languages Pascal, Modula, Chill, and Ada, Prentice-
Hall, Englewood Cliffs, NJ, 1983.
13. Watt, D. A., et al. Ada Language and Methodology, Prentice-Hall, Englewood Cliffs, NJ, 1987.
14. Zadeh, L. A. Fuzzy sets and information granularity, In M. M. Gupta, P. K. Ragade, R. R.
Yager, eds., Advances in Fuzzy Set Theory and Applications, North-Holland, Amsterdam, pp.
3–18, 1979.
15. Zadeh, L. A. Towards a theory of fuzzy information granulation and its centrality in human
reasoning and fuzzy logic, Fuzzy Sets and Systems. 19, 111–117, 1997.
16. Ada information clearance house, the web site: www.adaic.org/atwork/boeing.html
17. Booch, G. Software engineering with Ada, 2nd edition, Benjamin/Cummings Publishing,
Reading, MA, 1987.
18. Habermann, A. N. and Perry, D. E. Ada for Experienced Programmers, Addison-Wesley,
Reading, MA, 1983.
19. Welsh, J. and Lister, A. A Comparative Study of Task Communication in Ada, Software
Practice and Experience. 11, 257–290, 1981.
20. Mas project http://mars.jpl.nasa
21. Parnas, D. L. Is Ada Too Big? (letter) Communications of ACM. 29, 1155–1155, 1984.
22. Wichmann, B. A. Is Ada Too Big? A Designer Answers the Critics, Communications of ACM.
29, 1155–1156, 1984.
23. Struble, G. Assembler Language Programming, The IBM System/370 Family, Addison-
Wesley, Reading, MA, 1984.
328 T. Wu

24. Vax Ada Language Reference Manual, Digital Equipment Corporation, Maynard, MA, 1985.
25. Wu, T. Built-in reliability in the Ada programming language, Proceedings of IEEE 1990
National Aerospace and Electronics Conference, pp. 599–602, 1990.
26. Wu, T. An Accurate Computation of the Hypergeometric Distribution Function, ACM
Transactions on Mathematical Software. 19 (1), 33–43, 1993.
27. Wu, T. Ada programming language for Numerical Computation, Proceedings of IEEE 1995
National Aerospace and Electronics Conference, pp. 853–859, 1995.
28. Joint Task Force for Computing Curricula 2004, Overview Report. ACM and IEEE-CS, 2004.
Chapter 28
Analysis of Overhead Control Mechanisms
in Mobile AD HOC Networks

S. Gowrishankar, T.G. Basavaraju, and SubirKumarSarkar

Abstract In this chapter a description of several techniques that are proposed for
minimizing the routing overhead in ad hoc routing protocols is discussed. Differ-
ent algorithms are classified into several categories such as clustering, hierarchical,
header compression and Internet connectivity to mobile ad hoc networks based on
main objective of minimizing the routing overheads. Clearly, the selection of this
area in this paper is highly subjective. Besides, the routing overhead minimizing
schemes discussed in this chapter, there are dozen of research schemes that are cur-
rently the focus of the research community. With this tutorial, readers can have a
comprehensive understanding of different schemes that are employed to reduce the
routing overhead.

Keywords Routing overhead  minimizing  Mobile AD HOC Network  perfor-

mance comparison

28.1 Introduction

Mobile Ad hoc networks are autonomous systems formed by mobile nodes with-
out any infrastructure support. Routing in MANET is challenging because of the
dynamic nature of the network topology. Fixed network routing protocols can
assume that all the routers have a sufficient description of the underlying network,
either through global or decentralized routing tables. However, dynamic wireless
networks do not easily admit such topology state knowledge. The inherent random-
ness of the topology ensures that a minimum overhead is associated with the routing
in MANET and is greater than that of fixed networks. It is therefore of interest to

S. Gowrishankar (B)
Department of Electronics and Telecommunication Engineering, Jadavpur University,
Kolkata 700032, West Bengal, India,
E-mail: gowrishankarsnath@acm.org

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 329
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
330 S. Gowrishankar et al.

know how small the routing overhead can be made for a given routing strategy and
random topology [1].
To evaluate the performance of routing protocols in MANET, several perfor-
mance metrics such as packet delivery ratio, average end-to-end delay and routing
overhead are commonly used. Among these metrics routing overhead is the impor-
tant one as it determines the scalability of a routing protocol. Routing Overhead
means how many extra messages were used to achieve the acceptance rate of
improvement.
To evaluate the routing overhead in mobile ad hoc routing protocols different
methods are followed. They are (a) simulations, (b) physical experiments, and (c)
theoretical analysis [2].
In simulations a controlled environment is provided to test and debug many of
the routing protocols. Therefore, most of the literature [3–5] evaluates the control
overhead in routing protocols using software simulators like NS-2 [6], Glomosim
[7], Qualnet [8] and OPNET [9]. However, simulations are not foolproof method
and it may fail to accurately reveal some critical behaviors of routing protocols, as
most of the simulation experiments are based on simplified assumptions.
Physical experiments evaluate the performance of routing protocols by imple-
menting them in real environment. Some of the papers in the literature evaluate
routing overhead in real physical environment [10, 11]. But, physical experiments
are much more difficult and time consuming to be carried out.
Analysis of routing overhead from a theoretical point of view provides a deeper
understanding of advantages, limitations and tradeoffs found in the routing proto-
cols in MANET. Some of the papers in literature [12–15] have evaluated routing
protocols in MANET from theoretical analysis perspective.

28.2 Theoretical Analysis of Overhead in Hierarchical

Routing Scheme

Traditionally the routing schemes for ad hoc networks are classified into proactive
and reactive routing protocols. Proactive protocols like DSDV [16] and OLSR [17]
maintain routing information about the available paths in the network even if these
paths are not currently used. The drawback of such paths is that it may occupy a
significant part of the available bandwidth. Reactive routing protocols like DSR,
TORA [5] and AODV [18] maintain only the routes that are currently available.
However, when the topology of network changes frequently, they still generate large
amount of control traffic.
Therefore, the properties of frequent route breakage and unpredictable topology
changes in MANET make many of the routing protocols inherently not scalable with
respect to the number of nodes and the control overhead. In order to provide routing
scalability a hierarchy of layers is usually imposed on the network. Scalability issues
are handled hierarchically in ad hoc networks. Many hierarchical routing algorithms
28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 331

are adopted for routing in ad hoc wireless networks. For e.g., cluster based routing
and the dominating set based routing.
Sucec and Marsic provide a formal analysis of the routing overhead i.e., they
provide a theoretical upper bound on the communication overhead incurred by the
clustering algorithms that adopt the hierarchical routing schemes. There can be
many scalability performance metrics like hierarchical path length, least hop path
length and routing table storage overhead. Among these metrics, control overhead
per node (‰) is the most important one because of scarce wireless link capacity,
which has severe performance limitation. The control overhead ‰ is expressed
as a function of jVj, where V is the set of network nodes. It is shown that with
reasonable assumptions, the average overhead generated  per node per second is
only polylogarithmic in the node count i.e., ‰ D O log2 jV j bits per second per
node [19].
Communication overhead in hierarchically organized networks may result from
the following phenomenon: (a) Hello Protocols; (b) level-k cluster formation and
cluster maintenance messaging, k 2 f1; 2: : :Lg, where L is the number of levels
in the clustered hierarchy; (c) flooding of the cluster topology updates to cluster
members; (d) addressing information required in Datagram headers; (e) location
management events due to changes in the clustered hierarchy and due to node
mobility between clusters; (f) hand off or transfer of location management data; (g)
Location query events. Total communication overhead per node in hierarchically
organized networks is the sum of the above contributing elements.
The control overhead and network throughput under a cluster based hierarchi-
cal routing scheme is discussed in [20]. The authors claim that when the routing
overhead is minimized in a hierarchical design then there is a loss in the throughput
from the same hierarchical design. A strict hierarchical routing is assumed which is
not based on any specific routing protocol. In MANET, hierarchical routing proto-
cols do not require every node to know the entire topology information. Only a few
nodes called the cluster head nodes need to know about the topology information
and all the other nodes can simply send their packets to these cluster heads.
Hierarchical routing protocols reduce the routing overhead, as lesser nodes need
to know the topology information of an ad hoc network. The throughput  2of ad hoc
network with hierarchical routing scheme is smaller by a factor of O N N1 , where
N2 is the number of cluster head nodes and N1 is the number of all the nodes in
the network. Hence, the authors claim that there is a tradeoff between the gain from
the routing overhead and the loss in the throughput from the hierarchical design
of the ad hoc routing protocols.
The control overhead in a hierarchical routing scheme can be due to packet trans-
missions per node per second (), due to the maintenance of routing tables as well
as due to the address management or location management. Therefore the overhead
 required by hierarchical
 routing  is a polylogarithmic function of the network node
count (N) i.e., ˆ D ‚ log2 jN j packet transmissions per node per second. In this
equation, overhead due to hierarchical cluster formation and location management
are identified [15].
332 S. Gowrishankar et al.

28.3 Theoretical Analysis and Overhead Minimizing Techniques

for AD HOC Networks Using Clustering Mechanisms

The concept of dividing the geographical regions into small zones has been pre-
sented as clustering in the literature.
Clustering basically transforms a physical network into a virtual network of inter-
connected clusters or group of nodes. These clusters are dominated by clusterheads
(CH) and connected by gateways or border terminals as shown in Fig. 28.1. Any
node can be CH, if it has the necessary functionality such as the processing and
transmission power. The node registered with the nearest CH becomes the member
of that cluster. By partitioning a network into different clusters both the storage and
communication overhead can be reduced significantly.
Different clustering algorithms may use different clustering schemes but gener-
ally three different types of control messages are needed: (a) Beacon messages also
known as Hello messages are used to learn about the identities of the neighboring
nodes, (b) cluster messages are used to adapt to cluster changes and to update the
role of a node, (c) route messages are used to learn about the possible route changes
in the network [21].
The various types of control messages overhead are (a) Hello Overhead: To
reduce the hello overhead messages; the frequency of hello messages generated
by a node to learn about its neighboring node when a new link is formed should
be at least equal to the link generation rate. The link generation between any two
nodes can be notified by sending the hello messages and each of the nodes can hear
the hello message sent by the other node. (b) Cluster message overhead due to link
break between cluster members and their cluster heads: This event causes the node
to change its cluster or become a cluster head when it has no neighboring cluster-
ing heads. The cluster members send the cluster messages due to this type of link
changes. To minimize the control message overhead the ratio of such link breaks
to total link breaks should be equal to the ratio of links between the cluster mem-
bers and cluster heads divided by the total number of links in the entire network. (c)
Cluster message overhead due to link generation between two cluster heads: When

Fig. 28.1 Route establishment in clustering mechanism

28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 333

a link is generated between two cluster heads, one of the cluster heads needs to give
up its cluster head role, which is decided by the clustering algorithm. Every time a
link between two cluster heads appears, the number of cluster messages generated
is same as the number of nodes in the cluster that needs to undergo reclustering. (d )
Routing overhead: When a particular node in the cluster should be updated with the
route to other nodes in the cluster, the routing storage overhead is proportional to
the size of the cluster.
An analytical study on routing overhead of two level cluster based routing proto-
cols for mobile ad hoc networks is done in [2]. Routing protocols can be summarized
into generic proactive routing protocol and a generic reactive routing protocol. It’s
generic because there may be some different strategy employed for each of the
groups, but the underlying nature of the routing is similar.
In two level cluster based routing scheme, the routing is divided into two separate
parts, i.e., routing among different clusters (i.e., intercluster routing) and routing
within a cluster (i.e., intracluster routing). Since there are two types of routing
schemes i.e., proactive and reactive which can be employed in intercluster rout-
ing and intracluster routing, there are totally four types of two level hierarchical
routing schemes with different combinations of them. Hence we have proactive to
proactive, reactive to reactive, proactive to reactive and reactive to proactive routing
scheme.
When a proactive scheme is adapted for intercluster routing each cluster head
periodically collects its local cluster topology and then broadcasts it to its direct
neighbor cluster head via gateways. When a reactive routing scheme is used for
inter cluster routing, a route request for a route to the destination node that is in
another cluster will be broadcasted among cluster heads. When a proactive rout-
ing scheme is utilized for intracluster routing, each node broadcasts its local node
topology information, so the route to the destination within the same cluster will be
available when needed. When a reactive routing scheme is employed for intracluster
routing, a route request to the destination will be flooded within the same cluster.
Thus a proactive to proactive routing scheme will work as a proactive routing
protocol with a hierarchical structure. The proactive to proactive routing scheme
produces overhead due to periodical topology maintenance of
 
1 2 1
O N C N2
n kNc

where n is the total number of clusters in the network, N is the network size, k is the
cluster radius, Nc is the cluster size.
A reactive to reactive routing protocol operates as a purely reactive routing pro-
tocol with a hierarchical structure. Reactive to reactive routing protocol yields a
routing overhead due to route discovery of
 
1 2
O N :
k
334 S. Gowrishankar et al.

In a proactive to reactive routing scheme the cluster heads periodically exchange

topological information and a node always sends a route request to its cluster head
where there is no available route to an expected destination. Then the cluster head
will send a route reply packet to the node, which indicates that the destination is
within the local cluster or contains a route to the destination node, which is in
another cluster. Proactive to reactive routing protocol have a basic routing over-
head due to topology maintenance, cluster maintenance and route discovery which
is found to be
 
1 r 2
O N C
2
N where r is average route lifetime:
kNc nk

In a reactive to proactive routing scheme each node in the cluster will periodi-
cally broadcast local node topology information within the cluster. Thus, when the
destination is within the cluster, the route is immediately available. Otherwise, the
node will send a route request packet to its cluster head, which will be broadcasted
among the cluster heads. Reactive to proactive routing protocol have a basic routing
overhead due to cluster maintenance and route discovery, which is approximately
equal to  
1 2 1 2
O N C N :
n k
A mathematical framework for quantifying the overhead of a cluster based rout-
ing protocol (D-hop max min) is investigated by Wu and Abouzeid [13]. The authors
provide a relationship between routing overhead and route request traffic pattern.
From a routing protocol perspective, ‘traffic’, could be defined as the pattern by
which the source destination pairs are chosen. The choice of a source destination
pair depends on the number of hops along the shortest path between them. Also the
network topology is modeled as a regular two-dimensional grid of unreliable nodes.
It is assumed that an infinite number of nodes are located at the intersections of a
regular grid. The transmission range of each node is limited such that a node can
directly communicate with its four immediate neighbors only. It is reported that the
clustering does not change the traffic requirement for infinite scalability compared
to flat protocols, but reduces the overhead by a factor of
 
1
o where M is the cluster size:
m

Wu and Abouzeid show that the routing overhead can be attributed to events like
(a) route discovery, (b) route maintenance, and (c) cluster maintenance.
Route discovery is the mechanism where by a node i wishing to send a packet to
the destination j obtains a route to j. When a source node i wants to send a packet
to the destination node j, it first sends a route request (RREQ) packet to its clus-
ter head along the shortest path. The route reply (RREP) packet travels across the
shortest path back to the cluster head that initiated the RREQ packet. So the route
discovery event involves an RREP and RREQ processes. The overhead for RREQ
is generally higher than the RREP since it may involve flooding at the cluster head
28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 335

level. Therefore, the minimum average overhead of finding a new route is

     
f .k  3/ 6M C 6 C M 3
C f .k  2/ 4M 2 C 2  f .k  1/M M 2  1
3 .2M 2 C 2M C 1/ f .k  1/
Where, M is the radius of the cluster and changing the value of k controls the traffic
pattern.
Route maintenance is the mechanism by which a node i is notified that a link
along an active path has broken such that it can no longer reach the destination
node j through that route. When route maintenance indicates a link is broken, i may
invoke route discovery again to find a new route for subsequent packets to j. In
cluster based routing, the neighboring node sends a RERR packet to its cluster head
rather than the source node itself. The cluster head could patch a path locally with-
out informing the source node, if the failed node is not the destination node. This
is called local repair. In this case, the path is locally fixed. Also, the RERR packet
sent from a neighboring node to the cluster head is considered as the cluster main-
tenance overhead. Therefore the minimum overhead required for route maintenance
is 4C .f .k  2/  f .k  1// where, k > 3 and C is a constant.
Clustering incurs cluster maintenance overhead, which is the amount of control
packet needed to maintain the cluster membership information. The membership
in each cluster changes overtime in response to node mobility, node failure or new
node arrival. The average cluster maintenance overhead is

4 .M  1/ M .M C 1/
C 2M Where, M is the radius of the cluster:
3 .2M 2 C 2M C 1/
The work done in [14] by Zhou, provide a scalability analysis of the routing
overheads with regard to the number of nodes and the cluster size. Both the inte-
rior routing overhead within the cluster and the exterior routing overhead across the
clusters are considered. The routing protocol includes a mechanism for detecting,
collecting and distributing the network topology changes. The process of detecting,
collecting and distributing the network topology changes contribute to a total routing
overhead Rt . The routing overhead can be separated into interior routing overhead
(Ri ) and the exterior routing overhead (Re ). The interior routing overhead Ri is the
bit rate needed to maintain the local detailed topology. This includes the overhead of
detecting the link status changes by sending “HELLO” message, updating the clus-
ter head about the changes in link status and maintaining the shortest path between
the regular nodes to their cluster head.
Exterior routing overhead (Re ) is the bit rate needed to maintain the global owner-
ship topology, which includes the overheads of the distributing the local ownership
topologies among the cluster heads. Hence Rt D Ri C Re
336 S. Gowrishankar et al.

28.4 Minimizing Overhead in AD HOC Networks by Header

Compression

In literature it has been studied that [22] approximately half of the packets sent
across the Internet are 80 bytes long or less. This percentage has increased over the
last few years in part due to widespread use of real time multimedia applications.
The multimedia application’s packet size is usually smaller in size and these small
packets must be added with many protocol headers, while traveling through the
networks. In Ipv4 networks there can be at least 28 bytes (UDP) or 40 bytes (TCP)
overheads per packet. These overheads consume much of the bandwidth, which
is very limited in wireless links. Small packets and relatively larger header size
translates into poor line efficiency. Line efficiency can be defined as the fraction of
the transmitted data that is not considered overhead. Figure 28.2 shows some of the
common header chains and size of each component within the chain.
Ad hoc networks create additional challenges such as context initialization over-
head and packet reordering issues associated with node mobility. The dynamic
nature of ad hoc networks has a negative impact on header compression efficiency.
A context is established by first sending a packet with full uncompressed header
that provides a common knowledge between the sender and receiver about the static
field values as well as the initial values for dynamic fields. This stage is known as
context initialization. Then the subsequent compressed headers are interpreted and
decompressed according to a previously established context. Every packet contains
a context label. Here the context label indicates the context in which the headers are
compressed or decompressed.
A novel hop-by-hop context initialization algorithm is proposed in [23] that
depends on the routing information to reduce the overhead associated with the con-
text initialization of IP headers and uses a stateless compression method to reduce
the overhead associated with the control messages. Context initialization of IP head-
ers is done on a hop-by-hop basis because the headers need to be examined in an
uncompressed state at each of the intermediate nodes. The context initialization

20 28 40 48 60

1Pv4 UDP RTP

1Pv4/UDP/RTP
20 8 12

1Pv4/TCP 1Pv4 TCP

20 20

1Pv6 UDP RTP

1Pv6/UDP/RTP
40 8 12
1Pv6 TCP
1Pv6/TCP
40 20

Fig. 28.2 Common header chains and their sizes

28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 337

overhead is reduced by caching the address information that is transmitted in the

routing messages, in order to reduce the size of the context initialization headers.
Also a stateless header compression is proposed. It is stateless because the
state of the context is fixed and it does not change with time. Header compres-
sion improves the line efficiency by exploiting the redundancies between the header
fields within the same packet or consecutive packets belonging to the same stream.
The overall result is the reduced overhead, increased network capacity and line
efficiency even in the presence of rapid path fluctuations.
An ad hoc robust header compression protocol has been proposed (ARHC) in
[24]. ARHC protocol can be used to compress UDP, TCP and raw IP headers in ad
hoc network. The mechanism of ARHC is that when the first packet of a session
arrives, the compressor generates a unique number called context ID, which indexes
the quintuplet (source address, destination address, source port, destination port,
protocol) and all the constant fields. Compressor then records the context id, quintu-
plet and all the constant fields. Then the compressor will send the full packet header
along with the context ID. Upon receiving the very first packet the decompressor
records this information. When the subsequent packets arrive later, the compressor
and decompressor act as follows. The compressor will remove the constant fields
and the quintuplet from the header and transmits only the context ID. The decom-
pressor then retrieves the quintuplets and the constant fields from the context tables
indexed by the context ID, thereby restoring the original header.

28.5 Minimizing Overhead for AD HOC Networks Connected

to Internet

Today Internet has become the backbone of the wired and wireless communica-
tion. Also mobile computing is gaining in popularity. In order to meet the rapid
growing demand of mobile computing, many of the researchers are interested in the
integration of MANET with the Internet.
When a mobile node in MANET wants to exchange packets with the Internet,
first the node must be assigned a global IP address and then the available Internet
gateways has to be discovered to connect to the Internet as shown in Fig. 28.3. But,
this is achieved at the cost of higher control overhead.
For gateway discovery, a node depends on periodic gateway advertisement. To
make efficient use of this periodic advertisement, it is necessary to limit the adver-
tisement flooding area. For gateway discovery, a node depends on periodic gateway
advertisement. To make efficient use of this periodic advertisement, it is necessary
to limit the advertisement flooding area. A complete adaptive scheme to discover IG
in an efficient manner for AODV is given in [13]. In this approach both the periodic
advertisement and adaptive advertisement schemes are used. At a relatively long
interval each gateway sends the periodic advertisement messages. Periodic adver-
tisements performed at a widely spaced interval do not generate a great deal of
overhead but still provides the mobile nodes with a good chance of finding the short-
338 S. Gowrishankar et al.

D
Internet

Internet Gateway
Internet Gateway

S S
S

Fig. 28.3 MANET connected to Internet scenario

est path to a previously used gateway. The TTL of the periodic gateway message is
used as a parameter to adjust the network conditions. A heuristic algorithm called
“Minimal Benefit Average” [25] decides the next TTL to be used for the periodic
gateway advertisement messages.
The goal of the adaptive advertisement scheme is to send advertisement pack-
ets only when the gateway detects the movement of nodes, which would result
in the paths used by the source mobile nodes communicating with the gateway
to be changed. Adaptive advertisement is performed when needed, regardless of
the time interval used for periodic advertisement. [26]. In this approach there is
reduction in overhead messages since the periodic advertisements are sent at a long
time interval and perform adaptive advertisement only if there is mobility in the
network.
The various parameters that affect the control overhead created by interoperat-
ing the ad hoc routing protocols and IP based mobility management protocols is
addressed in [27]. Mobile IP is used as the baseline mobility management proto-
col and AODV is chosen as the ad hoc routing protocol. IP tunneling is used to
separate the ad hoc network from the fixed network. In mobile IP, a mobile node
can tell which router is available by listening to router advertisements, which are
periodically broadcasted by the routers.
A fixed access router is assigned the role of mobility agent and has connection
to at least one of the MANET nodes. Such router is referred to as Ad hoc Internet
Access Router (AIAR) and it maintains a list called ad hoc list, which keeps a list
of IP address of the mobile nodes that wish to have Internet connectivity. In an
integrated network the control overhead comprises of AIAR registration packets,
routing protocol control packets, mobile IP registration packets and mobile IP router
advertisement.
In mobile IP majority of the overhead is due to the route advertisement packets
that are being repeatedly and periodically forwarded among the mobile nodes. Also,
the router advertisement used by the mobility management protocol to carry network
28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 339

information is the major source of unnecessary control overhead within MANET.

Varying TTL value is an effective mechanism to control the amount of advertisement
packets [27].
A multihop router is developed in [28] for non-uniform route connectivity with
low routing overhead. To achieve efficient route discovery and route maintenance
a new routing scheme called hopcount based routing (HBR) protocol is developed.
HBR is an on demand routing protocol. When a source node needs to discover a
route to a node on the wired network, it initiates a route discovery to the nearest
access point by broadcasting a route request packet. By utilizing the hop counts
referring to access points, the route discovery region is localized to a small area.
On receiving the route request packet, the access point responds by sending a route
reply packet to the source node. Once a route is found, the source node begins to
forward data to the access point. After the access point receives these data packets
from the source node, it then forwards these packets through the wired line to the
destination node on the wired network using the routing protocol in the wired net-
work. By using the hop count information an attempt is made to reduce the number
of nodes to whom the route request is propagated. Thus in HBR routing protocol
to construct routes from mobile nodes to access points hop count information is
utilized to localize the route discovery within a limited area in order to reduce the
routing overhead.
In this chapter a description of several techniques that are proposed for min-
imizing the routing overhead in ad hoc routing protocols is discussed. Different
algorithms are classified into several categories such as clustering, hierarchical,
header compression and Internet connectivity to mobile ad hoc networks based
on main objective of minimizing the routing overheads. Clearly, the selection of
this area in this paper is highly subjective. Besides, the routing overhead minimiz-
ing schemes discussed in this chapter, there are dozen of research schemes that are
currently the focus of the research community.
With this tutorial, readers can have a comprehensive understanding of differ-
ent schemes that are employed to reduce the routing overhead. We hope that this
discussion can facilitate researchers to move in new direction and devise new meth-
ods to reduce the control overheads that are inherently associated with the routing
protocols.

References

1. R. C. Timo and L. W. Hanlen, “MANETs: Routing Overhead and Reliability”, IEEE, 2006.
2. Z.-H. Tao, “An Analytical Study on Routing Overhead of Two level Cluster Based Routing
Protocols for Mobile Ad hoc Networks”, IEEE, 2006.
3. J. Broch, D. A. Maltz, D. B. Johnson, Y.-C. Hu, and J. Jetcheva, “A Performance Compar-
ison of Multi-hop Wireless Ad Hoc Network Routing Protocols,” inProceedings of ACM
MobiCom, Oct. 1998.
4. S. R. Das, R. Castaneda, and J. Yan, “Simulation Based Performance Evaluation of Routing
Protocols for Mobile Ad Hoc Networks,” Mobile Networks and Applications, vol. 5, no. 3,
pp. 179–189, Sept. 2000.
340 S. Gowrishankar et al.

5. A. Boukerche, “Performance Evaluation of Routing Protocols for Ad Hoc Wireless Net-

works,” Mobile Networks and Applications, vol. 9, no. 4, pp. 333–342, Aug. 2004.
6. “The network simulator – NS-2,” http://www.isi.edu/nsnam/ns
7. X. Zeng, R. Bagrodia, and M. Gerla, “GloMoSim: A Library for Parallel Simulation of
Large-Scale Wireless Networks,” Proceedings of 12th Workshop on Parallel and Distributed
Simulations, May 1998.
8. T. H. Clausen et al., “The Optimized Link State Routing Protocol, Evaluation Through Exper-
iments and Simulation”, Proceedings of IEEE Symposium on Wireless Personal Mobile
Communications, 2001.
9. “OPNET Modeler,” http://www.opnet.com/
10. R. S. Gray, D. Kotz, C. Newport, N. Dubrovsky, A. Fiske, J. Liu, C. Masone, S. McGrath, and
Y. Yuan, “Outdoor Experimental Comparison of Four Ad Hoc Routing Algorithms,” inPro-
ceedings of ACM International Symposium on Modeling, Analysis and Simulation of Wireless
and Mobile Systems, Oct. 2004.
11. E. Borgia, “Experimental Evaluation of Ad Hoc Routing Protocols,”in Proceedings of IEEE
International Conference on Pervasive Computing and Communications Workshops, May
2005.
12. N. Zhou, H. Wu, and A. A. Abouzeid, “Reactive Routing Overhead in Networks with
Unreliable Nodes,” in Proceedings of ACM MobiCom, Sept. 2003.
13. H. Wu and A. A Abouzeid, “Cluster-Based Routing Overhead in Networks with Unreliable
Nodes,” in Proceedings of IEEE WCNC, Mar. 2004.
14. N. Zhou and A. A. Abouzeid, “Routing in Ad Hoc Networks: A Theoretical Framework with
Practical Implications,” in Proceedings of IEEE INFOCOM, Mar. 2005.
15. J. Sucec and I. Marsic, “Hierarchical Routing Overhead in Mobile Ad Hoc Networks”, IEEE
Transactions on Mobile Computing, vol. 3, no. 1, pp. 46–56, Jan.–Mar. 2004.
16. T. C. Chiang and Y. M. Huang, “A Limited Flooding Scheme for Mobile Ad Hoc Networks”,
IEEE, 2005.
17. C. E. Perkins and E. M. Royer, “Ad Hoc on Demand Distance Vector Routing”, Proceedings
of IEEE Workshop on Mobile Computing Systems and Applications , pp. 90–100, Feb. 1999.
18. C. E. Perkins and P. Bhagwat, “Highly Dynamic Destinatio-Sequenced Distance Vector
Routing (DSDV) for Mobile Computers”, Proceedings of ACM SIGCOMM, 1994.
19. J. Sucec and I. Marsic, “Hierarchical Routing Overhead in Mobile Ad Hoc Networks,” Clus-
tering Overhead for Hierarchical Routing in Mobile Ad Hoc Networks”, IEEE INFOCOM,
2002.
20. Y. Qin and J. He, “The Impact of Throughput of Hierarchical Routing in Ad Hoc Wireless
Networks”, IEEE, 2005.
21. X. Mingqiang, E. R. Inn-Inn and K. G. S. Winston, “Analysis of Clustering and Routing
Overhead for Clustered Mobile Ad Hoc Networks”, IEEE ICDCS, 2006.
22. Sprint. “IP Monitoring Project”. Feb. 6, 2004. http://ipmon.sprint.com/packstat/
packetoverview.php
23. A. Jesus, A. Syeed, and H. Daniel, “Header Compression for Ad Hoc Networks”, IEEE, 2004
24. H. Wang et al., “A Robust Header Compression Method for Ad hoc Network”, IEEE, 2006.
25. P. M. Ruiz and A. F. Gomez Skarmet, “Enhanced Internet Connectivity Through Adap-
tive Gateway Discovery”, 29th Annual IEEE International Conference on Local Computer
Networks, pp. 370–377, Nov. 2004.
26. R. Vanzara and M. Misra, “An Efficient Mechanism for Connecting MANET and Internet
Through Complete Adaptive Gateway Discovery”, IEEE, 2006.
27. K. A. Chew and R. Tafazolli, “Overhead Control in Interworking of MANET and Infras-
tructure Backed Mobile Networks”, International Conference on Wireless Ad Hoc Networks,
IEEE, 2004.
28. R. Teng, H. Morikawa, et al., “A Low Overhead Routing Protocol for Ad Hoc Networks with
Global Connectivity”, IEEE, 2005.
29. X. Li and L. Cuthbert,” Stable Node-Disjoint Multipath Routing with Low Overhead in Mobile
Ad Hoc Networks”, Proceedings of the 12th Annual International Symposium on Modeling,
Analysis, and Simulation of Computer and Telecommunications System, IEEE, 2004.
28 Analysis of Overhead Control Mechanisms in Mobile AD HOC Networks 341

30. A. M. Abbas and B. N. Jain, “Mitigating Overheads and Path Correlation in Node-Disjoing
Multipath Routing for Mobile Ad Hoc Networks”, IEEE, 2006.
31. R. C. Timo and L. W. Hanlen,” Routing Overhead and Reliability”, IEEE, 2006.
32. P. M. Ruiz and A. F. Gomez Skarmet, “Reducing Data Overhead of Mesh Based Ad Hoc
Multicast Routing Protocols by Steiner Trees”, IEEE, 2004.
33. S. Bansal, R. Shorey, and A. Misra, “Comparing the Routing Energy Overheads of Ad Hoc
Routing Protocols”, IEEE, 2003.
34. R. Teng, H. Morikawa, and T. Aoyama, “A Low Overhead Routing Protocol for Ad Hoc
Networks with Global Connectivity”, IEEE, 2005.
35. E. R. Inn Inn and K. G. S. Winston, “Clustering Overhead and Convergence Time Analy-
sis of the Mobility Based Multi hop Clustering Algorithm for Mobile Ad Hoc Networks”,
Proceedings of the 11th International Conference on Parallel and Distributed Systems, IEEE,
2005.
36. S.-H. Wang, M.-C. Chan, and T.-C. Hou, “Zone ased Controlled Flooding in Mobile Ad Hoc
Networks”, International Conference on Wireless Networks, Communications and Mobile
Computing, IEEE, 2005.
37. www.scalable-networks.com
Chapter 29
Context Aware In-Vehicle Infotainment
Interfaces

H. Sharma and A.K. Ramani

Abstract The current state-of-practice for developing automotive software for

Infotainment and Telematics systems offers little flexibility to accommodate such
heterogeneity and variation. This chapter presents a framework that enables devel-
opment of context aware interaction interfaces for in-vehicle infotainment applica-
tions. Essential framework concepts are presented with development and discussion
of a lightweight component model. Finally, the design of an example application is
presented and evaluated.

Keywords In-Vehicle Infotainment  Context Aware  Interface  Context Manage-

ment  Interaction Architecture

Today’s cars provide an increasing number of new functionalities that enhance the
safety and driving performance of drivers or raise their level of comfort. In addi-
tion, infotainment systems are innovating themselves, providing luxury facilities or
enabling modern communication mechanisms. With every new generation of cars,
there are more built-in software functions.
The current state-of-practice for developing automotive software for Infotain-
ment and Telematics systems offers little flexibility to accommodate such hetero-
geneity and variation. Currently, application developers have to decide at design
time what possible uses their applications will have and the applications do not
change or adapt once they are deployed on an infotainment platform. In fact,
In-vehicle infotainment applications are currently developed with monolithic archi-
tectures, which are more suitable for a fixed execution context.
This chapter presents a framework that enables development of context aware
interaction interfaces for in-vehicle infotainment applications. Essential framework
concepts are presented with development and discussion of a lightweight component
model. Finally, the design of an example application is presented and evaluated.
H. Sharma (B)
Software Engineer at Delphi Delco Electronics Europe GmbH, Bad Salzdetfurth, Germany,
E-mail: hemant.sharma@delphi.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 343
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
344 H. Sharma, A.K. Ramani

29.1 In-Vehicle Infotainment

With increased commuting times, vehicles are becoming a temporary mobile work-
place and home place for both drivers and passengers, implying significant addi-
tional content and tools related to depiction and interaction with information about
one’s own identity, entertainment, relationship-building/maintenance (e.g., com-
munication with others), self-enhancement (including education and health mon-
itoring), access to information resources (e.g., travel information, references, and
other databases), and commerce. Vehicles as communication, entertainment, and
information environments may involve distributed displays that depict virtual pres-
ence in other spaces, games played with other vehicles’ drivers and passengers,
and status or instruction concerning driving/parking maneuvers, repairs, parts re-
ordering, etc.
In-vehicle infotainment systems offer a new generation of ultra-compact embed-
ded computing and communication platforms, providing occupants of a vehicle with
the same degree of connectivity and the same access to digital media and data that
they currently enjoy in their home or office. These systems are also designed with
an all-important difference in mind: they are designed to provide an integrated,
upgradeable control point that serves not only the driver, but the individual needs
of all vehicle occupants.
Infotainment system manufacturers seeking to serve worldwide markets face a
complex matrix of 15 major OEMs (car companies) with 78 major marks (brands)
and perhaps as many as 600 vehicle models. Models may each have a number of trim
levels; serve multiple languages and regulatory regimes (countries). To compete, the
suppliers will also need to deliver new functionality; features and a fresh new look
on an annual basis. Building Human Machine Interface (HMI) for these complex
applications is difficult and code-intensive, consuming disproportionate amount of
resources and adding considerably to project risk.

29.1.1 Infotainment System Interaction

An In-vehicle Infotainment device is usually connected to the vehicle network as

well as to GPS network. Further it not only, may have WiFi or cellular connectiv-
ity, but also interface to various ad hoc networks using the infrared or Bluetooth
interfaces. The potential for interaction with its environment is great. However, the
system only provides limited HMI primitives for this. The result is that such devices
are still seen as stand-alone and independent system, which interacts mainly to offer
static services -interaction with their environment and peers is either not consid-
ered or is very limited. Thus, although physically mobile, they are logically static
systems.
This HMI interaction model in current infotainment systems has various disad-
vantages: There is little code sharing between applications running on the same
device. There is no framework providing higher level interoperability and commu-
nication primitives for HMI service applications running on different devices. HMI
29 Context Aware In-Vehicle Infotainment Interfaces 345

Applications are monolithic, composed of a single static interaction interface, which

makes it impossible to update part of HMI structure. The procedure needed to host
third party dynamic service applications is difficult.

29.2 Context Aware Interaction

Unlike fixed distributed systems, infotainment systems execute in an extremely

dynamic context. By context, it is meant everything that can influence the behavior
of an infotainment application, from resources within the physical device, such
as memory, power cycle, screen size, and processing power, to resources outside
the physical device, such as location, remote service available, network bandwidth
and latency. Infotainment applications need to be aware of changes occurring in
their execution context, and adapt to these changes, to continue to deliver a high
quality-of-service to their users.

29.2.1 Context Aware Design

In order to ease the development of context-aware applications, middleware layered

between the underlying software system and the HMI applications have to provide
application developers with powerful abstractions and mechanisms that relieve them
from dealing with low-level details. For example, applications must be able to spec-
ify, in a uniform way, which resources they are interested in and which behaviors
to adopt in particular contexts. The middleware, then, maintains, on behalf of the
applications, updated context information, detects changes of interest to them and
reacts accordingly.

29.2.1.1 Context Awareness by Reflection

It is argued that reflection is a powerful means to build mobile computing frame-

works that supports the development of context-aware applications. The key to
the approach is to make some aspects of the internal representation of the frame-
work explicit and, hence, accessible from the application, through a process called
reification. Applications are then allowed to dynamically inspect middleware behav-
ior (introspection) and also to dynamically change it (adaptation) by means of a
meta interface that enables runtime modification of the internal representation pre-
viously made explicit. The process where some aspects of the system are altered or
overridden is called absorption.
By definition, reflection allows a program to access, reason about and alter its
own interpretation. The key to the approach is to offer a meta-interface supporting
the inspection and adaptation of the underlying virtual machine (the meta-level).
346 H. Sharma, A.K. Ramani

29.2.1.2 Context Awareness by Logical Mobility

As a highly dynamic system, an infotainment system similar to mobile system

encounters, by definition, changes to its requirements to which it needs to adapt.
The acceptance of logical mobility (LM) techniques, or the ability to ship part
of an application or even a complete process from one host to another, has recently
been witnessed in automotive domain with upcoming infotainment applications. It
has been shown that providing the flexible use of LM primitives to mobile com-
puting applications through a middleware system, allow for a greater degree of
application dynamicity, and will provide new ways to tackle interoperability and
heterogeneity as well as ease deployment.
Infotainment system can associate the context information to the applications
and adopt a suitable policy, reflection or logical mobility or both, to achieve
context awareness. The infotainment applications need suitable mechanisms for
management of context information and application of context awareness policy.

29.2.2 Context Management

The evolving nature of automotive services makes complete formalization of all

context knowledge a challenging mission. Nevertheless, it is observed that certain
contextual entities (e.g., location, user, vehicle and computational entity, etc.) are
most fundamental for capturing the general condition about the context, as they not
only form the skeleton of overall situation, but also act as indices into associated
information.
There are several types of service components that are involved in the sensing,
interpreting, managing and utilizing context information in an intelligent automotive
environment. All these components are packaged in form of pluggable services such
that they can be remotely deployed and managed.

29.3 Interaction Architecture

29.3.1 Layered Model

The framework is designed according to layered architecture shown in the figure

below (Fig. 29.1). The architecture consists of components organized into four
layers.
The Application layer defines a standard framework to create and run context-
aware HMI applications on top of the framework’s application model.
Application management layer contains components that define context and QoS
policies for the HMI application. The layer is divided into Context Management and
29 Context Aware In-Vehicle Infotainment Interfaces 347

Fig. 29.1 Layered architecture of HMI framework

QoS Management components. The context management components are responsi-

ble for specification and management of HMI specific contexts.
The purpose of core layer of the framework, in general, is to provide higher
level interaction primitives than those provided by the vehicle infotainment network
and infotainment service system as a layer, upon which HMI applications are then
constructed.
The components of service layer offers application service patterns, design-
to-implementation mappings, platform profiles, and extensibility mechanisms for
interoperating with other infotainment applications.

29.3.2 Component Model

The term component model refers to a description of components and a compo-

nent infrastructure that abstracts from the details of a concrete implementation,
such as the exact format of the component executable. The goal in doing so is to
understand the choices that can be made in the design of component architecture,
without getting distracted by machine-specific or platform specific details. This is
especially important in infotainment systems, as it is believed that the diversity of
process and architectures will mean that many different implementations of any
given component model will be needed.
348 H. Sharma, A.K. Ramani

Fig. 29.2 Component meta model overview

29.3.2.1 Component Meta-Model Overview

The component metamodel, as shown in Fig. 29.2, is a Meta Object Facility (MOF)
compliant extension of the UML metamodel.
It builds upon and extends the UML concepts of Classifier, Node, Class, Inter-
face, Data Type, and Instance. The most novel aspect of the component model is the
way in which it offers distribution services to local components, allowing instances
to dynamically send and receive components at runtime.
The component metamodel is a local, or in process, reflective component meta-
model for HMI applications hosted on infotainment platforms. The model uses
logical mobility primitives to provide distribution services and offers the flexible
use of those primitives to the applications; instead of relying on the invocation of
remote infotainment services via the vehicle network. The HMI framework com-
ponents are collocated on the same address space. The model supports the remote
cloning of components between hosts, providing for system autonomy when appli-
cation service connectivity is missing or is unreliable. As such, an instance of HMI
framework, as part of HMI application, is represented as a collection of local com-
ponents, interconnected using local references and well-defined interfaces, deployed
on a single host. The model also offers support for structural reflection so that appli-
cations can introspect which components are available locally, choose components
to perform a particular task, and dynamically change the system configuration by
adding or removing components.
29 Context Aware In-Vehicle Infotainment Interfaces 349

29.3.2.2 Component Model Elements

This section describes the primitive elements of the component model.

Components

The framework components encapsulate particular functionality, such as, for

instance, a user interface, a service advertisement protocol, a service, a graphics
framework, or a widget library. The components separate interfaces and implemen-
tations. A component is implemented by one or several HMI framework classes. It
can implement one or more interfaces, called facades, with each facade offering any
number of operations. A metamodel for components that are going to be deployed
across autonomous domain boundaries needs to ensure that interfaces that have once
been defined cannot be changed.
Each framework component implements at least one facade, the Component
façade. The purpose of this facade is to allow an application to reason about the
component and its attributes. This permits access to the properties of the compo-
nent by retrieving, adding, removing, and modifying attributes. The component
facade also contains a constructor, which is used to initialize the component, and
a destructor, which is used when removing the component from the system.

Containers

The central component of every HMI application is the container component. A con-
tainer is a component specialization that acts as a registry of components installed
on the system. As such, a reference to each component is available via the container.
The container component implements a specialization of the component facade that
exports functionality for searching components that match a given set of attributes.
An adaptive system must also be able to react to changes in component availabil-
ity. For example, a media player interface for iPOD must be able to reason about
which streams it can decode. Hence, the container permits the registration of listen-
ers (represented by components that implement the ComponentListener facade) to
be notified when components matching a set of attributes given by the listener are
added or removed.
To allow for dynamic adaptation, the container can dynamically add or drop
components to and from the system. Registration and removal of components is del-
egated to one or more registrars. A registrar is a component that implements a facade
that defines primitives for loading and removing components, validating dependen-
cies, executing component constructors, and adding components to the registry.
350 H. Sharma, A.K. Ramani

29.3.2.3 Component Life Cycle

The HMI framework supports a very simple and lightweight component life cycle.
When a component is passed on to the container for registration by loading it from
persistent storage, using a Deployer, etc., the container delegates registration to a
registrar component. The registrar is responsible for checking that the dependencies
of the component are satisfied, instantiating the component using its constructor,
and adding it to the registry. Note that the component facade prescribes a single con-
structor. An instantiated component can use the container facade to get references
to any other components that it may require. A component deployed and instanti-
ated in the container may be either enabled or disabled. The semantics of those and
the initial state of the component depend on the component implementation. The
functionality needed to manipulate the state of the component is exported by the
component facade.

29.3.3 Framework Core

The aim of HMI Framework core in general is to provide higher level inter-
action primitives than those provided by the vehicle infotainment network and
infotainment service system as a layer upon which HMI applications are then
constructed. In doing so, the framework hides the complexities of addressing
distribution, heterogeneity, and failures (Fig. 29.3).
The framework core along with underlying middleware provides a number of
services to applications. The services themselves are seen as regular components
built on top of the container. As such, they can be dynamically added and removed.

Fig. 29.3 Overview of framework core

29 Context Aware In-Vehicle Infotainment Interfaces 351

29.3.4 Distribution and Logical Mobility

An HMI application built using the framework can reconfigure itself by using logical
mobility primitives. As different paradigms can be applied to different scenarios,
our metamodel does not build distribution into the components themselves, but it
provides it as a service; implementations of the framework metamodel can, in fact,
dynamically send and receive components and employ any of the above logical
mobility paradigms.
The framework considers four aspects of Logical Mobility: Components, Classes,
Instances, and Data Types; the last is defined as a bit stream that is not directly
executable by the underlying architecture. One such, the Logical Mobility Entity
(LME), is defined as an abstract generalization of a Class, Instance, or Data Type.
In the framework component metamodel, an LMU is always deployed in a
Reflective component. A Reflective component is a component specialization that
can be adapted at runtime by receiving LMUs from the framework migration ser-
vices. By definition, the container is always a reflective component, as it can receive
and host new components at runtime.

29.3.5 Application Model

The Application Model (Fig. 29.4) of HMI framework provides support for building
and running context-aware infotainment HMI applications on top of the framework
context model.
HMI applications access the framework functionality through an Application-
Interface: each time an application is started, an ApplicationSkeleton is created to
allow interactions between the framework and the application itself. In particular,
application interfaces allow applications to request services from the underlying
framework, and to access their own application context profile through a well-
defined reflective meta-interface. The HMI application is realized as composition
of components based on the component model of the HMIFramework.
The application model shall guide modeling of infotainment HMI application
components with the use of interfaces of component of the framework in an efficient
manner. The application model is also expected to ease the integration of context
awareness to applications by proper use of the framework interfaces.

29.4 Application Development

This section represents an implementation of a simple Traffic Message Notification

application using the HMI framework core. Components that implement Traffic
Message decoding and text presentation inherit message format from the radio
tuner running on the same platform. As such, the Traffic HMI application uses the
352 H. Sharma, A.K. Ramani

Fig. 29.4 Application model for the framework

notification service to be notified whenever the tuner façade component that has
Traffic Message attribute implemented is registered. Moreover, it uses the deployer
and the discovery components to premium Traffic message service that are found
remotely.

29.4.1 Application Model

Figure 29.5 presents the platform independent application level model for TMC
HMI application.
The Traffic Message HMI demonstrates an application that uses the container
to listen to the arrival of new components, adapting its interface and functionality
upon new TMC service component arrival. It also demonstrates reaction to con-
text changes, as the application monitors the discovery services for new service
components and schedules them for use as soon as they appear.

29.4.2 Evaluation

The TMC HMI application demonstrates an application that uses the container to
listen to the arrival of new components and then adapts its interface and function-
ality to reflect the arrival of a new component. It also demonstrates reaction to
context changes as the application monitors the TMC services for new messages
29 Context Aware In-Vehicle Infotainment Interfaces 353

Fig. 29.5 Traffic message application overview

and schedules them for navigation reroute as soon as they appear. The operation is
transparent to the end user.
Finally, it illustrates the ease with which existing code can be adapted to run
under the framework environment. The adapted code gains the ability to be deployed
dynamically at runtime and to be used as part of a component-based application,
such as the TMC HMI application. Hence, it can express dependencies on other
components or platforms, it can be required by other components, and a component-
based application can be built with it.

29.5 Summary

This chapter presented an approach for building context aware HMIs for In-vehicle
infotainment software applications by means of components. The HMI framework
enable this by using a logical mobility based component model. The lightweight
component metamodel is instantiated as a framework system for adaptable HMI
application and systems. The framework offers logical mobility primitives as first-
class citizens to achieve context awareness.
354 H. Sharma, A.K. Ramani

References

1. K. Henricksen, J. Indulska, and A. Rakotonirainy, “Modeling Context Information in Per-

vasive Computing Systems”, 1st International Conference on Pervasive Computing (Zurich,
Switzerland), Springer, pp. 167–180, Aug. 26–28, 2002.
2. A. L. Murphy, G. P. Picco, and G.-C. Roman, “Lime: A Middleware for Physical and Logi-
cal Mobility,” Proceedings of 21st International Conference Distributed Computing Systems
(ICDCS 21), pp. 368–377, May 2001.
3. F. Bachman, L. Bass, S. Buhman, S. Comella-Dorda, F. Long, R. C. Seacord, and K.
C. Wallnau, Technical Concepts of Component-Based Software Engineering. Tech. Rep.
CMU/SEI-2000-TR-008.
4. H. Sharma, Dr. R. Kuvedu-Libla, and Dr. A. K. Ramani, “Component Oriented Human
Machine Interface for In-Vehicle Infotainment Applications”, WEC 2008 at London, pp.
192–197, 2–4 July 2008.
Chapter 30
Facial Expression Analysis Using PCA

V. Praseeda Lekshmi, M. Sasikumar, Divya S. Vidyadharan, and S. Naveen

Abstract Face recognition has drawn considerable interest and attention from many
researchers. Generally pattern recognition problem rely upon the features inherent
in the pattern for efficient solution. Conversations are usually dominated by facial
expressions. A baby can communicate with its mother through the expressions on
its face. But there are several problems in analyzing communication between human
beings through non-verbal communication such as facial expressions by a computer.
In this chapter, video frames are extracted from image sequences. Using skin color
detection techniques, face regions are identified. PCA is used to recognize faces.
The feature points are located and their coordinates are extracted. Gabor wavelets
are applied to these coordinates and to the images as a whole.

Keywords Facial Expression  PCA  Face recognition  video frame  skin color
detection  Gabor wavelet

30.1 Introduction

Face recognition has drawn considerable interest and attention from many resear-
chers for the last 2 decades because of its potential applications, such as in the areas
of surveillance, secure trading terminals, control and user authentication. The prob-
lem of face recognition is either to identify facial image from a picture or facial
region from an image and recognize a person from set of images. A robust face
recognition system has the capability of distinguishing among hundreds, may be
even thousands of faces. For automated access control, most common and accepted
method is face recognition. A number of face recognition methods have been
proposed.

V. P. Lekshmi (B)
College of Engineering, Kidangoor, Kottayam, Kerala, India
E-mail: vplekshmi@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 355
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
356 P.V. Lekshmi et al.

Generally pattern recognition problem rely upon the features inherent in the
pattern for efficient solution. The challenges associated with face detection and
recognition problem are pose, occlusion, expression, varying lighting conditions,
etc. Facial expression analysis has wide range of applications in areas such as
Human Computer Interaction, Image retrieval, Psychological area, Image under-
standing, Face animation etc. Humans interact with each other both verbally and
non-verbally.
Conversations are usually dominated by facial expressions. A baby can com-
municate with its mother through the expressions on its face. But there are several
problems in analyzing communication between human beings through non-verbal
communication such as facial expressions by a computer because expressions are
not always universal. It varies with ethnicity. Further facial expressions can be
ambiguous. They have several possible interpretations. To analyze the facial expres-
sion, face regions have to be detected first. Next step is to extract and represent the
facial changes caused by facial expressions. In facial feature extraction for expres-
sion analysis, there are two types of approaches, namely Geometric based methods
and Appearance based methods. In Geometric based method, the shape and loca-
tion of facial features are extracted as feature vectors. In Appearance based method,
image filters are applied either to whole face or to specific regions of facial image
to extract facial features.
In this paper, video frames are extracted from image sequences. Using skin color
detection techniques, face regions are identified. PCA is used to recognize faces.
The feature points are located and their coordinates are extracted. Gabor wavelets
are applied to these coordinates and to the images as a whole.
Rest of the paper is organized as follows. Section 30.2 gives background and
related works. Section 30.3 discusses the proposed method. Results are given in
Section 30.4. Conclusion and future work are given in Section 30.5.

30.2 Background and Related Work

Most face recognition methods fall into two categories: Feature based and Holistic
[1]. In feature-based method, face recognition relies on localization and detection
of facial features such as eyes, nose, mouth and their geometrical relationships [2].
In holistic approach, entire facial image is encoded into a point on high dimensional
space. Images are represented as Eigen images. A method based on Eigen faces
is given in [3]. PCA [4] and Active Appearance Model (AAM) for recognizing
faces are based on holistic approaches. In another approach, fast and accurate face
detection is performed by skin color learning by neural network and segmentation
technique [5]. Facial asymmetry information can be used for face recognition [7].
In [8], ICA was performed on face images under two different conditions: In one
condition, image is treated as a random variable and pixels are treated as outcomes
and in the second condition which treated pixels as random variables and image as
outcome. Facial expressions are extracted from the detailed analysis of eye region
30 Facial Expression Analysis Using PCA 357

images is given in [9]. In the appearance based approaches given in [10], facial
images are recognized by warping of face images. Warping is obtained by automatic
AAM processing.
Another method of classification of facial expression is explained in [11] in
which the geometry and texture information are projected in to separate PCA spaces
and then recombined in a classifier which is capable of recognizing faces with differ-
ent expressions. Kernel Eigen Space method based on class features for expression
analysis is explained in [12]. In [13] facial expressions are classified using LDA, in
which the gabor features extracted using Gabor filter banks are compressed by two
stage PCA method.

30.3 Method

Our proposed method consists of training stage, face recognition and expression
classification stages.

30.3.1 Training stage

The face image sequences with certain degree of orientation, wearing glasses and
large variations in the facial expressions are considered. As a first step, frames
with peak expression called key frames are identified. Face regions from these key
frames are extracted using skin color detection method. Skin regions are identi-
fied and connected component labeling is done to classify the sub regions in the
image.
Faces are detected from these detected skin regions. Figure 30.1 shows detected
face regions from skin areas. Frontal looking faces with neutral expressions called
normal faces and faces with set of non-neutral expressions form the database. There
were ‘K’ frames with ‘N’ expressions for each face so that ‘K x N’ face images
were used as the database. Normalization is done to make the frames with uniform
scale. The normalized faces are shown in Fig. 30.2.
Facial expressions are highly reflected in eyes and mouth regions. Fourteen mark-
ers as mentioned in [11] are used to automatically select for registering important
facial features. A triangulation method is applied to fit the mask on the faces. The
marking points represented as white dots are shown in Fig. 30.3. The coordinates
are used for further verification.
358 P.V. Lekshmi et al.

Fig. 30.1 (a) Skin region identified, (b) connected component labeling, (c) detected face regions

Fig. 30.2 Normalized faces

Fig. 30.3 Marking points

30.3.2 Face Recognition and Expression Classification

Face recognition has long been a primary goal of computer vision, and it has turned
out to be a challenging task. The primary difficulty in attempting to recognize faces
from image gallery comes from immense variability of face appearance due to sev-
eral factors including illumination, facial expressions, and view points of camera,
poses, and occlusions. This method treated face recognition as a 2-D recognition
problem.
30 Facial Expression Analysis Using PCA 359

PCA is a useful statistical technique that has found applications in the fields such
as face recognition; image compression, etc. It is a common technique for finding
patterns in the data of high dimensions. In PCA, face images are projected into
feature space or face space. Weight vector comparison was done to get the best
match.
Let the training set of face images be X1 , X2 , X3 ,. . . ,Xn, then the average set or
mean of faces be defined as
Pn
Xi
i D1
mD (30.1)
n
Notice that the symbol m to indicate the mean of set X . The average distance of
each face from the mean of the data set is given by

Q1 D X1  mI Q2 D X2  m: : :Qn D Xn  m (30.2)

which is the standard deviation.

The covariance matrix is given by

C D A  A0 (30.3)

where A D ŒQ1 Q2 Q3 : : :: : :: : :: : :::Qn:

In order to reduce the dimensionality, co-variance can be calculated as

C D A0 A:

Eigen values and Eigen vectors are calculated for the covariance matrix. All the face
images in the database are projected in to Eigen space and weight for each image is
calculated.
Then image vectors for each face image is obtained as
X10
Image Vector D weight.i / Eigevector.i / (30.4)
i D1

Usually while using PCA, normal images are used as reference faces. To overcome
the large variations in transformations, mean image is used as the reference face.
Ten face image sequences are used here, each with five facial expressions as the
database. Mean faces of five key frames with peak expressions and Eigen faces are
shown in Figs. 30.4 and 30.5.
The probe image is also subjected to preprocessing steps before projecting it
into feature space. The weight vector is calculated to identify the image from the
database with closest weighting vector.
So far a reference face for each testing face is identified. After recognizing the
face, the coordinates of the facial features are extracted as explained in Section 30.1.
The coordinates of each testing face is compared with its reference face by calcu-
lating the mean square error between the testing face and all the prototypes of same
individual. This mean square error tells us how far the expression on the testing face
is from each type of expressions and thus can be used to classify expressions.
360 P.V. Lekshmi et al.

Fig. 30.4 Mean faces

Fig. 30.5 Eigen faces

As a second method, Gabor wavelet transformation is applied to the extracted

coordinates and to the image itself. The main characteristics of wavelet are the pos-
sibility to provide a multi-resolution analysis of the image in the form of coefficient
matrices. Strong arguments for the use of multi-resolution decomposition can be
found in psycho visual research, which offers evidences that the human visual sys-
tem processes the image in a multi scale way. Wavelets provide a time and frequency
decomposition at the same time. Computational complexity of wavelets is linear
with the number of computed coefficients. Thus the complexity is less compared to
other fast transformations.
The Gabor wavelet representation of images allows description of spatial fre-
quency structure in the image while preserving information about spatial relations.
The response image of Gabor filter can be written as a correlation of input image
I.x/ with Gabor kernel Pk .x/:
Z Z
ak.x0/ D I.x/P k.x  x0/dx (30.5)
    
k2 k2 2
pk .x/ D 2 exp  2 x 2 exp .i kx/  exp  (30.6)
 2 2

where k is the characteristic vector.

  vC2
k D kx ; ky D .kv cos w; kv sin w / and kv D 2 2 I w D w
8
Here v is the discrete set of different frequencies and w is the orientation.
The Gabor coefficients are projected to feature space or face space. The feature
vectors of each testing face is compared with its reference face by calculating the
mean square error between the testing face and all the prototypes of same individual.
30 Facial Expression Analysis Using PCA 361

30.4 Results

Ten face image sequences from FG-NET consortium with variations in lighting con-
ditions, small orientations, wearing glasses, heavy expressions, etc. selected. The
expressions used were ‘happy’, ‘sad’, ‘normal’, ‘surprise’, and ‘anger’. Frames with
peak expressions from color face image sequences were extracted. Face regions
were identified using skin color. Eigen faces were used for recognizing faces. These
faces were converted to gray scale and normalized to a size 192X192. Fourteen
points were marked at the highly expression reflected face regions. The face images
in the database and the test image were projected to face space for face recognition
and their coordinates were verified to identify which expression belongs to the test
face. The performance ratios are 100% for expression recognition from extracted
faces, 88% for expression recognition from frames and 88% for the combined
recognition. As a comparison two experiments were conducted on theses face
images. Gabor wavelet transformation is applied to all these coordinates and the
resultant Gabor coefficients were projected to feature space or PCA space.
As the second part, the whole face images were subjected to Gabor wavelet
transformation. Figure 30.6 shows Gabor faces. The high dimensional Gabor coef-
ficients were projected to PCA space to reduce he dimension. Mean square error

Fig. 30.6 Gabor faces

362 P.V. Lekshmi et al.

Table 30.1 Expression recognition from extracted face

Expressions I1 I2 I3 I4 I5 I6 I7 I8 I9 I10

Anger Y Y Y Y Y Y Y Y Y Y
Happy Y Y Y Y Y Y Y Y Y Y
Neutral Y Y Y Y Y Y Y Y Y Y
Sad Y Y Y Y Y Y Y Y Y Y
Surprise Y Y Y Y Y Y Y Y Y Y

Table 30.2 Expression recognition from frames

Expressions I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
Anger N Y Y Y Y Y Y Y Y Y
Happy Y N Y Y N Y Y Y Y Y
Neutral Y Y Y Y Y Y Y Y Y Y
Sad Y N N Y Y Y Y Y Y Y
Surprise N Y Y Y Y Y Y Y Y Y

Table 30.3 Combined performance

Expressions I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
Anger N Y Y Y Y Y Y Y Y Y
Happy Y N Y Y N Y Y Y Y Y
Neutral Y Y Y Y Y Y Y Y Y Y
Sad Y N N Y Y Y Y Y Y Y
Surprise N Y Y Y Y Y Y Y Y Y

Table 30.4 Expression recognition by applying Gabor wavelet on fudicial points

Expressions I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
Anger Y Y N Y Y Y Y Y Y Y
Happy N Y Y Y Y Y Y N Y Y
Neutral Y N Y Y N Y Y Y Y Y
Sad Y Y N Y Y Y Y Y Y Y
Surprise Y Y Y Y Y Y Y N Y Y

is calculated between feature vectors of test image and all the reference faces. The
performance ratio for the application of Gabor wavelets on extracted features is 86%
and to the whole image is 84%.
A comparison was made between the approaches for different databases are given
in the Tables 30.1, 30.2 and 30.3. Table 30.4 and 30.5 shows the expression recog-
nition performance by applying Gabor wavelets to the extracted fudicial points and
to the whole image.
30 Facial Expression Analysis Using PCA 363

Table 30.5 Expression recognition by applying Gabor wavelet to the whole image
Expressions I1 I2 I3 I4 I5 I6 I7 I8 I9 I10

Anger Y Y Y N Y Y N Y Y Y
Happy N Y Y Y Y Y N Y Y Y
Neutral Y Y Y Y Y Y Y N Y Y
Sad Y Y Y N Y Y Y Y Y Y
Surprise Y Y Y Y Y N Y Y Y N

30.5 Conclusion

In this paper, face recognition and expression classification from video image
sequences are explained. Frames are extracted from image sequences. Skin color
detection method is applied to detect face regions. A holistic based approach in
which whole face is considered for the construction of Eigen space. As a first step,
images are projected to PCA space for recognizing face regions. After recognizing
the face, our system could efficiently identify the expression from the face.
To compare the performance, two experiments are also conducted. Gabor wavelet
transformation is applied to the extracted fudicial points and to the images as a
whole. The Gabor coefficients are projected to feature space or face space for
comparison with that of test image.
This logic performs well for recognition of expressions from face sequences. The
computational time and complexity was also very small.

Acknowledgements We thank FG-NET consortium [13] for providing us the database for the
experiment.

References

1. Rama Chellappa, Charles L. Wilson and Saad Sirohey, “Human and Machine Recognition of
Faces: A Survey”, In Proceedings of IEEE, Vol. 83(5), 705–740, 1995
2. Stefano Arca, Paola Campadelli and Raffaella Anzarotti, “An Automatic Feature Based Face
Recognition System”, In Proceedings of the 5th International Workshop on Image Analysis
for Multimedia Interactive Services (WIAMIS 2004)
3. Matthew Turk and Alex Pentland, “Face Recognition Using Eigen Faces”, In Proceedings of
IEEE Conference on Computer Vision and Pattern Recognition, pp. 586–591, 1991
4. Sami Romdhani, Shaogang Gong and Alexandra Psarrou, A Multi-view Nonlinear Active
Shape Model Using Kernel PCA, BMVC99
5. Hichem Sahbi, Nozha, Boueimma, Tistarelli, J. Bigun and A.K. Jain (Eds), “Biometric
Authentication” LNCS2539, Springer, Berlin/Heidelberg, 2002, pp. 112–1206
6. Sinjini Mitra, Nicola A. Lazar and Yanxi Liu, “Understanding the Role of Facial Asymme-
try in Human Face Identification”, Statistics and Computing, Vol. 17, pp. 57–70, 2007. DOI
10.1007/s 1222–006–9004–9
364 P.V. Lekshmi et al.

7. Marian Stewrt Bartlett, R. Javier, Movellan and Terrence J. Seinowski, “Face Recognition by
Independent Component Analysis”, IEEE Transactions on Neural Networks, Vol. 113, No. 6,
Nov. 2002
8. Tsuyoshi Moriyama, Takeo Kanade, Jing Xiao and Jeffrey F. Cohn, “Meticulously Detailed
Eye Region Model and It’s Application to Analysis of Facial Images”, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 28, No. 5, May 2006
9. Hui Li, Hui Lin and Guang Yang, “A New Facial Expression Analysis System Based on Warp
Images”, Proceedings of Sixth World Congress on Intelligent Control and Automation, Dalian,
China, 2006
10. Xiaoxing Li, Greg Mori and Hao Zhang, “Expression Invariant Face Recognition with Expres-
sion Classification”, In Proceedings of Canadian Conference on Computer and Robot Vision
(CRV), pp. 77–83, 2006
11. Y. Kosaka and K. Kotani, “Facial Expression Analysis by Kernal Eigen Space Method Based
on Class Features (KEMC) Using Non-linear Basis for Separation of Expression Classes.”
International Conference on Image Processing (ICIP), 2004
12. Hong-Bo Deng, Lian-Wen Jin, Li-Xin Zhen and Ian-Cheng Huang, “A New Facial Expression
Recognition Method Based on Local Gabor Filter Bank and PCA Plus LDA”, International
Journal of Information Technology, Vol. 11, No. 11, 2005
13. Frank Wallhoff, Facial Expressions and Emotion Database http://www.mmk.ei.tum.de/
waf/fgnet/feedtum.html,Technische Universität München 2006
Chapter 31
The Design of a USART IP Core

A. H. El-Mousa, N. Anssari, A. Al-Suyyagh, and H. Al-Zubi

Abstract Embedded systems have drastically grown in importance and complexity

in recent years. Many systems are now designed using Field Programmable Gate
Array (FPGA) due to its size, flexibility, and resources. This chapter presents the
design and implementation of a flexible and user reconfigurable Universal Syn-
chronous Asynchronous Receive Transmit (USART) IP core suitable for use in
embedded systems and Systems on Chip (SoC). The design scheme employed,
allows the USART to be used in various modes of operation such as standalone and
9-bit addressable mode for multi-drop network of serial devices. It also supports
high speed data rates of up to 3 Mb/s. The design utilizes Hardware Description
Language (HDL) to describe the operation, ease implementation and allow cross
platform utilization. The chapter shows through a comprehensive testing method-
ology that the proposed design functions properly while consuming minimum
resources from the target FPGA.

Keywords Embedded systems  addressable USART  IP core  FPGA  multi-drop

networks  HDL

31.1 Introduction

Interest in embedded systems has grown drastically in recent years; the world mar-
ket for embedded software will grow from about $1.6 billion in 2004 to $3.5 billion
by 2009, at an average annual growth rate (AAGR) of 16%. On the other hand,
embedded hardware growth will be at the aggregate rate of 14.2% to reach $78.7
billion in 2009, while embedded board revenues will increase by an aggregate rate

A. H. El-Mousa (B)
University of Jordan Computer Engineering Department,
Faculty of Engineering & Technology, University of Jordan, Amman, Jordan
E-mail: elmousa@ju.edu.jo

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 365
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
366 A.H. El-Mousa et al.

of 10% [1]. At the same time, embedded systems are increasing in complexity and
more frequently they are also networked. As designs become more complex, embed-
ded systems based on FPGA are required to interact with software running on stock
commercial processors [2]. This interaction more often than not makes use of a
serial communications transmission link. Design techniques based on hardware-
software co-design are generally implemented on platforms that utilize FPGAs as
accelerators together with embedded CPU cores for control and operational proce-
dure definition [3]. Frequently these platforms also require high speed serial data
communication blocks.
USARTs have been around for years and they have become established for easy
and simple serial transmission. However, most of these take the form of hardwired
specialized ICs which make them unattractive and unsuitable for use in recent
embedded systems; especially those utilizing FPGA technology, since they can-
not be incorporated within the HDL design. Also, most are designed with limited
features and capabilities; for example: limited speed and no capability for use in
multi-drop networks. Attempts at the design of an HDL-based USART have been
reported in the literature. Many are just HDL implementation of the well known
industry standard 16550 UART without additional features [4–6].
This chapter presents the design, implementation and verification of a high speed
configurable USART suitable to be used on platforms that utilize FPGAs. The
architectural design allows serial communications in multi-drop networks using
9-bit operational mode using master-slave operation. It also features configurable
high speed transmission rates and transmission error detection and recovery. User
configuration is accomplished through specialized internal registers. The design is
suitable to be used for inter-chip, inter-processor, and inter-system communications
among others. The design implements both modes of operation synchronous and
asynchronous.
The rest of the chapter is organized as follows: Section 31.2 presents a general
description of USART theory of operation and describes the modes of operation.
Section 31.3 provides a detailed description of the specifications of the developed
USART. Section 31.4 discusses the design methodology followed and provides
a description and operational features of the various blocks used in the design.
Section 31.5 is concerned with the testing and verification procedures followed.

31.2 USART Theory of Operation

USARTs operate as parallel to serial converters at the transmitter’s side where data
is transmitted as individual bits in a sequential fashion whereas at the receiver’s
side, USARTs assemble the bits into complete data and thus act as serial to parallel
converters.
USARTs mandate preliminary configuration of data format, transfer rates and
other options specific to the design model. The user can send individual or a block
of data, whose size is determined by the design model, to the transmitter section of
31 The Design of a USART IP Core 367

the USART. Data is stored in the internal buffer and framed according to the transfer
mode and user defined options. Finally, the frame is sent to its destination one bit
at a time. On the receiver side, the data frame bits are received and sampled. The
extracted data from the received frame resides in the internal receiver buffer waiting
to be read. The receiver monitors the reception for possible errors and informs the
recipient of their existence should the proper control bit(s) be set. Most USARTs
offer a status monitoring mechanism via dedicated status register(s).

31.2.1 Modes of Operation

There are two major modes of operation in USARTs: synchronous and asyn-
chronous modes. The latter prevails in most applications.

31.2.1.1 Synchronous Mode

In synchronous mode, both the transmitter and receiver are synchronized to the same
clock signal which is usually generated by the transmitter and sent along with the
data on a separate link to the receiver. The receiver in turn utilizes the received
clock in extracting the timing sequence and determining the beginning and end of
the received bit interval. Therefore, the receiver knows when to read the bit’s value
and when the next bit in the sequence begins. However, when the line is idle (i.e. no
data is exchanged), the transmitter should send a fill character.
A synchronous communication session begins by sending one or more synchro-
nizing frames to indicate the beginning of transmission then the sequence of data
frames follow (a parity bit is appended to the end of the data frame if single error
detection is required), transmission is terminated by sending a stop frame after
which the line returns to the idle state.

31.2.1.2 Asynchronous Mode

An idle line remains at a predetermined level. The frame consists of a start bit
which differs in polarity to that of the line’s idle state, followed by the data bits and
a parity bit – if single error detection is used – and ends with at least one stop bit
which has the same polarity as that of an idle line. A stop bit might be followed by
another frame – back to back transmission – or an idle state of transmission. Both
the transmitter and receiver are preconfigured to run at the same fixed clock rate
which is an exact multiple of the required baud rate. Once the receiver recognizes
the transition from the idle state polarity to the opposing polarity, it waits a half
bit interval duration and verifies the presence of a start bit, if start bit arrival is
confirmed, the receiver reads the values of the bits every full bit-width interval until
the reception of a stop bit is confirmed denoting end of frame.
368 A.H. El-Mousa et al.

31.3 Specifications of the Developed USART

Performance oriented features include an interrupt driven approach and universal

eight bit addressability which make the USART ideal for industrial and control
applications. Eight-level buffering allows for data block transfers that is beneficial
considering the high speeds the USART can handle in data transfer which can reach
3 MHz.
Reliability oriented specifications include programmable odd/even parity bit with
the ability to detect parity, framing and overrun errors.
Table 31.1 lists the detailed specifications of the proposed USART.

31.4 USART System Design Methodology

The methodology adopted in the USART system design was based on systems and
software engineering approaches. It used a variation of both the waterfall and incre-
mental approaches suited to the design environment and constraints. The design
steps that the system went through are [7]:
1. System Requirements Definition. The requirements definition phase specified the
functionality as well as essential and desirable system properties. This involved
the process of understanding and defining services required from the system and
identifying constraints on its operation.
2. System/Subsystem Design. This phase was concerned with how system func-
tionality was to be provided by the components and how system specification
was converted into an executable system specification.
3. Subsystems Implementation and Testing. Here subsystems were implemented
and mapped into hardware code using HDL. In this stage, individual modules
were extensively tested for correct functional operation.
4. System Integration. During system integration, the independently developed
subsystems were merged together to build the overall USART system in an
incremental approach.
5. System Testing. The integrated system was subjected to an extensive set of tests
to assure correct functionality, reliability and performance.
Figure 31.1 shows the major parts of the USART system:
The input/output signals involved are:
Sys. clock: the oscillator clock Reset: master reset of the system
Serial out: transmitted data Serial in: received data
TXInt: transmitter interrupt RXInt: receiver interrupt
ErrInt: error interrupt
31 The Design of a USART IP Core 369

Table 31.1 Functional specifications of the USART

Specification Justification

Support for the following transmission modes Full duplex mode is employed in modern
(programmable): Asynchronous (full/half systems while half duplex mode is retained
duplex modes) Synchronous (full/half for compatibility with legacy systems
duplex modes)
Supports a wide range of High frequencies are essential for high speed
transmission/reception rates (from 50 Hz communication. Lower speeds are needed to
to 3 MHz) communicate with older USARTs. Moreover,
lower speeds can be used to minimize cross
talk if similar links are adjacent
Eight-level transmitter/receiver buffer To account for the high speeds of communication
that the USART can reach, blocks of data can
be received and buffered until read by the
user. Also, this allows the user to issue the
transmission of a block of eight-frame size in
a single operation. This will also reduce the
load on the module that controls the USART
operation in the system
Parity supported (programmable – Single error detection techniques might prove
enable/disable parity and odd/even parity) beneficial in noisy operation environments
Variable data lengths supported Byte communication is the modern norm. Five to
(programmable – 5–8 bits) seven bits data length is to retain
compatibility with legacy systems
Variable stop bits supported (asynchronous This is to comply with the RS232 standard where
mode) (programmable – 1 or 2 stop bits) two stop bits mode is used to accommodate
slightly different clocks in the transmitter and
receiver sides when USARTs from different
vendors are connected together
Error detection of the following errors: Parity Parity error detection provides a measure of the
error Overrun error Framing error reliability of communication. Framing error
(asynchronous mode) detection indicates the necessity of
reconfiguring the internal clocking sources at
both ends correctly. Finally, overrun error
informs that data has been lost and the need
to frequently read the received data
Supports interrupts (programmable – with Most modern systems are interrupt-driven for the
ability of global masking) reason that interrupt techniques save
processing clock cycles in comparison with
polling techniques and are essential in real
time applications
Supports addressability (8-bit universal – Used in industrial and applications in multi-drop
addresses up to 256 devices) while networks where a master can communicate
sending 9-bits with a certain other slave
370 A.H. El-Mousa et al.

Fig. 31.1 The USART module

Fig. 31.2 The transmitter module

31.4.1 Transmitter Module

Figure 31.2 shows the functional block diagram of the transmitter module.
It consists of the following sub modules:
– Transmitter buffer – Parity generation logic
– Bypass logic – Shift logic
– Transmit shift register (TSR) – TSR empty detection logic
In the following, due to space constraints, we only show the structure of the
transmitter buffer.

31.4.1.1 The Transmitter Buffer

The transmitting buffer is the memory where data to be sent are stored waiting for
the transmit shift register (TSR) to be empty. It becomes an essential component
when the inter arrival time of transmitted frames becomes very small. Moreover,
when the system is accessed using a processor/DSP that operates at a higher
31 The Design of a USART IP Core 371

Fig. 31.3 The input/output signals associated with the TX buffer

Fig. 31.4 Transmitting buffer dataflow diagram

frequency than the transmission baud clock, this buffer will reduce the number of
times the processor is interrupted to request for new data. The signal TDR empty is
generated to indicate that the buffer has at least one empty slot. Figure 31.3 shows
the input/output signals associated with the transmitting buffer while Fig. 31.4
shows the data flow diagram for it.
372 A.H. El-Mousa et al.

31.4.2 Receiver Module

In the asyn. mode, the receiver waits for a transition from a mark to space after an
idle line state or an expected stop bit to initiate the data reception logic provided that
the transition is not caused by a noise notch. This is ensured by sampling each of the
received bits at three different times and then using a majority detection circuit. The
end of asynchronous reception is detected by waiting for a stop bit at the end of the
frame.
In the synch. mode, instead of waiting for logic transition, the receiver waits for a
synchronizing character which if received after an idle state line, a start of reception
is detected. An end of reception is signaled if a certain stop character is received.
When data reception is detected at the serial in input, the internal receiver logic is
enabled and the received data bits are serially shifted into the Receiver Shift Register
(RSR). Meanwhile, the parity is calculated per each received bit for the received
word and finally compared to the received parity value. Parity and framing errors
are evaluated and stored along with the data in the receiver buffer. The buffer can
hold up to seven received words, if data are received while the buffer is full, the data
is dropped and an overrun error is indicated.
If 9-bit address detection mode is enabled, transmission is fixed at eight bits and
the ADD-Data bit is substituted for parity. The address of the receiving node must
be received with ADD-Data bit is set to “1” in order for the frame to be considered
an address frame.
In the synchronous addressable mode of operation, a synchronizing character
with ADD-Data bit value set to zero must be initially received, followed by a frame
containing the address of the receiving node but with ADD-Data bit value set to one,
followed by the data frames with ADD-Data bit reset again. Figure 31.5 shows the
functional block diagram of the receiver.

Fig. 31.5 Receiver module

31 The Design of a USART IP Core 373

The receiver module consists of the following sub-modules:

– Sampling and majority detection – Detect start bit logic
logic
– Rx shift register enable logic – Rx shift register (RSR)
– Rx shift register full detection logic – Parity error detection logic
– Framing error and stop bit detection
logic
– Rx buffer write with synch. and stop
character detect logic
– Rx buffer with overrun detection
logic
Due to space constraints, we only show the structure of the receiver buffer write
with synchronous and stop character detect logic sub-module as an example of the
methodology of design.

31.4.2.1 Rx Buffer Write with Synch and Stop Character Detect Logic

This sub-module is responsible for generating an internal signal to write the received
frames in RSR, together with their corresponding parity and framing information,
into RDR buffer only if certain conditions are met. In synchronous mode of opera-
tion, this involves checking for the reception of valid sync and stop characters that
delimit a block of consecutive frames. If 9th bit mode is used, all received frames
are dropped if the USART is not the targeted node, which is indicated by receiving a
valid address frame prior to receiving data. Figure 31.6 shows the data flow diagram
for this sub-module.

31.5 Testing and Verification Procedures

In general, the system testing process has two distinct goals:

1. To demonstrate to the developer and the customer that the system meets its
requirements
2. To discover faults or defects in the system where the behavior of the system is
incorrect, undesirable or does not conform to its specification [7]
The first goal leads to validation testing, where the system is expected to perform
correctly using a given set of test cases that reflect the system’s expected use. The
second goal leads to defect testing, where the test cases are designed to expose
defects. The test cases can be deliberately obscure and need not reflect how the
system is normally used. For validation testing, a successful test is one where the
system performs correctly. For defect testing, a successful test is one that exposes a
defect that causes the system to perform incorrectly.
374 A.H. El-Mousa et al.

Fig. 31.6 Receiver buffer write logic dataflow diagram

Fig. 31.7 Phases of the testing process

The proposed USART design was implemented and tested using firmware from
Xilinx Corporation. For software, the free ISE Webpack version 8.2i was used [8].
As for hardware, different development kits were used. These include: Digilab 2
XL (D2XL) Development Board [9], Digilent Spartan-3 System Board [10], and
Spartan-3E Starter Kit Board [11]. Since the design was entirely implemented using
a universal hardware description language, Verilog HDL, it is expected to be directly
interoperable with any environment provided by other vendors.
In general, the project went through several phases during the testing process as
illustrated in Fig. 31.7.
A unit test is a test of the functionality of a system module in isolation, and
the unit test should be traceable to the detailed design. A unit test consists of
a set of test cases used to establish that a subsystem performs a single unit of
31 The Design of a USART IP Core 375

functionality to some specification. Test cases should be written with the express
intent of uncovering undiscovered defects [12].
After the units of a system have been constructed and tested, they are integrated
into larger subcomponents leading eventually to the construction of the entire sys-
tem. The intermediate subsystems must be tested to make sure that components
operate correctly together. The purpose of integration testing is to identify errors in
the interactions of subsystems [12].
The primary goal of an acceptance test is to verify that a system meets its require-
ment specification. To demonstrate that the system meets its requirements, it must
be shown that it delivers the specified functionality, performance and dependabil-
ity, and that it does not fail during normal use. Ideally, the acceptance test plan is
developed with the engineering requirements and is traceable to them. Acceptance
testing is usually a black-box testing process where the tests are derived from the
system specification. The system is treated as a black box whose behavior can only
be determined by studying its inputs and the related outputs [7].
Before connecting the USART to an external system, it was tested at first by
connecting the serial output of its transmitter section to the serial input of its receiver
section to locate any potential errors within the system itself. Similarly, the clock
output of the transmitter section was connected to the clock input of the receiver
section in the synchronous mode.
The entire system was implemented on two Xilinx development boards. Both
were programmed with the same USART design. However, one was used to transmit
data, while the other was used to receive the sent data. Special temporary modifi-
cations to the internal design were implemented to allow certain internal signals to
be observed with a digital storage scope. Initially, data was exchanged with the PC
using the EPP mode of the parallel port. The PC was used to configure the USARTs
with the different communication options by sending the appropriate control words
to the respective registers and also to supply the data to be transmitted serially.
From this test, more indications were obtained that the system complies with its
requirements specification. Error conditions reflected the true state of the received
data when the two USARTs were deliberately configured with conflicting commu-
nications options. Moreover, the USARTs functioned as expected in the 9-bit mode
of operation.
Figure 31.8 illustrates a snapshot of a communication session that was estab-
lished between the two USARTs.
In performance testing, an effective way to discover defects is to design tests
around the limits of the system; that is, stressing the system (hence the name stress
testing) by making demands that are outside the design limits of the system until the
system fails.
During the test cases described previously, the USART design was subjected to
some extreme conditions to explore various aspects of the limits of its operation. For
example, the output of the baud rate generator was examined at the highest baud rate
possible, and the buffers were tested with increasing read and write speeds. More-
over, the operation of the entire USART was checked while operating at the highest
baud rate possible when two systems on separate boards were connected together,
376 A.H. El-Mousa et al.

Fig. 31.8 Asynchronous 9-bit address transmission and detection between two USARTs

as well as when the transmitter section was used to drive the receiver section of the
same system.

References

1. Ravi Krishnan, Future of Embedded Systems Technology, BCC Research Report, ID:
IFT016B, June 2005.
2. John A. Stankovic, Insup Lee, Aloysius Mok and Raj Rajkumar, Opportunities and Obliga-
tions for Physical Computing Systems, IEEE Computer Magazine, 11, pp. 23–31 (2005).
3. Wayne Wolf, High Performance Embedded Computing (Elsevier, Amsterdam, Netherlands
2007), pp. 383–387.
4. Mohd Yamani Idna Idris, Mashkuri Yaacob and Zaidi Razak, A VHDL Implentation of UART
Design with BIST Capabilty, Malaysian Journal of Computer Science, 19 (1), pp. 73–86
(2006).
5. Azita Mofidian, DesignWare Foundation DW 16550: A fine work of UART, Designware
Technical bulletin, Technical Forum for Design Automation Information, 4 3 (1999);
http://www.synopsys.com/news/pubs/dwtb/q499/dwtb art1.html
6. Shouqian Yu, Lili Yi, Weihai Chen and Zhaojin Wen, Implementation of a Multi-channel
UART Controller Based on FIFO Technique and FPGA, Proceedings of 2nd IEEE Conference
on Industrial Electronics and Applications ICIEA 2007, pp. 2633–2638 (2007).
7. Ian Sommerville, Software Engineering, 8th Edition (Addison-Wesley, England, 2007).
8. http://www.xilinx.com/ise/logic design prod/webpack.htm
9. Digilent D2XL System Board Reference Manual;Revision: (June 9, 2004); www.digilentinc.
com
10. Spartan-3 Starter Kit Board User Guide, version 1.1: (May 13, 2005); www.xilinx.com
11. Spartan-3E Starter Kit Board User Guide, version 1.0: (March 9, 2006); www.xilinx.com
12. Ralph M. Ford and Chris Coulston, Design for Electrical and Computer Engineers (McGraw-
Hill, New York, 2005).
Chapter 32
Multilayer Perceptron Training Optimization
for High Speed Impacts Classification

Angel Garcia-Crespo, Belen Ruiz-Mezcua, Israel Gonzalez-Carrasco,

and Jose Luis Lopez-Cuadrado

Abstract The construction of structures subjected to impact was traditionally car-

ried out empirically, relying on real impact tests. The need for design tools to
simulate this process triggered the development in recent years of a large number
of models of different types. Taking into account the difficulties of these methods,
poor precision and high computational cost, a neural network for the classification
of the result of impacts on steel armours was designed. Furthermore, the numerical
simulation method was used to obtain a set of input patterns to probe the capacity of
themodel development. In the problem tackled with, the available data for the net-
work designed include, the geometrical parameters of the solids involved - radius
and length of the projectile, thickness of the steel armour - and the impact velocity,
while the response of the system is the prediction about the plate perforation.

Keywords Impacts Classification  Multilayer Perceptron  Optimization  neural

network  simulation

32.1 Introduction

There are a wide number of systems which, during their service life, can suffer the
impact of objects moving at high speed (over 500 m/s). This phenomenon is named
impact ballistic and the most characteristic examples are found in the military field.
However over the last decades, this kind of problems has become of interest in civil

A. Garcia-Crespo, B. Ruiz-Mezcua, I. Gonzalez-Carrasco, and J. L. Lopez-Cuadrado

Department of Computer Science, Universidad Carlos III, Av. Universidad 30 - 28911,
Leganes (Madrid) Spain;
E-mails: acrespo@ia.uc3m.es,fbruiz,igcarras,jllopezg@inf.uc3m.es.
*
This research was done with the financial support of the Comunidad Autonoma de Madrid under
Project GR/MAT/0507/2004.

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 377
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
378 A. Garcia-Crespo et al.

applications. In them, the structural elements are required to absorb the projectile
energy so that it does not damage critical parts of the global system.
Due to this, there are new possibilities in similar fields, among which passive
safety in vehicles stands out. In this field, it is especially relevant the design of
structures whose mission is to absorb energy in crashes of Crashworthiness type
(200 m/s  speed  500 m/s), as well as those that can appear in road or railway
accidents, helicopters’ emergency landings, etc. Therefore, what is being sought is
to design structures capable of absorbing energy to avoid or lessen the damages to
the passengers of the concerned vehicles.
The construction of structures subjected to impact was traditionally carried out
empirically, relying on real impact tests, each one using a given configuration of
projectile and target. The mathematical complexity of solving the equations that rule
the impact phenomenon, and the relative ignorance of the mechanical behaviour of
the materials at high strain rates, discouraged any simulation of the problem.
The need for design tools to simulate this process triggered the development in
recent years of a large number of models of different types; all of them belong
to two families: analytical modelling and numerical simulation. Thus, the use of
expensive experimental tests has been postponed to the final stage of the design. All
the previous stages can be covered by the use of this kind of simulation tools.
Taking into account the difficulties of these methods, poor precision and high
computational cost, a neural network for the classification of the result of impacts
on steel armours was designed. Furthermore, the numerical simulation method was
used to obtain a set of input patterns to probe the capacity of the model development.
In the problem tackled with, the available data for the network designed include, the
geometrical parameters of the solids involved – radius and length of the projectile,
thickness of the steel armour – and the impact velocity, while the response of the
system is the prediction about the plate perforation.

32.2 Ballistic Impact

The ballistic terminology is defined as the scientific study of everything related to

the movement of the projectile. Within this science, the Terminal Ballistic or Effect
Ballistic discipline stands out. It is in charge of studying the results produced in the
body or object that suffers the impact of the bullet or projectile. Moreover it analyses
the way the projectile behaves when it reaches its target, how the projectile ends up,
how the transference of kinetic energy is carried out and what effects it has on the
objective, etc.
The conditions in which this projectile impacts on the objective depend most of
the times on different parameters, namely the materials, both of the projectile and
the target, the speed, the impact angle, the thickness of the target, the diameter of the
projectile and finally the shape and weight. The combination of these factors leads
to the existence of a vast variety of possible behaviours once the contact between
them is produced.
32 Multilayer Perceptron Training Optimization for High Speed Impacts Classification 379

The model developed was simplified to reduce the computational complexity

of the numeric simulations needed to recreate impact tests of this type. For this
purpose the obliquity of the projectile was not taking into account. Therefore, a
right angle between the axis of the projectile and the surface of the plate where it
impacts against, has been considered.
From a purely physique point of view, the impact of a projectile on a protection
can originate three different results: arrest, perforation or ricochet. Meanwhile the
process of the projectile’s entrance inside the protection during an impact is called
penetration. A perforation is produced when the projectile goes through the protec-
tion completely. On the other hand, an arrest takes place when the projectile goes
inside the protection but does not go through it. Finally if the angle of attack in
the impact comes into play, the projectile can rebound, what is known as ricochet
(Fig. 32.1).
The investigation carried out in this study analyses the behaviour of metallic
projectiles when the impact against metallic plates, both recreated with the same
SAE 1006 steel material. This material has been chosen because it is isotope, its
parameters are known and it allows carrying out further tests in the laboratory at a
low cost.

Fig. 32.1 Scheme of the projectile’s impact on steel armours

380 A. Garcia-Crespo et al.

32.3 Use of ANN in Impacts Situations

Artificial Neural Networks (ANNs) are statistical models of real world systems
which are built by tunning a set of parameters. These parameters, known as weights,
describe a model which forms a mapping from a set of given values, the inputs, to
an associated set of values, the outputs.
The mathematical model that the neural network builds is actually made up of
a set of simple functions linked together by the weights. The weights describe the
effect that each simple function (known as unit or neuron) will have on the overall
model [1].
Within the field of research of this article, the ANNs have been applied success-
fully within the limits of the fracture mechanic [2] to estimate material breakage
parameters such as concrete [3], and in no-destructive tests to detect breaks in more
fragile materials [4]. In all these applications, the input data needed for the training
has been obtained by means of experimentation and numeric simulation.
Nevertheless there are not enough studies dedicated to the application of ANN
to solve problems of ballistic impacts. At the present time the investigations made
have focused on low speed impacts or situations where it is necessary an energy
absorption (Crashworthiness) [5–7].
Therefore, and due to the mentioned experience on the mechanical problems and
on the good results shown on the previous researches [8], a MultiLayer Perceptron
(MLP) neural network with backpropagation algorithm has been defined. The most
important attribute of this kind of ANN is that it can learn a mapping of any
complexity or implement decision surfaces separating pattern classes.
An MLP has a set of inputs units whose function it is to take input values from the
outside, a set of outputs units which report the final answer, and a set of processing
hidden units which link the inputs to the outputs. The function of the hidden neurons
is to extract useful features from the input data which are, in turn, used to predict
the values of the output units.
MLPs are layered neural network, see Fig. 32.2, that means they are based on
several layers of adaptive weights and neurons. There are three different layers: the
input layer, at least one hidden layer and finally the output layer. Between the units
or neurons that compose the network there are a set of connections that link one to
each other. Each unit transmits signals to the ones connected with its output. Asso-
ciated with each unit there is a output signal or transference signal, that transform
current state of the unit into an output signal.
The feedforward connections between the units of the layers in a MLP represent
the adaptive weights that permit the mapping between the input variables and the
output variables [9].
Different authors have showed that MLP networks with as few as a single hidden
layer are universal approximators, in other words ANN are capable to approximate,
with accurate, arbitrary regions, if enough hidden units are available [10, 11].
32 Multilayer Perceptron Training Optimization for High Speed Impacts Classification 381

Fig. 32.2 Typical multilayer perceptron architecture

32.3.1 Impact Scenario Parameters

Within the limits this article deals with, there are different variables or parameters
that characterize the behaviour of the projectile when it impacts on an steel armour.
Therefore, there are various parameters to shape the input data to the network, and
these define the number of neurons of the input layer.
The available variables are: kinetic energy (K), velocity (V), radius (R), length
(L), mass (M) and quotient L/R, being all of them related to the projectile, and on
the other hand the thickness of the objective (H).
However, the use of all the available variables is not always necessary to carry
out the training. In some cases, an information overload or the existing connections
between the variables can saturate the prediction model that adjusts the output of the
network, therefore complicating the learning and reducing the generalization rates.
In this domain, K is correlated with V and M. On the other hand, the ratio L/R
and M are correlated with R and L because density is constant. So for the neural
network performance it is better to remove K and ratio L/R from the list of available
variables.
Moreover, in order to support the latter assertion, a series of studies were
designed to measure the influence that each variable has, separately and in con-
nection with the remainder, on the learning and the network generalization ability.
382 A. Garcia-Crespo et al.

32.3.2 Network Structure

For the network design, a three level MLP architecture was selected. According to
Lippman’s studies, this type of structure allows to solve most of the problems, being
able to form any complex random limit of decision [12].
The number of inputs in a MLP is determined by the number of available input
parameters in the problem dealing with. The number of outputs is the same as the
number of classes needed to solve the problem.
The number of hidden layers and the number of neurons of these layers have
to be chosen by the designer. There is not a method or rule that determines the
optimum number to solve a problem given. In some cases these parameters are
determined by test and error. However, there are current techniques for obtaining
automatic estimations [13], although this research follows the steps described by
Tarassenko [14].
The neural network was created with the commercial software Neurosolutions
for Excel v4 and its architecture is shown in Fig. 32.3.
Five neurons in the input layer linked with the five input variables (R,L,H,M,V).
The chosen transference function is the identity.
Four neurons in the hidden layer with hyperbolic tangent transference function.

Fig. 32.3 Structure of the neural network designed

32 Multilayer Perceptron Training Optimization for High Speed Impacts Classification 383

Two neurons in the output layer associated to the output variable plate perforation
(both outputs are complementary as it allows to improve the learning process).
The chosen transference function in this case is the hyperbolic tangent.

32.3.3 Input Data

It is important to make an estimation of the necessary information to train the net-

work in an appropriate way. For this purpose, it is essential to have a number of data
large enough to ensure that. While the network is trained, it is essential to recog-
nise and answer to the wide set of conditions that may appear during the testing
process. If the number of data is too small, the complete range of connections that
the neural network should learn will not be covered. In this sense, there are differ-
ent approximations to calculate the optimum number of input data, like [9] or [15].
The designed system has followed the guidelines set by Tarassenko to calculate the
number of examples needed [14].
Fifty impact examples were generated by numerical simulation. Some of them
were used for the training and testing process and the other for the validation of the
network after learning. The modelling and simulation were done with the commer-
cial software ABAQUS/Explicit v6.4.1 [16], a finite element explicit code widely
used in highly non-linear dynamic simulation.
The stopping criterion selected was to perform a specific number of epochs to
train the network, in this case, 10,000 epochs. Consequently the neural network
does not use validation data in the training process because the stopping criterion
is set by the epochs and not by the cross validation error. Therefore the ANN only
uses a train subset for the training process and a test set to validate the accurate of
the network.
The set of training patterns is repeatedly presented in random order to the net-
work and the weight update equations are applied after the presentation of each
pattern. When all the patterns have been presented once to the network, one epoch
of training is made. After the training process, the unseen patterns of the test set are
presented to the network. This demonstrates the generalization performance of the
trained network.

32.4 Solution Proposed

In the light of the results obtained by Garcia et al. in 2006 [8], the neural networks
present themselves as an alternative to be borne in mind to recreate impact prob-
lems. The network results reliable to predict the projectile arrest with the advantage,
opposing to the use of simulation tools, that it boasts a computational cost very
inferior once the training has been carried out.
384 A. Garcia-Crespo et al.

However, these conclusions lead to another question. Could there be a more effi-
cient ANN model? That is to say a model that would need less input variables and
input data to be able to approximate the output function correctly.
The research carried out to solve this question established two fundamental goals,
first, to minimize the number of training data without affecting the generalization
ability of the model; and second, to analyse the influence of the different input
variables on the network learning ability.
These input variables form the independent variables of the function that allows
to approximate the perforation function. The result of this answer function is a
dichotomic variable that depends on the input variables, and admits “Yes” or “No”
values.
Hence it is possible to infer that the easier the function that describes the net-
work’s behaviour is, the less costly will be generate the set of training data through
numeric simulation.
The type of heuristic selection of input variables and training set size selected is
well documented in the early literature of Neural Networks within other domains,
and the results obtained certifies the suitability of this election [17].

32.4.1 Randomness Elimination

The results obtained in the former research [8], present the uncertainty if they
depend on a possible randomness of the data intended for training and testing. In
other words, if the assignment of the available data influence on the predictions of
the network.
To carry out a research with conclusive results, it was established a number of
trials by far superior to the previous work. The study carried out is broken out in a
series of experiments, in which the number of patterns intended for training and test
varies in each series. This way, the minimum number of data needed to ensure an
acceptable percentage of correct answers will be found.
It has been accomplished 100 different trials in each experiment, exchanging in
each of them and in a random way the data destined to train and test the network.
Moreover, in two different trials of a same experiment the same training and test
data will never coincide, therefore increasing the reliability on the results obtained.
Thanks to this, the possible random data that are provided to the network, which
could lead to obtain not very reliable results, are eliminated.
The catalogue of experiments to be done in the study is the following: the first
one includes 100 trials with 50 training data and 10 testing data. From this point
onwards, in each of them the number of data intended to training is reduced to 5.
The last experiment consists of 100 trials with 15 training data and 10 testing data.
The result obtained in each one of these 100 trails is processed and the average
error porcentage is showed for each experiment. In this way, the results obtained
in each of the eight experiments can be compared. Thanks to this, it can be deter-
mined how many training data are necessary to achieve good generalization rates;
32 Multilayer Perceptron Training Optimization for High Speed Impacts Classification 385

and besides, it ensures that the possible random data of the input patterns does not
influence on these results.

32.4.2 Determination of the Influence of the Independent

Variables

The second part of this study is focused on the analysis of the influence of each
variable within the training process and, therefore, of the mathematical model that
regulates its output.
So the main question is to find the explanatory variables, in other words, to
include in the model just those variables that can influence on the answer, rejecting
those ones that do not provide information. It is also necessary look for the possible
interactions between the independent variables that affect the answer variable. For
example, this work will analyse the results obtained by including the thickness and
the velocity separately, as well as when they are both included.

32.5 Evaluation

The data shown in Table 32.1 are the average error percentages that the network
makes for each experiment. The different experiments are grouped in columns,
where it is indicated the parameters or variables used, e.g the column “All-H” points
out that the experiment uses all the variables except the parameter H. One result
example in Table 32.1 is the following one: in the model with all the variables, the
value 3,2 is the average error of the network for 100 trials with 50 data to train and
10 to testing.

Table 32.1 Error percentage obtained by varying the input variables and the number of training
data
Train. All All All All All All All All All
data -H -L -M -R -V -H&M -H&V -H,M&V

50 3,2 7,7 3,3 4 5,6 26,3 10,2 30,1 37,9

45 5,3 12 6,5 5,9 6,2 34,1 10 33 45,6
40 5,3 15,3 4 5,7 6,5 30,1 10,6 36,5 38,4
35 5,6 10 6,9 5,5 4,8 27,2 11,3 36,3 39,4
30 11 14,1 6,6 7,3 9,7 32,9 14,5 42,6 43,1
25 11,2 11,2 9,5 8,8 9,4 36,6 16,5 39 46,6
20 13,4 14,2 14,6 11,8 14,2 35,7 17,3 39,1 44,1
15 13,9 11,9 13,6 10,6 13,8 40,7 12,9 42,8 45,2
Average 8,61 12,05 8,13 7,45 8,78 32,95 12,91 37,43 42,54
386 A. Garcia-Crespo et al.

Regarding the objective sought, to find the network’s architecture that makes the
smallest error, when the data is analysed it could be established that the more input
variables the network has, the better it learns.
However, it can occur that the networks saturation, that is to say to introduce
all the available parameters, is not the best option. In this problem, this network
architecture has the lowest error probability only for 50 and 45 train data. In addition
to this, there are two designs that have lower average probability of error, the ones
without mass and without length. Specifically the design that omits the mass as a
network learning variable is the one that has the smallest average error.
On the other hand, for the average of the error obtained, it can be observed that
the velocity is a very relevant magnitude for the network learning. Its omission leads
the error probability to increase 342%, or what is the same, 4.4 times in relation
to the network’s design considered as the most efficient (architecture of network
without mass).
Finally as expected, the network with the worst results is the one with less infor-
mation, that is, with less input variables. However, it is advisable to highlight that
the network without the thickness and mass variables has quite acceptable results
(12.91%).
Taking the best configuration for the most architectures, 50 train and 10 testing,
the Figure 32.4 depicts the average error percentage made for each network archi-
tecture in which some variable is removed with regard to the network that holds all
of them.
Table 32.2 portrays the differences of the error percentages for each of the
network designs with regard to the one with the minimum error. As it has been

Fig. 32.4 Error made in each architecture for the configuration 50 train and 10 test
32 Multilayer Perceptron Training Optimization for High Speed Impacts Classification 387

Table 32.2 Deviation percentage made in each simulation in relation to the one that makes the
minimum error
Train. All All All All All All All All All
data -H -L -M -R -V -H&M -H&V -H,M&V
50 0 4,5 0,1 0,8 2,4 23,1 7 26,9 34,7
45 0 6,7 1,2 0,6 0,9 28,8 4,7 27,7 40,3
40 1,3 11,3 0 1,7 2,5 26,1 6,6 32,5 34,4
35 0,8 5,2 2,1 0,7 0 22,4 6,5 31,5 34,6
30 4,4 7,5 0 0,7 3,1 26,3 7,9 36 36,5
25 2,4 2,4 0,7 0 0,6 27,8 7,7 30,2 37,8
20 1,6 2,4 2,8 0 2,4 23,9 5,5 27,3 32,3
15 3,3 1,3 3 0 3,2 30,1 2,3 32,2 34,6
Total 13,8 41,3 9,9 4,5 15,1 208,5 48,2 244,3 285,2

mentioned, the network that includes all the available variables in the learning is
the one that makes the smaller error for 50 and 45 data; whilst the network without
mass is the one that boasts the smallest error for a smaller number of trials. Based
on these results, it can be concluded that the influence of the mass on learning is not
very relevant, and therefore the network that does not include it, it is considered the
most efficient.

32.6 Conclusions

In the light of the results and taking into account that perceptron is one of the sim-
plest topologies, this work shows clearly the possibilities of the neural networks in
the prediction of the material behaviour at high deformation speeds. Furthermore,
the architecture chosen presents a high reliability when predicting the result of a
projectile impact. Moreover, its computational cost once the training has started, is
smaller than the one of the simulations carried out with tools of finite elements.
It is crucial to highlight the small number of training data needed for the network
to learn with a relative small error in its predictions. With only 40 numeric simula-
tions of finite elements, the network obtains an error below 6% in both designs with
the best results (with all the variables and without the mass variable). In spite of the
small number of input data, the network does not present overlearning problems.
The experiments developed help to better understand the ballistic impact, analy-
sing the influence of the parameters that appear in the penetration problem. The
results obtained verify that the variable with the most influence is the velocity. Fur-
thermore, any network architecture where the velocity variable is removed obtains
error percentages quite high, what confirms this variable importance.
On the other hand, the research determines that the influence of the mass on learn-
ing is not very relevant. Therefore and taking into account its numeric simulation
cost, the network without this variable is considered the most efficient.
388 A. Garcia-Crespo et al.

The network with the worst results is the one with less information, that is to
say, with less input variables. However, it is advisable to highlight that the network
without thickness and mass variables has quite acceptable results taking into account
the little information it receives in comparison with the rest.
Finally, the knowledge acquired as a result of this research, can be spread out to
other fields of great practical interest. Among these, the design of structural compo-
nents of energy absorption stands out, being of great relevance in the field of passive
safety in vehicles.

References

1. K. Swingler. Applying Neural Networks: A Practical Guide. Morgan Kaufmann, San Fran-
cisco, CA 1996.
2. Z. Waszczyszyn and L. Ziemianski. Neural networks in mechanics of structures and materials.
New results and prospects of applications. Computers & Structures, 79 IS - 22-25:2261 EP–
2276, 2001.
3. R. Ince. Prediction of fracture parameters of concrete by artificial neural networks. Engineer-
ing Fracture Mechanics, 71(15):2143–2159, 2004.
4. S.W. Liu, J.H. Huang, J.C. Sung, and C.C. Lee. Detection of cracks using neural networks and
computational mechanics. Computer methods in applied mechanics and engineering (Comput.
methods appl. mech. eng.), 191(25-26):2831–2845, 2002.
5. W. Carpenter and J. Barthelemy. A comparison of polynomial approximations and artifi-
cial neural nets as response surfaces. Structural and multidisciplinary optimization, 5(3):166,
1993.
6. P. Hajela and E. Lee. Topological optimization of rotorcraft subfloor structures for crash
worthiness consideration. Computers and Structures, 64:65–76, 1997.
7. L. Lanzi, C. Bisagni, and S. Ricci. Neural network systems to reproduce crash behavior of
structural components. Computers structures, 82(1):93, 2004.
8. A. Garcia, B. Ruiz, D. Fernandez, and R. Zaera. Prediction of the response under impact of
steel armours using a multilayer perceptron. Neural Computing & Applications, 2006.
9. C. Bishop. Neural Networks for Pattern Recognition. Oxford University Press, USA, 1996.
10. G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of
Control, Signals, and Systems, 2:303–314, 1989.
11. K. Hornik and M. Stinchcombe. Multilayer feedforward networks are universal approxima-
tors. Neural Networks, 2(5):359–366, 1989.
12. R. Lippmann. An introduction to computing with neural nets. ASSP Magazine, IEEE [see also
IEEE Signal Processing Magazine], 4(2):4–22, 1987.
13. P. Isasi and I. Galvan. Redes de neuronas artificiales: un enfoque practico. Pearson Prentice
Hall, Madrid, 2004.
14. L. Tarassenko. A guide to neural computing applications. Arnol/NCAF, 1998.
15. B. Widrow. 30 years of adaptive neural networks: perceptron, madaline, and backpropagation.
Proceedings of the IEEE 78, 9:1415–1442, 1990.
16. USA. ABAQUS Inc., Richmond. Abaqus/explicit v6.4 users manual, 2003.
17. M. Gevrey, I. Dimopoulos, and S. Lek. Review and comparison of methods to study
the contribution of variables in artificial neural network models. Ecological Modelling,
160(16):249–264, 2003.
Chapter 33
Developing Emotion-Based Pet Robots

W.-P. Lee, J.-W. Kuo, and P.-C. Lai

Abstract Designing robots for home entertainment has become an important appli-
cation of intelligent autonomous robot. Yet, robot design takes considerable amount
of time and the short life cycle of toy-type robots with fixed prototypes and repetitive
behaviors is in fact disadvantageous. Therefore, it is important to develop a frame-
work of robot configuration so that the user can always change the characteristics of
his pet robot easily. In this paper, we present a user-oriented interactive framework
to construct emotion-based pet robots. Experimental results show the efficiency of
the proposed framework.

Keywords Pet robot  robot emotion  human-robot interaction  behavior coordina-

tion  neural network

33.1 Introduction

In recent years, designing robots for home entertainment has become an important
application of intelligent autonomous robot. One special application of robot enter-
tainment is the development of pet-type robots and they have been considered the
main trend of the next-generation electronic toys [1, 2]. This is a practical step of
expanding robot market from traditional industrial environments toward homes and
offices.
There have been many toy-type pet robots available on the market, such as Tiger
Electronics’ Furby, SONY’s AIBO, Tomy’s i-SOBOT and so on. In most cases, the
robots have fixed prototypes and features. With these limitations, their life cycle is
thus short, as the owner of pet robots may soon feel bored and no longer interested
in their robots. Sony’s robot dog AIBO and humanoid robot QRIO are sophisticated

W.-P. Lee (B)

Department of Information Management, National Sun Yat-sen University, Kaohsiung, Taiwan,
E-mail: wplee@mail.nsysu.edu.tw

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 389
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
390 W.-P. Lee et al.

pet robots with remarkable motion ability generated from many flexible joints [2,3].
But these robots are too expensive to be popular. Also the owners are not allowed
to reconfigure the original design. Therefore, it would be a great progress to have
a framework for robot configuration so that the user can always change the char-
acteristics of his robot according to his personal preferences to create a new and
unique one.
Regarding the design of pet robots, there are three major issues to be consid-
ered. The first issue is to construct an appropriate control architecture by which the
robots can perform coherent behaviors. The second issue is to deal with human-
robot interactions in which natural ways for interacting with pet robots must be
developed [4, 5]. And the third issue to be considered is emotion, an important
drive for a pet to present certain level of intelligence [6, 7]. In fact, Damasio has
suggested that efficient decision-making largely depends on the underlying mecha-
nism of emotions [8]. By including an emotional model, the pet robot can explicitly
express its internal conditions through the external behaviors, as the real living crea-
ture does. On the other hand, the owner can understand the need and the status of
his pet robot to then make appropriate interactions with it.
To tackle the above problems, in this paper we present an interactive framework
by which the user can conveniently design (and re-design) his personal pet robot
according to his preferences. In our framework, we adopt the behavior-based archi-
tecture ([9, 10]) to implement control system for a pet robot to ensure the robot
functioning properly in real time. A mechanism for creating behavior primitives and
behavior arbitrators is developed in which an emotion model is built for behavior
coordination. Different interfaces are also constructed to support various human-
robot interactions. To evaluate our framework, we use it to construct a control
system for the popular LEGO Mindstorms robot. Experimental results show that
the proposed framework can efficiently and rapidly configure a control system for a
pet robot. In addition, experiments are conducted in which a neural network is used
to learn a user-specified emotion model for behavior coordination. The results and
analyses show that an emotion-based pet robot can be designed and implemented
successfully by the proposed approach.

33.2 Building Emotion-Based Pet Robots

33.2.1 A User-Oriented Framework

Our aim is to develop a user-oriented approach that can assist a user to rapidly design
(and re-design) a special and personal robot. Robot design involves the configuration
of hardware and software. Expecting an ordinary user to build a robot from a set of
electronic components is in fact not practical. Therefore, instead of constructing a
personal pet robot from the electronic components, in this work we concentrate on
how to imbue a robot with a personalized control system.
33 Developing Emotion-Based Pet Robots 391

Fig. 33.1 The overall architecture of our system framework

Figure 33.1 illustrates the system framework of the proposed approach. As can be
seen, our system mainly includes three modules to deal with the problems in build-
ing a control system for a pet robot. The first module is to design an efficient control
architecture that is responsible for organizing and operating the internal control flow
of the robot. Because behavior-based control architecture has been used to con-
struct many robots acting in the real world successfully, our work adopts this kind
of architecture to design control systems for robots. The second module is about
human-robot interaction. As a pet robot is designed to accompany and entertain its
human partner in everyday life, interactions between the owner and his pet are essen-
tial. Here we develop three ways for human-robot interaction and communication,
including device-based (keyboard and mouse), voice-based and gesture-based meth-
ods. The third module is an emotion system that works as the arbitration mechanism
to resolve the behavior selection problem within the behavior-based control archi-
tecture. With the emotion system, a pet robot can act autonomously. It can choose
whether to follow the owner’s instructions, according to its internal emotions and
body states. Our system is designed to be user-oriented and has a modular struc-
ture. With the assist of the presented system, a user can build his robot according
to his preferences. If he is not satisfied with what the robot behaves, he can change
any part of the control software for further correction. Details of each module are
described in the following subsections.

33.2.2 Control Architecture

As in the behavior-based control paradigm [9, 10],the behavior system here takes
the structure of parallel task-achieving computational modules to resolve the
392 W.-P. Lee et al.

control problem. In order to achieve more complex tasks in a coherent manner, the
behavior modules developed have to be well organized and integrated. Inspired by
the ethological models originally proposed to interpret the behavior motivations of
animals, robotists have developed two types of architectures: the flat and the hierar-
chical ones. The former arranges the overall control system into two levels; and the
latter, multiple levels. In the flat type arrangement, all subtasks are independent and
have their own controllers. The independent outputs from the separate controllers
will be combined appropriately in order to generate the final output for the overall
task. As this work means to provide an interactive and easy-to-use framework for
the design and implementation of pet robots, the straightforward way (that involves
less task decomposition techniques), the flat architecture, is chosen to be the default
control structure for a pet.
Using behavior-based control architecture, one needs to deal with the corre-
sponding coordination (or action selection) problem. It is to specify a way to
combine the various outputs from the involved controllers. Our work here takes the
method of command switching that operates in a winner-take-all fashion. That is,
only one of the output commands from the involved behavior controllers is chosen
to take over the control at any given moment, according to certain sensory stimuli.
As mentioned above, we intend to build low cost toy-type robots as embodied
digital pets. Yet, since a toy-type robot only has limited computational power and
storage resources, it is generally not possible to perform all the above computation
on such a robot. Therefore, to implement the above architecture, an external com-
puting environment is needed to build robots. The experimental section will present
how we construct an environment for the LEGO robots.

33.2.3 User–Robot Interaction

To communicate with a pet robot, our framework provides two natural ways for
interacting with the robot by using oral or body languages. For the oral communi-
cation, we implement the popular speech-control method, command-and-control, to
communicate with the robot, and adopt the Microsoft Speech API to implement the
speech recognition mechanism in our Windows application. Though more sophis-
ticated language interface can be developed, here we simply parse a sentence into
individual words, distinguish whether it is an affirmative or negative sentence, and
then recognize the user commands from the words.
Gesture recognition is another way used in this work to support natural commu-
nication between human and robots. Gesture recognition is the process by which
the user’s gestures are made known to the system via appropriate interpretation. To
infer the aspects of gesture needs to sense and collect data of user position, configu-
ration, and movement. This can be done by directly using sensing devices attached
to the user (e.g., magnetic field trackers), or by indirectly using cameras and com-
puter vision techniques. Here we take a dynamic gesture recognition approach, and
use digital cameras with image processing techniques for gesture recognition.
33 Developing Emotion-Based Pet Robots 393

For simplicity, in our current implementation, we use light spots to represent

the hand positions of a pet robot owner, and extract the light track with the cubic
spline interpolation method to obtain a behavior curve. Then we take a curve-fitting
approach to match the curve produced by the continuous hand movement to the
ones recorded previously as user’s instructions in the database. Figure 33.2 shows
a typical example. A successful match means that the gesture has been recognized
and its corresponding instruction is then identified. As the traditional approach of
dynamic programming for curve fitting has inherent drawback caused by the curve
discretization-based distance measurement, in this work we employ the approach
reported in [11] that takes the underlying differential equation into account and finds
a continuous solution for curve fitting. The fitting result is then used to recognize
gestures.
Figure 33.3 illustrates our speech/gesture recognition mechanism that mainly
includes three parts. The first part is to develop and maintain a behavior library
within which different behavior modules are pre-constructed and recorded. The
second part is to capture and record an owner’s spoken-sentences/images, and to
extract the words/curves for further interpretation. The third part is a matching pro-
cedure that tries to match the extracted words/curves to the ones recorded in the
mapping table in which each command-word/curve represents a behavior controller.
If a match is found, the corresponding behavior module is then retrieved and acti-
vated from the behavior library. The behavior controller is executed and the control

Fig. 33.2 An example of extracting trajectory curve

Fig. 33.3 Speech/gesture-based control

394 W.-P. Lee et al.

command is used to drive the robot. Alternatively, if the behavior controllers can be
pre-uploaded to the on-board computer of the robot, the mapping result will send
an identification signal to trigger the corresponding controller on the robot. Here
the user is allowed to construct the mapping table to decide how to interact with
his robot.

33.2.4 Emotion Model

To coordinate different behavior controllers, our framework uses a special mecha-

nism that exploits emotions for selection of behavior. Emotions are often categorized
into two kinds: basic emotions and higher cognitive emotions. Basic emotions (such
as joy, anger, fear, etc.) are considered as universal and innate; they are encoded
in the genome. Higher cognitive emotions (such as love, shame, pride, etc.) are
universal too, but they exhibit more cultural variations. Basic emotions tend to
be reflex-like responses over which animals have little conscious control. They
involve less cognitive processes and are thus much faster in controlling motion than
those culturally determined experiences residing in the cognitive system (long term
memory). As our goal here is to establish a framework for constructing personal pet
robots, rather than to investigate the interactions and relationships between cogni-
tive and emotion systems, our work only models basic emotions to coordinate the
behave controllers the user has pre-chosen for his pet. To quantify the emotions, we
define each of the basic emotion as: Ei .t/ D Ei .t  1/ C ˛  Eventi , in which Ei .t/
represents the quantity of emotion Ei at any time t, ˛ is a user-specified weight,
and Eventi is a procedure describing how emotion Ei is affected by a set of events
pre-defined by the user. For example, a user can define an event to be the appear-
ance of a stranger. When this happens, “fear” (one kind of emotion of the robot) will
increase one unit accordingly. An experienced user can define a more sophisticated
procedure to alleviate the effect caused by the same event occurring during a short
period of time.
In addition to emotions, another set of internal variables is defined to describe a
robot’s body states (e.g., hungry). These variables represent the basic requirements
the robot has to be satisfied, and they must be regulated and maintained in certain
ranges. In our framework, the range of each internal variable can be defined by the
user, and the corresponding specification determines the innate characteristics of his
robot. Similar to the emotion functions defined above, an internal variable here is
described by a mathematical formula with a set of involved events, also specified by
the user.
Figure 33.4 shows the aspect of the emotion system used. The emotion model
determines the innate characteristics of a robot, which are highly related to its
abilities of learning and adaptation. With the above model, at any given time, the
emotions of the robot can be derived and used to trigger the behavior controller to
generate a specific action. After the action is performed, the external environment
conditions will thus change and that can further affect the emotion of the robot. The
33 Developing Emotion-Based Pet Robots 395

Fig. 33.4 Emotion-based behavior coordination

pet robot then makes new decision for behavior selection, based on the modified
quantities of emotions. For example, a pet dog in a hungry state may be angry, may
keep looking for food, and would eat anything as soon as the pet dog finds it. After
that, the dog may not be hungry any more (the body state has been changed). Then
it is happy (the emotion has been changed too) and may want to sleep or fool around
(now new behavior is selected). The above procedure is repeated and the emotion
model continuously works as a decision-making mechanism for behavior selection.
As can be observed, in the above operating model, the most important part is the
one for selecting appropriate behavior at any time. In this work, we use a feedfor-
ward neural network to implement such a mechanism that maps the emotions and
body states of a robot into a set of desired behaviors. To allow the user to determine
the characteristics of his pet, our work also provides an interface by which the user
can define examples to train the neural network to achieve the specified mapping of
emotions and behaviors. Once the neural network is obtained, it works as a behavior
selector to choose appropriate controllers for the robot.

33.3 Experiments and Results

33.3.1 Implementation

To evaluate our approach, two robots have been built and a distributed and net-
worked computing environment has been developed for the robots. The robot used in
the experiments is LEGO Mindstorms NXT 9797. It has light sensors for light detec-
tion, ultrasonic sensors for distance measurement, and micro-switches for touch
detection. In addition, to enhance its vision ability, we equip an extra mini-camera
on the head of the robot so that it can capture images for further interpretation.
Because the LEGO NXT only has limited computational power and memory
devices, a distributed and networked computing environment is thus constructed for
the robot, in which the individual pre-programmed behavior controllers are installed
in the NXT on-board memory, and two PCs are connected through the Internet for
other computation. Here, one PC is responsible for the computation of emotion
396 W.-P. Lee et al.

modeling, speech recognition, and communication management, and the other PC

is used for image/gesture recognition. The network-based computing environment
can be extended to deal with more computation when more robots are involved.
The emotion system is also executed in one of the host PCs. It interprets the
received signals to activate different events to change emotions and body states
accordingly, as described in Section 33.2.4. The newly derived values of emo-
tions and states are used to choose a behavior controller, according to the mapping
strategy that can be pre-wired manually or learnt automatically. Once the behavior
controller is selected, the host PC then sends an identification signal through the
Bluetooth channel to evoke the corresponding controller from the set of controllers
pre-stored on the robot.
After the above computing environment was established, experiment has been
conducted to show how our LEGO robots can achieve a user-specified task. As soon
as the owner asked one of the robots to push a red box, it started moving around to
look for the box. Once the robot found the box, it tried to push the box. As the box
was too heavy and the robot could not push it away alone, it then sent a message
to another robot through the built-in Bluetooth communication channel to ask for
help. The second robot came over, found the box and tried to push it too. As shown
in Fig. 33.5, after the two robots pushed the box together, the box was moved away
successfully.

33.3.2 Building Emotion-Based Behavior Selector

To verify our approach of emotion-based control, this section describes how the
emotion model can be built, trained and modified. In the experiments, the sim-
ple types of emotions are modeled. They are so-called “basic emotions”, including
“happy”, “angry”, “fear”, “bored”, “shock”, and “sad”. Also three variables, “hun-
gry”, “tired”, and “familiar” are defined to indicate the internal body states of the
robot. As mentioned in Section 33.2.4, the user is allowed to define event proce-
dures and the relevant weight parameters for the above emotions and body states
to describe how the quantities of different emotions vary over time for their own

Fig. 33.5 Two LEGO NXT robots push the red box together
33 Developing Emotion-Based Pet Robots 397

robots. For naı̈ve users, our framework also provides three default emotion models
(aggressive, gentle, and shy) and events as choices to represent different charac-
teristics of a robot. For example, a robot with an aggressive model will change its
emotions rapidly than others. Users can choose a default model from the interface
without extra settings. Experienced users can develop more sophisticated models to
guide the variations of emotions.
Currently, ten basic behaviors are built, including target seeking, barking, wan-
dering, shaking, sniffing, sleeping, escaping, scratching, wailing, and wagging. As
mentioned above, with the emotion model developed, a user can build a behavior
selector manually or automatically and use it to map the emotions and body states
into appropriate behavior controllers. At each time step, the internal emotions and
states of the robot change and the newly obtained values are used as the input of an
emotion model to select behavior controller at that moment. Figure 33.6 shows the
interface that presents the numerical values of the internal and emotion variables
over time during a trial. These values are illustrated to provide information about
the body state and emotion of the robot, so that the user can inspect the detailed
information related to his personal robot accordingly.
As can be seen in Fig. 33.6, the interface also includes a set of event buttons on
the right hand side to simulate the happening of different events. Each button here is
associated with an event procedure that describes how emotions and body states are
changed by this event. It corresponds to a situated event in reality. That is, when the
pre-conditions of an event procedure are satisfied in the real world, the same effect
will be given to change the emotions and body states of a robot. With the assistance

Fig. 33.6 The interface for showing the typical results

398 W.-P. Lee et al.

of event buttons, users can examine the correctness of the effect caused by the event
procedure he defined in an efficient way.
In addition, our framework offers a learning mechanism to train a feedforward
neural network from examples as a behavior selector. The above variables (emotions
and body states) are arranged as the input of the network, and it output is used to
determine which behavior to perform at a certain time. In the training phase, the user
is allowed to give a set of training examples and each specifies which behavior the
robot is expected to perform when the set of emotions and states reaches the values
he has assigned. The back-propagation algorithm is then used to derive a model for
the set of data examples. Figure 33.7 presents the interface through which a user can
edit the training set. Based on the training set provided by the user, the system then
tries to learn a mapping strategy with best approximation.
It should be noted that it is possible the examples provided by the user are incon-
sistent and consequently a perfect strategy cannot be obtained. In the latter case,
the user can use the interface shown in Fig. 33.7 to correct the training examples to
re-build the behavior selector (i.e., the neural network) again. If the model has been
derived successfully but the behavior of the robot did not satisfy the owner’s expec-
tation, he can still correct the robot behavior for any specific time step by editing
the output produced by the mapping strategy learnt previously through the inter-
faces shown in Fig. 33.7. Then the modified outputs can be used as new training
examples to derive a new strategy of behavior arbitration. In this way, the user can
easily and conveniently design (and re-design) the characteristics of his personal
robot.

33.4 Conclusions and Future Work

In this paper, we have described the importance of developing toy-type pet robots
as an intelligent robot application. We have also proposed to integrate knowledge
from different domains to build low-cost pet robots. To realize the development of
pet robot, a user-oriented interactive framework has been constructed with which
the user can conveniently configure and re-configure his personal pet robot accord-
ing to his preferences. Our system framework mainly investigates three issues:
robot control, human-robot interaction, and robot emotion. Different interfaces
have also been constructed to support various human-robot interactions. Most
importantly, an emotion-based mechanism has been developed in which different
emotions and internal states have been modeled and used to derive a behavior selec-
tor. The behavior selector is a neural network and the user is allowed to define
training examples to infer a behavior selector for his robot. To evaluate our frame-
work, we have used it to build LEGO NXT robots to achieve a cooperation task
successfully.
Based on the presented framework, we are currently trying different toy-type
robots with more sensors and actuators to evaluate our approach extensively. Also
33 Developing Emotion-Based Pet Robots 399

Fig. 33.7 The interface for preparing training data

we are implementing a new vision module for the robot so that it can recognize
human facial expressions and interact with people accordingly. In addition, we
plan to define a specific language and construct a message-passing channel through
which different types of robots can communicate with each other.
400 W.-P. Lee et al.

References

1. M. M. Veloso, Entertainment robotics, Communication of the ACM, 45(3), 59–63, 2002.

2. M. Fujita, On activating human communications with pet-type robot AIBO, Proceedings of
the IEEE, 92(11), 1804–1813, 2004.
3. T. Ishida, Y. Kuroki, and J. Yamaguchi, Mechanical system of a small biped entertainment
robot, Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems,
pp. 1129–1134, 2003.
4. S. Thrun, Toward a framework for human-robot interaction, Human-Computer Interaction,
19(1–2), 9–24, 2004.
5. D. Perzarcowski, A. Schultz, W. Adams, E. Marsh, and M. Bugajska, Building a multimodal
human-robot interface, IEEE Intelligent Systems, 16(1), 16–21, 2001.
6. J. E. LeDoux, The Emotional Brain, New York: Simon & Schuster, 1996.
7. J.-M. Fellous and M. Arbib (eds.), Who Needs Emotions? The Brain Meets the Robot, New
York: Oxford University Press, 2005.
8. A. R. Damasio, Descarte’s Error: Emotion, Reason, and Human Brain, New York: Grosset/
Putnam, 1994.
9. R. C. Arkin, Behavior-Based Robotics, Cambridge, MA: MIT Press, 1998.
10. R. A. Brooks, A robust layered control system for a mobile robot, IEEE Journal of Robotics
and Automation, 2(1), 14–23, 1986.
11. M. Frenkel and R. Basri, Curve matching using the fast marching method, Proceedings of
International Workshop on Energy Minimization Methods in Computer Vision and Pattern
Recognition, LNCS-2683, pp. 35–51, 2003.
Chapter 34
Designing Short Term Trading Systems
with Artificial Neural Networks

Bruce Vanstone, Gavin Finnie, and Tobias Hahn

Abstract There is a long established history of applying Artificial Neural Networks

(ANNs) to financial data sets. In this paper, the authors demonstrate the use of
this methodology to develop a financially viable, short-term trading system. When
developing short-term systems, the authors typically site the neural network within
an already existing non-neural trading system. This paper briefly reviews an existing
medium-term long-only trading system, and then works through the Vanstone and
Finnie methodology to create a short-term focused ANN which will enhance this
trading strategy. The initial trading strategy and the ANN enhanced trading strat-
egy are comprehensively benchmarked both in-sample and out-of-sample, and the
superiority of the resulting ANN enhanced system is demonstrated.

Keywords Trading System  Short-Term  Artificial Neural Network  trading

strategy  Vanstone and Finnie methodology

34.1 Introduction

There is a long established history of applying Artificial Neural Networks (ANNs)

to financial data sets, with the expectation of discovering financially viable trading
rules. Despite the large amount of published work in this area, it is still difficult
to answer the simple question, “Can ANNs be used to develop financially viable
stockmarket trading systems?” Vanstone and Finnie [1] have provided an empirical
methodology which demonstrates the steps required to create ANN-based trading
systems which allow us to answer this question.

B. Vanstone (B)
Faculty of Business, Technology and Sustainable Development Bond University Gold Coast,
Queensland, 4229 Australia
E-mail: bvanston@bond.edu.au

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 401
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
402 B. Vanstone et al.

In this paper, the authors demonstrate the use of this methodology to develop a
financially viable, short-term trading system. When developing short-term systems,
the authors typically site the neural network within an already existing non-neural
trading system. This paper briefly reviews an existing medium-term long-only trad-
ing system, and then works through the Vanstone and Finnie methodology to create
a short-term focused ANN which will enhance this trading strategy.
The initial trading strategy and the ANN enhanced trading strategy are compre-
hensively benchmarked both in-sample and out-of-sample, and the superiority of the
resulting ANN enhanced system is demonstrated. To prevent excessive duplication
of effort, only the key points of the methodology outlined are repeated in this paper.
The overall methodology is described in detail in Vanstone and Finnie [1], and this
methodology is referred to in throughout this paper as ‘the empirical methodology’.

34.2 Review of Literature

There are two primary styles of stockmarket trader, namely Systems traders, and
Discretionary traders. Systems traders use clearly defined rules to enter and exit
positions, and to determine the amount of capital risked. The strategies created
by systems traders can be rigorously tested, and clearly understood. The alterna-
tive, discretionary trading, is usually the eventual outcome of an individual’s own
experiences in trading. The rules used by discretionary traders are often difficult to
describe precisely, and there is usually a large degree of intuition involved. In many
cases, some of the rules are contradictory – in these cases, the discretionary trader
uses experience to select the appropriate rules. Despite these obvious drawbacks,
however, it is commonly accepted that discretionary traders produce better financial
results [2].
For the purposes of this paper, it is appropriate to have a simple, clearly defined
mathematical signal which allows us to enter or exit positions. This allows us to
accurately benchmark and analyze systems.
This paper uses the GMMA as the signal generator. The GMMA is the Guppy
Multiple Moving Average [3], as created and described by Daryl Guppy [4], a lead-
ing Australian trader. Readers should note that Guppy does not advocate the use of
the GMMA indicator in isolation (as it is used in this study), rather it is appropri-
ate as a guide. However, the GMMA is useful for this paper, as it is possible to be
implemented mechanically. In essence, any well defined signal generator could be
used as the starting point for this paper.
The GMMA is defined as:
00 1 0 11
ema.3/ C ema.5/ ema.30/ C ema.35/
GMMA D @@ Cema.8/ C ema.10/ A  @ Cema.40/ C ema.45/ AA (34.1)
Cema.12/ C ema.15/ Cema.50/ C ema.60/
34 Designing Short Term Trading Systems with Artificial Neural Networks 403

Creation of the ANNs to enhance this strategy involves the selection of ANN
inputs, outputs, and various architecture choices. The ANN inputs and outputs are
a cut-down version of those originally described in Vanstone [5]. The original list
contained 13 inputs, and this paper uses only five. These five variables, discussed
later in this paper, were selected as they were the most commonly discussed in
the main practitioners’ journal, ‘The Technical Analysis of Stocks and Commodi-
ties’. Similarly, the choices of output and architecture are described in the empirical
methodology paper. Again, these are only briefly dealt with here.
For each of the strategies created, an extensive in-sample and out-of-sample
benchmarking process is used, which is also further described in the methodology
paper.

34.3 Methodology

This study uses data for the ASX200 constituents of the Australian stockmarket.
Data for this study was sourced from Norgate Investor Services [6]. For the in-
sample data (start of trading 1994 to end of trading 2003), delisted stocks were
included. For the out-of-sample data (start of trading 2004 to end of trading 2007)
delisted stocks were not included. The ASX200 constituents were chosen primarily
for the following reasons:
1. The ASX200 represents the most important component of the Australian equity
market due to its high liquidity – a major issue with some previously published
work is that it may tend to focus too heavily on micro-cap stocks, many of which
do not have enough trading volume to allow positions to be taken, and many of
which have excessive bid-ask spreads.
2. This data is representative of the data which a trader will use to develop his/her
own systems in practice, and is typical of the kind of data the system will be used
in for out-of-sample trading.
Software tools used in this paper include Wealth-Lab Developer, and Neuro-Lab,
both products of Wealth-Lab Inc (now owned by Fidelity) [7]. For the neural net-
work part of this study, the data is divided into two portions: data from 1994 up
to and including 2003 (in-sample) is used to predict known results for the out-of-
sample period (from 2004 up to the end of 2007). In this study, only ordinary shares
are considered.
The development of an ANN to enhance the selected strategy is based on simple
observation of the GMMA signals. One of the major problems of using the GMMA
in isolation is that it frequently whipsaws around the zero line, generating spurious
buy/sell signals in quick succession.
One possible way of dealing with this problem is to introduce a threshold which
the signal must exceed, rather than acquiring positions as the zero line is crossed.
The method used in this paper, however, is to forecast which of the signals is most
likely to result in a sustained price move. This approach has a major advantage over
404 B. Vanstone et al.

the threshold approach; namely, in a profitable position, the trader has entered ear-
lier, and therefore, has an expectation of greater profit. By waiting for the threshold
to be exceeded, the trader is late in entering the position, with subsequent decrease
in profitability.
However, for the approach to work, the trader must have a good forecast of
whether a position will be profitable or not. This is the ideal job for a neural network.
In Fig. 34.1, there are a cluster of trades taken between June 2006 and September
2006, each open for a very short period of time as the GMMA whipsaws around
the zero line. Eventually, the security breaks out into a sustained up trend. What
is required is an ANN which can provide a good quality short-term forecast of the
return potential each time the zero line is crossed, to allow the trader to discard the
signals which are more likely to become whipsaws, thus concentrating capital on
those which are more likely to deliver quality returns.
The neural networks built in this study were designed to produce an output signal,
whose strength was proportional to expected returns in the 5 day timeframe. In
essence, the stronger the signal from the neural network, the greater the expectation
of return. Signal strength was normalized between 0 and 100.
The ANNs contained five data inputs. These are the technical variables deemed
as significant from the review of both academic and practitioner publications, and
details of their function profiles are provided in Vanstone [5]. The formulas used to
compute these variables are standard within technical analysis. The actual variables
used as inputs, and their basic statistical characteristics are provided in Table 34.1.
For completeness, the characteristics of the output target to be predicted, the 5
day return variable, are shown in Table 34.2. This target is the maximum percentage

Fig. 34.1 GMMA signals

Table 34.1 Technical variables: statistical properties

Variable Min Max Mean Std dev
ATR(3)/ATR(15) 0:00 3:71 1:00 0:30
STOCHK(3) 0:00 100:00 54:52 36:63
STOCHK(15) 0:00 100:00 64:98 27:75
RSI(3) 0:12 100:00 58:07 25:00
RSI(15) 32:70 98:03 58:64 8:48
34 Designing Short Term Trading Systems with Artificial Neural Networks 405

Table 34.2 Target variable: statistical properties

Variable Min Max Mean Std dev

Target 0.00 100.00 5.02 21.83

change in price over the next 5 days, computed for every element i in the input
series as:    !
highest closei C5:::j C1  closei
 100 (34.2)
closei
Effectively, this target allows the neural network to focus on the relationship
between the input technical variables, and the expected forward price change.
The calculation of the return variable allows the ANN to focus on the highest
amount of change that occurs in the next 5 days, which may or may not be the 5-
day forward return. Perhaps a better description of the output variable is that it is
measuring the maximum amount of price change that occurs within the next 5 days.
No adjustment for risk is made, since traders focus on returns and use other means,
such as stop orders, to control risk.
As explained in the empirical methodology, a number of hidden node archi-
tectures need to be created, and each one benchmarked against the in-sample
data.
The method used to determine the hidden number of nodes is described in the
empirical methodology. After the initial number of hidden nodes is determined, the
first ANN is created and benchmarked. The number of hidden nodes is increased
by one for each new architecture then created, until in-sample testing reveals which
architecture has the most suitable in-sample metrics. A number of metrics are avail-
able for this purpose, in this study, the architectures are benchmarked using the
Average Profit/Loss per Trade expressed as a percentage. This method assumes
unlimited capital, takes every trade signaled, and includes transaction costs, and
measures how much average profit is added by each trade over its lifetime. The
empirical methodology uses the filter selectivity metric for longer-term systems,
and Tharp’s expectancy [8] for shorter term systems. This paper also introduces the
idea of using overall system net profit to benchmark, as this figure takes into account
both the number of trades (opportunity), and the expected return of each trade on
average (reward).

34.4 Results

A total of 362 securities had trading data during the test period (the ASX200 includ-
ing delisted stocks), from which 11,897 input rows were used for training. These
were selected by sampling the available datasets, and selecting every 25th row as an
input row.
406 B. Vanstone et al.

Table 34.3 In-sample characteristics

Strategy (in-sample data) Overall net Profit/loss Holding
system per trade period
profit (%) (days)
Buy-and-hold naı̈ve approach 1; 722; 869:39 94:81 2; 096:03
GMMA alone 632; 441:43 1:09 35:30
ANN – 3 hidden nodes + GMMA 878; 221:68 2:32 46:15
ANN – 4 hidden nodes + GMMA 1; 117; 520:33 3:69 59:20
ANN – 5 hidden nodes + GMMA 353; 223:61 3:00 42:64

Table 34.3 reports the Overall Net System Profit, Average Profit/Loss per Trade
(as a percentage), and Holding Period (days) for the buy-and-hold naı̈ve approach
(first row), the initial GMMA method (second row), and each of the in-sample ANN
architectures created (subsequent rows). These figures include transaction costs of
$20 each way and 0.5% slippage, and orders are implemented as day C1 market
orders. There are no stops implemented in in-sample testing, as the objective is not
to produce a trading system (yet), but to measure the quality of the ANN produced.
Later, when an architecture has been selected, stops can be determined using ATR
or Sweeney’s [9] MAE technique.
The most important parameter to be chosen for in-sample testing is the signal
threshold, that is, what level of forecast strength is enough to encourage the trader
to open a position. This is a figure which needs to be chosen with respect to the indi-
viduals own risk appetite, and trading requirements. A low threshold will generate
many signals, whilst a higher threshold will generate fewer. Setting the threshold
too high will mean that trades will be signalled only rarely, too low and the trader’s
capital will be quickly invested, removing the opportunity to take higher forecast
positions as and when they occur.
For this benchmarking, an in-sample threshold of 5 is used. This figure is chosen
by visual inspection of the in-sample graph in Fig. 34.2, which shows a break-
down of the output values of a neural network architecture (scaled from 0 to 100)
versus the average percentage returns for each network output value. The percent-
age returns are related to the number of days that the security is held, and these
are shown as the lines on the graph. Put simply, this graph visualizes the returns
expected from each output value of the network and shows how these returns per
output value vary with respect to the holding period. At the forecast value of 5
(circled), the return expectation rises above zero, so this value is chosen.
As described in the empirical methodology, it is necessary to choose which ANN
is the ‘best’, and this ANN will be taken forward to out-of-sample testing. It is for
this reason that the trader must choose the in-sample benchmarking metrics with
care. If the ANN is properly trained, then it should continue to exhibit similar
qualities out-of-sample which it already displays in-sample.
From the above table, it is clear that ANN – four hidden nodes should be selected.
It displays a number of desirable characteristics – it shows the highest level of
Profit/Loss per Trade. Note that this will not necessarily make it the best ANN for
34 Designing Short Term Trading Systems with Artificial Neural Networks 407

Fig. 34.2 In-sample ANN function profile

Table 34.4 Number of trades signalled

Strategy (in-sample data) Number of trades signaled
Buy-and-hold naı̈ve approach 362
GMMA alone 11;690
ANN – 3 hidden nodes + GMMA 7;516
ANN – 4 hidden nodes + GMMA 6;020
ANN – 5 hidden nodes + GMMA 2;312

a trading system. Extracting good profits in a short time period is only a desirable
trait if there are enough opportunities being presented to ensure the traders capital
is working efficiently.
Therefore, it is also important to review the number of opportunities signalled
over the 10-year in-sample period. This information is shown in Table 34.4.
Here the trader must decide whether the number of trades signalled meets the
required trading frequency. In this case, there are likely to be enough trades to keep
an end-of-day trader fully invested.
This testing so far covered data already seen by the ANN, and is a valid indication
of how the ANN should be expected to perform in the future. In effect, the in-sample
metrics provide a framework of the trading model this ANN should produce.
Table 34.5 shows the effect of testing on the out-of-sample ASX200 data,
which covers the period from the start of trading in 2004 to the end of trading
in 2007. These figures also include transaction costs and slippage, and orders are
implemented as next day market orders.
This was a particularly strong bull market period in the ASX200.
408 B. Vanstone et al.

Table 34.5 Out-of-sample performance

Strategy (out-of-sample data) Overall net system profit Profit/loss per trade (%)

GMMA alone 622,630.01 3.88

ANN – 4 hidden nodes + GMMA 707,730.57 10.94

Table 34.6 ANOVA comparison

ANOVA GMMA GMMA + 4 hidden nodes
Number of observations 3,175 1,283
Mean 196.10 551.62
Std deviation 1,496.99 2,483.64
95% Confidence internal of the mean – lower bound 144.01 415.59
95% Confidence internal of the mean – upper bound 248.19 687.65

Although there appears a significant difference between the GMMA, and the
ANN enhanced GMMA, it is important to quantify the differences statistically. The
appropriate test to compare two distributions of this type is the ANOVA test (see
supporting work in Vanstone [5]). The results for the ANOVA test are shown in
Table 34.6 below.
The figures above equate to an F-statistic of 34.26 (specifically, F (14,456) =
34.261, p = 0.00 (p < 0.05)), which gives an extremely high level of significant
difference between the two systems.

34.5 Conclusions

The ANN out-of-sample performance is suitably close to the ANN in-sample per-
formance, leading to the conclusion that the ANN is not curve fit, that is, it should
continue to perform well into the future. The level of significance reported by the
ANOVA test leads to the conclusion that the ANN filter is making a statistically
significant improvement to the quality of the initial GMMA signals.
The trader now needs to make a decision as to whether this ANN should be
implemented in real-life.
One of the main reasons for starting with an existing successful trading strategy
is that it makes this decision much easier. If the trader is already using the signals
from a system, and the ANN is used to filter these signals, then the trader is still only
taking trades that would have been taken by the original system. The only difference
in using the ANN enhanced system is that trades with low expected profitability
should be skipped.
Often in trading, it is the psychological and behavioural issues which undermine
a traders success. By training ANNs to support existing systems, the trader can have
additional confidence in the expected performance of the ANN.
34 Designing Short Term Trading Systems with Artificial Neural Networks 409

Fig. 34.3 GMMA signals filtered with an ANN

Finally, Fig. 34.3 shows the same security as Fig. 34.1. The ANN has clearly met
its purpose of reducing whipsaws considerably, which has resulted in the significant
performance improvement shown in Tables 34.3 and 34.5.
Of course the result will not always be that all whipsaws are removed. Rather,
only whipsaws which are predictable using the ANN inputs will be removed.

References

1. Vanstone, B. and Finnie, G. (2007). “An Empirical Methodology for developing Stockmarket
Trading Systems using Artificial Neural Networks.” Expert Systems with Applications. In-Press
(DOI: http://dx.doi.org/10.1016/j.eswa.2008.08.019).
2. Elder, A. (2006). Entries & Exits: Visits to Sixteen Trading Rooms. Hoboken, NJ: Wiley.
3. guppytraders.com. “Guppy Multiple Moving Average.” Retrieved 04-05-2007, from
www.guppytraders.com/gup329.shtml
4. Guppy, D. (2004). Trend Trading. Milton, QLD: Wrightbooks.
5. Vanstone, B. (2006). Trading in the Australian stockmarket using artificial neural networks,
Bond University. Ph.D.
6. “Norgate Premium Data.” (2004). Retrieved 01-01-2004, from www.premiumdata.net
7. “Wealth-Lab.” (2005) from www.wealth-lab.com
8. Tharp, V. K. (1998). Trade Your Way to Financial Freedom. New York: McGraw-Hill.
9. Sweeney, J. (1996). Maximum Adverse Excursion: Analyzing Price Fluctuations for Trading
Management. New York, Wiley.
Chapter 35
Reorganising Artificial Neural Network
Topologies
Complexifying Neural Networks by Reorganisation

Thomas D. Jorgensen, Barry Haynes, and Charlotte Norlund

Abstract This chapter describes a novel way of complexifying artificial neural net-
works through topological reorganisation. The neural networks are reorganised to
optimise their neural complexity, which is a measure of the information-theoretic
complexity of the network. Complexification of neural networks here happens
through rearranging connections, i.e. removing one or more connections and plac-
ing them elsewhere. The results verify that a structural reorganisation can help to
increase the probability of discovering a neural network capable of adequately solv-
ing complex tasks. The networks and the methodology proposed are tested in a
simulation of a mobile robot racing around a track.

Keywords Artificial Neural Network  topological reorganization  information-

theoretic complexity  mobile robot racing  Complexification

35.1 Introduction

Artificial Neural Networks (ANNs) have been used in many different applications,
with varying success. The success of a neural network, in a given application,
depends on a series of different factors, such as topology, learning algorithm and
learning epochs. Furthermore all of these factors can be dependent or independent
of each other. Network topology is the focus of this research, in that finding the
optimum network topology can be a difficult process. Ideally all network topolo-
gies should be able to learn every given task to competency, but in reality a given
topology can be a bottleneck and constraint on a system. Selecting the wrong topol-
ogy can result in a network that cannot learn the task at hand [1–3]. It is commonly
known that a too small or too large network does not generalise well, i.e. learn a

T. D. Jorgensen (B)
Department of Electronic and Computer Engineering,University of Portsmouth, UK
E-mail: Thomas.Jorgensen@port.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 411
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
412 T.D. Jorgensen et al.

given task to an adequate level. This is due to either too few or too many parameters
used to represent a proper and adequate mapping between inputs and outputs.
This chapter proposes a methodology that can help find an adequate network
topology by reorganising existing networks by rearranging one or more connections,
whilst trying to increase a measure of the neural complexity of the network. Assum-
ing complex task solving requires complex neural controllers, a reorganisation that
increases the controller complexity can increase the probability of finding an ade-
quate network topology. Reorganising an existing network into a more complex one
yields an increased chance of better performance and thus a higher fitness.
There are generally four ways to construct the topology of an ANN [3–5]. (1)
Trial and Error is the simplest method. This essentially consists of choosing a topol-
ogy at random and testing it, if the network performs in an acceptable way, the
network topology is suitable. If the network does not perform satisfactory, select
another topology and try it. (2) Expert selection; the network designer decides
the topology based on a calculation or experience [3, 6]. (3) Evolving connections
weights and topology through complexification. Extra connections and neurons can
be added as the evolutionary process proceeds or existing networks can be reorgan-
ised [7–12]. (4) Simplifying and pruning overly large neural networks, by removing
redundant elements [3–5]. This chapter seeks to add another one to this list, as
reorganisation of existing networks can potentially help discover the appropriate
topology of a network.
One advantage of the proposed methodology, compared to complexification
through adding components, is that the computational overhead is constant because
no extra components and parameters are added. The time it takes to compute the out-
put of the network is effected, as well as the time it takes for the genetic algorithm to
find appropriate values for the connection weights. More parameters yield a wider
and slower search of the search space and additionally it yields more dimensions to
the search space.

35.2 Background

Most research in complexification has so far focused on increasing the structural

complexity, i.e. increasing the number of network components, of a neural net-
work, this is done to mimic natural evolution [13]. Different routes and techniques
have been proposed to continuously complexify neural network for a continuous
increase in fitness, most prominently is the NEAT framework [10]. In the NEAT
model, mechanisms are introduced to evolve network structure, either by adding
neurons or connections, in parallel with the normal evolution of weights. The results
of these experiments with complexification achieve, in some cases, faster learning
as well as a neural network structure capable of solving more complex tasks than
produced by normally evolved controllers. Other approaches do not cross breed
networks of different topology, but use mutation as the evolutionary operator that
evolves the network. Angeline [8] proposes networks that are gradually evolved by
35 Reorganising Artificial Neural Network Topologies 413

adding connections or neurons and new components are frozen, so that fitness is not
reduced. This is similar to the first method of topological complexification proposed
by Fahlman [7], which increased network size by adding neurons.
Research into the use of neural complexity in complexification to produce bio-
logically plausible structures is limited, this is due to the lack of proper calculation
tools and the variety of definitions and focus. Neural complexity is a measure of
how a neural network is both connected and differentiated [11]. It is measure of
the structural complexity as well as the connectivity of the network. The measure
was developed to measure the neural complexity of human and animal brains by
estimating the integration of functionally segregated modules. This measure reflects
the properties that fully connected networks and functionally segregated networks
have low complexity, whilst networks that are highly specialised and also well inte-
grated are more functionally complex. Yaeger [12] has shown, that when optimising
an artificial neural network with a fixed number of neurons for neural complexity,
the fitness increases proportionally, suggesting a link between neural and functional
complexity. The more complex a network, the greater the likelihood that it will be
capable of solving complex tasks and surviving in complex environments [9–12].

35.3 Neuroscientific Foundations

Complexification in artificial neural networks can prove to be as important, as it

is in the development of natural neural systems. It is important in artificial devel-
opment to unleash fitness potential otherwise left untouched and constrained by a
fixed neural topology. Complexification in neural networks is a vital process in the
development of the brain in any natural system [14]. Complexification in human
brains happens in several different ways, by growth, by pruning and by reorganisa-
tion. The first form of complexification happens from before birth and goes on up to
adulthood, as the brain is formed. During this period neurons and interconnections
grow and hence complexifies the network. The second form of complexification
happens through continuous pruning. Connections between neurons have to be used
for them not to fade away and eventually possibly disappear. This concept is called
Neural Darwinism, as it is similar to normal evolution, where the fittest, in this case
connections, survive [15]. The third form of complexification happens through reor-
ganisation. In some cases, for yet unknown reasons, connections detach themselves
from a neuron and reconnects to another. Mostly, reorganisations in natural systems
have a detrimental effect, but some might have unexpected positive effects. The
effects of reorganisation in artificial systems are investigated in this chapter.

35.4 Reorganising Neural Networks

Artificial neural network can be reorganised in several different ways and to dif-
ferent extents. The methodology proposed herein operates with two degrees of
reorganisation. Reorganising one connection is defined as a minor reorganisation,
414 T.D. Jorgensen et al.

whereas reorganising more connections is defined as a major reorganisation. Net-

works in this chapter are only reorganised once, where the objective is to increase
the neural complexity of the network.

35.4.1 Neural Complexity

The neural complexity measure is an information-theoretic measure of the complex-

ity of the neural network and not a measure of the magnitude of weights or of the
number of elements in the network [11]. The neural complexity measure uses the
correlation between neuron output signals to quantify the integration and the spe-
cialisation of neural groups in a neural network. Complex systems are characterised
highly specialised clusters, which are highly integrated with each other. Systems that
have very highly independent functional components or have very highly integrated
clusters will have a low complexity. X is a neural system with n neurons, represented
by a connection matrix. The standard information entropy H(X) is used to calcu-
late the integration between components [16]. The integration between individual
neurons can be expressed by:

X
n
I.X/ D H.xi /  H.X/ (35.1)
iD1

The integration I(X) of segregated neural elements equals the difference between
the sum of entropies of all of the individual components xi of the neural network
and the entropy of the network as a whole. In order to be able to give an estimate
of the neural complexity of a neural network, not only the integration between the
individual elements is needed, but also the integration of any neural clusters in the
network. It is very likely that neurons in an artificial neural network cluster together
form some sort of functional cluster. The average integration between functionally
segregated neural groups with k (out of n) elements is expressed with <I.X/> j is
an index indicating that all possible combinations of subsets with k components are
used. The average integration for all subsets with k components is used to calculate
the neural complexity:

X
n
CN .X/ D Œ.k=n/  I.X/  <I.Xkj /> (35.2)
kD1

The neural complexity CN of a neural system X is the sum of differences between

the values of the average integration <I.X/> expected from a linear increase for
increasing subset size k and the actual discrete values observed. This neural com-
plexity measure yields an estimate of the information-theoretic complexity of a
neural network by measuring the integration between individual components and
possible combinations of subsets.
35 Reorganising Artificial Neural Network Topologies 415

35.4.2 Using the Complexity Measure

The neural complexity measure is used to optimise the complexity of the neu-
ral network. A reorganisation only takes place if this complexity increases. The
reorganisation methodology proposed is summarised by the following algorithm:
1. Determine a starting topology of sufficient size and complexity. This network
can be chosen randomly or based on the experience of the designer.
2. The starting network is trained to proficiency given some predefined measure.
3. The network is now reorganised. The number of connections to be reorganised
decides the degree of reorganisation. A connection is chosen at random, removed
and reinserted elsewhere in the network.
4. If this reorganisation increases the neural complexity of the network, the reorgan-
isation is deemed valid and the network is retrained. If the reorganisation does
not increase the neural complexity the reorganisation has been unsuccessful and
it is undone. Another reorganisation can be attempted or the process stopped. In
the experiments conducted here, five reorganisation attempts are made before the
process stops.
5. If it is desired and previous reorganisations have been successful further reorgan-
isations can take place.
Ideally it would be preferable to remove and reinsert the connection in the place
that yields the largest possible increase in the complexity out of all possible reor-
ganisations. This requires knowledge of all possible topologies given this number of
connections and neurons, which is not computationally efficacious. Only one con-
nection is reorganised at any time, this could be increased to several connections if
desired.

35.4.3 The Simulated Track and Robot

The controllers evolved here are tested in a simulated environment with a robot. In
this environment a robot has to drive around a track, which consists of 32 sections.
The objective of this task is to complete 3 laps in the shortest amount of time. If a
robot fails to complete three laps, the distance covered is the measure of its perfor-
mance. The robot has to drive around the track covering all of the sections of the
track, it is not allowed to skip any sections. In total the robot has to complete three
laps, with 32 sections in each lap, all visited in the correct order. If the robot is too
slow at driving between two sections the simulation is terminated. The following
Fig. 35.1, illustrates the task to be completed and the robot and its perception of the
track:
Figure 35.1-left illustrates the track and the robot driving around it. The robot is
not limited in its movement, i.e. it can drive off the track, reverse around the track
or adapt to any driving patterns desired, as long at its drives over the right counter
clockwise sequence of sections. Figure 35.1-right illustrates the robot driving on the
416 T.D. Jorgensen et al.

Start

Fig. 35.1 Left is the track and the robot driving around it; right is a top-view of the robot and its
sensory perception of the track

track seen from above. The track sections have alternating colours to mark a clear
distinction between sections. The arrows illustrate the three sensors of the robot. The
front sensor measures the distance to the next turn and the two side sensor measures
the distance to the edge of the track. As illustrated, the simulated robot has three
wheels and not four, to increase the difficulty of evolving a successful controller.
The risk when driving this three wheeled robot, in contrast to a four wheeled vehicle,
is that it will roll over if driven to abruptly. A robot that has rolled over is unlikely
to be able to continue. The front wheel controls the speed as well as the direction.

35.5 Experiments and Results

A total of three sets of experiments have been conducted. One set of experiments,
with a randomly selected neural network, acts as a benchmark for further compar-
isons. This network only has its connection weights evolved, whereas the topology is
fixed. The second set of experiments starts with the benchmark network selected in
the first experiment, which is then reorganised. This reorganisation is only minor, in
that only one connection is removed and replaced in each network. The new network
resulting from the reorganisation is tested. The third and final set of experiments also
uses the network from experiment one as a starting condition. The network is then
reorganised more extensively, by reshuffling several connections to create a new
network based on the old network. The results from these experiments are used to
compare the different strategies of evolution.
35 Reorganising Artificial Neural Network Topologies 417

35.5.1 The Simulation Environment

The evolved neural network controllers are tested in a physics simulator to mimic
a real world robot subject to real world forces. The genetic algorithm has in all
tests a population size of 50 and the number of tests per method is 25. Uniformly
distributed noise has been added on the input and output values to simulate sensor
drift, actuator response, wheel skid and other real world error parameters. To give
the simulations similar attributes and effects as on real racing track, the track has
been given edges. Whenever the robot drives off the track it falls off this edge onto
another slower surface. This means, that if the robot cuts corners, it could potentially
have wheels lifting off the ground thus affecting stability and speed of the robot, due
to the edge coming back onto the track.

35.5.2 The Fitness Function

Fitness is rewarded according to normal motorsport rules and practice. Three laps
of the track have to be completed and the controller that finishes in the fastest time
wins the race, i.e. it is the fittest controller. If a controller fails to finish three laps,
the controller with the most laps or longest distance travelled wins. In the case that
two controllers have reached the same distance the racing time determines the fittest
controller. The following fitness function states that the longest distance covered
in the shortest amount of time yields the best fitness. Time is the time it takes to
complete the track. If a controller fails to finish this Time is set to 480 s, which is the
absolute longest time a controller is allowed to be in existence, before a simulation
is stopped. In the likely event that two controllers have covered the same distance,
the controller with the fastest time will be favoured for further evolving. The precise
version of the fitness function is:
Sections D .Laps  Track Length/
Fitness D (35.3)
Time
The fitness is equal to the distance divided by the time. The distance is equal to
the number of track sections covered in the current lap, plus the number of sections
covered in previous laps. Track length is the total number of sections, which is 32.
The minimum fitness obtainable is 1=480 0:002.

35.5.3 The Benchmark and the Reorganised Networks

The first set of experiments was conducted with a fixed structure network, where
the connection weights were evolved, is used as the benchmark for the other experi-
ments. The network used is a feed-forward connected network with three input
418 T.D. Jorgensen et al.

Table 35.1 Results form experiments

Method Minimum Average Maximum Standard Complexity
fitness fitness fitness deviation
Benchmark network 0.857 0.964 1.042 0.062 14.71
Minor reorganisation 0.866 0.998 1.138 0.112 15.03
Major reorganisation 1.002 1.058 1.120 0.047 15.40

#1 #4

#7
ation
Benchmark Network
Reo rganis #2 #5
Minor
#8
Left sensor
#1 #4 #3 #6
Speed
#7
Right sensor
#2 #5
Steer
#8
Front sensor #1 #4
#3 #6 Major
Reorg
anisa
tion #7

#2 #5

#8
#3 #6

Fig. 35.2 The benchmark neural network and the reorganised networks

neurons, three hidden layer neurons and two output neurons. The inputs are the
sensor values as described previously and the outputs are the direction and speed of
the front wheel. This network was trained to competency and the results are shown
in Table 35.1. The neural complexity of this network is 14.71, calculated with Eq.
(35.2). The network is shown to left in Fig. 35.2.
The second set of experiments was conducted with a network that has been reor-
ganised. The benchmark network has undergone a minor reorganisation, which is
shown in Fig. 35.2 upper right. The connection between neuron 6 and neuron 8 has
been rearranged and is now a recursive connection. The new network has increased
its neural complexity by the reorganisation to 15.03. Immediately after the reorgan-
isation the network loses some of its fitness, this fitness is regained by retraining.
The reorganised network was retrained with the same weights as before the reor-
ganisation and in all cases the network, as a minimum, regained all of its previous
fitness and behaviour. Additionally, all of the connection weights were re-evolved
in another experiment to see if the results and tendencies were the same, and as
expected the results were the same.
35 Reorganising Artificial Neural Network Topologies 419

The final set of experiments conducted used a network, which has had a major
reorganisation. The benchmark network was changed by removing a connection
between neuron 3 and 5 and between neuron 1 and 6. These connections are moved
to between neuron 5 and 4 and between neuron 8 and 6. As the benchmark network
is feed-forward connected only recursive connections are possible for this particular
network. The new network is shown in Fig. 35.2 lower right. The neural complexity
of the new network has risen to 15.40, which is a 5% increase. Similar to the pre-
viously reorganised network, this network was, after a reorganisation, subject to a
fitness loss, but it was retrained to competency. The controller increased its fitness
over the original network. Even re-evolving all of the connection weights yields a
better overall performance.

35.5.4 Evaluation and Comparison of the Proposed Methods

The results from all of the experiments show that all networks learn the task pro-
ficiently, however some networks seem to perform better than others. Figure 35.3
shows the route that controllers choose to drive around the track. The route reflects
the results which are summarised in Table 35.1. The race car starts in (0, 0) and
drives to (20, 0) where it turns. Hereafter it continues to (20, 11) where it turns and
continues to (2:5, 11) and from here it continues to (2:5, 0) and on to (0, 0). The
controller tries to align the car on the straight line between the points. Figure 35.3
illustrates the difference between an average lap of the benchmark networks and
of the reorganised networks, and it clearly illustrates the route that the robot takes
around the track.
Figure 35.3 illustrates the degree of overshoot when turning and recovering to
drive straight ahead on another leg of the track. Figure 35.3 clearly shows that the
controllers that have been reorganised overshot less than the benchmark networks.
Less overshot, ultimately means that the racing car is able to move faster, which
means it has a better fitness. Table 35.1 shows the fitness from the benchmark
network experiments, and the fitness regained by the new networks after a reor-
ganisation and retraining. To put these results into context, the best human driver on
the same race track has a record of 1.036, which is worse than the average major
reorganised controller.
The hypothesis that artificial neural networks that have undergone a minor reor-
ganisation, where the neural complexity is optimised, are statistically better than the
fixed structure network it originates from does not hold true for these experiments.
A t-test, with a 5% significance level, indicates that there is no statistical differ-
ence between the two methods, despite the higher minimum, average and maximum
values. The second hypothesis tested in this chapter, states that artificial neural net-
works that have undergone a major reorganisation, where the neural complexity is
optimised, are better than the networks they originate from, holds true. A t-test,
with a 5% significance level, indicates that there is a statistical difference between
the two methods. This can be due to the increased neural complexity of the new
420 T.D. Jorgensen et al.

12 Reorganised
Benchmark
10

8
Y - Position

0
-5 0 5 10 15 20 25
-2
X - Position

Fig. 35.3 The route of an average benchmark network and a reorganised network

network created by the reorganisation. Some of this increased performance can pos-
sibly be accredited to the fact that one of the networks has a recursive connection,
which is a common way to increase the neural complexity and performance of a
network, but the experiments indicate that this is only part of the explanation. The
experiment clearly indicates that increased neural complexity yield a higher proba-
bility of finding suitable and well performing neural networks, which is in line with
other research in the field [9].
The results from the experiments don’t indicate any significant difference in the
speed of learning produced by either methodology. This means that it takes about
the same number of iterations to learn a given task for any network topology used
in the experiments, this was expected as they all have the number of parameters.

35.6 Conclusion

This chapter has presented a new methodology for complexifying artificial neural
networks through structural reorganisation. Connections were removed and rein-
serted whilst trying to increase the neural complexity of the network. The evolved
neural networks learned to control the vehicle around the track and the results indi-
cate the viability of the newly reorganised networks. The results also indicate that it
might be necessary to rearrange more than one connection in order to achieve sig-
nificantly better results. This chapter indicates that neural complexity in conjunction
with reorganisation can help unleash potential and increase the probability of find-
ing neural network controllers of sufficient complexity to adequately solve complex
tasks. Furthermore, the results indicate that a reorganisation can substitute structural
35 Reorganising Artificial Neural Network Topologies 421

elaboration as a method for improving network potential, whilst keeping the compu-
tational overhead constant. These results are in line with previous research done in
the field and they reconfirm the importance of high neural complexity and structural
change.

References

1. A.S. Weigend, D.E. Rumelhart, and B.A. Huberman, Back-propagation, weight-elimination

and time series prediction, Proceedings of the 1990 Summer School on Connectionist Models
(1990).
2. J. Denker, D. Schwartz, B. Wittner, S. Solla, R. Howard, L. Jackel, and J. Hopfield, Large
Automatic Learning, Rule Extraction and Generalization, Complex Systems 1(5): 877–922
(1987).
3. S. Nolfi, and D. Floreano, Evolutionary Robotics; The Biology, Intelligence, and Technology
of Self-Organizing Machines (Cambridge, MA: MIT Press 2000).
4. X. Yao, and Y. Liu, A New Evolutionary System for Evolving Artificial Neural Networks,
IEEE Transactions on Neural Networks 8(3): 694–713 (1997).
5. X. Yao, Evolving Artificial Neural Networks, Proceedings of the IEEE 87(9) (1999).
6. R. Jacobs, and M. Jordan, Adaptive Mixtures of Local Experts, Neural Computation 3: 79–87
(1991).
7. S.E. Fahlman, and C. Lebiere, The Cascade-Correlation Learning Architecture, Advances in
Neural Information Processing Systems 2: 524–532 (1990).
8. P. Angeline, and J. Pollack, Evolutionary Module Acquisition, Proceedings of the Second
Annual Conference on Evolutionary Programming (1993).
9. O. Sporns, and M. Lungarella, Evolving Coordinated Behaviours by Maximizing Infor-
mational Structure, Proceedings of the Tenth International Conference on Artificial Life
(2006).
10. K.O. Stanley, and R. Miikkulainen, Continual Coevolution Through Complexification, Pro-
ceedings of the Genetic and Evolutionary Conference (2002).
11. G. Tononi, O. Sporns, and G.M. Edelman, A Measure for Brain Complexity: Relating Func-
tional Segregation and Integration in the Nervous System, Proceedings of the National
Academy of Science of USA (1994).
12. L.S. Yaeger, and O. Sporns, Evolution of Neural Structure and Complexity in a Computational
Ecology, Proceedings of the Tenth International Conference on Artificial Life (2006).
13. R. Dawkins, Climbing Mount Improbable (Reprint by Penguin Books, London, England,
2006).
14. G.N. Martin, Human Neuropsychology (London: Prentice Hall, 1998, Reprinted 1999).
15. G. Edelman, Neural Darwinism – The Theory of Neuronal Group Selection (New York: Basic
Books, 1989, Print by Oxford Press 1990).
16. C.E. Shannon, A Mathematical Theory of Communication, The Bell System Technical Journal
27: 379–423/623–656 (1948).
Chapter 36
Design and Implementation of an E-Learning
Model by Considering Learner’s Personality
and Emotions

S. Fatahi, M. Kazemifard, and N. Ghasem-Aghaee

Abstract Emotion, personality and individual differences are those effective param-
eters on human’s activities such as learning. People with different personalities show
different emotions in facing an event. In the case of teaching and learning, personal-
ity difference between learners plays an important role. In virtual learning projects
this point should consider that the learners’ personalities are various and the teach-
ing method used for each learner should be different from the other learners. In this
chapter, a new model presented according to the learning model based on emotion
and personality and the model of virtual classmate. Based on their knowledge base,
the virtual teacher and classmate express suitable behaviors to improve the process
of learning according to the learner’s emotional status.

Keywords E-Learning  Learners Personality  Emotion  Implementation  virtual

classmate

36.1 Introduction

E-learning is usually defined as a type of learning supported by information and

communication technology (ICT) that improves quality of teaching and learning.
E-learning system is a powerful tool for achieving strategic objectives of the univer-
sity (teaching, research and serving the society) [1]. E-Learning like all other tools
offers advantages such as: access to differentiated online resources, Self-directed
learning, and Learning matches learners’ lifestyles, etc. Despite of all the advan-
tages this kind of learning lacks the necessary attractiveness most of the time. It
seems that regarding the human characteristics and inserting them in virtual learning
environments, it would be possible to show these environments more real.

S. Fatahi (B)
Department of Computer Engineering, University of Isfahan, Isfahan, Iran
E-mail: fatahi somayeh@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 423
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
424 S. Fatahi et al.

Emotion, personality and individual differences are those effective parameters on

human’s activities such as learning. People with different personalities show differ-
ent emotions in facing an event. Difference in the characteristics of the individuals
is reflected in their daily activities and their works. In the case of teaching and learn-
ing, personality difference between learners plays an important role. The learner’s
personality will be effective in his learning style [2]. In virtual learning projects this
point should consider that the learners’ personalities are various and the teaching
method used for each learner should be different from the other learners.

36.2 Previous Works

In virtual learning systems created up to now, the learner’s emotions received much
more attention and the emotional agents were more employed. In a few of these
systems personality drew our attention as an independent parameter that some of
them are mentioned here:
In ERPA architecture by using ID3 algorithm, the learner’s emotional reaction
towards an event is predicted (for example, taking an exam score) [3]. Chaffar
and his colleagues used the Naı̈ve Bayes Classifier method to predict the learner’s
emotions [4]. In ESTEL architecture, the Naı̈ve Bayes Classifier method is used to
predict the optimized emotional status. In this architecture, in addition to emotion,
the learner’s personality is also considered. In this system, a module tries to create
and induce an optimized emotional state. For instance, when the learner enters the
system, after the identification of learner’s personality, for example extrovert, and
recognition of optimal emotional state, such as happiness, an emotion is induced to
that learner by showing various interfaces (e.g. music, picture, and etc.) to him [5].
In Passenger software designed by German researchers, cooperative learning meth-
ods are used. This software examines a series of emotions for the virtual teacher
that is present in the system based on OCC model, and invites the learners to group
work. The virtual teacher traces the learners’ activities and helps the learners who
are not able to do cooperative activities [6]. Abrahamian and his colleagues designed
an interface for computer learners appropriate for the type of their personality using
MBTI test and concluded that learning through this interface as a result of using
personality characters leads into developments in learning process [7]. In imple-
mentation performed by Maldonado and his colleagues, a virtual classmate agent
is used. This agent is placed beside the learners, and mostly plays the role of a co-
learner and a support. In this project each of the teacher, learner, and classmate has
own emotions and the learner’s emotions affected his/her classmate [8].

36.3 Psychological Principles

Emotion, personality and individual differences are those effective parameters on

human’s activities. Everybody needs special learning style according to his/her per-
sonality characteristics. Some tools are used to evaluate the different learning style
36 Design and Implementation of an E-Learning Model 425

to determine the learner’s learning style. MBTI is the well-known questionnaire

used for personality and learning style determination [9, 10]. According to MBTI
classification, every individual has a set of instinctive preferences that determine
how he or she behaves in different situations [9]. This questionnaire helps to iden-
tify personality characteristics and learning preferences of the individuals and to
elicit the suitable learning styles from these characteristics [11].
MBTI uses four two-dimensional functions according to the Jung’s theory.
According to the theory that Jung proposed the four functions of mind are thought,
emotion, comprehension, and exploration. These functions are the main ways for
understanding, and they explain the truth. These functions are related to each other
and they simulate one another. Nevertheless, one of the functions is often dom-
inant and that dominant function inclines the person to a state. The Jung theory
distinguished three dimensions of Extroversion/Introversion (E/I), Sensing/Intuition
(S/N) and Thinking/Feeling (T/F), but in MBTI another dimension of, Judging/
Perceiving (J/P) also was added [7, 11]. Irrational mental functions, Sensing (S)
or Intuition (N), relate to how an individual perceives information, while ratio-
nal mental functions, Thinking (T) or Feeling (F), provide insight into how one
makes judgments or decisions based upon their perceptions. Mental functions of
Extrovert/Introvert and Judgment/Perception are related to how individuals interact
with the external environment and the around world. Sixteen personality types are
resulted from mixing these 4 two-dimensional functions that each learner would be
placed in one group [11, 12].

36.4 Proposed Model

In this paper, a new model presented according to the learning model based on
emotion and personality [13] and the model of virtual classmate [14] in previous
our studies. The outline of the improved model is shown in Fig. 36.1. The model
contains six major modules:
Personality identification module: In first step, learner comes across MBTI ques-
tionnaire and his personality will be identified (for example ISFJ, ESTP, INTJ,
etc.).
Module of choosing a learning style commensurate with learner’s personality: Gen-
erally, there are three kinds of learning environment: individual, competitive and
collaborative [15]. System based on the identified personality of learner, put him
in one of three groups of independence, contribution with virtual classmate or
competition with virtual classmate [16].
Module of choosing virtual classmate agent: If the learner is put in the independence
group the process of learning and education will be started, otherwise the system at
first chooses a virtual classmate that matches the type of learner’s personality, then
the process of learning and education will get started. This module will explain at
the next section with more details.
426 S. Fatahi et al.

Fig. 36.1 Proposed model

Education module: In this module, lesson’s points are presented to learner as

exercises.
Module of emotion evaluation: When doing exercises and evaluating the extend
of learning, there are some emotion expressed in learner which are relevant to
level of learner’s learning and the events happen in the environment (as have lik-
ing for virtual classmate, be disappointed in doing exercises, etc.). According to
the performed studies, we found out only special emotions are effective in the
process of learning [17, 18]. Accordingly, the first and the third branch of emo-
tions in OCC model are used. The first branch of emotions in OCC model are the
effective emotions in process of learning and the third branch are those emotions
that a person shows them when facing the others (for example virtual classmate
agent).
Module of choosing behavior and updating the learning style: The module changes
the style of education according to the events happen in the environment that cause
changes to the learner’s emotion and also the learner’s personality characteristics.
Based on their knowledge base, the virtual teacher and classmate express suitable
behaviors to improve the process of learning according to the learner’s emotional
status.
36 Design and Implementation of an E-Learning Model 427

Fig. 36.2 Module of choosing virtual classmate agent

36.5 Virtual Classmate Model

In this section we explained module of choosing virtual classmate agent with more
details. This module is displayed in Fig. 36.2. The module includes three main parts,
each of them described below:
The part of personality generation: In this part, using the MBTI questionnaire, the
personality of the learner is recognized. In this paper we only considered two dimen-
sions of E/I and S/N which are important in learning process [7]. Considering two
dimensions, four types of personality that are EI, EN, IS, and IN would be resulted.
The part of classmate selection: In this part, a VCA appropriate for the learner’s
personality is selected. Selected VCA is completely opposite in his MBTI dom-
inant with learner. Based on research, the opposite personality displays a higher
performance than the similar personality [19–21]. The personality selection for the
VCA is so that it would result in improvements in learning process.
The part of classmate behavior selection: During the learning process, regarding the
events that happen in the environment and the learning situation of the individual,
the VCA exhibits appropriate behaviors. Tactics knowledgebase is used to interact
with the learner.
For two personality dimensions considered in this paper four parameters are
elicited. Independence and replying speed parameters for the E/I dimension and
detail-oriented attitude and attention for the S/N dimension, now based on the extent
of these parameters in each type of personality, the VCA exhibits a certain behav-
ior. These behaviors are shown separately for each dimension in Table 36.1 and
Table 36.2. Mixing the two dimensions’ functions, four personality types would
be resulted, that are IN, IS, EN, and ES. A sample of these tactics is presented in
Table 36.3.
 428

Table 36.1 The VCA behavior with E and I personality dimensions

Learner’s VCA’s Characteristics Independence Characteristic Replying speed Events and VCA
personality parameter parameter tactics (solving
problem)
I E The introvert person The E VCA decreases The introvert person The E VCA tries to 1. The E VCA tries to
mostly acts the independence takes a lot of time increase the speed cooperate with I 2.
independently and of I solving the of I The E VCA tries to
is inclined to do the problem, the activate I by
exercises alone. The extrovert person announcing the
extrovert person is mostly acts without remaining time so
interested in group thinking and replies that I answer the
work and rarely acts questions sooner
lonely
E I The I VCA increases The I VCA tries to 1. The I VCA do not
the independence decrease the speed cooperate with E so
of E of E that he act
independently 2.
The E VCA tries to
make I more relaxed
so that he thinks
more on problems
S. Fatahi et al.
Table 36.2 The VCA behavior with N and S personality dimensions
Learner’s VCA’s Characteristics detail-oriented Characteristic Attention Events and VCA
personality attitude parameter parameter tactics (solving
problem)
S N The S person mostly The N VCA tries to The N person never The N VCA pays 1. The N VCA tries to
pays attention to explain the relation pays attention to attention to S and shed some light on
details, while N between the your words and helps him if the relation
pays attention to problems for S predicts the words necessary between problems
the relation in the you want to say. for S, clarify the
problems, and the The S person if problem to S and
result is important there was any need explains the
to him to ask for help asks general points to
N and do not react the learner. 2. The
36 Design and Implementation of an E-Learning Model

negatively to his N VCA helps and

mistaken views cooperates with the
learner if he asks
for that
N S The S VCA tries to The S VCA tries to 1. The S VCA reminds
decrease the attract the N’s the details to N 2.
amount of attention to himself The S VCA tries to
attention N pays to convey his views
the details to N
429
430 S. Fatahi et al.

Table 36.3 VCA tactics

Learner’s VCA’s personality Events and VCA tactics (solving
personality problem)
IS EN 1. The E VCA tries to cooperate with I
2. The E VCA tries to activate I by
announcing the remaining time so that
I answer the questions sooner
3. The N VCA tries to shed some light on
the relation between problems and
explains the general points of the
exercise for S
4. If the learner asks for help, the N VCA
helps and cooperates with him

36.6 Simulation of Proposed Model

In our educational environment two agents – VTA and VCA – are used. These two
agents with identifying the learner’s emotions after each event, choose suitable tac-
tics when are face to face with the learner. In this environment the way of choosing
the tactics to face the learner is related to the learner’s type of group. Depending on
the point that which groups the learner belongs to, special tactic is chosen.
Knowledge base of this system contains 65 rules (Table 36.4): 16 rules are to
identify the learner’s group, 10 rules for independent learning group, 20 rules for
collaborative learning group and 19 rules for competitive group. Four examples of
these rules are in the following:
The first rule is an example of learner’s classifying into learning groups. Accord-
ing to the first rule, system groups the learner with ISFJ personality that is “Intro-
vert”, “Sensing”, “Feeling” and “Judging” groups in independent group. The second
rule is an example of the rules of teacher’s dealing in a situation that the learner is
in independent group. The third and fourth rules are examples of situations that the
learner is in collaborative and competitive groups respectively. As the rules shows
in these two situations relevant to the learner’s emotions the virtual teacher and
classmate use special tactics in interaction with learner.

36.7 Implementation

For implementing the e-learning environment, a series of English language exercises

were used. In this environment it was assumed that the learner has already learnt
the subject and refers to this software for solving the exercises. The exercises are
categorized in five levels by difficulties.
The learner begins to solve the exercises together with the VCA, and regarding
the events that happen in the environment interacts with the VCA. An image of this
environment is presented in Fig. 36.3. Visual C#.Net and SQL Server database were
36 Design and Implementation of an E-Learning Model 431

Table 36.4 Rules

Rule 1:
If student S1 has personality ISFJ
Then his/her group is independent
Rule 2:
IF Student Group IS Independent
AND Satisfaction IS High
OR Satisfaction IS Medium
AND Event IS Correct answer
THEN Teacher tactic1 IS Congratulate student
AND Teacher tactic2 IS Change student group to
competitive
Rule 3:
IF Student Group IS Collaborative
AND Like IS High
OR Like IS Medium
AND Disappointment IS High
OR Disappointment IS Medium
AND Event IS Wrong answer
THEN Classmate tactic1 IS Increase student self
ability
AND Classmate tactic2 IS Increase student effort
AND Classmate tactic3 IS Persuade student to
think more for
problem
Rule 4:
IF Student group IS Competitive
AND Like IS High
OR Like IS Medium
AND Fear IS High
OR Fear IS Medium
AND Virtual IS EN
classmate’s
personality
OR Virtual IS ES
classmate’s
personality
AND Event IS While student is
accomplishing to a
task
AND Student’s response IS Lower than threshold
speed
THEN Classmate tactic1 IS Increase student effort
AND Classmate tactic2 IS Notify student for
deadline
AND Teacher tactic1 IS Increase student self
ability
AND Teacher tactic2 IS Change student group to
collaborative
AND Classmate tactic2 IS Notify student for
deadline
432 S. Fatahi et al.

Fig. 36.3 Educational environment with classmate and teacher agent

Fig. 36.4 Learner satisfaction of learning environment

used for implementing. Also, to show the agents – VTA and VCA – we used two
present agents in Microsoft that are Merlin and Peedy respectively.

36.8 Results

We tested our model in real environment with 30 students. Students work with
our system, then they answered ten questions for evaluating our learning environ-
ment. We got rate of learner satisfaction Based on four question of our questioner
(Fig. 36.4).
36 Design and Implementation of an E-Learning Model 433

Fig. 36.5 Presence effect of VCA in learning environment

Results show that learners are satisfied of learning environment based on learner’s
emotion and personality.
We got rate of presence effect of VCA Based on six question of our questioner
(Fig. 36.5).
The results show that the presence of the VCA leads advancements in the learning
process and attractiveness of e-learning environment.

36.9 Conclusion and Future Works

In this paper a model for using in e-learning was presented. In this model some mod-
ules for personality recognition and selecting an appropriate VCA for the learner’s
personality, were considered to develop the interaction with the learner. The Behav-
ior of VCA saved in knowledgebase of system. The results show that placing the
learner beside an appropriate VCA, lead to improvement in learning and makes the
virtual learning environment more enjoyable.
In the future we will try to improve the system with considering the parameters of
culture, case based reasoning and agent’s learning and also makes the virtual teacher
and classmate agents more credible for the user.

References

1. N. Begičević, B. Divjak, and T. Hunjak, Decision Making Model for Strategic Planning of
e-Learning Implementation (17th International Conference on Information and Intelligent
Systems IIS, Varaždin, Croatia, 2006).
2. J. Du, Q. Zheng, H. Li, and W. Yuan, The Research of Mining Association Rules Between
Personality and Behavior of Learner Under Web-Based Learning Environment (ICWL, 2005),
pp. 406–417.
434 S. Fatahi et al.

3. P. Chalfoun, S. Chaffar, and C. Frasson, Predicting the Emotional Reaction of the Learner
with a Machine Learning Technique (Workshop on Motivaional and Affective Issues in ITS,
International Conference on Intelligent Tutoring Systems, Jhongli, Taiwan, 2006).
4. S. Chaffar, G. Cepeda, and C. Frasson, Predicting the Learner’s Emotional Reaction Towards
the Tutor’s Intervention (7th IEEE International Conference, Japan, 2007), pp. 639–641.
5. S. Chaffar and C. Frasson, Inducing Optimal Emotional State for Learning in Intelligent
Tutoring Systems (lecture notes in computer science, 2004), pp. 45–54.
6. B.F. Marin, A. Hunger, and S. Werner, Corroborating Emotion Theory with Role Theory
and Agent Technology: A Framework for Designing Emotional Agents as Tutoring Entities,
Journal of Networks (1), 29–40 (2006).
7. E. Abrahamian, J. Weinberg, M. Grady, and C. Michael Stanton, The Effect of Personality-
Aware Computer-Human Interfaces on Learning, Journal of Universal Computer Science (10),
27–37 (2004).
8. H. Maldonado, J.R. Lee, S. Brave, C. Nass, H. Nakajima, R. Yamada, K. Iwamura, and
Y. Morishima, We Learn Better Together: Enhancing e-Learning with Emotional Characters
(In Proceedings, Computer Supported Collaborative Learning, Taipei, Taiwan, 2005).
9. D.J. Pittenger, Measuring the MBTI . . . and Coming Up short, Journal of Career Planning and
Employment (54), 48–53 (1993).
10. M.D. Shermis, and D. Lombard, Effects of computer-based Test Administration on Test
Anxiety and Performance, Journal of Computers in Human Behavior (14), 111–123 (1998).
11. S. Rushton, J. Morgan, and M. Richard, Teacher’s. Myers-Briggs Personality Profiles: Identi-
fying Effective Teacher Personality Traits, Journal of -Teaching and Teacher Education (23),
432–441 (2007).
12. S.A. Jessee, P.N. Neill, and R.O. Dosch, Matching Student Personality Types and Learning
Preferences to Teaching Methodologies, Journal of Dental Education (70), 644–651 (2006).
13. N. Ghasem-Aghaee, S. Fatahi, and T.I. Ören, Agents with Personality and Emotional Fil-
ters for an e-Learning Environment (Proceedings of Spring Agent Directed Simulation
Conference, Ottawa, Canada, 2008).
14. S. Fatahi, N. Ghasem-Aghaee, and M. Kazemifard, Design an Expert System for Virtual
Classmate Agent (VCA) (Proceedings of World Congress Engineering, UK, London, 2008),
pp. 102–106.
15. S. Ellis, and S. Whalen, Cooperative Learning: Getting Started, Scholastic, New York (1996).
16. http://www.murraystate.edu
17. B. Kort, R. Reilly, and R.W. Picard, An Affective Model of Interplay between Emotions
and Learning: Reengineering Educational Pedagogy Building a Learning Companion (Pro-
ceedings IEEE International Conference on Advanced Learning Technology, Madison, 2001),
pp. 43–48.
18. S.M. Al Masum, and M. Ishizuka, An Affective Role Model of Software Agent for Effective
Agent-Based e-Learning by Interplaying between Emotions and Learning (WEBIST, USA,
2005), pp. 449–456.
19. K.S. Choi, F.P. Deek, and I. Im, Exploring the Underlying of Pair Programming: The
Impact of Personality, Journal of Information and Software Technology, doi: 10.1016/
j.infsof.2007.11.002 (2007).
20. L.F. Capretz, Implications of MBTI in Software Engineering Education, ACM SIGCSE
Bulletin – Inroads, ACM Press, New York, vol. 34, pp. 134–137 (2002).
21. A.R. Peslak, The impact of Personality on Information Technology Team Projects (Pro-
ceedings of the 2006 ACM SIGMIS CPR Conference on Computer Personnel Research:
Forty Four Years of Computer Personnel Research: Achievements, Challenges & the Future,
Claremont, California, USA, 2006).
Chapter 37
A Self Progressing Fuzzy Rule-Based System
for Optimizing and Predicting
Machining Process

Asif Iqbal and Naeem U. Dar

Abstract Researchers have put forward variety of knowledge acquisition methods

for automatic learning of the rule-based systems. It has been suggested that two
basic anomalies of rule-based system are incompleteness and incorrectness. In the
chapter, a fuzzy rule-based system has been presented that not only self-learns and
self-corrects but also self-expands. The chapter moves ahead with description of
the configuration of the system, followed by explanation of the methodology that
consists of algorithms for different modules. At the end, the operation of the self-
progressing fuzzy rule-based system is explained with the help of examples related
to optimization of machining process.

Keywords Fuzzy Rule-Based System  Self Progressing  Machining Process 

Optimizing

37.1 Introduction

Rule-based systems represent the earliest and most established type of AI systems
that tend to embody the knowledge of a human expert in a computer program [1].
Though, rule-based systems have been quite effectively utilized in approximat-
ing complicated systems – for which analytical and mathematical models are not
available – with considerable degree of accuracy, they still pose a very static pic-
ture. The foremost shortcoming, in their design, is the lack of dynamism. The price
paid by this inadequacy is their failure in coping with fast changing environments.
This describes the main reason of inability of rule-based system technology to find
its full fledge application at industrial levels.

A. Iqbal (B)
Department of Mechanical Engineering, University of Engineering & Technology,
Taxila, Pakistan
E-mail: asif.asifiqbal@gmail.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 435
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
436 A. Iqbal, N. U. Dar

It is predominantly important to keep the rule-base (or knowledge-base in gen-

eral) updated and corrected all the time in order to maintain the efficacy and viability
of the system. Obviously, it is not prudent to engage the services of a knowledge
engineer for this purpose, as the efficacy calls for rapid upgrading of the rule-
base. The dire need is to find the means for automatic accomplishment of this
requirement.
Most of the time required for the development of a rule-based system can be
attributed to the knowledge acquisition phase [2]. Researchers have put forward
variety of knowledge acquisition methods for automatic learning of the rule-based
systems. It has been suggested that two basic anomalies of rule-based system are
incompleteness and incorrectness [3]. It was further proposed that by integrating
machine learning (ML) techniques in the validation step of the rule-based systems’
evolutionary life cycle model, the knowledge-base can be refined and corrected
throughout a refinement cycle [3]. In other article [4] the authors presented an adap-
tive fuzzy learning algorithm, used for nonlinear system identification that provided
a new way for representing the consequent part of the production rule. In other
work [5], a knowledge factory has been proposed that allows the domain expert
to collaborate directly with ML system without needing assistance of a knowledge
engineer. In other paper [2] the authors have presented an inductive learning algo-
rithm that generates a minimal set of fuzzy IF-THEN rules from a set of examples. In
another work [6], the authors have presented a self-testing and self-learning expert
system, which is based on fuzzy certainty factor and it checks the rule-base for
correctness and completeness.
Few papers can be found that describe the application of machine learning in
manufacturing domain. In a paper [7] researchers have presented a fuzzy expert
system for design of machining conditions dedicated to the turning process. The
learning module contained by the system corrects the empirical relationships by
changing the fuzzy membership functions. Another paper [8] has presented an
approach for building knowledge-base from numerical data, which has proved to
be useful for classification purposes. In other article [9] the authors have utilized
Support Vector Regression, a statistical learning technique, to diagnose the condi-
tion of tool during a milling process. Finally, in another paper [10], the authors have
presented the comparison of three ML algorithms for the purpose of selection of the
right dispatching rule for each particular situation arising in a flexible manufacturing
system.
In the current research work a fuzzy rule-based system has been presented
that not only self-learns and self-corrects but also self-expands. Following are the
distinguishing features of the self-progressing fuzzy rule-based system:
1. For any physical process, it predicts the values of response (output) variables
based on the values of predictor (input) variables
2. Suggests the best values of predictor variables that maximize and/or minimize
the values of selected set of response variables
3. Automatically adjusts newly entered variable at any stage of development
4. Learns and corrects itself according to new set of data provided
37 A Self Progressing Fuzzy Rule-Based System 437

5. Automatically generates fuzzy sets for newly entered variables and regenerates
sets for other variables according to newly added data
6. Automatically generates rules for optimization and prediction rule-bases
7. Provides conflict resolution facility among contradictory rules
8. Updates interface of the system according to newly entered variables
The first two points depict the main objectives of the rule-based system, in con-
nection with any manufacturing process, while the others portray a picture of high
level automation required for the system to self-progress. The chapter moves ahead
with description of the configuration of the system, followed by explanation of the
methodology that consists of algorithms for different modules. At the end, the oper-
ation of the self-progressing fuzzy rule-based system is explained with the help of
examples related to optimization of machining process.

37.2 System Configuration

Figure 37.1 describes the configuration of the self-progressing rule-based system in

graphical form. The main purpose of the system is to self-generate the rule-bases,
from the experimental data, that could be used to: (1) optimize the settings of pre-
dictor variables (of any physical process) for the purpose of maximization and/or
minimization of set of selected response variables; (2) predict the values of response
variables according to the finalized settings of the predictor variables.
The pattern of the system consists of four parts: the knowledge-base, the shell,
the inference engine, and the self-development mode. The knowledge-base is the

Shell
Knowledge-Base
User
Facts User Interface

Optimization Prediction Forword

Rule-Base Rule-Base Chaining Data
Inference Fuzzifier
Engine

Self-Development Mode Working

Memory
Optimization Rule-Base Prediction Rule-Base
Development Module Development Module
Data
Defuzzifier
Data Acquisition Module Conflict Resolution
Module
Interface Recommendations
Fuzzy Sets Development Module & Predictions
Development

DoE Based Experimental Data

Fig. 37.1 The configuration of the self-progressing rule-based system

438 A. Iqbal, N. U. Dar

combination of facts, the optimization rule-base, and the prediction rule-base. The
functional details of optimization and prediction rule-bases are also available in ref-
erence [11]. The shell consists of a user interface through which the input is taken
from the user. The data fuzzifier fuzzifies the values of numeric predictor variables
according to the relevant fuzzy sets. The user-interface and the fuzzy sets are also
auto-developed by the system itself. For development of the knowledge-base con-
sisting of two rule-bases, the inference mechanism of forward chaining shell, Fuzzy
CLIPS (C Language Integrated Production Systems) was used [12].

37.3 The Self-Development Mode

The most important and distinguishing constituent of the system is the

self-development mode. This mode consists of following four modules: data acqui-
sition module; fuzzy sets development module; optimization rule-base development
module; and prediction rule-base development module integrated with the conflict
resolution sub-module. The detail is as follows.

37.3.1 Data Acquisition

This module facilitates the automation of intake, storage, and retrieval of data.
The data could be the specifications of a new variable or the values of input
and output variables obtained from experiments. Data related to specifications of
a new variable are stored in file Variable.dat, while that related to the new val-
ues/records are stored in Data.dat on hard disk. Figure 37.2 shows the flow chart
of data acquisition algorithm.

Fig. 37.2 The flow chart of the data acquisition module

37 A Self Progressing Fuzzy Rule-Based System 439

Membership
function
Low High

0
0 10 20 30 40 50 60 70 80 90 100
Desirability (%)

Fig. 37.3 Fuzzy sets for maximizing/minimizing output variable

37.3.2 Self-Development of Fuzzy Sets

This module covers three areas: (1) Rearranging the fuzzy sets for already entered
variables according to the newly entered data records; (2) development of fuzzy
sets for newly entered numeric variables; and (3) development of two fuzzy sets
(low & high) for each output variable that is included for optimization purpose.
The set low represents the minimization requirement and the other one represents
maximization. The design of sets for category (3) is fixed and is shown in Fig. 37.3,
while design of first two categories is dynamic and based upon the data values of
respective variables.
The desirability values shown in Fig. 37.3 are set by the user using the slider bar
available on interface of the rule-based system. Any value below 5% means desir-
ability is of totally minimizing the output variable (performance measure), and total
desirability of maximization is meant if the value is above 95%. The desirability of
50% means optimization of that output variable makes no difference.
Figure 37.4 shows the customized flow chart for the methodology used for self-
development of fuzzy sets. The user has to decide the maximum allowable number
of fuzzy sets for input as well as for output variables.
The logic involved in methodology is that a value of input variable, which has
higher frequency of appearance in the data records, has more right to be picked
up for allocation of a fuzzy set, while for output variable any value having greater
difference from its previous and next values in the list – termed as Neighbor Dis-
tance (Neighb dist in Fig. 37.4) – possesses more right for allocation of a fuzzy set.
Neighbor Distance can mathematically be represented as follows:
8
< ValueŒi C 1  ValueŒi I if .i D first/
Neighbor Distance D ValueŒi   ValueŒi  1I if .i D last/
:1
=2 .ValueŒi C 1  ValueŒi  1/ I otherwise
(37.1)

37.3.3 Self-Development of Prediction Rule-Base

This step consists of following two parts: (1) automatic development of rules, for
prediction of the manufacturing process’s performance measures, based on the data
440 A. Iqbal, N. U. Dar

Fig. 37.4 Customized flow chart for self-development of fuzzy sets

records provided by the users; and (2) conflict resolution among self-developed
contradictory rules.
Figure 37.5 provides the graphical description of the first part. The objective is
to convert each node of 2-D linked list Data output (including list of related values
of input variables Data input) into a rule. 2-D linked list is a list that expands in two
directions as shown in Fig. 37.5. The objective is achieved by finding and assigning
the most suitable fuzzy sets for all of the values involved per node of Data output.
The list Data output is navigated from first node to last and for all of its values
the closest values in fuzzy sets of respective variables are matched. If the match is
perfect then certainty factor (CF) of 1 is assigned to the match of the data value
and the fuzzy set. If the suitable match of any fuzzy set for a given data value is
not found then the data value is assigned the intersection of two closest fuzzy sets.
All the rules are stored in a 2-D linked list, named Rule Consequent, each node of
whose represents a rule. Each node contains the assigned fuzzy set of the output
variable and also a linked list (Rule antecedent) containing assigned fuzzy sets of
all the relevant input variables.
37 A Self Progressing Fuzzy Rule-Based System 441

Fig. 37.5 Framework for self-development of prediction rule-base

37.3.3.1 Conflict Resolution among Contradictory Rules

There is always a possibility that some anomalous data might be entered by the user
that could lead to self-development of some opposing rules. So it is necessary to
develop a mechanism that would detect such possible conflict and provide a way for
its resolution.
The mechanism of conflict resolution can be described as follows: Compare each
and every rule of the prediction rule-base to all the other rules of the same rule-base.
If, in the consequent parts of any two rules, following two conditions satisfy: (1) out-
put variables are same; and (2) assigned fuzzy sets are different, then check whether
the antecedent parts of both the rules are same (i.e., same input variables with same
fuzzy sets assigned). If yes, then these two rules form a pair of contradictory rules.
Next the user is inquired regarding which one of the two contradictory rules needs
to be abandoned. The CF value of the rule to be abandoned is set to zero. The same
procedure is continued for whole of the rule-base. At the completion of the process,
all the rules possessing the CF values greater than zero are printed to the CLIPS file.
442 A. Iqbal, N. U. Dar

Fig. 37.6 Framework for self-development of optimization rule-base

37.3.4 Self-Development of Optimization Rule-Base

This module generates a set of rules that is responsible for providing the opti-
mal settings of input variables that would best satisfy the maximization and/or
minimization of the selected output variables. Figure 37.6 describes the framework.
The idea exploited in development of this module is that for maximization of any
output variable select an ideal fuzzy set for each numeric input variable, which, on
average, would generate the maximum value of that output variable. For the mini-
mization purpose, select those fuzzy sets for respective input variables that would
result in the least possible value of the output variable, available in the data records.

37.4 Optimization and Prediction of Machining Process

In order to demonstrate the versatility and self-progressing capabilities of the

presented system, two application examples are presented.
Machining is the most wide-spread of all the manufacturing processes and
amount of investment being done in it is an indication of wealth of the nations
[13]. For demonstration of applicability of the self-progressing rule-based system
in optimization and prediction of the processes included in the machining domain,
the process of milling (a machining process in which tool rotates and workpiece
remains stationary) has been chosen. Table 37.1 presents a set of limited data related
37 A Self Progressing Fuzzy Rule-Based System 443

Table 37.1 Limited experimental data for rookie knowledge-base

No. Predictor variables Response variables
Speed (m/min) Rake ( ) o
Orientation Tool life (mm2 ) Cutting force (N)
1 150 8 Up 3,601 910
2 150 8 Down 5,500 643
3 150 5 Up 3,583 885
4 150 5 Down 4,912 532
5 250 8 Up 2,991 1,008
6 250 8 Down 4,004 769
7 250 5 Up 2,672 986
8 250 5 Down 3,609 704

1 1
Memb.

Memb.
func.

func.
S1 S2 S1 S2

0 0
100 150 200 250 300 -14.5 -7.5 -0.5 -6.5
Speed (m/min) Rake (degree)

1
Memb.
func.

S1 S2 S3 S4 S5 S6

0
2200 2900 3600 4300 5000 5700
Tool life (mm2)

1
Memb.
func.

S1 S2 S3 S4 S5 S6 S7 S8

0
472 572 672 772 872 972

Force (N)

Fig. 37.7 Self-developed fuzzy sets for numeric variables

to milling process. The first three variables, namely: speed, rake, and orientation
(milling-orientation) are predictor (input) ones, while the other two, tool life and
cutting force are response (output) variables. If the knowledge-base is developed
based entirely on the data presented in the table, it is very likely that the sys-
tem may provide anomalous results because of the fact that the other influential
milling parameters have not been taken into account, and thus, the self-progressed
knowledge-base can be termed as “Rookie Knowledge-Base”.
Suppose the system is asked to develop its rule-bases and update its interface
based on the data provided and it is also asked to include tool life, but not the cut-
ting force, as output variable for optimization. Figure 37.7 shows the detail of the
triangular fuzzy sets of numeric variables, developed itself by the rule-based system,
in addition to the sets of the objective, as shown in Fig. 37.3. Following is the detail
of the six rules, self-generated by the self-development mode and to be operated by
the optimization module of the rule-based system:
444 A. Iqbal, N. U. Dar

Table 37.2 Self-generated prediction rule-base

Rule No. Antecedents Consequents
Speed Rake Orientation Tool life CF Force CF
1 S1 S1 Up S3 0:88 S6 1
2 S1 S1 Down S6 1 S2 1
3 S1 S2 Up S3 1 S5 1
4 S1 S2 Down S5 1 S1 1
5 S2 S1 Up S2 1 S8 1
6 S2 S1 Down S4 1 S4 1
7 S2 S2 Up S1 1 S7 1
8 S2 S2 Down S3 0:82 S3 1

Rule 1: IF Objective Tool life is High AND Speed is not fixed THEN Speed is S1.
Rule 2: IF Objective Tool life is High AND Rake is not fixed THEN Rake is S1.
Rule 3: IF Objective Tool life is High AND Orientation is not fixed THEN
Orientation is Down.
Rule 4: IF Objective Tool life is Low AND Speed is not fixed THEN Speed is S2.
Rule 5: IF Objective Tool life is Low AND Rake is not fixed THEN Rake is S2.
Rule 6: IF Objective Tool life is Low AND Orientation is not fixed THEN
Orientation is Up.
Out of these six rules the first three perform the maximization operation, while
the others perform minimization. Table 37.2 presents the detail of eight rules, self-
generated by the rule-based system and to be operated by its prediction module.
Figure 37.8 shows the interface of the system related to the rookie knowledge-
base. The slider bar, shown at middle of the figure, prompts the user whether to
maximize or minimize the selected output variable and by how much desirability.
Suppose the rule-based system is provided with following input:
Objective: maximize tool life with desirability of 98%.
Rake angle of tool prefixed to 0ı .
Cutting speed and milling-orientation: open for optimization.
Pressing the Process button starts the processing and following results are displayed
in the information pane:
The recommended orientation is down-milling cutting speed is 154.2 m/min.
The predicted tool life is 4; 526:4 mm2 and cutting force is 649.68 N.
Suppose the same rule-based system is provided with more experimental data, cov-
ering effects of all the influential parameters of milling process. When the system
is asked to generate knowledge-base from that set of data, the resulting knowledge-
base would be a veteran knowledge-base. As more and more data will be provided
to the rule-based system it will keep improving its accuracy of optimization and
prediction processes.
Figure 37.9 presents the interface of the rule-based system from the experimental
data provided in papers [14, 15] in addition to that provided in Table 37.1.
37 A Self Progressing Fuzzy Rule-Based System 445

Processing Button for self

button development

Slider bar for optimization

Check box for
numerical variable

Choice box for

categorical variable

Fig. 37.8 Self-generated interface of rule-based system utilizing rookie knowledge-base

Fig. 37.9 Self-generated interface of rule-based system utilizing veteran knowledge-base

446 A. Iqbal, N. U. Dar

37.5 Conclusion

The chapter presents a unique approach for designing mechanism of a self-

progressing fuzzy rule-based system. The system possesses abilities to manage new
variables, to self-develop fuzzy sets, to self-generate rules for optimization and pre-
diction modules, to resolve the conflict among contradictory rules, and to keep its
interface updated. The discussion shows the distinctiveness of the presented rule-
based system along with its high degree of applicability to process of optimizing the
machining process. Brisk and automatic development of knowledge-base makes it
well adaptable to ever-changing industrial environment.

References

1. L. Monostori, AI and machine learning techniques for managing complexity, changes, and
uncertainty in manufacturing, Eng. Appl. Artif. Intell., 16: 277–291 (2003)
2. J.L. Castro, J.J. Castro-Schez, and J.M. Zurita, Use of a fuzzy machine learning technique in
the knowledge-acquisition process, Fuzzy Sets Syst., 123: 307–320 (2001)
3. H. Lounis, Knowledge-based systems verification: A machine-learning based approach,
Expert Syst. Appl., 8(3) 381–389 (1995)
4. K.C. Chan, A comparative study of the MAX and SUM machine-learning algorithms using
virtual fuzzy sets, Eng. Appl. Artif. Intell., 9(5): 512–522 (1996)
5. G.I. Webb, Integrating machine learning with knowledge-acquisition through direct interac-
tion with domain experts, Knowl-Based Syst., 9: 253–266 (1996)
6. A. Lekova and D. Batanov, Self-testing and self-learning fuzzy expert system for technological
process control, Comput. Ind., 37: 135–141 (1998)
7. Y. Chen, A. Hui, and R. Du, A fuzzy expert system for the design of machining operations,
Int. J. Mach. Tools Manuf., 35(12): 1605–1621 (1995)
8. B. Filipic and M. Junkar, Using inductive machine learning to support decision making in
machining processes, Comput. Ind., 43: 31–41 (2000)
9. S. Cho, S. Asfour, A. Onar, and N. Kaundinya, Tool breakage detection using support vector
machine learning in a milling process, Int. J. Mach. Tools Manuf., 45: 241–249 (2005)
10. P. Priore, D. De La Fuente, J. Puente, and J. Parreno, A comparison of machine learning
algorithms for dynamic scheduling of flexible manufacturing systems, Eng. Appl. Artif. Intell.,
19: 247–255 (2006)
11. A. Iqbal, N. He, L. Li, and N.U. Dar, A fuzzy expert system for optimizing parameters and
predicting performance measures in hard-milling process, Expert Syst. Appl., 32(4): 1020–
1027 (2007)
12. R.A. Orchard, Fuzzy CLIPS, V6.04A Users’ Guide, NRC, Canada (1998)
13. T. Childs, K. Maekawa, T. Obikawa, and Y. Yamane, Metal Machining: Theory and Applica-
tions, Arnold, London (2000)
14. A. Iqbal, N. He, L. Li, W.Z. Wen, and Y. Xia, Influence of tooling parameters in high-speed
milling of hardened steels, Key Eng. Mater. (Advances Machining & Manufacturing and
Technology 8), 315–316: 676–680 (2006)
15. A. Iqbal, N. He, L. Li, Y. Xia, and Y. Su, Empirical modeling the effects of cutting parameters
in high-speed end milling of hardened AISI D2 under MQL environment. Proceedings of 2nd
CIRP Conference on High Performance Cutting, Vancouver, Canada, 2006
Chapter 38
Selection of Ambient Light for Laser Digitizing
of Quasi-Lambertian Surfaces

D. Blanco, P. Fernández, E. Cuesta, C. M. Suárez, and N. Beltrán

Abstract Present work deals with the influence of ambient light on the quality
of surfaces digitized with a laser triangulation probe. Laser triangulation systems
project a laser beam onto the surface of the workpiece, so that an image of this pro-
jection is captured in a photo-sensor. The system is then able to establish vertical
position for each point on the projection, processing the information contained in
the mentioned image. As the sensor does not only capture the light projected by the
laser, but also captures the ambient light emitted in the same wavelength of the laser
beam, ambient light becomes a potential error source. A methodology for testing
different light sources under the same digitizing conditions has been developed and
applied to the digitizing of a 99% quasi-lambertian reflectance standard. Tests have
been carried out for six different types of ambient light sources and the resulting
point clouds have been set for comparison. Three different criteria have been used
to analyze the quality of the point cloud: the number of captured points, the average
dispersion of the test point cloud with respect to a reference point cloud, and the dis-
tribution of such geometric dispersion across the whole surface. Results show that
best quality is obtained for low pressure sodium lamps and mercury vapor lamps.

Keywords Laser triangulation  digitizing  influence of ambient light

38.1 Introduction

In laser triangulation (LT), the projection of a laser beam onto a surface is cap-
tured as an image in a charged coupled device (CCD). Applying image processing
techniques and the triangulation principle, 3D coordinates of the surface points are

D. Blanco (B)
Department of Manufacturing Engineering, University of Oviedo,
Campus de Gijón, 33203 Gijón, Spain
E-mail: dbf@uniovi.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 447
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
448 D. Blanco et al.

acquired (Fig. 38.1). If the distance between a particular point P and the CCD
matches exactly the value of the reference distance (stand-off), its image in the
CCD will be placed exactly in a reference point P 0 . Otherwise, if the point onto
the surface were further away a distance H in the direction of the laser beam, its
image on the CCD will be placed a distance h from the reference point. This way,
it is possible to determine the spatial position of every single point from its image
position on the sensor [1]. When digitizing a part, the laser beam projection sweeps
the target surface, capturing a set of digitized points (point cloud) from its surface.
Accurate calculation of the spatial position for each point of the laser stripe
depends on the accurate calculation of the centroid of its light distribution in the
sensor [2]. If the intensity of the light distribution captured in the sensor is too weak,
the system can not properly calculate the position of the points. Otherwise, if laser
intensity is too high, the sensor will turn into saturation, so that the system could not
calculate the position of points. For intermediate situations, the light distribution is
analysed to determine its centroid position, which corresponds to distance h mea-
sured from reference point. Consequently, the light distribution affects the accuracy
of distance H calculation (Fig. 38.1).
The result of the scanning process depends on the LT system characteristics,
the geometry and quality of the surface and the environmental conditions [3]. These
elements determine the shape and contrast of the laser stripe onto the surface and the
image captured by the sensor. Since surface quality is an important influence factor,
most LT systems allow for adjusting laser intensity according to surface colour and
roughness requirements, achieving an improvement in the sharpness of the laser
beam projection.

CCD Sensor
-h
Laser +h
Diode
P’P’
Lens
Reference Distance
(Stand off)

+H P

-H

Fig. 38.1 Scheme of the laser triangulation principle

38 Selection of Ambient Light for Laser Digitizing of Quasi-Lambertian Surfaces 449

The ambient light present at the scanning process is one of the possible environ-
mental influences [4]. Usually, LT systems incorporate optical filters to reduce or
eliminate the influence of the ambient light. These filters accept only those wave-
lengths in the laser emission band. Commercial light sources emit light in a wide
spectrum of frequencies. Since ambient light emitted in the laser emission band will
not be filtered, it will become part of the information captured by the sensor and will
be used in the calculation of point position.

38.2 Objectives

In this work, ambient light influence on the quality of digitized point clouds is evalu-
ated. The tests have been set for the digitizing of a quasi-lambertian surface. In this
type of surfaces, the reflexion is ideally diffused, as the whole incident energy is
reflected in a uniform way in all spatial directions. The quasi-lambertian surface has
been chosen as its behaviour is accepted to provide the best results for LT digitizing.
Digitizing tests have been carried out to compare results under different ambient
light conditions. Although there is a wide range of commercial light sources, the
present work deals with the most commonly used lamps. Tests have been carried
out under laboratory conditions, where illumination for each experiment is reduced
to a single light source. For each test configuration, nature and importance of the
uncertainty introduced by ambient light have been established. Results for each light
source have been afterwards compared to elaborate usage suggestions.

38.3 Configuration of the Tests

Tests have been carried out using a LT stripe commercial system from Metris (model
Metris LC50) which has been mounted on a Brown & Sharpe Global CMM (model
Image). The LC50 uses a laser beam emitting in the red visible spectrum with a
wavelength emission band between 635 and 650 nm. The maximum peak power is
1 mW.
A reflectance standard from Labsphere has been used as the surface to be dig-
itized [5]. This surface (aka reference surface) is 99% reflectance certified. This
means that its surface is quasi-lambertian, so 99% of the received energy is reflected.
Orientation of the light source with respect to the reference surface and the LT sys-
tem has been selected so that the light direction theoretically makes an angle of 45ı
(' D 45ı ) with the standard surface.
Moreover, theoretical direction of the incident light is orthogonal to the sweep
direction. For every point, the system calculates the z coordinate value taking into
account distance H (Fig. 38.1) and sensor position and orientation. Therefore,
influence of ambient light affects the calculated vertical position of the points.
450 D. Blanco et al.

Distance
d

Latitude φ

φ=0 θ=0
Azimut θ

Fig. 38.2 Spherical coordinate system used to orientate the light source

MetrisLC 50

Reflectance
Standards
Light
Source

Fig. 38.3 Scheme of the main elements used in the test

Figure 38.2 shows the spherical coordinate system used for the light sources ori-
entation. The origin of this coordinate system is the centre of the reference surface.
In order to incorporate the influence of light source intensity to this work, tests have
been carried out with two different positions for the light sources: 200 mm (ı1) or
400 mm (ı2) from the origin of the coordinate system.
Light sources and reference surface are mounted in a test-bench designed ad
hoc. This test-bench provides a proper position and orientation of the light source
according to the spherical coordinate system (Fig. 38.3). This mounting allows for
comparing point clouds obtained under different test conditions. The set formed by
the test-bench, the light source and the reference surface has been installed on the
CMM table.
The light sources used for this work are among the most usual types on a metro-
logical laboratory or a workshop. This way, three types of incandescent lamps (clear,
38 Selection of Ambient Light for Laser Digitizing of Quasi-Lambertian Surfaces 451

Table 38.1 Different lamps used in the experiments

K Model Manufacturer Lamp type Lumens

1 CLAS A CL OSRAM Clear incandescent 1,360

2 DECOR A OSRAM Blue incandescent –
3 64476 BT OSRAM Halogen 1,600
4 PLE-T PHILIPS Fluorescent 1,500
5 HQL 50 SDL OSRAM Mercury vapour 1,600
6 SOX 18 OSRAM Low pressure sodium 1,800

tinted in blue and halogen), one fluorescent lamp, one low pressure sodium lamp and
one mercury vapour lamp constitute the final selection.
Although there is a wide variety of light sources for each class that can be tested
(attending to power or shape), the selected lamps have similar values for the lumi-
nous flux (lumens). This selection criterium is based on finding alternatives that
offer the operator a similar visual comfort, when performing long-time running
digitizing processes.
Commercial references for the light sources used in this work are in Table 38.1.

38.4 Experimental Procedure

Test where the only illumination comes from the laser itself will not be altered
by any external energy. Assuming this, a point cloud that has been digitized in the
absence of light will suffer no distortions. Hence, digitizing in the dark appears to be
the most appropriated way for scanning surfaces, although working in the absence
of light is an unapproachable situation for human operators. Nevertheless, a point
cloud obtained in the absence of light can be used as a reference when evaluating
the quality of point clouds digitized under normal ambient lighting.
Experimental procedure used in this work allows for comparing the results
obtained when digitizing under particular light sources with the results obtained
in the absence of ambient light.
Although a single reference point cloud for all test may seem a suitable option,
in practice, this approach is not recommended. Sensitivity of the internal geometry
of the sensor to thermal variations (related to the time the laser remains switched
on) must be taken into account. The fact that the manufacturer of the sensor rec-
ommends a minimum 40 min warm-up period since switching on the laser until a
proper stability is reached, confirms the importance of this effect. Therefore, instead
of using a single reference cloud for all tests, specific reference clouds have been
used in each test comparison. These reference clouds must be digitized immediately
after the capture of each single test cloud. This procedure will minimize the possible
alteration of the sensor internal geometry due to thermal drift.
Thus, the first cloud N kı is obtained by digitizing the reference surface under
a particular type of light source (k), placed at a given distance from the origin of
452 D. Blanco et al.

the coordinate system (ı). Immediately after the first one, a second point cloud,
known as reference point cloud P kı , is obtained by digitizing the same surface in
the absence of ambient light.
Comparison between these two clouds requires each point in the test cloud to
have its equivalent in the reference cloud. To ensure this relationship, a computer
application has been implemented, capable of selecting and classifying a group of
400 points (20  20) in the central area of the reference surface. Matrix constructed
in this way allows for the comparison between each point and its equivalent.
LT system test parameters (as laser light intensity) have been adjusted to avoid
loos of points due to saturation in the reference cloud. Distance between digitized
points has been set to be 1.5 mm in both X and Y directions of the CMM coordinate
system.

38.5 Analysis Criteria

Three criteria have been used for comparing the quality of point clouds obtained
under different sources of light.
The first criterium evaluates the influence of lighting on the number of points
captured in the test cloud. As discussed previously, an excessive input of light energy
turns the sensor to saturation. Therefore, it becomes impossible to calculate a proper
value for the z coordinate of the saturated points.
The saturated points are not included in the cloud N kı as the system rejects them.
The parameter used to characterize this criterium is the number of valid points (nkı )
on the cloud.
The second criterium evaluates the influence of lighting in the proper calculation
of z coordinate value for each point. Improper values will cause the points of the
test cloud to appear in a higher or lower place than they really are.
The absolute difference between the z values for each equivalent pair of valid
points in both the test cloud N kı and its reference cloud P kı is calculated (38.1).
ˇ ˇ ˇ ˇ
ˇ kı ˇ ˇ Nkı ˇ
ˇdi ˇ D ˇzi  ziPkı ˇ (38.1)

The standard deviation  kı of the calculated differences dikı has been used as the
characteristic parameter for this second criterium (38.3).

1 X kı
n
 kı
D di (38.2)
n
i D1
v
u n
u1 X 2
 kı Dt dikı  kı (38.3)
n
i D1
38 Selection of Ambient Light for Laser Digitizing of Quasi-Lambertian Surfaces 453

The last criterium used in this work is qualitative. It consists on a graphical rep-
resentation of deviations dikı for each digitized point cloud. This representation
shows how the ambient light modifies the position of each single point. It allows for
determining whether light influence is equal across the whole surface or not.

38.6 Results Discussion

Attending to first criterium (nkı ), the results of the tests in Table 38.2, illustrate how
certain types of light sources cause a high percentage of points to become saturated.
Thus, in three of the tests (N 11 , N 12 y N 31 ), no point has been captured due to
saturation caused by the great amount of energy in the laser wavelength band. The
sensor can not obtain properly the z coordinate value of these points, therefore the
point cloud is empty.
A partial loose of points has only occurred in one case (test N 21 ). However, the
same light source placed on a further position (test N 22 ) provides a complete point
cloud. The rest of the tests provide complete point clouds, so that proper information
from surface geometry can be easily obtained.
An order of preference between different light sources can be established by
using the second criterium (explained in previous section) referring to the standard
deviation ( kı /.
From these tests it can be concluded that the best results for both testing positions
are obtained for the low pressure sodium lamp. This light source causes a lower
distortion for the test cloud over the reference cloud. Comparison between pairs of
clouds, all obtained in the absence of light, provides a medium value of 1.2 m as
the systematic error attributable to the system itself. Then, the result obtained for

Table 38.2 Results of the tests for the number of valid points and his standard deviation with
respect to the reference cloud
K ı N Kı nKı  Kı .m/ di mi n .m/ di max .m/

1 1 N 11 0 – – –
2 N 12 0 – – –
2 1 N 21 44 8.36 29:24 17:33
2 N 22 400 4.64 21:49 21:12
3 1 N 31 0 – – –
2 N 32 400 4.83 19:4 25:15
4 1 N 41 400 4.33 7:33 21:97
2 N 42 400 2.14 8:97 9:89
5 1 N 51 400 2.15 10:75 9:21
2 N 52 400 1.88 7:75 9:89
6 1 N 61 400 1.99 18:19 13:79
2 N 62 400 1.13 5:92 5:07
454 D. Blanco et al.

the sodium lamp and the furthest position ( kı D 1:13 m) indicates a moderate
influence of this light source on points positions.
The behaviour of the fluorescent lamp is clearly worse than the mercury vapour
one for the closest position, while for the furthest one the difference is not so evident.
For the tinted blue incandescent lamp and the furthest position, the standard devi-
ation is approximately 4.1 times greater than the value calculated for the sodium
lamp. For the closest position, its standard deviation is extremely high, but it must
be remarked that this value has been calculated considering a small number of valid
points, as most of the theoretical points in the cloud have not been captured due to
saturation.
In the case of the halogen lamp, the results for deviation  kı are the worst of all.
At the furthest position, the deviation is the maximal observed, approximately 4.3
times greater than the sodium lamp. This result was predictable as this lamp causes
all the points to become saturated for the distance of 200 mm.
Finally, the third criterium (graphic representation of the parameter dikı along
the whole area) shows how the distribution of these values depends on the type of
light source and is clearly non-uniform (Fig. 38.4 and 38.5).
Thus, when testing the blue incandescent lamp at a distance of 200 mm from the
origin of the coordinate system (Fig. 38.4), the graph shows how the lack of points
due to saturation is produced in points of the reference surface that are close to the
light source.
On the other hand, valid points are only registered in a narrow strip placed in
the area where the points of the surface are further from the light source. In certain
areas of each test cloud, some of the points are located in an upper vertical position
from their equivalent ones in the reference cloud, whereas points in other areas are
located in a lower position.
However, this distortion in the test clouds does not seem to be related to the prox-
imity of points to the light source, even when such a relationship can be established
for the saturation of points.
By contrast, the distribution of peaks and valleys in Fig. 38.4 and 38.5 shows a
parallel orientation to the laser stripe projection onto the surface.
This result does not fit with any of the previous assumptions. The effect may
be related to local irregularities on the reference surface properties. This should be
confirmed by later work.
The appearance of the differences plotted in Fig. 38.4 confirms the conclusions
obtained for the second criterium. Since sodium lamp generates less distortion in
the cloud, it provides the best performance among the tested lamps.
This distortion is increased for mercury lamp and for fluorescent lamp. On the
other hand, the blue incandescent lamp causes a huge distortion effect.
When testing the lights on the furthest position (ı D 400 mm), five completely-
full clouds have been obtained (Fig. 38.5). As it was set for the closest distance,
distribution of dikı shows a non-uniform behaviour. Furthermore, the parallelism
between the preferential direction of peaks and valleys in the graphs and the
orientation of the laser stripe can be noticed as in previous Fig. 38.4.
38 Selection of Ambient Light for Laser Digitizing of Quasi-Lambertian Surfaces 455

Fig. 38.4 dikı for a distance ı D 200 mm

Moreover, the result in terms of the level of distortion for the test clouds follows
the same previously established pattern according to the type of light source.
The sodium lamp is again the source that has a lower influence in the distortion
introduced in the test cloud. For this position, the distortion introduced by mercury
456 D. Blanco et al.

Fig. 38.5 dikı for a distance ı D 400 mm

and fluorescent lamps is very similar, when a better behaviour of the mercury one
has been established for the closest distance.
The results for the incandescent lamps (both the blue one and the halogen) are
the worst among the tested lamps. The differences (both positive and negative) are
significantly higher than for the rest of lamps.
38 Selection of Ambient Light for Laser Digitizing of Quasi-Lambertian Surfaces 457

38.7 Conclusions

This work has established the influence of ambient light on the quality of laser
triangulation digitized surfaces. Results of the tests show how the influence of the
light sources affects the digitizing in different ways. Sources that introduce a huge
amount of energy on the laser wavelength band will cause some of the points to
saturate. In severe cases this affects the whole point cloud and no information will
be obtained. It has been also demonstrated that different types of sources cause
different results when calculating vertical position for every point of the cloud.
The experimentation carried out confirms that laser digitizing of surfaces in com-
plete absence of external sources of light provides the best results. In the usual
case that this requirement can not be satisfied, results lead to recommend using
those sources of light that cause less distortion of the point cloud: low pressure
sodium lamps and the mercury vapour lamps. Sodium lamps emit in the orange
range (589 nm) of the visible spectrum, which is especially annoying when working
for long time periods, as it disables the operator for distinguishing different colors.
This leads to recommend mercury vapour lamps as the most appropriate election.

Acknowledgements The authors wish to thank the Spanish Education and Science Ministry and
the FEDER program for supporting this work as part of the research project MEC-04-DPI2004-
03517.

References

1. Hüser-Teuchert, D., Trapet, E., Garces, A., Torres-Leza, F., Pfeifer, T., Scharsich, P., 1995. Per-
formance test procedures for optical coordinate measuring probes final project report. European
Communities.
2. Hüsser, D., Rothe, H., 1998. Robust averaging of signals for triangulation sensors. Measure-
ment Science and Technology. Vol. 9, pp.1017–1023.
3. Boehler, W., Bordas Vicent, M. Marbs, A., 2003. Investigating Laser Accuracy. In: CIPA XIXth.
International Symposium, 30 Sept.–4 Oct., Antalya, Turkey, pp. 696–701.
4. Blais, F. 2003. A Review of 20 Years of Ranges Sensor Development, SPIE Proceedings,
Electronic Imaging, Videometrics VII. Santa Clara, CA, USA. Vol. 5013, pp. 62–76.
5. Forest, J., Salvi, J., Cabruja, E., Pous, C., 2004. Laser Stripe Peak Detector for 3D Scanners. A
FIR Filter Approach. Proceedings of the 17th International Conference on Pattern Recognition.
Vol. 3, pp. 646–649.
Chapter 39
Ray-Tracing Techniques Applied
to the Accessibility Analysis for the Automatic
Contact and Non Contact Inspection

B. J. Álvarez, P. Fernández, J. C. Rico, and G.Valiño

Abstract Accessibility analysis represents one of the most critical tasks in inspec-
tion planning. The aim of this analysis is to determine the valid orientations of the
inspection devices used in probe operations and non-contact scanning operations
on a Coordinate Measuring Machine (CMM). A STL model has been used for dis-
cretizing the inspection part in a set of triangles, which permits the application of
the developed system to any type of part, regardless of its shape and complexity.
The methodology is based on the application of computer graphics techniques such
as ray-tracing, spatial subdivision and back-face culling. This analysis will take
into account the real shape and geometry of the inspection device and the con-
straints imposed by the CMM on which it is mounted. A simplified model has
been developed for each inspection device component using different basic geo-
metrical shapes. Finally, collision-free orientations are clustered for minimizing the
orientation changes during the inspection process.

Keywords Accessibility  CMM  inspection  probe  laser stripe  scanning

39.1 Introduction

Among the activities of an automatic process planning system for inspection on a

Coordinate Measuring Machine (CMM), the determination of inspection device ori-
entation with regard to the part stands out. These inspection device orientations are
obtained from a methodology based on the accessibility analysis [1] and the applica-
tion of ray-tracing algorithms [2]. Moreover, different computer graphics techniques

B. J. Álvarez (B)
Department of Manufacturing Engineering, University of Oviedo, Campus de Gijón, 33203 Gijón,
Spain
E-mail: braulio@uniovi.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 459
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
460 B.J. Álvarez et al.

like space partitioning and back-face culling have been applied in order to speed up
the searching of valid orientations.
The methodology has been applied to inspection processes based on a touch-
trigger probe [3] and to non-contact scanning processes based on a laser stripe [4].
In both cases, different constraints have been considered: real shape and dimen-
sions of the inspection devices, process parameters and possible orientations of
the motorized head of the CMM where the inspection device has been mounted.
This motorized head (PH10MQ) provides 720 feasible orientation of the inspection
device by rotating at a resolution of 7:5ı about both horizontal and vertical axes
(A, B).

39.2 Analysis Methodology

The accessibility analysis for a touch-trigger probe deals with determining all the
feasible probe orientations that allow for performing the part inspection avoiding
collisions with the part or any other obstacle in the environment of the inspection
process. Moreover, the valid orientations of the non-contact scanning device will be
obtained to guarantee the visibility of the surface to be scanned. The methodology
applied in both cases will be exactly the same. In both cases several simplifications
have to be made. In a first stage, the inspection devices have been abstracted by
means of infinite half-lines. In a second stage, the real shape and dimensions of the
inspection devices are taken into account.
Another simplification is related to the model of the part being inspected. The
CAD model of the part, that may include complex surfaces, is converted to a STL
model, where each surface is discretized by a set of simple triangular facets. For
solving the accessibility analysis, the naı̈ve approach consists on testing the intersec-
tion between the part triangles and any feasible orientation of the inspection device.
In order to reduce the calculation time, different computer graphics techniques like
space partitioning based on kd-tree, ray traversal algorithm, back-face culling and
ray-triangle intersection tests have been also applied [5].

39.3 Analysis Considering Inspection Devices as Infinite

Half-Lines

Figure 39.1 shows the abstractions made to contact and non contact inspection
devices. In the case of touch-trigger probe, each probe orientation is replaced by
an infinite half-line with the same direction of the orientation being analyzed. In the
case of the laser triangulation system, two infinite half-lines have been used. One
represents the laser beam while the second represents the vision space of the CCD
sensor. Then, the analysis is divided in two phases: local and global analysis.
39 Ray-Tracing Techniques Applied to the Accessibility Analysis 461

Fig. 39.1 Inspection devices as infinite half-lines

First, the local analysis only takes into account the possible interference between
the infinite half-line and the part surface being inspected (probed or digitized)
whereas the possible interferences with the rest of part surfaces or any other obstacle
are ignored. Hence, valid orientations l will be those which make an angle between
0 and =2 with the surface normal vector n .
In a second phase, the orientations obtained in the local analysis will be checked
in order to determine if they have or not collision with the rest of part surfaces or
any other obstacle. This analysis is called global analysis and it is complex and
expensive from a computational point of view because it involves the calculation
of multiple interferences tests. To make this calculation easier, a STL model of the
part is used, where each surface is discretized by a set of triangles. Thus the global
accessibility analysis is reduced to determine if there exist interferences between
orientations l obtained in the local analysis and the triangles that compose the STL
model of either the part or the obstacle.
The use of space partitioning structures like kd-trees allows for reducing the
number of the triangles to test in the global analysis because it involves checking
intersection exclusively with triangles which can potentially be traversed by each
inspection device orientation. The part is partitioned in regions bounded by planes
(bounding boxes) and each part triangle is assigned to the region within which it is
located. Then, regions traversed by each inspection device orientation are selected
by means of the ray-traversal algorithm that was first developed and applied by
Kaplan [2].
Before checking intersection between each inspection device orientation and all
triangles included in the traversed regions previously determined, the number of
intersection tests can be reduced even more by applying a back-face culling algo-
rithm [5]. Thus, from the initial set of triangles included in the traversed regions, a
subset is extracted that do not include those triangles whose visibility according to
an analyzed inspection device orientation is completely blocked by other triangles.
462 B.J. Álvarez et al.

Finally, the intersection test between each undiscarded triangle and the inspec-
tion device orientation l is carried out. This verification is based on the algorithm
developed by Möller and Trumbore [6]. If any intersection is detected then the ori-
entation is rejected. If there is no interference with any triangle, the orientation will
be considered as valid.

39.4 Interference Analysis Considering Real Dimensions

of the Inspection Device

The orientations obtained in previous sections are based on an ideal representation

of the inspection device as an infinite half-line. To check if these orientations are
really valid will be necessary to take into account the real shape and dimensions
of the inspection device. Therefore, intersection between triangles of the STL part
model and each of the inspection device components must be checked. Figure 39.2
shows the components for each of the inspection devices that have been considered
in this work:
1. Touch-trigger probe: column, head, touch probe, stylus and tip (Fig. 39.2a)
2. Laser stripe system: column, machine head, adapter and laser head (Fig. 39.2b)
The intersection analysis is speeded up by using a simplified model of each inspec-
tion device component (Fig. 39.2). The column and laser head have been modeled
by straight prisms, the machine head by a sphere and the laser adapter, touch probe
and stylus by a capsule.
Similarly to Section 39.2, a kd-tree algorithm has been implemented in order to
test exclusively the interference between each of the inspection device components
and the triangles included in the part regions that they traverse. To carry out this task
effectively, each inspection device component is enclosed in a bounding volume and
only the part regions that overlap with that volume are analyzed. Then, intersections
between part triangles included in these regions and the component are checked.
Since several geometrical shapes have been used to model the components of
each inspection device, different algorithms for checking intersections are applied:
1. Sphere-triangle intersection algorithm to analyze interferences with the machine
head [7]. This algorithm simply calculates the minimum distance between a point
(sphere centre) and a triangle. If this distance is smaller than the radius of the
sphere, intersection will occur.
2. Prism-triangle intersection algorithm to analyze interferences with the column.
This algorithm derives from the separating axis theorem which was initially
developed to determine if two convex polyhedral are disjoint [8]. The theorem
states that two convex polyhedra, A and B, are disjoint if they can be separated
along either an axis parallel to a normal of a face of either A or B, or along an
axis formed from the cross product of an edge from A with an edge from B.
The application of this theorem to a prism-triangle test involves checking their
relative position with regard to different potential separation axes [9].
39 Ray-Tracing Techniques Applied to the Accessibility Analysis 463

Fig. 39.2 Components of the inspection device and their simplified models

3. Capsule-triangle intersection algorithms to analyze interferences with the touch

probe and the stylus-tip [10]. In this case, intersection analysis is based on finding
the minimum distance between each edge of the triangle and the capsule line
segment, as well as the minimum distance between each extreme point of the
capsule segment and the triangle. If any of these distances is smaller than the
radius of the capsule, then intersection will occur.
464 B.J. Álvarez et al.

39.5 Clustering

From the previous analysis, the inspection device orientations that are valid for
probing each point or scanning each part triangle have been determined. These
orientations are mathematically represented by means of a binary matrix Ai .q; r/
where each element corresponds to a combination of discrete values of A and B
angles:
8 9
ˆ
ˆ 1 if .A D aq ; B D br / >
>
< =
is a valid orientation for point Pi
Ai .q; r / D (39.1)
ˆ 0 if .A D aq ; B D br / >
>
:̂ ;
is a not valid orientation for point Pi

To reduce the process operation time related to device orientation changes, orienta-
tions (aq , br / common to the greatest number of points to probe (clusters of points)
or triangles to scan (clusters of triangles) must be found. The classification of points
or triangles continues until no intersection can be found between the final clusters.
The algorithm used is similar to that developed by Vafaeesefat and ElMaraghy
[11]. Next, the algorithm is explained for an inspection process using a touch-trigger
probe.
Each point Pi .i D 1; 2; : : :; n/ to be probed is associated to a binary matrix of
valid orientations Aki .i D 1; 2; : : :; n/. Initially (k D 1), the clusters will be the
same as each of the points to be probed: Cik D Pi .
Starting from the clusters Cik and from the binary matrices Aki , a new matrix
CI k .i; j / D Aki \ Akj is built showing the common probe orientations to the
clusters two against two.
With the purpose of creating clusters whose points are associated with the great-
est number of valid orientations, the algorithm searches the indices (s, t/ of CI k
that correspond to the maximum number of common valid orientations. After that,
clusters Csk and Ctk associated to points Ps and Pt respectively will be regenerated
as follows:
Csk D Csk [ Ctk and Ctk D Ø (39.2)
and the binary matrix of valid orientation associated to the new cluster Csk will be
Aks D Aks \ Akt .
With these new clusters, matrix CI k is updated to:

Aki \ Akj for i D t and j D s

CI k .i; j / D (39.3)
Ø for i D t and j D t

The clustering process finishes when the number of common orientations corre-
sponding to all the elements above the main diagonal of matrix CI k have become
zero. A similar process is used to determine the clusters of triangles Ti (i D
1; 2; : : :; n) for a part to be scanned.
39 Ray-Tracing Techniques Applied to the Accessibility Analysis 465

39.6 Application Results

39.6.1 Inspection Process by Means of a Touch-Trigger Probe

In order to analyze the application results of accessibility and clustering algorithms

to the inspection process, the part shown in Fig. 39.3 has been considered. For this
part, the STL model contains 2,438 triangles and 38 inspection points located on
three different surfaces (s7, s19 and s39). As an example, Fig. 39.3 shows the acces-
sibility maps for the inspection point P4s39 (point P4 on surface s39) considering
the different geometrical abstractions of the probe. As it can be seen, when real
dimensions and different probe components are taken into account, the accessibility
map is substantially reduced. For the rest of inspection points, similar accessibility
maps can be obtained.
Furthermore, the application of the clustering algorithm allows for obtaining the
clusters shown in Table 39.1. In this case seven clusters have been found.

Fig. 39.3 Accessibility map for the point P4 on the surface s39
466 B.J. Álvarez et al.

Table 39.1 Cluster of points and common orientations in these clusters

Cluster Points/surface Orientations (A, B)

1 P1s19, P2s19, P12s19, P4s7, P4s39, (37.5, 67.5) (45, 67.5)

P7s39, P11s39, P13s39
2 P3s19, P1s39, P5s39, (67.5, 60)
P6s39
3 P4s19, P5s19 (97.5, 90) (97.5, 82.5) (97.5, 75)
(97.5, 67.5) (97.5, 60)
4 P6s19, P8s7 (30, 165) (30, 180) (37.5, 172.5)
(37.5, 165) (37.5, 157.5) (37.5,
150) (37.5, 142.5) (37.5, 135)
(37.5, 157.5) (37.5, 165) (37.5,
172.5) (37.5, 180) (45, 172.5)
(45, 165) (45, 157.5) (45,
150) (45, 142.5) (45, 135)
(45, 127.5) (45, 120) (45, 180)
(52.5, 172.5) (52.5, 157.5)
(52.5, 150) (52.5, 142.5) (52.5,
135) (52.5, 127.5) (52.5, 120)
(52.5, 112.5)
5 P7s19, P8s19, P9s19, P10s19, P11s19, (15, 120) (15, 127.5) (22.5, 112.5)
P6s7, P2s39, P3s39, P8s39, P9s39, (22.5, 120) (22.5, 127.5) (22.5, 135)
P10s39, P12s39, P14s39, P15s39, (22.5, 142.5) (22.5, 150) (22.5, 165)
P16s39, P17s39, P18s39 (30, 112.5) (30, 120) (30, 127.5)
(30, 135) (30, 142.5) (30, 150) (30,
165)
6 P2s7, P3s7 (7.5, 15)
7 P7s7 (7.5, 165) (7.5, 157.5) (7.5, 150)
(7.5, 165) (7.5, 172.5) (7.5, 180)
(15, 172.5) (15, 165) (15,
157.5) (15, 150) (15, 135) (15,
142.5) (15, 150) (15, 157.5) (15,
165) (15, 172.5) (15, 180) (22.5,
172.5) (15, 165) (15, 157.5)
(15, 172.5) (15, 180) (30, 172.5)

39.6.2 Scanning Process by Means of a Laser Stripe System

Apart from the inspection process, the developed methodology allows for determin-
ing the orientations of a laser stripe system to scan a part. In this type of scanning
systems a laser stripe of known width is projected onto the part surface to be scanned
and the reflected beam is detected by a CCD camera. Therefore, not only the incident
laser beam orientation has been taken into account for the accessibility analysis but
also the possible occlusion due to the interference of the reflected laser beam with
the part.
Figure 39.4 shows the laser head orientation map associated to the local and
global accessibility analysis for triangle 558 of the example part. The darkest color
39 Ray-Tracing Techniques Applied to the Accessibility Analysis 467

Fig. 39.4 Local and global accessibility maps for triangle 558

Fig. 39.5 Different stages to determine the global visibility map for a triangle and an incident laser
beam orientation

represents the head orientations (A, B) that coincide or are closest to the normal
direction of the triangle analyzed. These are considered as optimal orientations.
Grey colors represent head orientations far from the optimal value which lead to
worse scanning quality. White color represents head orientations that do not enable
to scan the considered triangles. For a better visualization of the orientation map,
increments of 15ı have been considered for angles A and B.
For triangle 558 an incident laser beam orientation (A D 30ı , B D 180ı/ has
been selected in order to show the part regions (bounding boxes) that it traverses
(Fig. 39.5). Triangles partially or totally enclosed in these bounding boxes are shown
in the figure before and after applying back-face culling.
468 B.J. Álvarez et al.

Fig. 39.6 Clusters associated to the example part

Figure 39.6 shows in a grey scale the ten triangle clusters obtained for the
example part.

39.7 Conclusions

Most of the accessibility analysis presented in other works only deal with a lim-
ited number of the inspection device orientations, simple parts with only planar
surfaces or specific geometrical shapes, simplified device representations or a short
number of points in the case of inspection by touch-trigger probe. However, in this
paper, a new methodology for accessibility analysis is presented which allows for
overcoming the previous limitations:
1. The methodology has been extended to the inspection process by means of a
touch-trigger probe and the scanning process by means of a laser stripe system.
2. All the possible orientations (720) of the inspection device are taken into consid-
eration.
3. The use of the STL model permits the application of the developed system to any
type of part, regardless of its shape and complexity.
4. The real shape and dimensions of the inspection device are considered for the
analysis.
5. The implemented algorithms based on Computer Graphics reduce computation
time and consequently can deal with a high number of inspection points and
complex surfaces.
6. Moreover, a clustering algorithm is applied that efficiently groups the inspection
points and triangles of the STL of the part to be scanned in order to reduce the
number of probe orientation changes.
The developed system has been applied to different parts with satisfactory results
demonstrating the application in practice. Future research will concentrate on devel-
oping new algorithms that further reduce computation time, and on generating the
inspection paths from the orientations obtained in the clustering process.
39 Ray-Tracing Techniques Applied to the Accessibility Analysis 469

Acknowledgements This paper is part of the result of a Spanish State Commission for Science
and Technology Research Project (DPI2000-0605) and was supported by this Commission and
FEDER.

References

1. Spyridi, A. J. and Requicha, A. A. G., 1990, Accessibility analysis for the automatic inspection
of mechanical parts by coordinate measuring machines, Proceedings of the IEEE International
Conference on Robotics and Automation. Vol. 2, 1284–1289.
2. Kaplan, M., 1985, Space-tracing: A constant time ray-tracer, Proceedings of the
SIGGRAPH’85, 149–158.
3. Limaiem, A. and ElMaraghy, H. A., 1999, CATIP: A computer-aided tactile inspection
planning system, Int. J. Prod. Res. 37(3): 447–465.
4. Lee, K. H. and Park, H. -P., 2000, Automated inspection planning of free-form shape parts by
laser scanning, Robot. Comput. Integr. Manuf. 16(4): 201–210.
5. Foley, J. D., Van Dam, A., Feiner, S. K., and Hughes, J. F., 1996, Computer Graphics,
Principles and Practice, 2nd ed., Addison-Wesley, Reading, MA, pp. 663–664.
6. Möller, T. A. and Trumbore, B., 1997, Fast, minimum storage ray/triangle intersection,
J. Graph. Tools 2(1): 21–28.
7. Schneider, P. J. and Eberly, D. H., 2003, Geometrical Tools for Computer Graphics, Morgan
Kaufmann, San Francisco, CA, pp. 376–382.
8. Gottschalk, S., Lin, M. C., and Manocha, D., 1996, OBBTree: A hierarchical structure for
rapid interference detection, Proceedings of the SIGGRAPH’96, 171–180.
9. Möller, T. A., 2001, Fast 3D triangle-box overlap testing, J. Graph. Tools 6(1): 29–33.
10. Eberly, D. H., 2000, 3D Game Engine Design. A Practical Approach to Real-Time Computer
Graphics, Morgan Kaufmann, San Diego, CA, pp. 53–57.
11. Vafaeesefat, A. and ElMaraghy, H. A., 2000, Automated accessibility analysis and measure-
ment clustering for CMMs, Int. J. Prod. Res. 38(10): 2215–2231.
Chapter 40
Detecting Session Boundaries to Personalize
Search Using a Conceptual User Context

Mariam Daoud, Mohand Boughanem, and Lynda Tamine-Lechani

Abstract Most popular Web search engines are carachterized by “one size fits all”
approaches. Involved retrieval models are based on the query-document matching
without considering the user context, interests ang goals during the search. Per-
sonalized Web search tackles this problem by considering the user interests in the
search process. In this chapter, we present a personalized search approach which
adresses two key challenges. The first one is to model a conceptual user context
across related queries using a session boundary detection. The second one is to per-
sonalize the search results using the user context. Our experimental evaluation was
carried out using the TREC collection and shows that our approach is effective.

Keywords Personalized search  Session Boundaries  Conceptual User Context 

search engine

40.1 Introduction

Most popular web search engines accept keyword queries and return results that are
relevant to these queries. Involved retrieval models use the content of the document
and their link structure to assess the relevance of the document to the user query.
The major limitation of such systems is that information about the user is completely
ignored. They return the same set of results for the same keyword query even though
these latter are submitted by different users with different intentions. For exam-
ple, the query python may refer to python as a snake as well as the programming
language.
Personalized Web search aims at tailoring the search results to a particular user
by considering his profile, which refers to his interests, preferences and goals during

M. Daoud (B)
IRIT, University of Paul Sabatier, 118, Route de Narbonne, Toulouse, France
E-mail: daoud@irit.fr

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 471
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
472 M. Daoud et al.

the search. One of the key challenges in personalized web search are how to model
accurately the user profile and how to use it for an effective personalized search.
User profile could be inferred from the whole search history to model long term
user interests [1] or from the recent search history [2] to model short term ones.
According to several studies [2], mining short term user interests is more effective
for disambiguating the Web search than long term ones. User profile representation
model has also an impact on the personalized retrieval effectiveness. Involved mod-
els could be arranged from a very simple representation based on bags of words to
complex representation based on concept hierarchy, namely the ODP [3, 4].
This chapter focuses on learning short term user interests to personalize search.
A short term user interest is represented by the user context in a particular search
session as a set of weighted concepts. It is built and updated across related queries
using a session boundary identification method. Search personalization is achieved
by re-ranking the search results of a given query using the short term user context.
The chapter is organized as follows. Section 40.2 presents some related works
of search personalization approaches and session boundary identification proposi-
tions. Section 40.3 presents our contribution of search personalization and session
boundary detection. In Section 40.4, we present an experimental evaluation, discus-
sion and results obtained. In the last section, we present our conclusion and plan for
future work.

40.2 Related Works

Personalized Web search consists of two main components: the user profile model-
ing and the search personalization processes.
User profile is usually represented as a set of keyword or class of vectors [5,6] or
by concept hierarchy issued from the user search history [7, 8]. As we build the user
context using the ODP ontology, we review some related works based on the same
essence. The ODP ontology is used in [3] to learn a general user profile applicable
to all users represented by a set of concepts of the first three levels of the ontology.
Instead of using a set of concepts, an ontological user profile in [9] is represented
over the entire ontology by classifying the web pages browsed by the user into its
concepts. Similar to this last work, an ontological user profile is described in [4]
where the concept weights are accumulated using a spreading activation algorithm
that activates concepts through the hierarchical component of the ontology.
The user profile is then exploited in a personalized document ranking by means
of query reformulation [10], query-document matching [5] or result re-ranking [9].
Mining short term user interests in a session based search requires a session bound-
ary mechanism that identifies the most suitable user interests to the user query. Few
studies have addressed this issue in a personalized retrieval task. We cite the UCAIR
system [2] that defines a session boundary detection based on a semantic similarity
measure between successive queries using mutual information.
40 Detecting Session Boundaries to Personalize Search 473

Unlike previously related works, our approach has several new features. First, we
build a semantic user context as a set of concepts of the ODP ontology and not as an
instance of the entire ontology [9]. The main assumption behind this representation
is to prune non relevant concepts in the search session. Second, we build the user
profile across related queries, which allow using the most suitable user interest to
alleviate an ambiguous web search.

40.3 Detecting Session Boundaries to Personalize Search

Our general approach for search personalization consists of modeling a concep-

tual user context built across a search session to improve the web search accuracy of
related queries. We define a session boundary recognition method that allows group-
ing a sequence of related queries in the same search session. Involved approach is
detailed in next sections by: (a) modeling the user context and use it to personalize
search, (b) describing the session boundary recognition mechanism.

40.3.1 A Conceptual User Context Modelling for Personalizing

Search

We present in this section how to build and update the user context in a search
session and how to exploit it in a personalized document re-ranking.

40.3.1.1 Building the User Context Using the ODP Metadata

We build a short-term user context that refers generally to the user’s topics of inter-
ests during a search session. It is built for each submitted query using the relevant
documents viewed by the user and the ODP ontology. The first step consists of
extracting the keyword user context K s for a submitted query q s . Let Drs be the
set of documents returned with respect to query q s and judged as relevant by the
user, each represented as a single term vector using the tf*idf weighting scheme.
The keyword user context K s is a single term vector where the weight of term t is
computed as follows:
1 X
K s .t/ D wt d (40.1)
jDrs j s
d 2Dr

wt d is the weight of term t in document d . The concept-based user context is built

by first mapping it on the ODP ontology, then disambiguating the obtained concept
set using a sub-concept aggregation scheme.
(A) Mapping the Keyword User Context on the Ontology
The user context K s is matched with concepts of the ODP ontology using the cosine
similarity measure. Each concept of the ODP is related to sub-concepts with “is-a”
474 M. Daoud et al.

relations and is associated to a set of web pages classified under that concept. We
represent each concept by a single term vector cj where terms are extracted from
all individual web pages classified under that concept as detailed in a previous
work [11]. Given a concept cj , its similarity weight sw.cj / with Ks is computed
as follows:
sw.cj / D cos.cj ; Ks / (40.2)
We obtain a set of concepts that contain relevant concepts at different levels and
having different weights. In the next section, we proceed by disambiguating the
concept set in order to rank the most relevant concepts of general level in the top of
the user context representation.
(B) Disambiguating the Concept Set
We aim at disambiguating the obtained concept set using a sub-concept aggrega-
tion scheme. We retain the level three of the ontology to represent the user context.
Indeed, the level two of the ontology is too general to represent the user interests,
and leaf nodes are too specific to improve the web search of related queries.
Then, we recomputed the weight of each weighted concept by summing the
weights of its descendants. We are based on the assumption that the most rele-
vant concepts are those having a greater number of descendant concepts weighted
according to the ontology. As shown in Fig. 40.1a, we identify a cluster of weighted
concepts having a common general depth-three concept; we assign to this latter
a relevance score computed by adding the weights of its descendant concepts as
shown in Fig. 40.1b. The weight of a general concept cj , having a set of n related
descendant concepts S.cj /, is computed as follows:

1 X
sw.cj / D  sw.ck / (40.3)
n
1kn^ck 2S.cj /

We finally represent the user context C s performed at time s as a set of depth-three

weighted concepts, noted < cj ; p.cj / >.

To To

Art computer Art

computer

depth 0.1 0. 0. 0. 0.7 depth

depth 0. 0.2 0. depth

depth depth
0.4
Aggregate sub-concepts under their general Compute depth-three concept scores by summing
depth-three concept its descendant concepts weights
(a) (b)

Fig. 40.1 Disambiguating the user context

40 Detecting Session Boundaries to Personalize Search 475

40.3.1.2 A Session-Based Context Updating

We update the user context across queries identified as related using a session
boundary recognition mechanism described in the next section. Let C s1 and C s be
respectively the user contexts for successive and related queries. Updating method
is based on the following principles: (1) enhance the weight of possible common
concepts that can appear in two successive user contexts, (2) alter the weight of
non-common concepts using a decay factor ˇ. This allows taking into account the
most recent concepts of interests to the user in the search session. The new weight
of a concept cj in the user context C s is computed as follows:

ˇ  swC s1 .cj / C .1  ˇ/  swC s .cj / if cj 2 C s1

swC s .cj / D (40.4)
ˇ  swC s .cj / otherwise

where swC s1 .cj / is the weight of concept cj in context C s1 , swC s .cj / is the
weight of concept cj in context C s .

40.3.1.3 Personalizing Search Using the User Context

We personalize search results of query q sC1 related to the user context C s repre-
sented as an ordered set of weighted concepts < cj ; sw.cj / > by combining for each
retrieved document dk , its initial score Si and its contextual score Sc as follows:

Sf .dk / D   Si .q; dk / C .1   /  Sc .dk ; C s /

0< <1 (40.5)

Contextual score Sc of document dk is computed using the cosine similarity

measure between dk and the concepts of the user context C s as follows:
1 X
Sc .dk ; C s / D  sw.cj /  cos.dk ; cj / (40.6)
h
j D1::h

Where cj is a concept in the user context, score.cj / is the weight of concept cj in

the user context C s .

40.3.2 How to Detect Session Boundaries

We propose a session boundary recognition method using the Kendall rank correla-
tion measure that quantifies the conceptual correlation I between the user context
C s and the query q sC1 . We choose a threshold  and believe the queries are from
the same session if the correlation is above the threshold.
476 M. Daoud et al.

Here, the term-based query vector qsC1

t (where terms are weighted according to
their frequency in the query) is matched with concepts of the ontology in order to
represent the concept vector qsC1
c . We adopt a context-sensitive query weighting
scheme by introducing the query frequency .QF / in the current search session, in
order to rank concepts in the top of qsC1
c when they are close to the user context.
Indeed, query vector qsC1
c D< w1 ; w2 ; ::; wi ; ::: > is computed as follows:

wi D C W .q sC1 ; ci /  QF .ci / (40.7)

where the query frequency .QF / and the concept weight (CW) are formally
defined as:

jqjS
QF .ci / D S
; C W .q sC1 ; ci / D cos.qsC1
t ; ci / (40.8)
< jqj ; ci >

jqjS is the total number of related queries submitted in search session S , <
jqjS ; ci > is the number of user contexts built in the current search session S and
containing concept ci .
Thus, the similarity I gauge the changes in the concept ranks between the
query and the user context as follows.
P P
0 Scc 0 .q/  Scc 0 .C /
s
I D .qoC / D qP Pc c
s
P P (40.9)
c c 0 S 2
cc 0 .q/  c c 0 S 2
cc 0 .C s
/

v.ci /  v.cj /
Sci cj .v/ D sign.v.ci /  v.cj // D ˇ ˇ
ˇv.ci /  v.cj /ˇ
Where, ci and cj are two concepts issued from both the query and the user context,
qsC1 s sC1
c .ci / (resp. C .ci /) is the weight of the concept ci in qc (resp. Cs ).

40.4 Experimental Evaluation

Our experimental evaluation is designed to evaluate empirically the accuracy of the

session boundary recognition measure and the effectiveness of the search personal-
ization approach.

40.4.1 Experimental Data Sets

The experiments were based on the test data provided by TREC collection especially
from disks 1&2 of the ad hoc task that contains 741670 documents. We particu-
larly tested topics from q51  q100 presented in Table 40.1. The choice of this test
40 Detecting Session Boundaries to Personalize Search 477

collection is due to the availability of a manually annotated domain for each query.
This allows us, on one hand, to enhance the data set with simulated user interests
associated for each TREC domain. On the other hand, we can define a search session
as a set of related queries annotated in the same domain of TREC.

40.4.2 Experimental Design and Results

Our experimental methodology is designed to evaluate three particular topics:

The accuracy of our proposed session boundary recognition mechanism
The effectiveness of our personalized search approach comparatively to typical
search

40.4.2.1 Evaluating the Session Boundary Recognition Method

The goals of the session boundary recognition experiments are: (A) analyzing query
– context correlation values (I ) according to the Kendall coefficient measure, (B)
computing the accuracy of the session identification measure and identifying the
best threshold value ().
For this purpose, we apply a real evaluation scenario that consists of choosing
a query sequence holding six successive sessions related to six domains of TREC
listed in Table 40.1. For testing purpose, we build the user context for each query
using 30 of its relevant documents listed in the TREC assessment file and update it
using formula (8) across related queries of the same domain using ˇ D 0:2.
(A) Analyzing Query-Context Correlations
In this experiment, we computed the query-context correlations values between a
particular query and the user context built across previous and queries related to the
same TREC domain with respect to the query sequence. Figure 40.2 shows a query
sequence holding six search sessions presented on the X-axis and the query–context
correlation values on the Y-axis. A fall of the correlation curve means a decrease
of the correlation degree with the previous query and possible session boundary
identification. A correct session boundary is marked by a vertical line according to

Table 40.1 TREC domains with annotated queries

Domains Queries
Environment 59 77 78 83
Military 62 71 91 92
Law and government 70 76 85 87
International relations 64 67 69 79 100
US economics 57 72 84
International politics 61 74 80 93 99
478 M. Daoud et al.

Enviornment Military Law&Gov Int Relations US.Eco Int Politics

0,00

69
79
71
91

72

61

80
93
99
0
62

76
78
83

84
92
70

74
85

64
67

57
77

87

10
-0,10
Kendall correlation values

-0,20

-0,30

-0,40

-0,50

-0,60

-0,70

Queries

Fig. 40.2 Kendall correlation values computed across a query sequence

the annotated queries in each TREC domain. We can notice that correlation values
vary between queries in the same domain. Indeed, in the domain “Environment” of
TREC, some queries are related to the environmental concepts of the ODP, while a
specific query (q59 ) is related to the “Weather” topic that has no match with the set
of the environmental concepts.
Based on the range of correlation values Œ0:61; 0:08, we identify in the next
paragraph the best threshold cut-off value ().
(B) Measuring the Session Boundary Measure Accuracy
The goal of this experiment is to evaluate the accuracy of the session boundary
detection measure. It is computed for each threshold value  as follows:

jCQj
P ./ D (40.10)
jQj

jCQj is the number of queries identified as correctly correlated to the current user
context along the query sequence, and jQj is the total number of correlated queries
in the query sequence.
We show in Fig. 40.3 the accuracy of the Kendall correlation measure with vary-
ing the threshold in the range of (Œ0:61; 0:08. The optimal threshold value is
identified at 0:58 achieving the optimal accuracy of 70%. We can conclude that
the Kendall measure achieves significant session identification accuracy. Indeed, it
takes into account the concept ranks in both the query and the user context represen-
tations, which makes it tolerant for errors of allocating related queries in different
search sessions.
40 Detecting Session Boundaries to Personalize Search 479

%
80

70
Kendall measure accuracy

60

50

40

30

20

10

0
0

0
,6

,5

,3

,3
,5

,5

,4

,4

,4

,3
,5

,5

,4

,4

,3

,3
-0

-0

-0

-0
-0

-0

-0

-0

-0

-0
-0

-0

-0

-0

-0

-0
Threshold values

Fig. 40.3 Kendall correlation accuracy with varying the threshold value

40.4.2.2 Retrieval Effectiveness

Our experimental design for evaluating the retrieval effectiveness consists of com-
paring the personalized search performed using the query and the suitable user
context to the standard search performed using only the query ignoring any user
context.
We conducted two sets of controlled experiments: (A) study the effect of the
re-ranking parameter  in the re-ranking formula (9) on the personalized precision
improvement, (B) evaluate the effectiveness of the personalized search compara-
tively to the standard search.
We used “Mercure” as a typical search engine where the standard search is based
on the BM25 scoring formula retrieval model. We measure the effectiveness of re-
ranking search results in terms of Top-n precision (P5, P10) and Mean average
precision (MAP) metrics.
The evaluation scenario is based on the k-fold cross validation explained as
follows:
For each simulated TREC domain in Table 40.1, divide the query set into k
equally-sized subsets, and using k  1 training subsets for learning the user
context and the remaining subset as a test set.
For each query in the training set, an automatic process generates the associated
keyword user context based on its top n relevant documents listed in the TREC
assessment file provided by TREC using formula (1), and then maps it on the
ODP ontology to extract the semantic user context.
Update the user context concept weights across an arbitrary order of the queries
in the training set using ˇ D 0:2 in formula (8) and use it for re-ranking the
search results of the queries in the test set using h D 3 in formula (9).
480 M. Daoud et al.

Percentage improvment of personalized search 20,00

15,00

10,00

% P@5
5,00 % P@10
% MAP

0,00
0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9

-5,00

-10,00
values
γ

Fig. 40.4 Percentage of improvement achieved by personalized search as a result of varying in

the re-ranking formula (9)

(A) Effects of Re-ranking Parameter  on the Retrieval Effectiveness

We present in Fig. 40.4 the precision improvement graph obtained for the person-
alized search compared to the standard search at each cutoff of P5, P10 and MAP
averaged over the queries belonging to the same domain. In this experiment, we fix
the number of relevant documents per query used to represent the user context ran-
domly at 30 for testing purpose. We see that the setting ( D 0:3) produces the best
improvement in personalized search since it produces higher precision improvement
at P5 (11.63%). This proves that favoring contextual score returned by the system in
the result ranking using small values of  ( < 0:5) against the original score gives
better performance for improving the web search ranking.
(B) Personalized Retrieval Effectiveness
In this experiment, we evaluated the effectiveness of the personalized search over
various simulated domains. We have computed the percentage of improvement
of personalized search comparatively to the standard search computed at P5, P10
and MAP and averaged over the queries belonging to the same domain. Pre-
cision improvement (Impro) is computed as follows: Impro D .Ppersonalized 
Pbaseline /=Pbaseline  100. We fixed  at the best value occurring at 0:3 in the re-
ranking formula (9). Then, we ran experiment to identify the best number of the
relevant document used to represent the user context per query and get 20 at the best
value. Results are presented in Table 40.2.
We see that personalized search improves the retrieval precision of almost the
queries in the six simulated domains. However, the precision improvement varies
between domains. This is probably due on one hand, to the accuracy level of the
user context representation, and on the other hand to the correlation degrees between
queries of the same domain. Indeed some queries annotated in the same domain of
TREC may not share concepts of the ontology, then re-ranking search results with
40 Detecting Session Boundaries to Personalize Search 481

Table 40.2 Result effectiveness of the personalized search

The baseline model Our personalized retrieval model

Domain P@5 P@10 MAP P@5 Impro P@10 Impro MAP Impro
Environment 0.25 0.32 0.18 0.35 40% 0.37 15:38% 0.19 1:73%
Military 0.25 0.27 0.05 0.35 40% 0.32 18:18% 0.07 46:46%
Law & Gov. 0.40 0.42 0.12 0.50 25% 0.45 5:88% 0.14 12:33%
Inter. Rel. 0.16 0.12 0.01 0.16 0% 0.16 33:33% 0.02 36:59%
US Eco. 0.26 0.30 0.09 0.33 25% 0.36 22:22% 0.10 8:35%
Int. Pol. 0.16 0.10 0.05 0.20 25% 0.16 60% 0.07 42:26%

not related concepts influences the precision improvement and probably reduce the
retrieval performance especially for Law & Gov TREC domain.

40.5 Conclusion and Outlook

In this chapter, we described our approach for a session-based personalized search.

It consists of learning short term user interests by aggregating concept-based user
contexts identified within related queries. We proposed a session boundary iden-
tification based on the Kendall rank correlation mesure, which tracking changes
in the dominant concepts between a new query and the current user context.
Our experimental results show that the session boundary identification based on
the Kendall measure achieves significant accuracy. Moreover, our experimental
evaluation shows a significant improvement of personalized retrieval effectiveness
compared to the typical search.
In future work, we plan to enhance the user context representation with seman-
tically related concepts according to the ontology. We plan also to evaluate the
accuracy of the user context representation as well as its impact on the retrieval
effectiveness. Moreover, we intend to evaluate the session boundary recognition
measure using real user data provided by web search engine log file, which can
reveal comparison with time-based session boundary recognition approaches.

References

1. Bin Tan, Xuehua Shen, and ChengXiang Zhai. Mining long-term search history to improve
search accuracy. In KDD ’06: Proceedings of the 12th ACM SIGKDD international conference
on Knowledge discovery and data mining, pages 718–723, New York, NY, USA, 2006. ACM.
2. Smitha Sriram, Xuehua Shen, and Chengxiang Zhai. A session-based search engine. In
SIGIR’04: Proceedings of the International ACM SIGIR Conference, 2004.
3. Fang Liu, Clement Yu, and Weiyi Meng. Personalized web search for improving retrieval
effectiveness. IEEE Transactions on Knowledge and Data Engineering, 16(1):28–40, 2004.
482 M. Daoud et al.

4. Ahu Sieg, Bamshad Mobasher, and Robin Burke. Web search personalization with ontological
user profiles. In Proceedings of the CIKM’07 conference, pages 525–534, New York, NY,
USA, 2007. ACM.
5. Lynda Tamine-Lechani, Mohand Boughanem , Nesrine Zemirli. Personalized document rank-
ing: exploiting evidence from multiple user interests for profiling and retrieval. to appear. In
Journal of Digital Information Management, vol. 6, issue 5, 2008, pp. 354–366.
6. John Paul Mc Gowan. A multiple model approach to personalised information access. Master
thesis in computer science, Faculty of science, Universit de College Dublin, February 2003.
7. Alessandro Micarelli and Filippo Sciarrone. Anatomy and empirical evaluation of an adaptive
web-based information filtering system. User Modeling and User-Adapted Interaction, 14(2–
3):159–200, 2004.
8. Hyoung R. Kim and Philip K. Chan. Learning implicit user interest hierarchy for context in
personalization. In Proceedings of IUI ’03, pages 101–108, New York, NY, USA, 2003. ACM.
9. Susan Gauch, Jason Chaffee, and Alaxander Pretschner. Ontology-based personalized search
and browsing. Web Intelli. and Agent Sys., 1(3–4):219–234, 2003.
10. Ahu Sieg, Bamshad Mobasher, Steve Lytinen, Robin Burke. Using concept hierarchies
to enhance user queries in web-based information retrieval. In The IASTED International
Conference on Artificial Intelligence and Applications. Innsbruck, Austria, 2004.
11. Mariam Daoud, Lynda Tamine-Lechani, and Mohand Boughanem. Using a concept-based
user context for search personalization. to appear. In Proceedings of the 2008 International
Conference of Data Mining and Knowledge Engineering (ICDMKE’08), pages 293–298.
IAENG, 2008.
Chapter 41
Mining Weather Information in Dengue
Outbreak: Predicting Future Cases Based
on Wavelet, SVM and GA

Yan Wu, Gary Lee, Xiuju Fu, Harold Soh, and Terence Hung

Abstract Dengue Fever has existed throughout the contemporary history of mankind
and poses an endemic threat to most tropical regions. Dengue virus is transmitted to
humans mainly by the Aedes aegypti mosquito. It has been observed that there are
significantly more Aedes aegypti mosquitoes present in tropical areas than in other
climate regions. As such, it is commonly believed that the tropical climate suits
the life-cycle of the mosquito. Thus, studying the correlation between the climatic
factors and trend of dengue cases is helpful in conceptualising a more effective
pre-emptive control measure towards dengue outbreaks. In this chapter, a novel
methodology for forecasting the number of dengue cases based on climactic factors
is presented. We proposed to use Wavelet transformation for data pre-processing
before employing a Support Vector Machines (SVM)-based Genetic Algorithm to
select the most important features. After which, regression based on SVM was used
to perform forecasting of the model. The results drawn from this model based on
dengue data in Singapore showed improvement in prediction performance of dengue
cases ahead. It has also been demonstrated that in this model, prior climatic knowl-
edge of 5 years is sufficient to produce satisfactory prediction results for up to
2 years. This model can help the health control agency to improve its strategic plan-
ning for disease control to combat dengue outbreak. The experimental result arising
from this model also suggests strong correlation between the monsoon seasonality
and dengue virus transmission. It also confirms previous work that showed mean
temperature and monthly seasonality contribute minimally to outbreaks.

Keywords Dengue fever modelling  wavelet  SVM  infectious disease  data

mining

Y. Wu (B)
Agency for Science, Technology and Research
20 Biopolis Way, #07–01 Centros (A*GA), Singapore 138668
E-mail: yan wu@scholars.a-star.edu.sg

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 483
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
484 Y. Wu et al.

41.1 Introduction

In 2006, the World Health Organization reported that “dengue is the most rapidly
spreading vector-borne disease”. Between 1985 and 2006, in Southeast Asia alone,
an estimated 1,30,000 people were infected annually with dengue virus, which is
transmitted to humans mainly by the Aedes aegypti mosquito [1]. The Dengue virus
is transmitted from a female Aedes mosquito to a human or vice versa during the
feeding process, generally referred to as horizontal transmission. The Dengue virus
may also be passed from parent mosquitoes to offspring via vertical transmission.
However, mosquitoes can only survive in certain climate conditions. For exam-
ple, the Aedes aegypti mosquito cannot survive below the freezing temperature of
0ı C [2]. However, climatic conditions worldwide have displayed enormous short-
term abnormalities, which increases the likelihood of mosquitoes surviving longer
and penetrating temperate regions that were previously free from mosquito-borne
diseases [2].
In addition to climate conditions, other factors contribute to the time-series trans-
mission profile of the dengue virus, e.g. geographical expansion of Aedes mosquito
population and its density, the demographic distribution, mobility and susceptibility
of the human population. This dynamic relationship is illustrated in the Fig. 41.1
below.

Fig. 41.1 The dynamics of dengue virus spread

41 Mining Weather Information in Dengue Outbreak 485

The ideal method of modelling the dynamics of the dengue outbreak is to

precisely capture each and every contributing element in the chain. However, associ-
ating factors such as mobility of human population to an exact function in the model
remains a challenging problem. In this chapter, we proposed a method to simplify
the problem in a small relatively well-defined area.
Previous work by Fu et al. [3] incorporated time-lags of weather features into
their model and selected contributing features with a genetic algorithm. However,
the work assumed the usefulness of all levels of detail, in terms of time resolu-
tions/time lags, including noise that was likely present in every feature. Although
the method was able to detect seasonal factors that contributed to the outbreak of
dengue fever, the resulting model possessed a large number of features, which raised
the curse of dimensionality [4]. The work also ignored the use of regression learning
for testing the generality of its feature selection model for case forecasting.
The following sections of this chapter describe how a simple yet effective model
can be constructed to accommodate the above limitations. Briefly, the proposed
method pre-processes the data using Wavelet Transformation (WT) to separate
information at different time resolutions superposed in the time-series. Genetic
Algorithms (GA) with Support Vector Machines (SVM) are then applied to select
features. Following this, regression based on SVM is used to perform forecasting
using the model. Conclusions drawn from the empirical modelling using the dengue
cases in Singapore are studied, discussed and compared against [3].

41.2 Model Construction

41.2.1 Data Representation

Data selected for the model should represent the many variables affecting the
spread of the dengue virus. Prior studies have shown that the population size of
the mosquito depends largely on the presence of suitable breeding sites and climate
conditions [5]. As such, we can make the following assumptions of the input data:
Variations in the weather data represent the breeding patterns of the Aedes
mosquitoes and
The time-series of dengue cases in a relatively confined area represents the
ensemble of dengue spread information such as susceptibility of human popu-
lation to the virus
The coding pattern of the above two pieces of information is unknown but likely
non-linear. Moreover, not all information at different time resolutions superposed in
the time-series of an input feature, such as temperature, contributes to the fluctuation
of dengue cases. Perhaps only one or a few seasonal levels of detail in a particular
input feature contribute to the outbreak. Hence, the use of an input feature without
eliminating irrelevant levels of details induces unnecessary noise into the model.
486 Y. Wu et al.

In our model, the input features consisted of the past weekly dengue cases and
eight daily weather features, namely rainfall intensity, cloudiness (maximum, mean
and minimum), temperature and humidity. As the dengue cases were only recorded
when a patient was diagnosed and the incubation period of dengue virus is known
to differ across individuals, daily dengue cases do not accurately represent the out-
break. Thus, summing the daily cases into weekly buckets reduces such inaccuracy.
As such, all daily weather factors were averaged into weekly means after performing
wavelet decomposition as described in Section 41.2.2 below.

41.2.2 Wavelet Decomposition

In order to separate the different levels of details encoded in the input elements of
the system, wavelet decomposition of the time-series inputs was chosen. WT are
similar to Fourier transforms. The fundamental difference is that instead of deriv-
ing the spectral density of a give signal, WT projects the signal onto functions
called wavelets with dilation parameter a and translation parameter b of their mother
wavelet:  
1 t b
a;b .t/ D p (41.1)
a a
subject to the condition of:

Z
C1 2
j‰.!/j
‰o D d! is bounded (41.2)
j!j
0

where ‰.!/ is the Fourier transform of .t/

The continuous WT is defined as follows:
Z
C1

Ta;b .f / D f .t/ .a; b/dt or
1
˝ ˛
Ta;b .f / D f .t/; a;b .t/ (41.3)

where h; i denotes the scalar product in the space of f .t/

Hence the reconstructed signal from the WT is:

Z
C1 Z
C1
1
f .t/ D Ta;b .f / .a; b/da=a2 dc (41.4)
‰Ø
1 0

If a finite number of terms are taken to represent the original signal, Eq. (41.4) can
be rewritten as a multi-resolution approximation [4]:
41 Mining Weather Information in Dengue Outbreak 487

^ X
N
f .t/ D ‰i .ai ; bi / (41.5)
i D1

This can be interpreted as an ensemble of N levels of details fD1 : : : DN g and a level

of approximation fAN g which is illustrated in Eq. (41.6).

^ X
N
f .t/ D A0 D AN C Dn ; where An D AnC1 C DnC1 (41.6)
nD1

While the decomposed signals resemble the fundamental ideas of the use of time-
lags, it separates signals at different levels which in turn separate potential noise
from signals.

41.2.3 Genetic Algorithm

After Wavelet decomposition, each signal input into the model generates a series of
daughter Wavelets inputs. As mentioned earlier, not all levels of detail of every input
signal may be relevant. Thus, it is important to filter sub-signals that are irrelevant
to the output.
For this purpose, we implemented a biologically-inspired, robust, general-
purpose optimisation algorithm: the genetic algorithm (GA). In the GA, each input
feature is represented as an allele in a chromosome [6]. The evolution of chromo-
somes begins from the original (randomly initialized) population and propagates
down generationally. During each generation, the fitness of all chromosomes in the
population is evaluated. Then, competitive selection and modification are performed
to generate a new population. This process is repeated until a predefined limit
or convergence is achieved. In order to confirm and validate the feature selection
model and to generate a better representative set of contributing factors, a 10 cross-
validation technique was introduced into the methodology. The converged solution
of weights in zeros and ones represented the non-selection/selection of a particular
sub-signal by the GA.

41.2.4 Support Vector Machines

The core component of GA is to construct the learning classifier. Among the super-
vised learning techniques used to correlate input and output data, the Support Vector
Machine (SVM) stands firmly as one of the most successful methods. The SVM
maps input vectors to a higher dimensional space where a maximum hyperplane
is drawn to separate classes [7]. Given a set of input f.xi ; yi /g, where yi is the
corresponding output f  1; 1g to the ith input feature vector xi fxi 2 Rn g. ˆ.xi /
488 Y. Wu et al.

maps xi into a higher dimensional space where a possible linear estimate function
can be defined:
!
X
f .x/ D w ˆ.x/ C b D sgn
T
˛i yi K .xi ; x/ C b (41.7)
i "SVs

where K.xi ; x/ D ˆ.x/T ˆ.xi / is the kernel function, SVs are the Support Vectors,
˛ is the Lagrange multipliers and b is the bias. The SVM problem can be formulated
as a quadratic programming problem by optimising ˛ as:
The SVM problem can be formulated as a quadratic programming problem by
optimising:
X l
1 XX
l l
min ˛i  ˛i ˛j yi yj K.xi ; xj / (41.8)
˛i
iD1
2 iD1 jD1

subject to the condition of:

X
0  ai  C 8i and ai yi D 0
i

where C is the capacity control penalty term.

41.2.5 Regression Learning

After the SVM-based GA selects the most relevant features, the input is fed into a
regression learning algorithm which generates a prediction model. An SVM regres-
sion (SVR) method based on the material discussed in Section 41.2.4 above can be
used for non-linear learning by introducing a threshold " analogous to the function
of C in SVM Classification [8].

41.2.6 Model Assessment Criteria

To assess the quality of the final model, we compared predicted results to the actual
data using two simple metrics: the Mean Squared Error (MSE) and the Coefficient
of Determination (R2 ). A small MSE and large R2 .0 <D R2 <D 1/ suggest a better
modelling of the data.

41.3 Results and Discussions

Because Dengue fever occurs throughout the year in Singapore, we gathered the
dengue case and weather data in Singapore from 2001 to 2007. As shown in
Fig. 41.2 below, this set of dengue data includes three significant peaks at Weeks
190, 245 and 339.
41 Mining Weather Information in Dengue Outbreak 489

Fig. 41.2 Weekly dengue cases in Singapore from 2001 to 2007

Dengue cases over these years in Singapore are reported weekly while climate
data from weather stations were collected daily. We corrected the resolution dif-
ference by first performing Wavelet transform of the weather data before sorting
the samples into weekly bins. Additionally, we discarded the finest three levels of
detail generated by Wavelet transforms because the down-sampling by seven makes
log2 .7/ 3 levels of detail redundant. The mother Wavelet employed here is the
Daubechies [9] 6 Wavelet which was selected because reconstruction of the signal
after suppression of the finest three levels of detail showed the best approximation
of the original signal compared to Daubechies 2 and 4.
Before the GA and SVM were applied, all the input and output vectors were
normalised. The normalisation process ensured that none of the variables over-
whelmed the others in terms of the distance measure, which was critical for proper
feature selection. In our experiment, a zero-mean unit-variance normalisation was
employed to give equal distance measures to each feature and levels of detail.
The current dengue cases and decomposed current weather series were used as
input features with future dengue cases as output into GA with 10 cross-validations
to generate a general model of the data represented. A feature was finally selected
if and only if it was selected in at least seven of the ten validation cycles. The
method employed in GA was a tournament with the aim of minimising the root
mean squared error. A recursive method was used to tune the parameter C for best
performance (C D 1 in this case). There are only 13 terms selected by GA out of
the 63 input signals, tabulated in Table 41.1 below. As compared to the 50 features
selected by the model in [3], the dimensionality required to describe the number
of dengue cases was significantly smaller in this model. In the Table, the first term
of a feature denotes its most general trend .AN ) while the second to seventh term,
490 Y. Wu et al.

Table 41.1 GA selected features

Features Terms selected
Current dengue cases Entire series
Cloudiness Third, fourth, seventh
Maximum temperature Fourth
Mean temperature None
Minimum temperature Fifth
Maximum humidity Fifth, seventh
Mean humidity Fifth
Minimum humidity Second
Rainfall intensity Second, fifth

DN to D1 respectively, denotes different levels of detail from the coarsest to the

finest.
From Table 41.1, we observed that either the fourth or fifth term was selected
in almost all the input features. These levels of detail corresponded to the averaged
details between 2–4 months. In Singapore, the climate can be divided into two main
seasons, the Northeast Monsoon and the Southwest Monsoon season, separated by
two relatively short inter-monsoon periods [10]. Each of these periods has a span of
2–4 months, which correlates well with our finding and increases the likelihood of
a close relationship between dengue outbreak and seasonality in Singapore, i.e. the
seasonal change of temperature, humidity and rainfall. From the literature reviewed
in [2], it had been concluded that the rise in temperature does accelerate the life-
cycle of the mosquito. This agrees with the findings from our model, i.e. the first
terms of all temperature features were not selected.
In [3], Fu suggested that time-lags of (0, 2, 6, 8, 12) weeks are important. In con-
trast, our model selected a time-lag of 4 weeks which corresponded to an averaged
detail of 4 weeks, i.e. sixth term in our feature vectors. From Table 41.1, none of the
sixth terms were selected in this model. Taken together, both models suggest that the
contribution of monthly fluctuations to dengue virus transmission may not be signif-
icant. Moreover, both models also suggest that mean temperature does not contribute
to the prediction models at any level. The general trend (seventh term) present in
Cloudiness and Maximum Humidity suggests that some positive correlation exists
between the growth of dengue cases and these two features.
In our previous work [11], we demonstrated that for future dengue outbreak pre-
diction, SVR was likely a better model than linear regression due to its robustness
even in the event of over-fitting. In [11], to investigate our claim that the coding
pattern of weather data into dengue cases was non-linear, we performed regression
with a Radial Basis Function (RBF) SVR with 5 years of training data and 1 year of
testing input. The width  used in RBF was set to be the inverse of the sample size
(312). The range of data used for training and testing were taken from 2001–2006.
To confirm the validity of the model, we carried out a similar experiment with
a different range of training and testing data (2002–2007) while keeping the model
variables unchanged. Instances of the results from these two sets of data are tabu-
41 Mining Weather Information in Dengue Outbreak 491

Table 41.2 Performance of an instance of RBF SVR approach

RBF SVR With GA MSE R2 Figure
Training data: 01–05 No 0:27 0.75 3a
Testing data: 06 Yes 0:091 0.92 3b
Training data: 02–06 No 0:37 0.71 3c
Testing data: 07 Yes 0:085 0.91 3d

Fig. 41.3 Prediction results of SVR with and without GA (a) SVR without GA (2001–2006),
(b) SVR with GA (2001–2006), (c) SVR without GA (2002–2007), (d) SVR with GA (2002–2007)

lated in Table 41.2 and plotted in Fig. 41.3 (red and blue lines denote actual and
predicted No. of cases respectively). The performance statistics show that mapping
the weather data into the RBF space (which is non-linear) produced a high corre-
lation between the input data and the actual dengue cases, reinforcing our claim of
non-linearity.
Statistics in Table 41.2 show that on average, SVR with GA yields a 70% reduc-
tion in MSE and a 25% increase in correlation with the actual data as compared
to that of SVR without GA. From this observation, we can infer that a significant
492 Y. Wu et al.

Table 41.3 Performance of different sample combinations

RBF SVR with GA MSE R2
Training: 01–05 Testing: 06 0.091 0.92
Training: 01–05 Testing: 06 & 07 0.094 0.90
Training: 02–06 Testing: 07 0.085 0.91

number of irrelevant input features were removed by the GA, supporting our claim
that not all levels of detail influenced the spread of dengue.
In order to test the stability of our proposed model, a further experiment was
carried out. We extended the testing samples to two consecutive years (2006–2007)
using the same training data (2001–2005). The result is tabulated in Table 41.3 in
comparison with the previous two experiments.
The performance statistics in Table 41.3 clearly show that the difference in MSE
is minimal .< 0:01/ and the correlation coefficient is within 2% of each other. These
empirical results suggest that the proposed model is capable of the followings:
1. Producing relatively reliable predictions for up to 2 years ahead
2. Constructing a stable model instance using only 5 years of data

41.4 Conclusion

In this chapter, a novel model for predicting future dengue outbreak was presented.
We proposed the use of Wavelet decomposition as a mean of data pre-processing.
This technique, together with SVM-based GA feature selection and SVR learning,
proved to be effective for analysing a special case of infectious diseases – dengue
virus infection in Singapore. Our empirical results strongly suggest that weather fac-
tors influence the spread of dengue in Singapore. However, the results also showed
that mean temperature and monthly seasonality contribute minimally to outbreaks.
Analysis of the results in this chapter gives rise to further epidemiological research
questions on dengue transmissions:
As low frequency (higher order) terms denote general trends of the signal, the
results tabulated in Table 41.1 indicate that dengue incidence seems to be closely
related to the rainfall and humidity trends. Can this be proven epidemiologically?
Although the rise in average temperature is thought to accelerate mosquito’s
breeding cycle, why was there no low frequency terms of temperature repre-
sented in the model? Moreover, why are some moderate frequencies (between
2 and –4 months) of both maximum and minimum temperatures selected in the
model?
As Singapore is an area surrounded by sea, does this research finding fit well into
other regions with dengue prevalence?
41 Mining Weather Information in Dengue Outbreak 493

To provide answers to these questions, further work has to be performed, particularly

in the collection of more precise data over a longer period of time. Predictions of
longer durations ahead, such as 2 weeks onwards, can also be constructed to measure
the performance of the model.
Because outbreaks are stochastic events, it is not possible to precisely predict
the specific realization of one particular outbreak. However, models, such as the
one proposed in our work, are useful for providing quantitative risk assessments of
disease spread in a well-defined area, which is essential information for constructing
effective public-health strategies.

Acknowledgements Our heartfelt gratitude is expressed to William C. Tjhi from Institute of

High Performance Computing, Singapore for reviewing the Chapter. We also wish to acknowledge
NOAA of the USA for releasing daily weather data free-of-charge for research and education.

References

1. World Health Organisation: Dengue Reported Cases. 28 July 2008. WHO <http://www.searo.
who.int/en/Section10/Section332 1101.htm>
2. Andrick, B., Clark, B., Nygaard, K., Logar, A., Penaloza, M., and Welch, R., “Infectious
disease and climate change: detecting contributing factors and predicting future outbreaks”,
IGARSS ’97: 1997 IEEE International Geoscience and Remote Sensing Symposium, Vol. 4,
pp. 1947–1949, Aug. 1997.
3. Fu, X., Liew, C., Soh, H., Lee, G., Hung, T., and Ng, L.C. “ Time-series infectious disease data
analysis using SVM and genetic algorithm”, IEEE Congress on Evolutionary Computation
(CEC) 2007, pp. 1276–1280, Sept. 2007.
4. Mallat, S.G., “Multiresolution approximations and wavelet orthonormal bases of L2 .R/”,
Transactions of the American Mathematical Society, Vol. 315, No. 1, pp. 69–87, Sept. 1989.
5. Favier, C., Degallier, N., Vilarinhos, P.T.R., Carvalho, M.S.L., Yoshizawa, M.A.C., and Knox,
M.B., “Effects of climate and different management strategies on Aedes aegypti breeding
sites: a longitudinal survey in Brası́lia (DF, Brazil)”, Tropical Medicine and International
Health 2006, Vol. 11, No. 7, pp. 1104–1118, July 2006.
6. Grefenstette, J.J., Genetic algorithms for machine learning, Kluwer, Dordrecht, 1993.
7. Burges, C.J.C., “A tutorial on support vector machines for pattern recognition”, Data Mining
and Knowledge Discovery, Vol. 2, No. 2, pp. 955–974, 1998.
8. Drucker, H., Burges, C.J.C., Kaufman, L., Smola, A., and Vapnik, V., “Support Vector Regres-
sion Machines”, Advances in Neural Info Processing Systems 9, MIT Press, Cambridge,
pp. 155–161, 1996.
9. Daubechies, I. “Orthonormal Bases of Compactly Supported Wavelets.” Communications on
Pure and Applied Mathematics, Vol. 41, pp. 909–996, 1988.
10. National Environment Agency, Singapore: Climatology of Singapore. 20 Aug. 2007. NEA,
Singapore. <http://app.nea.gov.sg/cms/htdocs/article.asp?pid=1088>
11. Wu, Y., Lee, G., Fu, X., and Hung, T., “Detect Climatic Factors Contributing to Dengue Out-
break based on Wavelet, Support Vector Machines and Genetic Algorithm”, World Congress
on Engineering 2008, Vol. 1, pp. 303–307, July 2008.
12. Bartley, L.M., Donnelly, C.A., and Garnett, G.P., “Seasonal pattern of dengue in endemic
areas: math models of mechanisms”, Transactions of the Royal Society of Tropical Medicine
and Hygiene, pp. 387–397, July 2002.
13. Shon, T., Kim, Y., Lee, C., and Moon, J., “A machine learning framework for network anomaly
detection using SVM and GA”, Proceedings from the Sixth Annual IEEE SMC Information
Assurance Workshop 2005, pp. 176–183, June 2005.
494 Y. Wu et al.

14. Nakhapakorn, K. and Tripathi, N. K., “An information value based analysis of physical and cli-
matic factors affecting dengue fever and dengue haemorrhagic fever incidence”, International
Journal of Health Geographics, Vol. 4, No. 13, 2005.
15. Ooi, E., Hart, T., Tan, H., and Chan, S., “Dengue seroepidemiology in Singapore”, The Lancet,
Vol. 357, No. 9257, pp. 685–686, Mar 2001.
16. Ministry of Health, Singapore: Weekly Infectious Diseases Bulletin. 28 July 2008. M.O.H.
Singapore. <http://www.moh.gov.sg/mohcorp/statisticsweeklybulletins.aspx>
17. Gubler, D.J., “Dengue and dengue hemorrhagic fever”, Clinical Microbiology Reviews,
Vol. 11, No. 3, pp. 480–496, July 1998.
Chapter 42
PC Tree: Prime-Based and Compressed Tree
for Maximal Frequent Patterns Mining

Mohammad Nadimi-Shahraki, Norwati Mustapha, Md Nasir B Sulaiman,

and Ali B Mamat

Abstract Knowledge discovery or extracting knowledge from large amount of data

is a desirable task in competitive businesses. Data mining is an essential step in
knowledge discovery process. Frequent patterns play an important role in data min-
ing tasks such as clustering, classification, and prediction and association analysis.
However, the mining of all frequent patterns will lead to a massive number of pat-
terns. A reasonable solution is identifying maximal frequent patterns which form the
smallest representative set of patterns to generate all frequent patterns. This research
proposes a new method for mining maximal frequent patterns. The method includes
an efficient database encoding technique, a novel tree structure called PC Tree and
PC Miner algorithm. Experiment results verify the compactness and performance.

Keywords Maximal frequent pattern  frequent pattern  prime number  database

encoding

42.1 Problem of Maximal Frequent Patterns Mining

Since the introduction of the Apriori algorithms [1], frequent patterns mining plays
an important role in data mining research for over a decade. Frequent patterns are
itemsets or substructures that exist in a dataset with frequency no less than a user
specified threshold.
Let L D fi1 ; i2 : : : in g be a set of items. Let D be a set of database transactions
where each transaction T is a set of items and jDj be the number of transactions
in D. Given P D fij : : : ik g be a subset of L (j  k and 1  j, k  n) is called a

M. Nadimi-Shahraki (B)
Department of Computer Engineering, Islamic Azad University, Najafabad branch, Iran and PhD
student of Computer Science, Faculty of Computer Science and Information Technology, Putra
University of Malaysia, 43400 UPM, Selangor, Malaysia
E-mail: admin1@iaun.ac.ir

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 495
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
496 M. Nadimi-Shahraki et al.

pattern. The support of a pattern P or S (P) in D is the number of transactions in D

that contains P. Pattern P will be called frequent if its support is no less than a user
specified support threshold min sup .0    jDj/.
In many real applications especially in dense data with long frequent patterns
enumerating all possible 2L – two subsets of an L length pattern is infeasible [2].
A reasonable solution is identifying a smaller representative set of patterns from
which all other frequent patterns can be derived [3]. Maximal frequent patterns
(MFP) form the smallest representative set of patterns to generate all frequent pat-
terns [4]. In particular, the MFP are those patterns that are frequent but none of their
supersets are frequent. The problem of maximal frequent patterns mining is finding
all MFP in D with respect to .
This research introduces a new method to find all MFP using only one database
scan. The method includes an efficient database encoding technique, a novel tree
structure called Prime-based encoded and Compressed Tree or PC Tree and also
PC Miner algorithm. The database encoding technique utilizes prime number the-
ory and transforms all items from each transaction into only a positive integer. The
PC Tree is a novel and simple tree structure but yet efficient to capture whole of
transactions information. The PC Miner algorithm traverses the PC Tree efficiently
using pruning techniques to find the MFP.

42.2 Related Work

Many efficient algorithms have been introduced to solve the problem of maximal
frequent pattern mining more efficiently. They are almost based on three funda-
mental frequent patterns mining methodologies: Apriori, FP-growth and Eclat [5].
Mostly, they traverse the search space to find MFP. The key to an efficient traversing
is the pruning techniques which can remove some branches in the search space. The
pruning techniques can be categorized into two groups:
SUBSET FREQUENCY PRUNING : T HE ALL SUBSETS OF ANY FREQUENT PATTERN
ARE PRUNED BECAUSE THEY CAN NOT BE MAXIMAL FREQUENT PATTERN .
SUPERSET INFREQUENCY PRUNING : T HE ALL SUPERSETS OF ANY INFREQUENT
PATTERN ARE PRUNED BECAUSE THEY CAN NOT BE FREQUENT PATTERN .

The Pincer-Search algorithm [6] uses horizontal data layout. It combines a bottom-
up and a top down techniques to mine the MFP. However search space is traversed
without an efficient pruning technique. The MaxMiner algorithm [4] uses a breadth-
first technique to traverse of the search space and mine the MFP. It makes use of
a look ahead pruning strategy to reduce database scanning. It prunes the search
space by both subsets frequency and supersets infrequency pruning. The Depth
Project [7] finds MFP using a depth first search of a lexicographic tree of patterns,
and uses a counting method based on transaction projections. The DepthProject
demonstrated an efficient improvement over previous algorithms for mining MFP.
The Mafia [2] extends the idea in DepthProject. It uses a search strategy has been
42 PC Tree 497

improved by an effective pruning mechanisms. The MaxMiner, DepthProject and

Mafia use Rymon’s set enumeration [8] to enumerate all the patterns. Thus these
algorithms avoid having to compute the support of all the frequent patterns. The
Flex [9] is a lexicographic tree designed in vertical layout to store pattern P and
list of transaction identifier where pattern P appears. Its structure is restricted test-
and-generation instead of Apriori-like is restricted generation-and-test. Thus nodes
generated are certainly frequent. The Flex tree is constructed in depth-first fash-
ion. Recently a new two-way-hybrid algorithm [10] for mining MFP uses a flexible
two-way-hybrid search method. The two-way-hybrid search begins the mining pro-
cedure in both the top-down and bottom-up directions at the same time. Moreover,
information gathered in the bottom-up can be used to prune the search space in the
other top down direction.

42.3 Proposed Method

This research proposes a new method to mine all MFP in only one database scan
efficiently. The method introduces an efficient database encoding technique, a novel
tree structure called PC Tree to capture transactions information and PC Miner
algorithm to mine MFP.

42.3.1 Database Encoding Technique

The presentation and encoding of database is an essential consideration in almost all

algorithms. The most commonly database layout is the horizontal and vertical layout
[11]. In both layouts, the size of database is very large. The database encoding is a
useful technique which can reduce the size of database. Obviously, reducing of the
size of database can enhance performance of mining algorithms. Our method uses a
prime-based database encoding technique to reduce the size of transaction database.
It transforms each transaction into a positive integer called Transaction Value (TV)
during of the PC Tree construction as follows: Given transaction T D .tid; X/ where
tid is the transaction-id and X D fij : : : ik g is the transaction-items or pattern X.
While the PC Tree algorithm reads transaction T, the encoding procedure considers
a prime number pr for each item ir in pattern X, and then TVtid is computed by
Eq. (42.1). Therefore, all transactions can be represented in new layout using this
encoding technique.

Y
k
TV t id D pr T D .tid; X/ (42.1)
j

X D fij : : : ik g and ir is presented by pr /

498 M. Nadimi-Shahraki et al.

Table 42.1 The transaction database DB and its Transaction Values

TID Items Encoded TV
1 A, B, C, D, E 2, 3, 5, 7, 11 2,310
2 A, B, C, D, F 2, 3, 5, 7, 13 2,730
3 A, B, E 2, 3, 11 66
4 A, C, D, E 2, 5, 7, 11 770
5 C, D, F 5, 7, 13 455
6 A, C, D, F 2, 5, 7, 13 910
7 A, C, D 2, 5, 7 70
8 C, D, F 5, 7, 13 455

The encoding technique utilizes Eq. (42.1) based on simple following definitions.
A positive integer N can be expressed by unique product
N D p1m1 p2m2 : : : prmr where pi is prime number,
p1 p2    pr and mi is a positive integer, called the multiplicity of pi [12].
For example, N D 1; 800 D 23 32 52 . Here we restrict the multiplicity to mi D 1
because there is no duplicated item in transaction T.
To facilitate the process of the database encoding technique used in our method,
let’s examine it through an example. Let item set L D fA, B, C, D, E, Fg and the
transaction database, DB, be the first two columns of Table 42.1 with eight transac-
tions. The fourth column of Table 42.1 shows TVtid computed for all transactions.

42.3.2 PC Tree Construction

Using tree structure in mining algorithms makes two possibilities to enhance the
performance of mining. Firstly, data compressing by well-organized tree structures
like FP-tree. Secondly, reducing search space by using pruning techniques. Thus
the tree structures have been considered as a basic structure in previous data mining
research [5, 9, 13]. This research introduces a novel tree structure called PC Tree
(Prime-based encoded and Compressed Tree). The PC Tree makes use of both
possibilities data compressing and pruning techniques to enhance efficiency.
A PC Tree includes of a root and some nodes that formed sub trees as children
of the root or descendants. The node structure consisted mainly of several different
fields: value, local-count, global-count, status and link. The value field stores TV
to records which transaction represented by this node. The local-count field set by
1 during inserting current TV and it is increased by 1 if its TV and current TV
are equal. The global-count field registers support of pattern P which presented by
its TV.
In fact during of insertion procedure the support of all frequent and infrequent
patterns is registered in the global-count field. It can be used for interactive mining
where min sup is changed by user frequently [13]. The status field is to keep track-
42 PC Tree 499

{} {} {}

2310:1 2730:1 2310:1 2730:1 2310:1 2730:1

TID TV

1 2310 66:1 66:1 770:1 455:1

2 2730 (b) After insert TIDs 1-2 (c) After insert TID 3 (d) After insert TIDs 4-5
3 66

4 770 {} {} {}
5 455

6 910
2310:1 2730:1 2310:1 2730:1 2310:1 2730:1
7 70

8 455
66:1 770:1 910:1 66:1 770:1 910:1 66:1 770:1 910:1
(a) TVs of the DB

455:1 70:1 455:1 70:1 455:2

(e) After insert TID 6 (f) After insert TID 7 (g) After insert TID 8

Fig. 42.1 Construction of PC Tree

ing of traversing. When a node visited in the traversing procedure the status field is
changed from 0 to 1. The link field is to form sub trees or descendants of the root.
Figure 42.1 shows step by step construction of PC Tree for transactions shown
in Table 42.1 which summarized in Fig. 42.1a.
The construction operation mainly consists of insertion procedure that inserts
TV(s) into PC Tree based on definitions below:
Definition 42.1. I F TV OF NODE nr AND ns IS EQUAL T HEN r D s. I NSER -
TION PROCEDURE INCREASES LOCAL – COUNT FIELD OF NODE nr BY 1 IF THE
CURRENT TV IS EQUAL WITH TV OF nr .

Definition 42.2. R D .ni ; ni 1 : : : nj ; root/ IS A D ESCENDANT IFF TV OF NODE

nr 2 R.i  r  j / CAN DIVIDE ALL TV S KEPT IN NODES D ESCENDANT
Rr D .nrC1 ; nrC2 ; : : : nj ; root/.

Definition 42.3. TV OF THE ROOT IS ASSUMED NULL AND CAN BE DIVIDED

BY ALL TV S .

Definition 42.4. N ODE nr CAN BE BELONGED TO MORE THAN ONE DESCENDANT.

Figure 42.1b shows insertion of the first and second transactions. The second TV
with value 2,730 can not be divided by the first TV with value 2,310 and it creates
a new descendant using definition 2 and 3. Transactions 3–6 are inserted into their
descendant based on definition 2 shown in Fig. 42.1c–e. Insertion of the seventh
transaction applies definition 4 when TV 70 is belonged to two descendants (second
descendant shown in the red and bold line) shown in Fig. 42.1f. The TV of eighth
500 M. Nadimi-Shahraki et al.

transaction with value 455 is equal with the fourth TV, then the local-count field
of forth TV is increased by 1 using definition 1 shown in Fig. 42.1g (shown in
underline).
Each TV in PC Tree represents a pattern P and the support of pattern P or S (P)
is registered in the global-count field. Given pattern P and Q have been presented by
TVP and TVQ respectively, the PC Tree has some nice below properties:
Property 42.1. The S (P) is computed by traversing only descendants of TVP.

Property 42.2. P and Q belong to descendant R and S (P) < S (Q) iff TVP can be
divided by TVQ.

Property 42.3. S (P)  and S (P’s fathers in its descendants)  IFF P is a

Maximal Frequent Pattern.

Property 42.4. Nodes are arranged according to TV order, which is a fixed global
ordering. In fact the PC Tree is an independent-frequency tree structure.

Property 42.5. important procedures are almost done only by two simple math-
ematic operations product and division. Obviously using mathematic operation
enhances the performance instead of string operation.

42.3.2.1 PC Miner Algorithm

As explained in previous section, during of insertion each TV in the PC Tree, the

following procedures are done.
(a) Item-frequency counting
(b) Local-counting
(c) Global-counting
(d) Descendant constructing
The PC Miner algorithm traverses the completed PC Tree to discover the MFP in
top-down direction. There is no need to database scan again, because all informa-
tion about items and patterns are stored in the PC Tree. However Fig. 42.1g as a
completed PC Tree didn’t show some information like global-count stored in the
tree. The miner algorithm makes use of a combined pruning strategy including both
superset infrequency and subset frequency pruning. As a result the search space
is reduced, which dramatically reduces the computation and memory requirement
and enhances the efficiency. The superset infrequency pruning is assisted by the
frequency of items computed during the PC-Tree construction. Table 42.2 shows
the item frequency and considered prime number for transaction database shown in
Table 42.1.
To illustrate the pruning strategy used in the PC Miner algorithm let’s examine
it by some examples. Figure 42.2 shows the completed PC Tree using TVs and
patterns respectively. Given  D 5, according to Table 42.2, infrequent item set
(IFI) consists of fB, E, Fg then all superset of IFI like node (A, B, C, D, E) can be
42 PC Tree 501

Table 42.2 Frequency of items and considered prime numbers

Item Prime number Item frequency

A 2 6
B 3 3
C 5 7
D 7 7
E 11 3
F 13 4

{}

{} (A, B, C, D, E): 1 (A, B, C, D, F): 1

2310:1 2730:1
(A, B, E ): 1 (A, C, D, E): 1 (A, C, D, F): 1

66:1 770:1 910:1

70:1 455:2 (A, C, D): 1 (C, D, F): 2

(a) The completed PC_Tree using TVs (b) The completed PC_Tree using patterns

Fig. 42.2 The completed PC Tree

pruned – the superset infrequency pruning. Given  has been changed from 5 to 2
then all items are frequent. While the PC Miner traverses left sub tree or descendant
of node (A, B, C, D, E), node (A, C, D, E) is found as a maximal frequent – the
property 3. Then all its subsets (here is only node (A, C, D)) are pruned – the subset
frequency pruning.

42.4 Experimental Results

In this section, the accuracy and performance of the method is evaluated. All the
experiments are performed on PC with CPU Intel P4 2.8 GHz, 2 GB main mem-
ory, and running Microsoft Windows XP. All the algorithms are implemented using
Microsoft Visual CCC6:0.
In first experiment, the Synthetic data used in our experiments are generated
using IBM data generator which has been used in most studies on frequent patterns
mining. We generate five datasets with number of items 1,000, average transaction
length ten and number of transaction 1,000–10,000 that called D1, D2, D4, D8,
and D10 respectively. In this experiment, the compactness of our database encoding
technique is verified separately. In fact all TVs transformed are stored in a flat file
502 M. Nadimi-Shahraki et al.

Size of dataset (100 KB)

6 Original dataset
Encoded File
5
4
3
2
1
0
D1 D2 D4 D8 D10

Fig. 42.3 The compactness of the database encoding technique

2500
Numbers of MFP

2000

1500

1000

500

0
5 4 3 2 1
Min_sup %

Fig. 42.4 The number of MFP discovered for T10.I6.D10k

called encoded file then size of this file is compared with the size of original dataset.
It shows good results and reducing the size of these datasets more than half as shown
in Fig. 42.3.
In second experiment, we show accuracy and correctness of the method. The
test dataset T10.I6.D10K is also generated synthetically by the IBM data generator.
Figure 42.4 shows the numbers of maximal frequent patterns discovered for the tests
at varying min sup on this dataset.
In third experiment, in order to evaluate the scalability of our new algorithm,
we applied it as well as Apriori to four IBM dataset generated in experiment 1.
Figure 42.5 shows the performance of two algorithms as a function of the num-
bers of transactions for min sup 2%. When number of transaction is less than 5,000
Apriori slightly outperforms the PC Miner in execution time. When the numbers of
transactions are increased, the execution time of Apriori degraded as compared to
the PC Miner.
Fourth experiment is to compare the performance of the PC-Miner with the Apri-
ori and Flex algorithms on the dataset mushroom, which is a real and dense dataset
contains characteristics of various species of mushrooms [14]. In dataset mushroom,
there are 8,124 transactions, the number of items is 119 and the average transaction
length is 23. Figure 42.6 shows the PC Miner algorithm outperforms both Apriori
and Flex algorithms.
42 PC Tree 503

8
Apriori
7
PC_Miner

Time (100 sec)

6
5
4
3
2
1
0
D1 D2 D4 D8 D10

Fig. 42.5 The scalability of PC Miner vs. Apriori on datasets D1-D10

1.2 Flex
1 PC_Miner
Time (100 sec)

0.8 Apriori

0.6
0.4
0.2
0
10 8 6 4 2 1
Min_sup %

Fig. 42.6 The performance of PC-miner vs. Flex on dataset mushroom

42.5 Conclusion and Future Works

In this paper, we proposed a new method to discover maximal frequent patterns effi-
ciently. Our method introduced an efficient database encoding technique, a novel
tree structure called Prime-based encoded and Compressed Tree or PC Tree and
also PC Miner algorithm. The experiments verified the compactness of the database
encoding technique. The PC Tree presented well-organized tree structure with nice
properties to capture transaction information. The PC Miner reduced the search
space using a combined pruning strategy to traverses the PC Tree efficiently. The
experimental results showed the PC Miner algorithm outperforms the Apriori and
Flex algorithms. For interactive mining where min sup can be changed frequently,
the information kept in the PC Tree can be used and tree restructuring is no needed.
In fact this research introduced a number-theoretic method to discover MFP
that made use of prime number theory and simple computation based on division
operation and a combined pruning strategy. There are some directions for future
improvement optimal data structures, better memory management and pruning
method to enhance the efficiency. This method also can be extended for incremental
frequent patterns mining where transaction database is updated or minimum support
threshold can be changed [13].
504 M. Nadimi-Shahraki et al.

References

1. R. Agrawal and R. Srikant, Fast algorithms for mining association rules, 20th International
Conference. Very Large Data Bases, VLDB, 1215: 487–499 (1994).
2. D. Burdick, M. Calimlim, and J. Gehrke, Mafia: A maximal frequent itemset algorithm for
transactional databases, 17th International Conference on Data Engineering: pp. 443–452
(2001).
3. S. Bashir and A.R. Baig, HybridMiner: Mining Maximal Frequent Itemsets Using Hybrid
Database Representation Approach, 9th International Multitopic Conference, IEEE INMIC:
pp. 1–7 (2005).
4. R.J. Bayardo Jr, Efficiently mining long patterns from databases, ACM SIGMOD International
Conference on Management of Data: pp. 85–93 (1998).
5. J. Han, H. Cheng, D. Xin, and X. Yan, Frequent pattern mining: Current status and future
directions, Data Mining and Knowledge Discovery, 15(1): 55–86 (2007).
6. D.I. Lin and Z.M. Kedem, Pincer-Search: A New Algorithm for Discovering the Maximum
Frequent Set, Advances in Database Technology–EDBT’98: 6th International Conference on
Extending Database Technology, Valencia, Spain (1998).
7. R.C. Agarwal, C.C. Aggarwal, and V.V.V. Prasad, Depth first generation of long patterns,
Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining:
pp. 108–118 (2000).
8. R. Rymon, Search through systematic set enumeration, Third International Conference on
Principles of Knowledge Representation and Reasoning: pp. 539–550 (1992).
9. N. Mustapha, M.N. Sulaiman, M. Othman, and M.H. Selamat, FAST DISCOVERY OF LONG
PATTERNS FOR ASSOCIATION RULES, International Journal of Computer Mathematics,
80(8): 967–976 (2003).
10. F. Chen, and M. Li, A Two-Way Hybrid Algorithm for Maximal Frequent Itemsets Mining,
Fourth International Conference on Fuzzy Systems and Knowledge Discovery, FSKD 2007.
11. M.J. Zaki, Scalable algorithms for association mining, IEEE Transactions on Knowledge and
Data Engineering, 12(3): 372–390 (2000).
12. T.T. Cormen, C.E. Leiserson, and R.L. Rivest, Introduction to algorithms: MIT Press,
Cambridge, MA (1990).
13. M. Nadimi-Shahraki, N. Mustapha, M.N. Sulaiman, and A. Mamat, Incremental updating of
frequent pattern: basic algorithms, Second International Conference on Information Systems
Technology and Management (ICISTM 08) (1): 145–148 (2008).
14. C.B.a.C. Merz, UCI Repository of Machine Learning, 1998.
Chapter 43
Towards a New Generation of Conversational
Agents Based on Sentence Similarity

Karen O’Shea, Dr. Zuhair Bandar, and Dr. Keeley Crockett

Abstract The Conversational Agent (CA) is a computer program that can engage in
conversation using natural language dialogue with a human participant. Most CAs
employ a pattern-matching technique to map user input onto structural patterns of
sentences. However, every combination of utterances that a user may send as input
must be taken into account when constructing such a script. This chapter was con-
cerned with constructing a novel CA using sentence similarity measures. Examining
word meaning rather than structural patterns of sentences meant that scripting was
reduced to a couple of natural language sentences per rule as opposed to poten-
tially 100s of patterns. Furthermore, initial results indicate good sentence similarity
matching with 13 out of 18 domain-specific user utterances as opposed to that of the
traditional pattern matching approach.

Keywords Conversational Agent  Sentence Similarity  word meaning

43.1 Introduction

The concept of ‘intelligent’ machines was first conceived by the British mathemati-
cian Alan Turing [1]. The imitation game, known as the ‘Turing Test’, was devised
to determine whether or not a computer program was ‘intelligent’. This led to the
development of the Conversational Agent (CA) [1] – a computer program that can
engage in conversation using natural language dialogue with a human participant.
CAs can exist in two forms: ‘Embodied’ agents [2] possess an animated huma-
noid body and exhibit attributes such as facial expressions and movement of eye
gaze. ‘Linguistic’ agents [3, 4] consist of spoken and/or written language without

K. O’Shea (B)
The Intelligent Systems Group, Department of Computing and Mathematics, Manchester
Metropolitan University, Chester Street, Manchester M1 5GD
E-mail: k.oshea@mmu.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 505
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
506 K. O’Shea et al.

embodied communication. One of the earliest text-based CAs developed was ELIZA
[3]. ELIZA was capable of creating the illusion that the system was actually listen-
ing to the user simply by answering questions with questions. This was performed
using a simple pattern matching technique, mapping key terms of user input onto a
suitable response. Further advancements on CA design led to PARRY [4], capable
of exhibiting personality, character, and paranoid behavior by tracking its own inter-
nal emotional state during a conversation. Unlike ELIZA, PARRY possessed a large
collection of tricks, including: admitting ignorance by using expressions such as “I
don’t know” in response to a question; changing the subject of the conversation or
rigidly continuing the previous topic by including small stories about the theme [4].
CAs can also engage in social chat and are capable of forming relationships with a
user. ALICE [5], an online chatterbot and InfoChat [6] are just two such examples.
By conversing in natural language these CAs are able to extract data from a user,
which may then be used throughout the conversation.
Considerable research has been carried out on the design and evaluation of
embodied CAs [2,7]; however, little work appears to have been focused on the actual
dialogue. This paper will concentrate on text-based CAs and the development and
evaluation of high-quality dialogue.
Most text-based CA’s scripts are organized into contexts consisting of a number
of hierarchically organized rules. Each rule possesses a list of structural patterns
of sentences and an associated response. User input is then matched against the
patterns and the pre-determined response is sent as output. InfoChat [6] is one
such CA capable of interpreting structural patterns of sentences. However, every
combination of utterances must be taken into account when constructing a script –
an evidently time-consuming, high maintenance task, which undoubtedly suggests
scope for alternative approaches. It is, therefore, envisaged that the employment of
sentence similarity measures could reduce and simplify CA scripting by using a few
prototype natural language sentences per rule.
Two successful approaches to the measurement of sentence similarity are: Latent
Semantic Analysis (LSA) [8] and Sentence Similarity based on Semantic Nets and
Corpus Statistics [9]. LSA is a theory and method for extracting and representing
the contextual-usage meaning of words by statistical computations applied to a large
corpus of text [8]. A word by context matrix is formed based on the number of
times a given word appears in a given set of contexts. The matrix is decomposed
by Singular Value Decomposition (SVD) into the product of three other matrices,
including the diagonal matrix of singular values [10]. This dimension reduction step
collapses the component matrices so that words that occurred or did not occur in
some contexts now appear with a greater or lesser frequency [8]. Reconstruction of
the original matrix enables LSA to acquire word knowledge among large numbers of
contexts. Although LSA makes no use of syntactic relations, it does, however, offer
close enough approximations of people’s knowledge to underwrite and test theories
of cognition. Sentence Similarity based on Semantic Nets and Corpus Statistics will
be employed as the measure in this research and will be described in further detail
in Section 43.2.
43 Towards a New Generation of Conversational Agents Based on Sentence Similarity 507

This chapter is organized as follows: Section 43.2 will describe and illustrate the
sentence similarity measure; Section 43.3 will describe a traditional CA scripting
approach; Section 43.4 will describe CA scripting using sentence similarity; Sec-
tion 43.5 will present an experimental analysis of the two approaches; Section 43.6
will evaluate the results and Section 43.7 will conclude and highlight areas for
further work.

43.2 Sentence Similarity Measure

Sentence Similarity based on Semantic Nets and Corpus Statistics [9] is a mea-
sure that focuses directly on computing the similarity between very short texts
of sentence length. Through the use of a lexical/semantic knowledge-base such
as WordNet [11], the length of separation between two words can be measured,
which in turn, can be used to determine word similarity. The synset – a collection of
synonyms – at the meeting point of the two paths is called the subsumer. The depth
of the subsumer is similarly measured by counting the levels from the subsumer to
the top of the hierarchy. Li et al. [9, 12] proposed that the similarity between two
words be a function of the attributes: path length and depth. The algorithm initiates
by combining the two candidate sentences (T1 and T2) to form a joint word set
using only distinct words. For example:
T1 = Mars is a small red planet
T2 = Mars and Earth orbit the sun
A joint word set ‘T’ is formed where:
T = Mars is a small red planet and earth orbit the sun
As a result, each sentence is represented by the use of the joint word set with no
surplus information. Raw semantic vectors are then derived for each sentence using
the hierarchical knowledge-base WordNet [11], in order to determine the separation
between words. Taking a non-linear transfer function as an appropriate measure,
the following formula denotes a monotonically decreasing function of l, where l D
path length between words and α is a constant.

f .l/ D e ˛l (43.1)

As for the depth of the subsumer, the relationship of words at varying levels of the
hierarchy must be taken into consideration. For example, words at the upper layers
are far more general and less semantically similar than words at lower layers [9].
Therefore, subsuming words at upper layers must be scaled down whereas words at
lower layers must be scaled up, resulting in a monotonically increasing function of
h, where h D depth of subsumer and ˇ is a constant.

f .h/ D .e ˇ l  e ˇ h /=.e ˇ l C e ˇ h / (43.2)

508 K. O’Shea et al.

As such, the raw similarity s(w1, w2) between two words is calculated as:

s.w1; w2/ D e ˛l :.e ˇ l  e ˇ h /=.e ˇ l C e ˇ h / (43.3)

where α D 0:2 and ˇ D 0:45.

Each word is then weighted, that is, assigned an information content value, based
on its significance and contribution to contextual information. By combining the
raw semantic vector s(w1, w2) with the information content of each word, I.w1/
and I.w2/, semantic vectors are created:

si D s.w1; w2/:I.w1/:I.w2/ (43.4)

Finally, the semantic similarity Ss between two sentences, s1 and s2, is calcula-
ted as: p p
S s D si1:si 2= si1= si 2 (43.5)
where si1 is the resultant semantic vector of sentence 1 and si2 is the resultant
semantic vector of sentence 2.
Word order also plays an active role in sentence similarity. Each word is assigned
a unique index number which simply represents the order in which the word appears
in the sentence. For example, take the following sentences denoted T1 and T2:
T1 = The cat ran after the mouse
T2 = The mouse ran after the cat
A joint word set ‘T’ is formed where:
T = The cat ran after the mouse
Each sentence is than compared to that of the joint word set. If the same word
is present – or if not, the next most similar word – then the corresponding index
number from T1 will be placed in the vector, r1. As such, the word order vectors r1
and r2 for the example sentence pair T1 and T2 would be formed as follows:

r1 D f123456g
r2 D f163452g

Therefore, word order similarity Sr is calculated as:

p p
S r D 1  .r1  r2/= .r1  r2/ (43.6)

Finally, the sentence similarity is derived by combining both semantic similarity

and word order similarity. The overall sentence similarity between two sentences
S.T 1; T 2/ is calculated as:

S.T 1; T 2/ D ıS s C .1  ı/S r (43.7)

where ı takes into account that word order plays rather a less significant role when
determining sentence similarity.
43 Towards a New Generation of Conversational Agents Based on Sentence Similarity 509

43.3 Traditional CA Scripting

Traditional approaches [6] interpret structural patterns of sentences by using scripts

consisting of rules organized into contexts. A context may be described as a col-
lection of rules relating to a particular topic. Each context contains a number of
hierarchically organized rules each possessing a list of structural patterns of sen-
tences and an associated response. A user’s utterance is then matched against the
patterns and the associated response is ‘fired’ (selected) and sent as output. The
following steps 1–3 illustrate the procedure.
1. Natural language dialogue from the user is received as input and is matched to a
pattern contained in a rule.
2. Match-strength is calculated based on various parameters, including the activa-
tion level of each rule.
3. The pattern with the highest strength is thus ‘fired’ and sent as output.
Scripts are constructed by first assigning each rule a base activation level, a number
between 0 and 1. The purpose of the activation level is to resolve conflicts when
two or more rules have patterns that match the user’s input [13]. The scripter must
then decide which patterns a user may send in response to output. Each pattern is
assigned a pattern-strength value, typically ranging between 10 and 50. For example,
a rule may be constructed as follows:
<Rule 01>
a:0.5
p:50 *help*
p:50 I do not *<understand-0>*
r: How can I help you
where a = activation level, p = pattern strength/pattern, r = response.
Patterns can also contain wildcard elements * which will match with one or more
consecutive characters. In addition, the macro < understand-0 > enables the scripter
to incorporate stock patterns into a rule [6]. Writing such scripts is a time-consuming
and highly skilled craft [14]. For example, a script typically consists of a number
of contexts each denoting a particular topic of conversation. Each context contains
a hierarchically organized list of rules each possessing a collection of structural
patterns of sentences. However, modifying one rule or introducing a new rule into
the script invariably has an impact on the remaining rules. As such, a reassessment
of the entire script would be warranted, without which would render the CA futile.
The scripter is, therefore, required to remember the rankings of the rules and predict
how the introduction of new rules will interact with existing rules [13]. The huge
overhead and maintenance of this type of scripting undoubtedly suggests scope for
an alternative approach.
510 K. O’Shea et al.

43.4 CA Scripting Using Sentence Similarity

The new proposed approach using sentence similarity will maintain the same script
as that of the traditional approach; however, all patterns will be replaced with nat-
ural language sentences. This considerably reduces the burden and skill required to
produce CA scripts. Through the use of a sentence similarity measure [9], a match
is determined between the user’s utterance and the natural language sentences. The
highest ranked sentence is fired and sent as output. Figure 43.1. and the following
steps 1–3 illustrate the procedure.
1. Natural language dialogue is received as input, which forms a joint word set with
each rule from the script using only distinct words in the pair of sentences. The
script is comprised of rules consisting of natural language sentences.
2. The joint word set forms a semantic vector using a hierarchical semantic/lexical
knowledge-base [11]. Each word is weighted based on its significance by using
information content derived from a corpus.
3. Combining word order similarity with semantic similarity the overall sentence
similarity is determined. The highest ranked sentence is ‘fired’ and sent as output.
The proposed scripts are simply constructed by assigning a number of prototype
natural language sentences per rule. For example, one such rule may be constructed
as follows:
<Rule 01>
I need help
I do not understand
r: How can I help you
where s = sentence and r = response.
The precise number of sentences per rule will start at one and increase to ‘n’
where ‘n’ is determined by experimental analysis. However, it is expected that the
value of ‘n’ will be small and significantly less than the number of patterns used in
traditional scripting methodologies.

Sentence
Semantic
Similarity

User Highest
Joint
input ranked
Word
and sentence
Set
script fired
Word
Order
Similarity

Fig. 43.1 New proposed CA algorithm

43 Towards a New Generation of Conversational Agents Based on Sentence Similarity 511

43.5 Experimental Methodology

43.5.1 Domain

The real world domain is concerned with advising students at University on debt
management and the payment of tuition fees. For the purpose of experimentation,
one script, which consists of 18 rules, was taken from a substantially extensive script
developed by Convagent Ltd. [6]. These rules were representative of the domain and
its complexity.

43.5.2 Experiments

Two sets of experiments were undertaken to compare the traditional scripted CA

and the sentence similarity based CA. The first experiment examined the traditional
approach using structural pattern of sentences [6]. The rules consisted of patterns,
which were in some cases consolidated with macros. This accumulated the count of
patterns into the 100s. In comparison, the second experiment examined the new pro-
posed approach, which used a sentence similarity measure and was scripted using
natural language sentences. Through the use of a sentence similarity measure, the
level of scripting was reduced to a couple of generic prototype sentences. Table 43.1
illustrates the scripting by the two approaches for the same rule.
Approach one consists of structural patterns of sentences consolidated with
macros. The macro <confused-0> contains 16 patterns. Similarly, the macros
<confusing-0>, <sure-neg-0>, <sure-neg-1>, and <understand-0> contain a fur-
ther 8, 21, 10 and 13 additional patterns respectively. This accumulates the final
number of patterns, including the patterns *help* and *not* to 70. Approach two,

Table 43.1 Example scripting by two approaches to CA design

Approach one Approach two
Traditional pattern New proposed
scripting scripting
<Rule 01> <Rule 01>
a: 0.5 s: I need help
p: 50 *<confused-0> s: I do not understand
p: 50 *<confused-0>* s: This is confusing
p: 50 *<sure-neg-0>* r: How can I help you
p: 50 *<sure-neg-1>*
p: 50 *help*
p: 50 *not *<understand-0>*
r: How can I help you
512 K. O’Shea et al.

however, replaces the above patterns for three generic natural language sentences:
“I need help”, “I do not understand” and “This is confusing”.

43.6 Results and Discussion

The first experiment examined the traditional approach using structural patterns of
sentences [6], while the second approach examined the new proposed approach
using natural language sentences. The experiments entailed sending as input 18
domain-specific user utterances. The 18 selected user utterances were deemed to be
a good sample of the domain. The resulting output, that is the fired pattern/sentence,
for the 18 cases are displayed in Table 43.2.

Table 43.2 Results of user input for two approaches to CA design

Utterance Approach one Approach two
User input Traditional pattern New proposed
scripting Fired pattern scripting Fired sentence
1. I am having trouble with * I have a problem with my
mybenefactor sponsor
2. I want assistance * Will pay * I need help
3. I have not quit my course *I* not *quit* course I have not received my
funding
4. Could I pay a tiny quantityof Could I * I would like to pay a small
the cost amount of the fee
5. I have no finance * No * I have no funding
6. I have already paid the fee * Have paid * I could pay part of the fee
7. I have a different reason * Have a * It is none of those reasons
8. I have not sent any payment * Not sent * payment * Payment has not been sent
9. I am no longer studying atthe * No * I am still attending my
University course
10. I have to wait for my career * Wait * loan I am still waiting for my
development loan draft loan
11. I have not sent any payment * However * Payment has not been sent
however I have not quit in the post
12. Could you repeat the * Please repeat the option
choices
13. I have not yet obtained my * Student loan * I have not received my
student loan student loan
14. My local education * I have not received my local
authority appraisal has been education authority
delayed assessment
15. My hardship finance has * Hardship * I have not received hardship
failed to arrive funding
16. I am having trouble with * Direct debit * I have direct debit problems
my direct debit
17. I am broke * I am not at the University
18. I sent you the cash weeks * Sent * Payment was sent in the
ago post to the University last
week
43 Towards a New Generation of Conversational Agents Based on Sentence Similarity 513

The results of the user utterances are: The outputs generated after the input of
user utterances 3, 6, 8, 10, 13, 15, 16, and 18 indicate a correct firing by approach
one. As a result, approach one appears to have found a structurally comparable
match. The outputs generated after the input of user utterances 1, 2, 4, 5, 7, 8, 10,
12, 13, 14, 15, 16, and 18 indicate a correct firing by approach two. As a result,
approach two appears to have found sufficient semantic similarity between the user
utterances and the corresponding natural language sentences.
The outputs generated after the input of user utterances 1, 2, 4, 5, 7, 9, 11, 12,
14, and 17 indicate a miss-firing by approach one. As a result, approach one appears
to have failed to find an identical or comparable match to that of the user utterance.
The outputs generated after the input of user utterances 3, 6, 9, 11, and 17 indicate
a miss-firing by approach two. As a result, approach two appears to have failed to
identify sufficient semantic similarity between the user utterances and the natural
language sentences.
In the cases where approach one miss-fired, this was due to the script not pos-
sessing an identical or comparable structural match. This, however, may be rectified
by incorporating the missing patterns into the script. In the cases where approach
two miss-fired, this was due to insufficient sentences representative of that specific
rule. This, however, can be rectified by incorporating additional natural language
sentences into the script. Furthermore, the sentence similarity measure could be
adjusted so as to consider other parts-of-speech.
In totality, approach one correctly matched 8 out of 18 user utterances, whereas
approach two correctly matched 13 out of 18 user utterances. Typically the number
of patterns per rule for the traditional pattern script was between 50 and 200. In
contrast, the average number of sentences per rule for the natural language script
was three. Moreover, a considerable amount of time is required to write and maintain
the scripts of approach one as opposed to that of approach two in which scripting is
greatly simplified.

43.7 Conclusions and Further Work

Most CAs employ a pattern-matching technique to map user input onto structural
patterns of sentences. However, every combination of utterances that a user may
send as input must be taken into account when constructing such a script. This paper
was concerned with constructing a novel CA using sentence similarity measures.
Examining word meaning rather than structural patterns of sentences meant that
scripting was reduced to a couple of natural language sentences per rule as opposed
to potentially 100s of patterns. Furthermore, initial results indicate good sentence
similarity matching with 13 out of 18 domain-specific user utterances as opposed to
that of the traditional pattern matching approach.
Further work will entail considerable development of the new proposed approach.
The aim will be to incorporate the use of context switching whereby each context
defines a specific topic of conversation. The CA will be robust, capable of tolerating
514 K. O’Shea et al.

a variety of user input. The new approach will become easily adaptable via auto-
matic or other means. It is intended that a user evaluation of the two approaches
to CA design will be conducted. Firstly, each approach would be subjected to a
set of domain-specific utterances. Each CA would then compute a match between
the user utterance and the rules within the scripts, firing the highest strength pat-
tern/sentence as output. A group of human subjects would evaluate the scripts and
their corresponding outputs in order to judge whether the correct pattern/sentence
had been fired. This would provide a means for evaluating the opposing approaches
and their scripting methodologies.

References

1. A. Turing, “Computing Machinery and Intelligence” Mind, Vol. 54 (236), 1950, pp. 433–460.
2. D. W. Massaro, M. M. Cohen, J. Beskow, S. Daniel, and R. A. Cole, “Developing and
Evaluating Conversational Agents”, Santa Cruz, CA: University of California, 1998.
3. J. Weizenbaum, ELIZA – “A Computer Program for the Study of Natural Language Com-
munication between Man and Machine”, Communications of the Association for Computing
Machinery, Vol. 9, 1966, pp. 36–45.
4. K. Colby, “Artificial Paranoia: A Computer Simulation of Paranoid Process”. New York:
Pergamon, 1975.
5. R. S. Wallace, ALICE: Artificial Intelligence Foundation Inc. [Online]. Available:
http://www.alicebot.org (2008, February 01).
6. D. Michie and C. Sammut, Infochat TM Scripter’s Manual. Manchester: Convagent, 2001.
7. G. A. Sanders and J. Scholtz, “Measurement and Evaluation of Embodied Conversational
Agents”, in Embodied Conversational Agents, Chapter 12, J. Cassell, J. Sullivan, S. Prevost
and E. Churchill, eds., Embodied Conversational Agents, MIT Press, 2000.
8. T. K. Landauer, P. W. Foltz, and D. Laham, “Introduction to Latent Semantic Analysis”.
Discourse Processes, Vol. 25 (2–3), 1998, pp. 259–284.
9. Y. Li, D. McLean, Z. A. Bandar, J. D. O’Shea, and K. Crockett, “Sentence Similarity
Based on Semantic Nets and Corpus Statistics”, IEEE Transactions on Knowledge and Data
Engineering, Vol. 18 (8), 2006, pp. 1138–1149.
10. D. Landauer, D. Laham, and P. W. Foltz, “Learning Human-Like Knowledge by Singular
Value Decomposition: A Progress Report”, in Advances in Neural Information Processing
Systems 10, M. I. Jordan, M. J. Kearns, and S. A. Solla, eds., Cambridge, MA: MIT Press,
1998, pp. 45–51.
11. G. A. Miller, “WordNet: A Lexical Database for English”, Communications of the Association
for Computing Machinery, Vol. 38 (11), 1995, pp. 39–41.
12. Y. Li, Z. A. Bandar, and D. Mclean, “An Approach for Measuring Semantic Similarity between
Words using Multiple Information Sources”, IEEE Transactions on Knowledge and Data
Engineering, Vol. 15 (4), 2003, pp. 871–881.
13. C. Sammut, “Managing Context in a Conversational Agent”, Electronic Transactions on
Artificial Intelligence, Vol. 3 (7), 2001, pp. 1–7.
14. D. Michie, “Return of the Imitation Game”, Electronic Transactions in Artificial Intelligence,
Vol. 6 (2), 2001, pp. 203–221.
Chapter 44
Direction-of-Change Financial Time Series
Forecasting Using Neural Networks:
A Bayesian Approach

Andrew A. Skabar

Abstract Conventional neural network training methods find a single set of values
for network weights by minimizing an error function using a gradient descent-based
technique. In contrast, the Bayesian approach infers the posterior distribution of
weights, and makes predictions by averaging the predictions over a sample of net-
works, weighted by the posterior probability of the network given the data. The
integrative nature of the Bayesian approach allows it to avoid many of the diffi-
culties inherent in conventional approaches. This paper reports on the application
of Bayesian MLP techniques to the problem of predicting the direction in the
movement of the daily close value of the Australian All Ordinaries financial index.
Predictions made over a 13 year out-of-sample period were tested against the null
hypothesis that the mean accuracy of the model is no greater than the mean accu-
racy of a coin-flip procedure biased to take into account non-stationarity in the data.
Results show that the null hypothesis can be rejected at the 0.005 level, and that the
t-test p-values obtained using the Bayesian approach are smaller than those obtained
using conventional MLP methods.

Keywords Direction-of-change forecasting  financial time series  Markov Chain

Monte Carlo  neural networks

44.1 Introduction

Financial time series forecasting has almost invariably been approached as a regres-
sion problem; that is, future values of a time series are predicted on the basis of past
values. In many cases, however, correctly predicting the direction of the change
(i.e., up or down) is a more important measure of success. For example, if a trader

A.A. Skabar
Department of Computer Science and Computer Engineering, La Trobe University, Australia,
E-mail: a.skabar@latrobe.edu.au

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 515
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
516 A.A. Skabar

is to make buy/sell decisions on the basis of forecast values, it is more important to

be able to correctly predict the direction of change than it is to achieve, say, a small
mean squared error. We call this direction-of-change forecasting, and its importance
has been acknowledged in several recent papers [1–4].
Clearly, if one has made a prediction for the future value of a time-series, then that
prediction can trivially be converted into a direction-of-change prediction by simply
predicting up/down according to whether the forecast value is greater/smaller than
the present value. There may, however, be advantages in predicting the direction of
change directly (i.e., without explicitly predicting the future value of the series). For
example, traders are likely to base their trading decisions primarily on their opinion
of whether the price of a commodity will rise or fall, and to a lesser degree on the
extent to which they believe that it will rise or fall. Speculatively, this may create in
financial systems an underlying dynamic that allows the direction of change to be
predicted more reliably than the actual value of the series. Therefore in this paper we
conceptualize direction-of-change forecasting not as a regression problem, but as a
binary classification problem, and use multilayer perceptrons (MLPs) to perform
the classification.
Whereas conventional (i.e., gradient descent-based) MLP training methods
attempt to find a single set of values for the network weights by minimizing an
error function, Bayesian MLP methods infer the posterior distribution of the weights
given the data. To perform a classification, samples are drawn from this distribu-
tion, each sample representing a distinct neural network. The final prediction is
obtained by averaging the predictions of the individual sampled networks weighted
by the posterior probability of the network given the training data. Thus, whereas
the conventional approach optimizes over parameters, the Bayesian approach inte-
grates over parameters, allowing it to avoid many of the difficulties associated with
conventional approaches.
This paper reports on the application of Bayesian MLP methods [5–7] to the
financial prediction problem and expands on preliminary results that have been pre-
sented in Skabar [8]. The paper is structured as follows. Section 44.2 outlines the
operation of MLPs and highlights some of the difficulties inherent in using gradient
descent-based MLP training methods on financial forecasting tasks. Section 44.3
then describes how Bayesian methods can be applied to MLPs. Section 44.4 pro-
vides empirical results from applying the technique to predicting the direction
of change in the values of the Australian All Ordinaries Index over a 13 year
out-of-sample period, and Section 44.5 concludes the paper.

44.2 MLPs for Financial Prediction

A multilayer perceptron (MLP) is a function of the following form:

!
X
N1 X
N0
f .x n / D h.u/ where u D wkj g wj i xin (44.1)
j D0 i D0
44 Direction-of-Change Financial Time Series Forecasting Using Neural Networks 517

where N0 is the number of inputs, N1 is the number of units in a hidden layer, wji is
a numerical weight connecting input unit i with hidden unit j , and wkj is the weight
connecting hidden unit j with output unit k. The function g is either a sigmoid
(i.e., g.x/ D .1 C exp.x//1 / or some other continuous, differentiable, nonlinear
function. For regression problems h is the identity function (i.e., h.u/ D u/, and for
classification problems h is a sigmoid.
Thus, an MLP with some given architecture, and weight vector w, provides a
mapping from an input vector x to a predicted output y given by y D f .x, w).
Given some data, D, consisting of n independent items (x1 , y1 /, . . . , (xN , yN /, the
objective is to find a suitable w. In the case of financial prediction, the raw data
usually consists of a series of values measured at regular time intervals; e.g: the
daily close value of a financial index such as the Australian All Ordinaries. The
input vector, xN , corresponds to the N th value of the series, and usually consists of
the current value, in addition to time-lagged values, or other quantities which can be
derived from the series, such as moving averages.
The conventional approach to finding the weight vector w is to use a gradient
descent method to find a weight vector that minimizes the error between the net-
work output value, f (x, w), and the target value, y. For regression problems, it is
usually the squared error that is minimized, whereas for binary classification prob-
lems it is usually more appropriate to minimise the cross-entropy error, which will
result in output values that can be interpreted as probabilities (See Bishop [9] for
a detailed discussion of error functions). This approach is generally referred to as
the maximum likelihood (ML) approach because it attempts to find the most proba-
ble weight vector, given the training data. This weight vector, wML , is then used to
predict the output value corresponding to some new input vector xnC1 .
Because of the very high level of noise present in financial time series data, there
is an overt danger that applying non-linear models such as MLPs may result in
overfitting – a situation in which a model performs well on classifying examples on
which it has been trained, but sub-optimally on out-of-sample data. While the risk of
overfitting can be reduced through careful selection of the various parameters that
control model complexity (input dimensionality, number of hidden units, weight
regularization coefficients, etc.), this usually requires a computationally expen-
sive cross-validation procedure. Bayesian methods provide an alternative approach
which overcomes many of these difficulties.

44.3 Bayesian Methods for MLPs

Bayesian methods for MLPs infer the posterior distribution of the weights given the
data, p.wjD/. The predicted output corresponding to some input vector xn is then
obtained by performing a weighted sum of the predictions over all possible weight
vectors, where the weighting coefficient for a particular weight vector depends on
p.wjD/. Thus, Z
yO n D f .xn ; w/ p.wjD/ d w (44.2)
518 A.A. Skabar

where f .xn , w) is the MLP output. Because p.wjD/ is a probability density func-
tion, the integral in Eq. (44.2) can be expressed as the expected value of f .xn , w)
over this density:
Z
f .xn ; w/p.wjD/ d w D Ep.wjD/ Œf .xn ; w/
(44.3)
1 X
N
' f .xn ; w/
N
i D1

Therefore the integral can be estimated by drawing N samples from the density
p.wjD/, and averaging the predictions due to these samples. This process is known
as Monte Carlo integration.
In order to estimate the density p.wjD/ we use the fact that p.wjD/ /
p.Djw/p.w/, where p.Djw/ is the likelihood, and p(w) is the prior weight dis-
tribution. Assuming that the target values, t n , are binary, then the likelihood can be
expressed as
X
p.Djw/ D exp ft n ln f .xn ; w/ C .1  t/n ln.1  f .xn ; w//g: (44.4)
n

The distribution p(w) can be used to express our prior beliefs regarding the com-
plexity of the MLP. Typically, p(w) is modelled as a Gaussian with zero mean and
inverse variance ˛:
!
 ˛ m=2 ˛X 2
m
p.w/ D exp  wi (44.5)
2 2
i D1

where m is the number of weights in the network [7]. The parameter ˛ therefore
controls the smoothness of the function, with large values of ˛ corresponding to
smoother functions, and hence weights with smaller magnitudes. Because we do
not know what variance to assume for the prior distribution, it is common to set a
distribution of values. As ˛ must be positive, a suitable form for its distribution is
the gamma distribution [7]. Thus,

.a=2/a=2 a=21
p.˛/ D ˛ exp.˛a=2/ (44.6)
.a=2/
where the a and  are respectively the shape and mean of the gamma distribution,
and are set manually. Because the prior depends on α, Eq. (44.2) should be modified
such that it includes the posterior distribution over the ˛ parameters:
Z
yO D f .xn ; w/ p.w; ˛jD/ dwd˛
n
(44.7)

where
p.w; ˛jD/ / p.Djw/p.w; ˛/: (44.8)
44 Direction-of-Change Financial Time Series Forecasting Using Neural Networks 519

The objective in Monte Carlo integration is to sample preferentially from the region
where p.w; ˛jD/ is large, and this can be done using the Metropolis algorithm [10].
This algorithm operates by generating a sequence of vectors in such a way that each
successive vector is obtained by adding a small random component to the previous
vector; i.e., wnew D wold C", where " is a small random vector. Preferential sampling
is achieved using the criterion:

if p.wnew jD/ > p.wold jD/ accept

p.wnew jD/
if p.wnew jD/ < p.wold jD/ accept with probability
p.wold jD/
A problem with applying this sampling method to estimate integrals for MLPs is
that the strong correlations in the posterior weight distribution will lead to many
candidates in the sampling chain being rejected due to the fact that they lead to a
decrease in p.wjD/ [9]. However, this problem can be alleviated by using gradient
information to reduce the random walk behaviour, and the resulting algorithm is
known as the Hybrid Monte Carlo algorithm [11]. While the Hybrid Monte Carlo
algorithm allows for the efficient sampling of parameters (i.e., weights and biases),
the posterior distribution for ˛ should also be determined. In this paper we use
Neal’s approach [6], and use Gibbs sampling [12] for the ˛ s.

44.4 Empirical Results

The experimental results reported in this section were obtained by applying the tech-
nique described above to the daily close values of the Australian All Ordinaries
Index (AORD) for the period from November 1990 to December 2004.

44.4.1 Input Pre-processing

Almost invariably, successful financial forecasting requires that some pre-processing

be performed on the raw data. This usually involves some type of transformation of
the data into new variables more amenable to learning from. Thus, rather than using,
for example, the past prices of a stock, transformed variables might include the abso-
lute change in the price (pt –pt 1 /, the rate of change of price (pt –pt 1 /=pt 1 , or
the price or change in price relative to some index or other baseline such as a moving
average. The advantage of using variables based on changes in price (either relative
or absolute) is that they help to remove non-stationarity from the time series.
In this study, the input variables used are the relative change in close value from
the previous day to the current day (r/, and the 5, 10, 30 and 60 day moving averages
(ma5 , ma10 , ma30 , ma60 /. The moving averages are calculated by simply averaging
the x previous closing values, where x is the period of the moving average. Thus,
520 A.A. Skabar

the input to the network is the vector

.r.t/; ma5 .t/; ma10 .t/; ma30 .t/; ma60 .t// where
r.t/ D .pt  pt 1 / =pt 1 (44.9)

and
1X
t
man .t/ D pt (44.10)
n t n
Note that there would almost certainly exist some other combination of inputs and
pre-processing steps that might lead to better performance than that which can
be achieved using this particular combination. However, in this paper we are not
concerned with optimizing this choice of inputs; rather, we are concerned with
comparing the performance of the Bayesian approach with that of the ML approach.

44.4.2 Hypothesis Testing

Assuming that a return series is stationary, then a coin-flip decision procedure for
predicting the direction of change would be expected to result in 50% of the pre-
dictions being correct. We would like to know whether our model can produce
predictions which are statistically better than 50%. However, a problem is that many
financial return series are not stationary, as evidenced by the tendency for commod-
ity prices to rise over the long term. Thus it may be possible to achieve an accuracy
significantly better than 50% by simply biasing the model to always predict up.
A better approach is to compensate for this non-stationarity, and this can be done
as follows. Let xa represent the fraction of days in an out-of-sample test period for
which the actual movement is up, and let xp represent the fraction of days in the
test period for which the predicted movement is up. Therefore, under a coin-flip
model the expected fraction of days corresponding to a correct upward prediction is
(xa  xp /, and the expected fraction of days corresponding to a correct downward
prediction is .1  xa /  .1  xp /. Thus the expected fraction of correct predictions is

aexp D .xa  xp / C ..1  xa /  .1  xp // (44.11)

We wish to test whether amod (the accuracy of the predictions of our model) is
significantly greater than aexp (the compensated coin-flip accuracy). Thus, our null
hypothesis may be expressed as follows:

Null Hypothesis W H0 W amod  aexp H1 W amod > aexp

We test this hypothesis by performing a paired one-tailed t-test of accuracies

obtained using a collection of out-of-sample test sets. Specifically, we take the
period from January 1992 to December 2004, and divide this into 120 30-day test
periods. Predictions for each 30-day test period are based on a model constructed
44 Direction-of-Change Financial Time Series Forecasting Using Neural Networks 521

using the 200 trading days immediately preceding this test period. The training and
prediction windows are each then advanced by 30 days. For each 30-day prediction
period we calculate amod and aexp . We then use a paired t-test to determine whether
the means of these values differ statistically.

44.4.3 Setting the Priors

The Bayesian approach requires that we specify the prior weight distribution, p(w).
Recall that p(w) is assumed to be Gaussian, with inverse variance ˛, and that ˛
is assumed to be distributed according to a Gamma distribution with shape a and
mean ., which remain to be specified. In order to gain some insight into the range of
values may be suitable for these parameters, we conducted a number of trials using
the ML approach, with weight optimization performed using the scaled conjugate
gradients algorithm [13].
Table 44.1 shows the training and test accuracies corresponding to various values
of ˛. Accuracies are averaged over the 120 30-day prediction windows. The values
in the fourth column of the table represent the p-values obtained through performing
the t-test. Italics show best performance.
Figure 44.1 plots training and test set accuracies against the ˛ value. Low values
for ˛, such as 0.01, impose a small penalty for large weights, and result in overfit-
ting; i.e., high accuracy on training data, but low accuracy on test data. In this case,
the null hypothesis cannot be rejected at the 0.05 level. In contrast, when ˛ is very
high (e.g., 10.0), large weights will penalised more heavily, leading to weights with
small magnitudes. In this case the MLP will be operating in its linear region and the
MLP will display a strong bias towards predictions that are in the same direction
as the direction of the majority of changes on the training data. Thus, if the num-
ber of upward movements on the training data is greater than the number of negative
movements, the MLP will be biased towards making upwards predictions on the test
data; however, this is not likely to lead to a rejection of the null hypothesis because
the null hypothesis takes the non-stationarity of the data into account. It can be seen

Table 44.1 Train accuracy, test accuracy and p-value for various ˛ values
˛ value Train. acc. Test. acc. p-value (Ho)
0.010 0.725 0.490 0.402
0.100 0.701 0.501 0.459
0.500 0.608 0.516 0.169
0.750 0.593 0.520 0.070
1.000 0.585 0.524 0.007
1.500 0.570 0.521 0.038
2.000 0.562 0.518 0.587
5.000 0.549 0.526 0.528
10.00 0.542 0.525 0.479
522 A.A. Skabar

0.75
Average Training Accuracy
Average Testing Accuracy
0.70

0.65
Accuracy

0.60

0.55

0.50

0.45
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
alpha (inverse variance of weights)

Fig. 44.1 Training and test set accuracy averaged over 120 training/test set pairs. A local maximum
test accuracy corresponds to an ˛ value of approximately 1.0

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3
alpha

Fig. 44.2 Gamma distribution with mean a = 1.0 and shape parameter  D 10

from Fig. 44.1 that a local maximum for the test set accuracy occurs for an ˛ value
of approximately 1.0, and in this case the null hypothesis can clearly be rejected at
the 0.01 level.
The range of ˛ values for which the null hypothesis can be rejected is very nar-
row, and ranges from a lower ˛ value in the vicinity of 0.5–0.75, to an upper ˛ value
in the vicinity of 1.52.0. After visualizing the pdf for Gamma distributions with
mean 1.0 and various values for the shape parameter, a shape parameter of 10 was
chosen. The pdf is shown in Fig. 44.2. Note that the pdf conforms roughly to the ˛
value identified in the Table 44.1 as leading to a rejection of the null hypothesis.
44 Direction-of-Change Financial Time Series Forecasting Using Neural Networks 523

Table 44.2 Train accuracy, test accuracy and p-value for Bayesian MLPs (MCMC) and conven-
tionally trained (ML) MLPs (a D 1:0,  D 10)
Method Train. acc. Test. acc. p-value (Ho)
MCMC 0.571 0.528 0.0011
ML 0.585 0.524 0.0068

44.4.4 MCMC Sampling

We now describe the application of the Bayesian approach, which relies on MCMC
sampling to draw weight vectors from the posterior weight distribution.
Monte Carlo sampling must be allowed to proceed for some time before the sam-
pling converges to the posterior distribution. This is called the burn-in period. In the
results presented below, we allowed a burn-in period of 1,000 samples, following
which we then saved every tenth sample until a set of 100 samples was obtained.
Each of the 100 samples was then applied to predicting the probability of an upwards
change in the value of the index on the test examples, and the probabilities were then
averaged over the 100 samples. The resulting accuracies and p-values are shown
in the first row of Table 44.2. The second row shows results obtained using the
maximum-likelihood approach. Note that the p-values for MCMC are much smaller
than those for the maximum likelihood approach, indicating increased confidence
in the rejection of the null hypotheses. Also note that the average test accuracy
for MCMC (52.8%) is greater than that for ML (52.4%), while the average train-
ing accuracy for the MCMC (57.1%) is less than that for ML (58.5%), thereby
supporting the claim that the Bayesian approach is better at avoiding overfitting.

44.5 Conclusions

The superior performance of the Bayesian approach can be attributed to its integra-
tive nature: each individual weight vector has its own bias, and by integrating over
many weight vectors, this bias is decreased, thus reducing the likelihood of inferior
generalization performance resulting from overfitting on the training data.
The most important decision to be made in using the Bayesian approach is the
choice of prior. In this study, we used a relatively narrow distribution for ˛, the
parameter controlling the degree of regularization present in the error function.
This choice was made based on experimenting with different ˛ values within the
ML approach. The criticism could be made that this prior distribution was selected
based on the same data that we had used for testing, and hence that the signifi-
cance of our results may be overstated; however, this is unlikely to be the case, as
the prior depends heavily on factors such as the degree of noise in the data, and
this is relatively constant over different periods of the same time series. Moreover,
the fact that we need only select parameters describing the distribution of ˛, rather
524 A.A. Skabar

than a specific value for ˛, further diminishes the possibility that our prior is biased
towards the particular dataset that we have used.
One of the advantages of the Bayesian approach is its inherent ability to avoid
overfitting, even when using very complex models. Thus, although the results
presented in this paper were based on MLPs with six hidden units, the same per-
formance could, in principle, have been achieved using a more complex network. It
is not clear, however, whether we should expect any coupling between the number of
hidden layer units and the prior distribution. For this reason, we would recommend
preliminary analysis using the ML approach to ascertain the appropriate range for ˛
and selection of priors based on values which lead to significant predictive ability.

References

1. Y. Kajitani, A.I. McLeod, K.W. Hipel, Forecasting Nonlinear Time Series with Feed-forward
Neural Networks: A Case Study of Canadian Lynx data, Journal of Forecasting, 24, 105–117
(2005).
2. J. Chung, Y. Hong, Model-Free Evaluation of Directional Predictability in Foreign Exchange
Markets, Journal of Applied Econometrics, 22, 855–889 (2007).
3. P.F. Christoffersen, F.X. Diebold, Financial Asset Returns, Direction-of-Change Forecasting,
and Volatility Dynamics, Penn Institute for Economic Research PIER Working Paper Archive
04-009, 2003.
4. S. Walczak, An Empirical Analysis of Data Requirements for Financial Forecasting with
Neural Networks, Journal of Management Information Systems, 17(4), 203–222 (2001).
5. D.J.C. MacKay, A Practical Bayesian Framework for Back Propagation Networks, Neural
Computation, 4(3), 448–472 (1992).
6. R.M. Neal, Bayesian Training of Backpropagation Networks by the Hybrid Monte Carlo
Method, Department of Computer Science, University of Toronto Technical Report CRG-
TR-92-1, 1992.
7. R.M. Neal, Bayesian Learning for Neural Networks (Springer-Verlag, New York, 1996).
8. A. Skabar, Application of Bayesian Techniques for MLPs to Financial Time Series Fore-
casting, Proceedings of 16th Australian Conference on Artificial Intelligence, 888–891
(2005).
9. C. Bishop, Neural Networks for Pattern Recognition (Oxford University Press, Oxford, 1995).
10. N.A. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A. Teller, and E. Teller, Equation of
State Calculations by Fast Computing Machines, Journal of Chemical Physics, 21(6), 1087–
1092 (1953).
11. S. Duane, A.D. Kennedy, B.J. Pendleton, and D. Roweth, Hybrid Monte Carlo, Physics Letters
B, 195(2), 216–222 (1987).
12. S. Geman, G. Geman, Stochastic Relaxation, Gibbs Distributions and the Bayesian Restora-
tion of Images, IEEE Transactions on Pattern Analysis and Machine Intelligence, 6, 721–741
(1984).
13. M.F. Moller, A scaled conjugate gradient algorithm for fast supervised learning, Neural
Networks, 6, 525–533 (1993).
Chapter 45
A Dynamic Modeling of Stock Prices
and Optimal Decision Making Using
MVP Theory

Ramin Rajabioun and Ashkan Rahimi-Kian

Abstract In this paper, first a precise mathematical model is obtained for four
competing or cooperating companies’ stock prices and then the optimal buy/sell
signals are ascertained for five different agents which are trading in a virtual market
and are trying to maximize their wealth over 1 trading year period. The model is
so that gives a good prediction of the next 30th day stock prices. The companies
used in this modeling are all chosen from Boston Stock Market. Genetic Program-
ming (GP) is used to produce the predictive mathematical model. The interaction
among companies and the effect imposed by each of five agents on future stock
prices are also considered in our modeling. Namely, we have chosen eight compa-
nies in order that there is some kind of interrelation among them. Comparison of
the GP models with Artificial Neural Networks (ANN) and Neuro-Fuzzy Networks
(trained by the LoLiMoT algorithm) shows the superior potential of GP in predic-
tion. Using these models; five players, each with a specific strategy and all with one
common goal (wealth maximization), start to trade in a virtual market. We have also
relaxed the short-sales constraint in our work. Each of the agents has a different
objective function and all are going to maximize themselves. We have used Parti-
cle Swarm Optimization (PSO) as an evolutionary optimization method for wealth
maximization.

Keywords Stock market model  price prediction  Genetic Programming  mean-

variance portfolio selection  Particle Swarm Optimization (PSO)

R. Rajabioun (B)
School of Electrical and Computer Engineering, Control and Intelligent Processing Center of
Excellence, University of Tehran, Tehran, Iran
E-mail: r.rajabioun@ece.ut.ac.ir

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 525
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
526 R. Rajabioun, A. Rahimi-Kian

45.1 Introduction

Forecasting the change in market prices and making correct decisions is one of the
most principal needs of anyone who economical environments concerns him. Time
series are the most common methods used in price prediction [1–3]. But the pre-
dominant defect of these methods is that they use only the history of a company’s
price to do a prediction. Recently, there has been growing attention to the models
that concern the interaction among companies in modeling and the use of game the-
ory [4–6] in decision making because of providing more realistic models. Because of
complexity of the mutual effects of each company on the others, methods like Artifi-
cial Neural Networks (ANN), Neuro-Fuzzy Networks and State Space (SS) models
are used more often for the stock price modeling. In references [7–10] Neural
Network is used to model the stock market and make prediction. In reference [8],
Genetic algorithm (GA) is incorporated to improve the learning and generalizability
of ANNs for stock market prediction. In reference [11] the difference between the
price and the moving average, highest and lowest prices is used as inputs for one-
day-ahead price prediction. More over, volume of transactions, market indicators
and macro economic data are also considered as input variables [12]. There are also
some studies being performed on the fluctuations and the correlations in stock price
changes in physics communities, using the concepts and methods in physics [13,14].
In reference [15] a neuro-genetic stock prediction system is introduced, which uses
genetic algorithm to select a set of informative input features for a recurrent neural
network. In references [16, 17], the neuro-genetic hybrids for stock prediction are
proposed. The genetic algorithm is used to optimize the weights of ANN.
Producing the right buy/sell signals are also important for those who trade in
the stock markets. In reference [18], two simple and popular trading rules including
moving average and trading range breakout are tested in the Chilean stock market.
Their results were compared with the buy-and-hold strategy, and both trading rules
produced extra returns compared to the buy-and-hold strategy.
Genetic Programming (GP) is a symbolic optimization technique, developed by
Koza [19]. It is an evolutionary computational technique based on the so-called
“tree representation”. This representation is extremely flexible because trees can
represent computer programs, mathematical equations, or complete models of pro-
cess systems [20]. In reference [21] GP is used to produce a one-day-ahead model
to predict stock prices. This model is tested for a fifty consecutive trading days of
six stocks and has yielded relatively high returns on investment.
In this paper, we use the GP to find the best mathematical models for the four
companies’ stock prices under study. Our GP models are able to predict these stock
prices for up to the next 30 days with acceptable prediction errors in the market.
Because, the GP is a well known algorithm we will not present it in details. However,
reference [22] provides a good review of the GP algorithm.
The modeling is done for four companies in Boston Stock Market [23]. Selected
companies include: Advanced Micro Devices (AMD), Ericsson (ERIC), Sony
(SNE), Philips (PHG), International Business Machines (IBM), Intel Corporation
(INTC), Microsoft (MSFT) and Nokia (NOK). These companies are assumed to
45 A Dynamic Modeling of Stock Prices and Optimal Decision 527

have a relationship like competition or cooperation and so their stock prices could
affect on each other. Letters allocated in parentheses are the symbols using which
one can access the price data of each company. We use the price history of these
eight companies as inputs to predict our four objective companies’ prices including:
Ericsson (ERIC), International Business Machines (IBM), Sony (SNE) and Philips
(PHG). Obtained four models precision is compared with two traditional methods:
(1) Multi Layer Perceptron (MLP) and (2) Neuro-Fuzzy Network trained by Locally
Linear Model Tree (LoLiMoT) method.
After modeling the four companies’ stock prices, we create five agents who trade
in a virtual market in order to maximize their wealth. These agents (players) will buy
or sell their in hand stocks according to their uniquely defined strategies. Each player
has a unique objective function. The Buy/Sell actions of each player are obtained so
as to maximize its objective function in each trading period. The maximization is
done using the Particle Swarm Optimization (PSO) method [24].
At rest of the paper, in Section 45.2, modeling and prediction is discussed.
Section 45.3 demonstrates the virtual stock market and argues its constraints and
presumptions. Then in Section 45.4 the results of our simulations are shown and
finally the conclusion is done in Section 45.5.

45.2 Modeling and Prediction

As stated earlier, our primary goal is to obtain a predictive model that is able to
predict the future stock prices precisely. The companies that we are going to pre-
dict their stocks include: Ericsson (ERIC), International Business Machines (IBM),
Sony (SNE) and Philips (PHG). We presume that these companies have some kind
of interrelations with four other companies including: Advanced Micro Devices
(AMD), Intel Corporation (INTC), Microsoft (MSFT) and Nokia (NOK). So we
downloaded these eight companies’ price data from the Boston Stock Market [23].
The downloaded data encompasses some information like: daily opening price, daily
closing price, daily highest price, daily lowest price and exchange volume. In this
paper, we predict the average of daily opening and closing prices. Our data set con-
tains sampled price data for the interval of 2001/07/08 to 2006/03/11. The criterion
used to evaluate the models is the Normalized Mean Square Error (NMSE), which
is defined as follows:
Pn
.yi  yO i /2
i D1
NMSE D (45.1)
P
n
y2i
i D1

where yi and yO i are the original and predicted price values, respectively. Figure 45.1
shows the NMSE values for the train data set (the first 70%) using GP, ANN
(MLP) and Neuro-Fuzzy networks (trained using LoLiMoT). Figure 45.2 depicts
this comparison for the test data set (the last 30%).
528 R. Rajabioun, A. Rahimi-Kian

NMSE for train data

0.012
MLP
0.01 NEUROFUZZY
GP
0.008
NMSE

0.006

0.004

0.002

0
IBM ERIC SNE PHG

Fig. 45.1 The prediction error (NMSE) for all companies (using train data)

NMSE for test data

0.5
MLP
NEUROFUZZY
0.4 GP

0.3
NMSE

0.2

0.1

0
IBM ERIC SNE PHG

Fig. 45.2 The prediction error (NMSE) for all companies (using test data)

The GP-based stock price models were initialized with some functions and ter-
minals. The terminals included random number generators together with integers
from 1 to 16. The functions included: fC; ; ; ; log x; ex ; xy g. The population
sizes were set to 600 except for ERIC, which was set to 400. Meanwhile the number
of iterations was set to 80. As it can be seen from Figs. 45.1 and 45.2, the GP-
based price model prediction errors are acceptable for the training data set and less
than both of the MLP and Neuro-Fuzzy models for test data set. The only draw-
back of the GP algorithm is its time-consuming modeling characteristics, which is
acceptable comparing to its precise modeling.
Until now we have modeled the interactions of eight different companies that
affect the future price values. But due to the fact that buyers/sellers also affect future
stock prices of the companies, it is essential to include such interactions in the mod-
eling. Therefore, we augment a new term to our price models in order to include
the effects of the market players’ actions (buy/sell weights) into the future price
45 A Dynamic Modeling of Stock Prices and Optimal Decision 529

Table 45.1 The PSO model parameters

Parameter range [1, 1]
Maximum optimization iterations each day 200
Population size 180
Acceleration constant 1 2
Acceleration constant 2 2

changes. Since there are not much available data on how the buy/sell volumes of the
market players affect the future prices, we decided to add a new term to show these
effects in our price prediction models as follows:

augmented term D   W  a  Price vector (45.2)

where:
γ : is a weighting coefficient that regulates the impact of the augmented term on
the models. When γ is large the augmented term makes the model deviate from
the trend of the time-series market historical data.
W: is a weight vector that its elements show each company’s stock trade impact
on future prices. The elements of this vector are between 0 and 1.
a: is the action vector of all players. Its elements are between -1 and 1 that show
the buy/sell rates of the stocks.
Price vector: contains the current stock price values in the market.
The best value for the γ factor obtained to be 0.1. The W-vector was chosen as
follows: W = [0.1 0.05 0.1 0.2 0.2 0.2 0.05 0.1]. The corresponding companies’
symbol vector is: (AMD ERIC IBM INTC MSFT NOK PHG SNE).
The augmented term makes it possible for us to see the effect of each player’s
market decision on the stock prices and other players’ wealth (similar to a non-
cooperative game).
Our objective in the next section would be to find the best market actions (sell/buy
of each stock) of each player so as to maximize its expected objective function
(wealth) in the market. Our market simulation studies are done in a Virtual Stock
Market and by means of an evolutionary optimization algorithm (PSO). In our simu-
lations a common PSO with inertia was used. Table 45.1 shows the parameters used
in the PSO optimization.

45.3 Virtual Stock Market

We assume five players (agents) in the stock market. We also assume that these
players have no stocks at the beginning of the market. They just have US$5,000,000
and intend to buy stocks that would maximize their expected wealth in the market.
The players are free to buy and sell stocks in each round of the market. There are
530 R. Rajabioun, A. Rahimi-Kian

1,000,000 stocks assumed to be available from each company in our virtual stock
market (4,000,000 stocks in total). The only limitation imposed by the market is
the maximum number of stocks each player can buy or sell each day. This buy/sell
volume is limited to 1,000 stocks trading per day for each company. This constraint
is essential because if there is no limitation the whole stocks might be bought at
the beginning of the trading period by one of the players. This way there will be
no chance for other players to buy some of the stocks. Through the augmented term
added to the stock price models we can see the effect of each agent’s action (sell/buy
stocks) on the future prices and other agents’ wealth in the market.
We assume five players (agents) with different objective functions and different
strategies in the market, but we assume that all the agents have access to the stock
price models (developed in Section 45.2) symmetrically.
The players’ strategies are as follows:

Strategy of player 1:
This player buys (sells) the maximum number of allowed stocks when the prediction
shows an increase (decrease) in next 30 day prices compared to the average prices
of the last 10 days.

Strategy of player 2:
This player uses the Mean-Variance Analysis (MVA). He chooses the standard devi-
ation of the expected return (rp / as a measure of risk (p /. He plots the opportunity
set (efficient frontier) for a four-asset portfolio and takes an average risk and for an
average return each day. A sample opportunity set for a four-asset portfolio is shown
in Fig. 45.3.

Strategy of player 3:
This player believes in Random Walk Theory. He believes that the stock prices are
unpredictable and therefore, he buys and sells stocks randomly.

Fig. 45.3 A sample oppor-

tunity set for a four-asset
portfolio (dotted line: the
opportunity set with A and B
assets only)
45 A Dynamic Modeling of Stock Prices and Optimal Decision 531

Mean-Variance Efficient Frontier and Random Portfolios

0.08
Working Region of Player 5
0.07
Working Region of
0.06 Player 2
Expected Return

0.05

0.04
Some of Selectable
0.03 Points

Working
0.02
Region of
Player 4
0.01
0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
Risk (Standard Deviation)

Fig. 45.4 The working regions of players 2, 4 and 5 on the risk-return efficient frontier (the red
points can be selected in each trading day)

Strategy of player 4:
This player acts just like player 2. The only difference is in his risk averse behavior.
To reach the minimum risk, in each stage, he selects the buy/sell weights on the knee
of efficient frontier curve.

Strategy of player 5:
This player also acts like player 2 with the difference that this player is risk lover.
Therefore, in each stage this player selects the buy/sell weights with the maximum
risk and maximum expected return.
The working regions of players 2, 4 and 5 on the risk-return efficient frontier are
shown in Fig. 45.4.
These five players buy and sell in the virtual stock market. In the related litera-
ture it is usually seen that short-sales are disregarded when optimizing the players’
objective functions and the optimization is just done through stock purchases. How-
ever, in this paper we have relaxed this constraint and allowed the players to buy
and sell their stocks when needed.
In the following, we define the objective functions for all players and demonstrate
their optimization process. For players 2, 4 and 5 that the risk values are important
in their decisions, we define their objective function as follows:

Ei D E.rpi /  .1  /pi ; i D 2; 4; 5 (45.3)

where Ei : is the Expected return of player i , : is a constant between 0 and 1.

In fact, this is a weight that shows the relative importance of the expected return
532 R. Rajabioun, A. Rahimi-Kian

(E.rp // versus the risk (p / of player-i , for  D 1 the risk term disappears form
the objective function of player-i: Ei D E.rpi /. In our market simulation studies, we
chose  D f0; 0:5; 1g for players f2, 4, 5g respectively according to their defined
risk behaviors in the market. The players’ objective functions were optimized with
respect to their decision variables (stock sell/buy actions) using the Particle Swarm
Optimization method and the results are presented and analyzed in the following
section.

45.4 The Market Simulation Results

The market simulation results for the five players are presented and analyzed in
this section. Figures 45.5 and 45.6 show the optimal buy/sell actions for players 1
and 5 for each company’s stock (ERIC, IBM, PHG and SNE). The optimal buy/sell
actions for players 2, 3 and 4 are shown in the appendix. In these figures, the buy
actions are positive, sell actions are negative and no-actions are zero.
If the buy action gets +1 (or 1), then the player, should buy (or sell) the max-
imum number of stocks allowed for that company. In Fig. 45.7, the wealth of each
player is shown for 1 trading year period. The wealth is measured as the values of
the in-hand stocks added to the cash amount in-hand for each player.
As can be seen from Fig. 45.7, players 1 and 5 have done better than the rest
of them in terms of the wealth maximization for 1 year stock trading. Figure 45.8
shows the expected risk values in each trading day for each player. As we expected,
player 1 has the minimum expected risk over the trading period and has also
obtained the maximum return from the market.

Player 1 Trade weights for each portfolio assets

1
ERIC
0
-1
0 100 200 300 400
1
IBM
0
-1
0 100 200 300 400
1
PHG
0
-1
0 100 200 300 400
1
SNE
0
-1
0 100 200 300 400
Trading year

Fig. 45.5 The optimal trading actions of player 1 for all companies’ stocks
45 A Dynamic Modeling of Stock Prices and Optimal Decision 533

Player 5 Trade weights for each portfolio assets

1
ERIC
0
-1
0 100 200 300 400
0.5
IBM
0
-0.5
0 100 200 300 400
1
PHG
0
-1
0 100 200 300 400
1
SNE
0
-1
0 100 200 300 400
Trading year

Fig. 45.6 The optimal trading actions of player 5 for all companies’ stocks

x 107 Treasure profile for each player

2.5
player 1
player 2
2
Treasure in hand ($)

player 3
player 4
1.5 player 5

0.5

0
0 100 200 300 400
Trading year

Fig. 45.7 The wealth of all players for all days of the trading year

Its strategy was to buy/sell maximum stocks with respect to the comparison of
the predicted future-prices’ trends with those of the moving average 10-day-before
prices. Since the GP prediction models had small prediction errors for the test data,
this player did well in the market by relying on the prediction results.
Among players 2, 4 and 5 (referring to Fig. 45.7), we can see that player 5 with
the maximum risk level has made the most wealth (expected returns) and stands in
the second rank (after player 1) in terms of market returns. Player 3’s strategy was
to buy and sell randomly; by referring to Figs. 45.7 and 45.8, one can see that his
expected returns are similar to those for player 2 (see the Fig. 45.9, 45.10 and 45.11
in the Appendix) but, his expected risks values are more than other players.
534 R. Rajabioun, A. Rahimi-Kian

Risk Values (%)

40
P1
20
0
0 100 200 300 400
40
P2

20
0
0 100 200 300 400
40
P3

20
0
0 100 200 300 400
40
P4

20
0
0 100 200 300 400
40
P5

20
0
0 100 200 300 400
Trading Year

Fig. 45.8 The expected risk values for each player over the 1 trading year (P1 to P5 from top to
bottom graphs)

45.5 Conclusion

In this paper, precise price predictive models were obtained for four companies’
stocks using the GP. This model incorporated the effects of the players’ actions on
the stock prices and other players’ wealth. After the GP model was verified (using
the test data), it was used for making sell/buy decisions by five agents that traded
in a virtual stock market. The trading period was considered 1 year for our market
simulation studies. Five different strategies and objective functions were defined
for the market trading agents (with different risk attitudes). The PSO algorithm was
used to obtain the optimal buy/sell actions for each player in order to maximize their
objective functions (expected returns). The players’ strategies and their expected
risk-returns were obtained and analyzed for the 1 year trading period. Our market
simulation studies showed that the player (P1) was the most successful one in our
virtual stock market.
45 A Dynamic Modeling of Stock Prices and Optimal Decision 535

Appendix

Player 2 Trade weights for each portfolio assets

1
ERIC
0
-1
0 100 200 300 400
1
IBM
0

-1 0 100 200 300 400

1
PHG
0
-1
0 100 200 300 400
1
SNE
0
-1
0 100 200 300 400
Trading year

Fig. 45.9 The optimal trading actions of player 2 for all companies’ stocks

Player 3 Trade weights for each portfolio assets

0.5
ERIC
0
-0.5
0 100 200 300 400
1
IBM
0
-1
0 100 200 300 400
0.5
PHG
0
-0.5
0 100 200 300 400
0.5
SNE
0
-0.5
0 100 200 300 400
Trading year

Fig. 45.10 The optimal trading actions of player 3 for all companies’ stocks
536 R. Rajabioun, A. Rahimi-Kian

Player 4 Trade weights for each portfolio assets

1
ERIC
0
-1
0 100 200 300 400
0.5
IBM
0
-0.5
0 100 200 300 400
1
PHG
0
-1
0 100 200 300 400
1
SNE
0
-1
0 100 200 300 400
Trading year

Fig. 45.11 The optimal trading actions of player 4 for all companies’ stocks

References

1. R. Thalheimer and M. M. Ali, Time series analysis and portfolio selection: An application to
mutual savings banks, Southern Economic Journal, 45(3), 821–837 (1979).
2. M. Pojarliev and W. Polasek, Applying multivariate time series forecasts for active portfolio
management, Financial Markets and Portfolio Management, 15, 201–211 (2001).
3. D. S. Poskitt and A. R. Tremayne, Determining a portfolio of linear time series models,
Biometrika, 74(1), 125–137 (1987).
4. T. Basar and G. J. Olsder, Dynamic Noncooperative Game Theory. (San Diego, CA:
Academic, 1995).
5. J. W. Weibull, Evolutionary Game Theory. (London: MIT Press, 1995).
6. D. Fudenberg and J. Tirole, Game Theory. (Cambridge, MA: MIT Press, 1991).
7. S. Lakshminarayanan, An Integrated stock market forecasting model using neural network,
M. Sc. dissertation, College of Engineering and Technology of Ohio University (2005).
8. K. J. Kim and W. B. Lee, Stock market prediction using artificial neural networks with optimal
feature transformation, Journal of Neural Computing & Applications, Springer, 13(3), 255–
260 (2004).
9. K. Schierholt and C. H. Dagli, Stock market prediction using different neural network classi-
fication architectures, Computational Intelligence for Financial Engineering, Proceedings of
the IEEEIAFE Conference, pp. 72–78 (1996).
10. K. Nygren, Stock Prediction – A Neural Network Approach, M.Sc. dissertation, Royal
Institute of Technology, KTH (2004).
11. E. P. K. Tsang, J. Li, and J. M. Butler, EDDIE beats the bookies, International Journal of
Software, Practice and Experience, 28(10), 1033–1043 (1998).
12. D. S. Barr and G. Mani, Using neural nets to manage investments, AI EXPERT, 34(3), 16–22
(1994).
13. R. N. Mantegna and H. E. Stanley, An Introduction to Econophysics: Correlation and
Complexity in Finance (Cambridge University Press, Cambridge, MA, 2000).
14. J. P. Bouchaud and M. Potters, Theory of Financial Risks: From Statistical Physics to Risk
management (Cambridge University Press, Cambridge, MA, 2000).
15. Y. K. Kwon, S. S. Choi, and B. R. Moon, Stock Prediction Based on Financial Correlation, In
Proceedings of GECCO’2005, pp. 2061–2066 (2005).
45 A Dynamic Modeling of Stock Prices and Optimal Decision 537

16. Y. K. Kwon and B. R. Moon, Daily stock prediction using neuro-genetic hybrids, Proceedings
of the Genetic and Evolutionary Computation Conference, pp. 2203–2214 (2003).
17. Y. K. Kwon and B. R. Moon, Evolutionary ensemble for stock prediction, Proceedings of the
Genetic and Evolutionary Computation Conference, pp. 1102–1113 (2004).
18. F. Parisi and A. Vasquez, Simple technical trading rules of stock returns: evidence from 1987
to 1998 in Chile, Emerging Market Review, 1, 152–64 (2000).
19. J. Koza, Genetic Programming: On the Programming of Computers by Means of Natural
Evolution. (Cambridge, MA: MIT Press, 1992).
20. J. Madar, J. Abonyi, and F. Szeifert, Genetic Programming for the identification of nonlinear
input-output models, Industrial Engineering Chemistry Research, 44, 3178–3186 (2005).
21. M. A. Kaboudan, Genetic Programming prediction of stock prices, Computational Economics,
16(3), 207–236 (2000).
22. S. Sette and L. Boullart, Genetic Programming: principles and applications, Engineering
Applications of Artificial Intelligence, 14, 727–736 (2001).
23. Boston Stock Group web page (August 1, 2007); http://boston.stockgroup.com
24. J. Kennedy and R. Eberhart, Particle Swarm Optimization, Proceedings of the IEEE Interna-
tional Conference on Neural Networks, Perth, Australia, pp. 1942–1945 (1995).
Chapter 46
A Regularized Unconstrained Optimization
in the Bond Portfolio Valuation and Hedging

Yury Gryazin and Michael Landrigan

Abstract In this chapter, a numerical approach to the valuation and hedging of a

portfolio of bond and options is presented in the case of strong dependency of bond
principal on the market interest rate metric. Collateralized Mortgage Obligations
(CMO) represents one of the important classes of such bond type. To improve the
CMO valuation methodology and to develop a robust procedure for the construction
of the optimal hedge for CMO, we introduce an optimization approach to mini-
mize the dispersion of the portfolio spread distribution by using available on market
options. In doing so, we design an optimal hedge with respect to the set of available
benchmarks and we obtain the two new valuation metrics that represent the quality
of hedge with respect to this set. Our two main outputs are the mean and the stan-
dard deviation of the individual spreads for the optimal portfolio. These metrics can
be used in comparison analysis in addition to the standard OAS valuation.

Keywords Regularized Unconstrained Optimization  Bond Portfolio  Valuation 

Hedging  Collateralized Mortgage Obligation

46.1 Introduction

This paper presents a numerical approach to the valuation and hedging of a portfolio
of bond and options in the case of strong dependency of bond principal on the market
interest rate metric. Collateralized Mortgage Obligations (CMO) represent one of
the important classes of such bond type. CMOs can have a high degree of variability
in cash flows. Because of this, it is generally recognized that a yield to maturity of

Y. Gryazin (B)
Department of Mathematics, Idaho State University, Pocatello, ID 83209, USA,
E-mail: gryazin@isu.edu

The work of this author was supported in part by Mexican Consejo Nacional de Ciencia y
Tecnologia (CONACYT) under Grant # CB-2005-C01-49854-F.

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 539
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
540 Y. Gryazin, M. Landrigan

static spread calculation is not a suitable valuation methodology. Since the 1980s
Option-Adjusted Spread (OAS) has become a ubiquitous valuation metric in the
CMO market. There have been many criticisms of OAS methodology, and some
interesting modifications have focused on the prepayment side of the analysis, e.g.,
[1,2]. One of the problems with using OAS analysis is the lack of information about
the distribution of the individual spreads, which in turn leads to the difficulties in
the construction of the hedging portfolio for CMO.
To improve the CMO valuation methodology and to develop a robust procedure
for the construction of the optimal hedge for CMO, we introduce an optimization
approach to minimize the dispersion of the portfolio spread distribution by using
available on market options. In doing so, we design an optimal hedge with respect
to the set of available benchmarks and we obtain the two new valuation metrics that
represent the quality of hedge with respect to this set. Our two main outputs are the
mean and the standard deviation of the individual spreads for the optimal portfolio.
These metrics can be used in comparison analysis in addition to the standard OAS
valuation.
This new methodology can lead to quite different conclusions about CMO than
OAS does. In particular, in comparing two bonds, the more negatively convex bond
may look cheaper on an OAS basis, but be more attractive according to our analysis
since we provide a way to estimate the quality of available hedge.
The main difficulty in implementing our new methodology is in the minimization
of the spread-variance functional. The difficulty is partly because the optimization
problem is ill-conditioned, and in many situations, requires introduction of some
type of regularization. Our approach is to use the standard Tikhonov regularization
[3], which has the strong intuitive appeal of limiting the sizes of benchmark weights
used in hedging.
A word about static versus dynamic hedging may be in order. Our methodology
is to set up a static hedge and conduct bond valuation based on the static optimal
portfolio. It may be argued that an essentially perfect hedge may be created dynam-
ically. However, a dynamically-hedged portfolio will not have the OAS as a spread.
Moreover, ones hedging costs will be related to the amount of volatility in the future
so can be quite uncertain. Our methodology greatly reduces dynamic hedging costs
by setting up an optimized static hedge, and thus reduces the uncertainty in dynamic
hedging costs.
The rest of the paper is organized as follows. In the second section we briefly
review OAS and point out in more detail the problems we see with it. In the third
section we explicitly define our hedging methodology. In the forth section we give a
brief description of the regularized numerical method. In the fifth section we present
some details on the term-structure model used; on our prepayment assumptions and
summarize numerical results from our analysis.
46 A Regularized Unconstrained Optimization 541

46.2 Option-Adjusted Spread Analysis

Standard OAS analysis is based on finding a spread at which the expected value
of the discounted cash flows will be equal to the market value of a CMO. This is
encapsulated in Eq. (46.1):

X
L
M VCMO e
DE d.ti ; s/  c.ti /: (46.1)
i D1

Here M VCMO is the market value of CMO, E e denotes expectation with respect to
the risk neutral measure, d.ti ; s/ is the discount factor to time ti with a spread of
s, and c.ti / is the cash flow at time ti , for i D 1; 2; : : : ; L. Note that in the OAS
framework, a single spread term is added to the discounted factors to make this
formula a true equality, this spread, s, is referred to as the OAS.
The goal of OAS analysis is to value a CMO relative to liquid benchmark
interest rate derivatives, and, thus, the risk-neutral measure is derived to price
those benchmarks accurately. To calculate expected values in practice one uses
Monte-Carlo simulation of a stochastic term-structure model. We use the two-factor
Gaussian short-rate model G2++ as in [4] calibrated to U.S. swaption prices, but
other choices may be suitable. In terms of numerical approximation, it will give us
Eq. (46.2):
( L )
1 XN X Y t
1
k
M Vbench D k
cfbench .n; t/ C Err k : (46.2)
N nD1 t D1
1 C ti  r.n; ti /
i D1

Here ti D ti  ti 1 , i D 1; : : : ; L, t0 D 0, M Vbenchk

, k D 1; : : : ; M are the
market values of the benchmarks, cfbench .n; ti /, n D 1; : : : ; N; i D 1; : : : ; L, k D
k

1; : : : ; M are the future cash flows of the benchmarks, N is the number of generated
trajectories, L is the number of time intervals until expiration of all benchmarks
and CMO, and M is the number of benchmarks in the consideration. The last term
Err k represents the error term. Though, the detailed consideration of calibration
procedure is outside the scope of this presentation, it is worth to mention that the
absolute value of the Err k term is bounded in most of our experiments by five basis
points.
The second step in the OAS analysis of CMOs is to find the spread term from the
Eq. (46.3):
8 9
XN <X i Y
i =
1 1
M VCMO D cf .n; ti / : (46.3)
N : 1 C tj  .r.n; tj / C s/ ;
nD1 i D1 j D1

Here cf .n; ti /, n D 1; : : : ; N , i D 1; : : : ; L, are cash flows of the CMO. These

cash flows come from the structure of the CMO, a perfectly known quantity, and a
prepayment model, a more subjectively known quantity. The parameter s, the OAS,
542 Y. Gryazin, M. Landrigan

is an indicator as to whether the CMO is underpriced or overpriced: If the OAS is

positive then the CMO is underpriced, if it is negative then the CMO is overpriced.
Not only its sign, but also the magnitude of the OAS commonly quoted as a measure
of cheapness of a CMO.
An important issue is managing a portfolio of CMOs is how to hedge the portfolio
by using actively traded benchmarks. We return to this question little bit later but for
now let’s assume that we somehow have found the list and corresponding weights of
benchmarks that would provide a sufficient hedge for our portfolio. To evaluate this
portfolio, we will extend the OAS analysis to include the calculation of spread for
the portfolio of CMO and the benchmarks. This straightforward extension of OAS
approach will result in Eq. (46.4):
( L " #
XM
1 X X
N XM
M VCMO C w M Vbench D
k k
cf .n; ti / C wk  cfbench .n; ti /
k
N nD1
kD1 i D1 kD1
9
Yi
1 =
 : (46.4)
1 C tj  .r.n; tj / C s/ ;
j D1

We will see in more detail later how consideration of the OAS of various portfolios
makes some serious drawbacks of OAS analysis apparent. Some of the drawbacks
are:
1. It matches a mean to market value, but provides no information on the distribu-
tion of the individual discounted values.
2. One can use it to calculate some standard risk metrics, but it gives no way to
design a refined hedge.
3. It is sensitive to positions in your benchmarks.
In our approach an affective hedging strategy is proposed and new valuation metrics
are considered.

46.3 The Spread Variance Minimization Approach

In our proposed approach, instead of using the same spread for all paths, we are
looking at the individual spreads for every path of the portfolio of CMO and bench-
marks. The dispersion of the distribution can be considered a measure of risk or
alternately as a measure of the quality of a hedge. In a classical mean-variance
portfolio optimization problem (see, e.g., [5]) there are two parts of the formula-
tion usually considered: (1) the task of the maximizing the mean of the portfolio
return under a given upper bound for the variance; (2) the task of minimizing
the variance of the portfolio return given a lower bound on the expected portfolio
return. We approach the problem of hedging CMOs with swaptions from a simi-
lar point of view, but focus on the distribution of spreads. The goal is to apply an
46 A Regularized Unconstrained Optimization 543

optimization algorithm to find weights so that the variance of the s.n; w1 ; : : : ; wM /

is a minimum.
The spread for path n is the value s.n; w1 ; : : : ; wM / such that Eq. (46.5) satisfied:

X
M
M VCMO C k
wk M Vbench D
kD1
" #
X
L X
M
k
cf .n; ti / C wk  cfbench .n; ti / (46.5)
i D1 kD1

Y
i
1
:
1 C ti  .r.n; tj / C s.n; w1 ; : : : ; wM //
j D1

Here wk ; k D 1; : : : ; M are the weights of the individual benchmarks (a negative

indicates a short position).
The goal of our analysis is to minimize the variance of the individual spreads. In
this form the problem is not well defined; the weights may exhibit some instability
and tend to drift to infinity. A common approach to the solution of this type of
problem is to add regularization term to the target functional. Instead of minimizing
only variance, we will use Tikhonov regularization [3]. We introduce this in more
detail in the next section.

46.4 Numerical Method

As mentioned before, we are considering that for every interest rate trajectory the
individual spread depends on the weights for benchmarks in the portfolio. Then the
target functional is defined in Eq. (46.6):

1 X
N
f .w1 ; : : : ; wn / D s.n; w1 ; : : : ; wn /2  2 : (46.6)
N nD1
P
Here  D N1 N nD1 s.n; w1 ; : : : ; wM /. The Jacobian and Hessian of the functional
are given by Eq. (46.7):

2 X
N
@f @s.n; w1 ; : : : ; wn /
D .s.n; w1 ; : : : ; wn /  / ; i D 1; : : : ; M;
@wi N nD1 @wi
(46.7)
@2 f 2 X
N
@s @s @2 s
D  C .s  / (46.8)
@wi @wl N nD1 @wi @wl @wi @wl
( ) ( )
1 X @s 1 X @s
N N
2   ; i; l D 1; : : : ; M:
N nD1 @wi N nD1 @wl
544 Y. Gryazin, M. Landrigan

Using implicit differentiation one can find @s=@wi ; @2 s=.@wi @wl /, i; l D 1; :::; M .
But this functional in general is ill-conditioned. To ensure the convergence of the
optimization method, we introduce standard Tikhonov regularization. The modified
target functional could be presented as in Eq. (46.9):

fQ.w1 ; : : : ; wM / D f .w1 ; : : : ; wM / C ˛jjwjj22 : (46.9)

In most situations this guaranties the convergence of the numerical method to the
unique solution. In our approach we are using the optimization technique based
on the combination of the Broyden-Fletcher-Goldfarb-Shanno (BFGS) (see, e.g.,
[6]) and Newton methods. The BFGS approach is applied on the early iterations to
ensure the convergence of Newton’s method to a minimum. When the l2 norm of
the gradient of the target function becomes less then 1010 we apply the Newton
method assuming that the approximation is already close enough to the solution and
the quadratic convergence rate of the Newton method can be achieved.
As we already mentioned, the regularization keeps the value of the target func-
tional small and stabilizes the optimization method by preventing the weights of the
benchmarks in the portfolio wi ; i D 1; : : : ; M from becoming too large. To keep the
condition number of the Hessian bounded, one has to use fairly large regularization
parameter ˛. On the other hand, in order to keep the regularized problem reasonably
close the original one, we expect that ˛ needs to be small. In addition to these pure
mathematical requirements, in the case of managing a portfolio, one has to take into
consideration the cost of the hedging. Since regularization term prevents the optimal
weights from drifting to infinity, the regularization parameter becomes a desirable
tool in keeping hedging cost under control. In our experiments we found that the
parameter ˛ D 109 represents a good choice.

46.5 Numerical Results

To illustrate our proposed methodology we consider hedging an unstructured trust

IO (interest only strip) with a basket of European swaptions of different strikes and
expiration dates, all with 10-year tenor. To generate the trajectories of the short
rate, we use the two-additive-factor Gaussian model G2++ [4]. The dynamics of the
instantaneous-short-rate process under the risk-neutral measure are given by r.t/ D
x.t/Cy.t/C'.t/, r.0/ D r0 , where the processes fx.t/ W t 0g and fy.t/ W t 0g
satisfy Eq. (46.10):

dx.t/ D ax.t/dt C d W1 .t/; x.0/ D 0;

(46.10)
dy.t/ D by.t/dt C d W2 .t/; y.0/ D 0:

Here .W1 ; W2 / is a two-dimensional Brownian motion with instantaneous corre-

lation W d W1 .t/d W2 .t/ D dt. The parameters a; b; ; ; are defined in the
calibration procedure to match the prices of a set of actively traded European
46 A Regularized Unconstrained Optimization 545

swaptions. The l2 norm of the difference vector of the model swaption prices and
the market prices in our experiments is less than five basis points.
A key component in valuing mortgage bonds is to calculate the principal bal-
ance. This calculation is based on normal amortization and prepayment. In order to
calculate amortization rates, we assume that the loans backing the bond have some
specific parameters. Those parameters can take a wide range of values, but here we
assume the following. The loan age (WALA) is 60 months, the interest rate on the
loans is (WAC) is 6%. We also assume that the coupon on the bond is 5.5%, and it
has a price of $22.519. The bond under consideration has the following parameters:
WALA = 60 months, WAC = 6%, coupon = 5.5%, and a price of $22.519. We are
using very simple prepayment model defined by linear interpolation of the data from
Table 46.1.
For optimization we use a basket of 64 payer and receiver European swaptions
with yearly expiration dates from years 1 to 16, including at the money (ATM) pay-
ers and receivers, C50 basis points out of the money payers and 50 basis points
out of the money receivers. The regularization parameter used is 109 , and the num-
ber of trajectories in all experiments was 500. The SVM numerical method was
implemented in Matlab with using the BFGS standard implementation available in
the Matlab optimization toolbox. The computer time for the Matlab optimization
routine was 42 s on the standard PC with 2 Hz processor frequency.
Figure 46.1 represents the convergence history of the iterations in the optimiza-
tion method, showing the log of the norm of the gradient.
Figure 46.2 shows the square root of the target functional, which is approximately
the standard deviation of the spread distribution. One can see that as a result of the
application of the new methodology, the standard deviation of the spread distribu-

Table 46.1 Prepayment rates

Interest rates (%) 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0
CPR (%) 70 50 30 15 10 8 4 3 3

Fig. 46.1 Convergence

history
546 Y. Gryazin, M. Landrigan

Fig. 46.2 Convergence of

target functional

Fig. 46.3 Evolution of

weights

tion is reduced significantly. This can be considered as reduction in riskiness of the

portfolio of CMO and benchmarks.
Figure 46.3 presents the evolution of weights in our portfolio.
Figure 46.4 presents the distribution of the spreads for the original portfolio on
the left and the optimal portfolio on the right.
In the next series of experiments, we illustrate the quality of the hedge con-
structed using different baskets of swaptions. Table 46.2 presents the results of these
experiments. In our experiments we used the IO strip in combination with different
numbers of swaptions. For reference we include results with no hedge at all; this
is the common OAS analysis (but we also include the mean spread). For our SVM
approach, we first use just two ATM swaptions, one payer and one receiver, both
with expiration date of 1 year. Then we use eight swaptions with expiration dates
of 1 and 2 years. For each of these expirations we include an ATM payer, an ATM
receiver, an ATM C50 bps payer, and an ATM-50 bps receiver.
The test with twenty options uses expiration dates 1–5 years, again with four
swaptions per expiration date. In the last experiment we use swaptions with
46 A Regularized Unconstrained Optimization 547

Fig. 46.4 Spread distributions

Table 46.2 Hedging results

Number of swaptions 0 2 8 20 64
Cost of hedge($) 0 –0.002 5.628 60.14 11.361
 (bsp) 900 614 407 224 113
Mean spread (bsp) –71 –18 37 49 33
OAS (bsp) 127 101 95 68 40

expiration dates 1–16. In all of these experiments we used one unit of the trust
IO with the cost of $22.519.
As we can see from the first line of the table, the cost of the hedge is an increasing
function of the number of options. But the most importantly, we manage to signif-
icantly decrease the standard deviation of the spread distribution by using spread
variance minimization methodology.
Notice that, when using just two options, the cost of our hedging portfolio is
essentially zero, we are buying the ATM receiver swaption and selling an equal
amount of the ATM payer swaption. This is effectively entering into a forward rate
agreement, and so this case could be considered as a hedging strategy of offsetting
the duration of the CMO. The experiments with the larger sets of swaptions refine
this strategy and take into account more detail about the structure of the cash flows
of the bond. It proves to be a very successful approach to the hedging of the path-
dependent bond.
The last two rows of the table present two different metrics: The mean spread and
the OAS of the portfolio of bond and hedges. We can see that they become close as
standard deviation decreases. In fact, they would be the same if the standard devia-
tion becomes zero. With no swaptions, or with very few of them, these metrics can
result in drastically different conclusions about the cheapness of the bond. Contrary
to a common interpretation of OAS, it does not represent the mean spread of an
unhedged bond. In fact, it can be expected to be close to the mean spread only when
548 Y. Gryazin, M. Landrigan

Table 46.3 Results for short receiver, long payer

Number of payer/ 0 1 10 100 500
receiver pairs
OAS (bps) 122 123 125 149 267
 (bps) 893 907 968 1,636 5,363
OASopt 44 44 44 44 45
sNopt 37 37 38 38 38
opt 133 133 133 133 138

considering a refined hedging portfolio of swaptions. Though the discussion of the

advantages of the SVM analysis is outside of the scope of this paper, it is worth
mentioning that it caries significant information about the path-dependent bond and
is worth taking into account alongside the standard OAS parameter.
In the last set of experiments, we see that we can increase the OAS of a portfolio,
even without changing its cost. We include in our portfolio the IO, a long position
in a payer, and a short position in a receiver, with the swaptions being ATM and
having an expiration of 1 year. Table 46.3 presents the results of experiments with
the portfolio of an IO strip and different amounts of the swaptions. We can see
that we can actually increase the OAS (second row in Table 46.3) by increasing the
number of fairly priced payer/receiver combinations, but it comes with the price of
increasing the standard deviation (third row in the Table 46.3) of the spreads for this
portfolio. Usually this information is missing in OAS analysis, and it could give a
skewed impression about the cheapness of a portfolio. It should be expected that a
more risky portfolio will have a better return, but in normal OAS analysis there is no
assessment of the tradeoff between risk and return. So the idea is to try first decrease
the standard deviation of the spread distribution as much as possible by using in the
hedge the same type of swaptions as in original portfolio, and only then use the OAS
or/and mean spread constant as a metric for the cheapness of the portfolio.
As indicated in the second row of Table 46.3, the OAS of the portfolio may
increase dramatically, even though we are not changing the cost of setting up or
portfolio, i.e., strictly speaking there is no leveraging. On the other hand, rows three,
four and five show that the optimized portfolio of swaptions, from SVM analysis
remains stable. We can see that OAS and the mean of spreads stay the same in
all experiments and serves as a better metric for the cheapness or richness of the
portfolio.

46.6 Conclusions

In this paper we presented a regularization approach to the construction of an opti-

mal portfolio of CMO and swaptions. The standard Tikhonov regularization term
serves as an important tool for preventing the weights of the benchmarks in the
optimal portfolio drift to the infinity and so keeps the hedging procedure practical.
46 A Regularized Unconstrained Optimization 549

The optimization method demonstrates excellent convergent properties and could be

used in practical applications for hedging a portfolio of CMO. The numerical results
demonstrate the effectiveness of the proposed methodology. The future development
may include the construction of new target functional based on the combination
of the spread variance and cost function. This modification might improve the
efficiency of the developed hedging strategy.
There appears to be a wide variety of application of our SVM approach, and in
general of systematic path-by-path analysis of securities with stochastic cash flows.
We plan to apply our analysis to other CMO structures directly. In addition, our
analysis lends itself to comparing structured CMO with liquid strip interest only
and principal only bonds.

References

1. Cohler, G., Feldman, M. & Lancaster, B., Price of Risk Constant (PORC): Going Beyond OAS.
The Journal of Fixed Income, 6(4), 1997, 6–15.
2. Levin, A. & Davidson, A., Prepayment Risk and Option-Adjusted Valuation of MBS. The
Journal of Portfolio Management, 31(4), 2005.
3. Tikhonov, A. N., Regularization of incorrectly posed problems. Soviet Mathematics Doklady,
4, 1963, 1624–1627.
4. Brigo, D. & Mercurio, F., Interest Rate Models: Theory and Practice. Berlin Heidelberg,
Springer-Verlag, 2001, pp. 132–165.
5. Korn, R. & Korn, E., Option Pricing and Portfolio Optimization. AMS, Providence, Rhode
Island, V. 31, 2000.
6. Vogel, C. R., Computational methods for inverse problems. SIAM, Philadelphia, PA, 2002.
Chapter 47
Approximation of Pareto Set in Multi Objective
Portfolio Optimization

I. Radziukyniene and A. Zilinskas

Abstract In this chapter we experimentally investigate several evolutionary multi

objective optimization methods and compare their efficiency in problems of portfo-
lio selection with the efficiency of specially tailored method of adjustable weights.
Test problems were based on standard portfolio quality criteria, and data on stocks
of ten Lithuanian companies. We do not concern here in mach between analytical
properties of the criteria functions and such properties favorable for the consid-
ered methods; we believe, however that general (global) structure of multi-criteria
portfolio selection problem will be invariant with respect to switching from criteria
defined by simple analytical formula to criteria defined by complicated numerical
methods.

Keywords Multi Objective Portfolio Optimization  Pareto Set  Approximation 

portfolio selection  evolutionary optimization

47.1 Introduction

Many computational finance problems ranging from asset allocation to risk man-
agement, from option pricing to model calibration can be solved efficiently using
modern optimization techniques. The question of optimal portfolio allocation has
been of long-standing interest for academics and practitioners in finance. In 1950s
Harry Markowitz published his pioneering work where he has proposed a simple
quadratic program for selecting a diversified portfolio of securities [1]. His model
for portfolio selection can be formulated mathematically either as a problem of max-
imization of expected return where risk, defined as variance of return, is (upper)

I. Radziukyniene (B)
Department of Informatics, Vytautas Magnus University, 8 Vileikos str.,
Kaunas LT 44404, Lithuania,
E-mail: i.radziukyniene@if.vdu.lt

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 551
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
552 I. Radziukyniene, A. Zilinskas

bounded or as a problem of minimization of risk where expected return is (lower)

bounded. The classical Markowitz approach to portfolio selection reduces the prob-
lem of two criteria optimization to a one criterion optimization where the second
criterion is converted to a constraint. Reduction of a multi-criteria problem to one
criterion problem not always is the best method to solve multi-criteria problems
especially in the case of vague a priory comparability of criteria. In such problems
some evaluation of a whole Pareto set is of interest. It is the case for a portfolio
selection problem formulated as a multi-criteria optimization problem where com-
promise between criteria crucially depends on the subjective priorities of a decision
maker. Therefore Pareto set approximation is an important method facilitating ratio-
nal portfolio selection. We have implemented a method of adjustable weights for
Pareto set approximation where the classical scalarization idea of weighted summa-
tion of criteria is enhanced using branch and bound type procedure. To attack the
problem of Pareto set approximation evolutionary optimization methods are claimed
very promising; see e.g. [2–7]. In this paper we experimentally investigate several
evolutionary multi objective optimization methods and compare their efficiency in
problems of portfolio selection with the efficiency of specially tailored method of
adjustable weights. Test problems were based on standard portfolio quality criteria,
and data on stocks of ten Lithuanian companies. We do not concern here in mach
between analytical properties of the criteria functions and such properties favorable
for the considered methods; we believe, however that general (global) structure of
multi-criteria portfolio selection problem will be invariant with respect to switching
from criteria defined by simple analytical formula to criteria defined by compli-
cated numerical methods. The paper is organized as follows. In Section 47.2 the
multi-objective portfolio optimization problem is outlined, Sections 47.3 and 47.4
describe the considered multi objective optimization methods and their characteris-
tics. In Sections 47.5 and 47.6 we discuss the used performance metrics and the
experimental results. The paper is completed with conclusions on properties of
the considered methods relevant to their applicability to multi objective portfolio
selection.

47.2 Multi-Objective Portfolio Optimization Problem

Risk plays an important role in modern finance, including risk management, capital
asset pricing and portfolio optimization. The problem of portfolio selection can be
formulated as the problem to find an optimal strategy for allocating wealth among a
number of securities (investment) and to obtain an optimal risk-return trade-off. The
portfolio optimization problem may be formulated in various ways depending on
the selection of the objective functions, the definition of the decision variables,
and the particular constraints underlying the specific situation [7–10]. Beyond the
expected return and variance of return, like in Markowitz portfolio model [1],
the additional objective function can include number of securities in a portfo-
lio, turnover, amount of short selling, dividend, liquidity, excess return over of
47 Approximation of Pareto Set in Multi Objective Portfolio Optimization 553

a benchmark random variable and other [7]. In the bank portfolio management,
the additional criteria such as the prime rate, processing cost, expected default
rate, probability of unexpected losses, quantity of the long-term and short-term
can be considered [8]. For example, the multi-objective portfolio selection problem
can include the following objectives [9]: (to be maximized) portfolio return, divi-
dend, growth in sales, liquidity, portfolio return over that of a benchmark, and (to
be minimized) deviations from asset allocation percentages, number of securities
in portfolio, turnover (i.e., costs of adjustment), maximum investment proportion
weight, amount of short selling. We considered two multi-objective portfolio prob-
lems. The first problem was based on a simple two objectives portfolio model
including the standard deviation of the returns and mean of the returns, where the
return Ri is 1 month return of stock i ; return means percentage change in value. The
second problem included three objectives, where annual dividend yield is added
to two above mentioned objectives. For the experiment we used a data set of ten
Lithuanian companies’ stock data from Lithuanian market.
Financial portfolio selection is one of real world decision making problems with
several, often conflicting, objectives which can be reduced to multi objective opti-
mization. There are many methods to attack multi objective optimization problems.
In some cases a solution to a multi objective problem can be defined as a special
point of Pareto set by means of scalarization, i.e. by converting the initial multi
objective problem to a single criterion problem. However, in many cases decision
can not be made without some information about the whole Pareto set. The the-
oretical problem to find the whole Pareto set normally (e.g. in case of continuum
cardinality of Pareto set) can not be solved algorithmically, thus the theoretical
problem of finding of whole Pareto set should reformulated to algorithmic construc-
tion of an appropriate its approximation. In the present paper we consider several
algorithms of Pareto set approximation including the newly proposed method of
adjustable weights and some evolutionary algorithms well assessed in recent publi-
cations. Over the past decades evolutionary algorithms have received much attention
owing to its intrinsic ability to handle optimization problems with both single and
multiple objectives including problems of financial optimization [2–7]. Let us note
that financial optimization problems were attacked also by means of other heuris-
tics, e.g. by simulated annealing in [3], and by Tabu search in [8]. Comparisons of
the performance of different heuristic techniques applied to solve one criterion port-
folio choice problems are given by Chang et al. [3]. However, to the best knowledge
of the authors’ similar comparative analysis of performance of recently proposed
evolutionary multi-criteria algorithms has not been yet reported.

47.3 Multi-Objective Optimization by the Method

of Adjustable Weights

In this paper we investigated approximation of Pareto set by means of a finite set of

points uniformly distributed in close vicinity of the Pareto set; the terms “uniformly”
and “close vicinity” are defined more precisely in the section on experimental
results.
554 I. Radziukyniene, A. Zilinskas

First method that attracted our attention was the widely used one scalarization
method of weighted criteria summation. Besides of the status of “classics” of multi
objective optimization, an important argument to consider the method of weighted
criteria summation is simplicity of its implementation. Since a multi criteria prob-
lem is converted into a single criterion optimization problem where the objective
function is a weighted sum of the criteria functions, weights can be used to express
the relative significance of different criteria. For some multi objective optimization
problems it is convenient to take into account in this way the preferences of decision
makers. However, we do not assume availability of information about importance
of criteria, and for us weights play the role of parameters defining points in the
Pareto set corresponding to the solutions of parametric single criteria problems. The
mathematical model of the weighted sum method takes the form of:

X
m
min f .x/ ; f .x/ D !i fi .x/ ; (47.1)
i D1
P
where !i is the weight of i th criterion, 0  !i  1; i D 1; :::; m, and m i D1 !i .
Assume that all fi .x/ are convex functions; then for every point of Pareto set
there exist weights !i ; i D 1; :::; m such that this Pareto point is a minimizer of
f .x/ defined by Eq. (47.1). The violation of the convexity assumption can imply
not existence of weights implying location of the minimizer of the corresponding
composite objective function f .x/ in some subsets of Pareto set. However, in the
problems of portfolio selection objectives fi .x/ normally are defined by rather sim-
ple analytical expressions, and checking of the validity of convexity assumption is
easy. The fact of existence of the correspondence between points in Pareto set and
minimizers of the parametric minimization problem (Eq. (47.1)) theoretically jus-
tifies further investigation of properties of this correspondence. In many cases it is
not an easy task to choose the set of weights defining solutions of Eq. (47.1) that
would be well (in some sense uniformly) distributed over the Pareto set. To generate
such a subset of the Pareto set by means of repeatedly solution of Eq. (47.1) with
different weights the latter should be chosen in a special but a priory unknown way.
We propose a branch and bound type method for iterative composing of a set of
weights implying the desirable distribution of solutions of Eq. (47.1). The feasible
region for weights
( )
Xm
 D .!1 ; :::; !m / W 0  !i  1; i D 1; :::; m; !i D 1 ; (47.2)
i D1

is a standard simplex. Our idea is to partition  into sub simplices whose vertices are
mapped to the Pareto set via Eq. (47.1). The sequential partition is controlled aiming
to generate such new sub simplices that mapping of their vertices would generate
points uniformly distributed in Pareto set. The partition procedure is arranged as a
branching of a tree of nods corresponding to sub simpices. The original standard
simplex is accepted as the root of the tree. Branching means partition of a sim-
plex into two sub simplices where the new vertex is the midpoint of the favorable
47 Approximation of Pareto Set in Multi Objective Portfolio Optimization 555

edge defined later. Simplicial partitioning procedure has been applied to construct
algorithms for single criterion global optimization in [11, 12]; this procedure can be
applied also in multi objective optimization if only preconditions for selection of a
simplex for partition would be reconsidered.
We aim to subdivide the original simplex into sub simplices in such a way that
solutions of Eq. (47.1) corresponding to the vertices of these sub simplices would
be well distributed over the Pareto set. Let us consider the current set of sub sim-
plices in the space of weights, and points in the criteria space corresponding to the
vertices of these simplices. The eligibility of a simplex to partition is defined by the
longest distance between the points in the criteria space corresponding to the ver-
tices of the considered simplex. Branching strategy is based on depth first search,
and the selected simplices are partitioned until the longest distance between the
corresponding points in criteria space is reduced up to predefined tolerance.
The method of adjustable weights (AdjW) was implemented for two and three
criteria cases in MATLAB using fminconstr for solving minimization problem
(Eq. (47.1)). To illustrate the advantage of the proposed method of adjustable
weights over the standard weighted sum method both methods were applied for the
construction of Pareto set of two criteria portfolio optimization with the data men-
tioned above. Figure 47.1 shows the distributions of the Pareto points found by both
methods. In the standard method weights have been changed with the step for 0.05.
Because of space limitation we do not describe in detail the tolerance for branching
of the method of adjustable weights; it have been set aiming at generation of similar
number of points as generated by weighting method. In this experiment the number
of points generated by the weighting method was equal to 21, and that generated by
our method was equal to 26.

Fig. 47.1 Points in Pareto set generated by the standard weighted sum method and by the method
of adjustable weights
556 I. Radziukyniene, A. Zilinskas

47.4 Evolutionary Methods for Multi-Objective Optimization

Three methods: Fast Pareto genetic algorithm (FastPGA) [2], Multi-Objective Cel-
lular genetic algorithm (MOCeLL) [3], and Archive-based hybrid Scatter Search
algorithm [4] were proposed during the last 2 years. Their efficiency for vari-
ous problems has been shown in original papers, but their application to portfolio
optimization problem was not yet explored. NSGA-II [5], the state-of-the-art evo-
lutionary method, was chosen following many authors who use it as a standard for
comparisons.
FastPGA. Eskandari and Geiger [2] have proposed framework named fast Pareto
genetic algorithm that incorporates a new fitness assignment and solution ranking
strategy for multi-objective optimization problems where each solution evaluation
is relatively computationally expensive. The new ranking strategy is based on the
classification of solution into two different categories according to dominance. The
fitness of non-dominated solutions in the first rank is calculated by comparing each
non-dominated solution with one another and assigning a fitness value computed
using crowding distance. Each dominated solution in the second rank is assigned
a fitness value taking into account the number of both dominating and dominated
solutions. New search operators are introduced to improve the proposed method’s
convergence behavior and to reduce the required computational effort. A population
regulation operator is introduced to dynamically adapt the population size as needed
up to a user-specified maximum population size, which is the size of the set of non-
dominated solutions. FastPGA is capable of saving a significant number of solution
evaluations early in the search and utilizes exploitation in a more efficient manner
at later generations.
Characteristics of FastPGA: the regulation operator employed in FastPGA
improves its performance for fast convergence; proximity to the Pareto optimal set,
and solution diversity maintenance.
MOCeLL. Nebro et al. [3] presented MOCeLL, a multi-objective method based
on cellular model of GAs, where the concept of small neighborhood is intensively
used, i.e., population member may only interact with its nearby neighbors in the
breeding loop. MOCell uses an external archive to store the non-dominated solutions
found during the execution of the method, however, the main feature characterizing
MOCell is that a number of solutions are moved back into the population from the
archive, replacing randomly selected existing population members. This is carried
out with the hope of taking advantage of the search experience in order to find a
Pareto set with good convergence and spread.
MOCell starts by creating an empty Pareto set. The Pareto set is just an additional
population (the external archive) composed of a number of the non-dominated solu-
tions found. Population members are arranged in a two-dimensional toroidal grid,
and the genetic operators are successively applied to them until the termination
condition is met. Hence, for each population member, the method consists of
selecting two parents from its neighbourhood for producing an offspring.
An offspring is obtained applying operators of recombination and mutation. After
evaluating of the offspring as the new population member it is inserted in both the
47 Approximation of Pareto Set in Multi Objective Portfolio Optimization 557

auxiliary population (if it is not dominated by the current population member) and
the Pareto set. Finally, after each generation, the old population is replaced by the
auxiliary one, and a feedback procedure is invoked to replace a fixed number of ran-
domly chosen population members of the population by solutions from the archive.
In order to manage the insertion of solutions in the Pareto set with the goal to
obtain a diverse set, a density estimator based on the crowding distance has been
used. This measure is also used to remove solutions from the archive when this
becomes full.
Characteristics of MOCeLL: the method uses an external archive to store the
non-dominated population members found during the search; the most salient fea-
ture of MOCeLL with respect to the other cellular approaches for multi-objective
optimization is the feedback of members from archive to population.
AbYSS. This method was introduced by Nebro et al. [4]. It is based on the scat-
ter search using a small population, known as the reference set, whose population
members are combined to construct new solutions. Furthermore, these new popula-
tion members can be improved by applying a local search method. For local search
the authors proposed to use a simple (1 + 1) Evolution Strategy which is based on
a mutation operator and a Pareto dominance test. The reference set is initialized
from an initial population composed of disperse solutions, and it is updated by tak-
ing into account the solutions resulting from the local search improvement. AbYSS
combines ideas of three state-of-the-art evolutionary methods for multi criteria opti-
mization. On the one hand, an external archive is used to store the non-dominated
solutions found during the search, following the scheme applied by PAES [5], but
using the crowding distance of NSGA-II [6] as a niching measure instead of the
adaptive grid used by PAES; on the other hand, the selection of solutions from
the initial set to build the reference set applies the density estimation used by
SPEA2 [4].
Characteristics of AbYSS: it uses an external archive to store the non-dominated
population members found during the search; salient features of AbYSS are the feed-
back of population members from the archive to the initial set in the restart phase of
the scatter search, as well as the combination of two different density estimators in
different parts of the search.
NSGA-II. The evolutionary method for multi-criteria optimization NGSA-II
contains three main operators: a non-dominated sorting, density estimation, and
a crowded comparison [6]. Starting from a random population the mentioned
operators govern evolution whose aim is uniform covering of Pareto set.
Non-dominated sorting maintains a population of non dominated members: if
a descendant is dominated, it immediately dies, otherwise it becomes a member
of population; all members of parent generation who are dominated by descen-
dants die. The density at the particular point is measured as the average distance
between the considered point and two points representing the neighbour (left and
right) population members. The crowded comparison operator defines selection for
crossover oriented to increase the spread of current approximation of Pareto front.
Population members are ranked taking into account “seniority” (generation number)
and local crowding distance.
558 I. Radziukyniene, A. Zilinskas
 
The worst-case complexity of NSGA-II algorithm is O mN 2 , where N is the
population size and m is the number of objectives [6].
Characteristics of NSGA-II: this method is of the lower computational complexity
than that of its predecessor NSGA; elitism is maintained, no sharing parameter
needs to be chosen because sharing is replaced by crowded-comparison to reduce
computations.

47.5 Description of Experimental Investigation

Theoretical assessment of available multi objective optimization methods with

respect to their applicability to portfolio optimization does not seem possible.
Therefore such an assessment has been performed experimentally. There are some
publications on experimental comparison of different performance aspects of evo-
lutionary methods. For example, in the study presented in [13] some general
conclusions are drawn, however they could not be directly applied to a specific
problem. For assessing of the considered methods several different performance
measures can be taken into account: the distance between the approximated Pareto
set generated by the considered method and the true Pareto set, the spread of the
solutions, and computational time. To determine the first measure the true Pareto set
should be known. In this experimental investigation we didn’t know true Pareto sets.
Therefore, the best approximation found by means of combining results of all con-
sidered methods was used instead of true Pareto set. Although, computational time
is one of the most important performance measures in comparison of optimization
methods, it has not been included in our study. This can be justified only by the
severe differences in implementations of the algorithms (e.g. in C and in MATLAB)
making direct comparison of running times unfair. We compared methods according
to three performance measures:
1. Generational distance (GD) shows how far the approximation is from true Pareto
set [14].
2. Inverted generational distance (IGD) [15]. This quality indicator is used to mea-
sure how far the elements are in the Pareto optimal set from those in the set of
non-dominated vectors found.
3. Hypervolume (HV) [14]. This quality indicator calculates the volume (in the
objective space) covered by members of a non-dominated set of solutions.
Methods with larger values of HV are desirable.
Before evaluating the fitness function in FastPGA, MOCeLL, AbYSS, and NSGAII
the proportions of stocks in the portfolio were normalized as in reference [7]. These
methods were run with different parameters that were recommended by the authors
[14, 16–18].
They were:
1. FastPGA Maximum population = 100, initial population = 100, crossover prob-
ability = 1.0
47 Approximation of Pareto Set in Multi Objective Portfolio Optimization 559

2. MOCeLL Population = 100, archive = 100, crossover probability = 0.9

3. AbYSS Population = 20, archive = 100, crossover probability = 1.0, setref1 = 10,
setref2 = 10
4. NSGAII Population = 100, crossover probability = 0.9
To evaluate each of these methods, while solving two objectives portfolio opti-
mization problem we performed three series of experiments. First, we ran all the
methods for 15,000 function evaluations, and then repeated them with the execu-
tion of 25,000 and 35,000 function evaluations as the stopping condition. In the
case of three objectives portfolio, two series of experiments with 25,000 and 50,000
function evaluations have been performed. For each problem we have executed 50
independent runs.

47.6 Discussion on Experimental Results

Experiments were performed to compare performance of the selected evolutionary

methods and the method of adjustable weights (AdjW). The evolutionary meth-
ods are randomized, but the method of adjustable weights is deterministic. Because
of this difference the experimental results in tables below are presented corre-
spondingly. The averages (Avg.) and standard deviations (Std.) of all performance
measures for each evolutionary method were calculated from the samples of 50 runs.
The best results in tables are printed in boldface.
It follows from the Table 47.1 that MOCeLL outperforms the other methods
in all cases except of comparison with respect to Inverted general distance where
number of functions evaluations was equal to 25,000; in the latter case NSGAII was
considerably better. The differences in performance weaken with increase of number
of evaluations as it is well illustrated by the results of the last block of results in the
Table 47.1.
The upper part of the Pareto set shown in Fig. 47.1 is most difficult to reveal for
all evolutionary methods. Approximation of the Pareto set with a curve of changing
curvature shows that the curve is flattening at its upper part. Similar dependence
between flatness of the Pareto front and decrease of quality of its approximation
using evolutionary methods is mentioned also by the other authors. For illustration
of this phenomenon the points generated by all methods in the mentioned part of
Pareto set are presented in Fig. 47.2; maximal number of function evaluations was
fixed equal to 25,000. It can be noticed that the method of adjustable weights does
not suffer from flattening of the Pareto front.
The experimental results for three criteria problem are given in Table 47.2. These
results show that the best of the evolutionary methods according GD and HV mea-
sures is MOCeLL. NSGAII and FastPGA are the best methods with respect to IGD.
It can be noticed that GD value of the method of adjustable weights is equal to
zero; this means that all solutions lie precisely in Pareto set. The approximations of
Pareto sets obtained by evolutionary methods with 50,000 function evaluations and
the adjustable weights method were visualized using color stereo display. Since such
560 I. Radziukyniene, A. Zilinskas

Table 47.1 Performance metrics for two objective problem; the performance measures for AdjW
were GD D 0:0, IGD D 2:97e-5, HV D 0:886; F denotes the number of function evaluations
by the evolutionary methods
Method GD IGD HV
Avg. Std. Avg. Std. Avg. Std.
F D 15; 000
AbYSS 5.47e-4 4.09e-4 1.86e-3 2.14e-3 8.55e-1 3.31e-2
FastPGA 4.12e-4 2.81e-4 2.28e-3 1.64e-3 8.56e-1 2.38e-2
MOCeLL 3.10e-4 2.31e-4 1.23e-3 1.65e-3 8.69e-1 2.12e-2
NSGAII 4.17e-4 2.75e-4 1.40e-3 1.24e-3 8.69e-1 1.42e-2
F D 25; 000
AbYSS 2.06e-4 5.95e-5 1.60e-4 5.59e-4 8.82e-1 5.50e-3
FastPGA 2.26e-4 7.04e-5 4.88e-4 9.77e-4 8.79e-1 1.26e-3
MOCeLL 9.29e-5 1.88e-5 1.11e-4 2.92e-4 8.84e-1 1.90e-3
NSGAII 2.36e-4 2.87e-5 9.90e-5 1.28e-5 8.82e-1 3.00e-4
F D 35; 000
AbYSS 1.63e-4 3.57e-5 7.20e-5 2.40e-6 8.83e-1 3.20e-4
FastPGA 2.11e-4 2.72e-5 1.04e-4 1.15e-4 8.83e-1 4.70e-4
MOCeLL 6.34e-5 8.90e-6 6.70e-5 9.00e-7 8.84e-1 8.50e-5
NSGAII 2.41e-4 3.02e-5 9.70e-4 4.70e-6 8.82e-1 1.80e-4

Fig. 47.2 Pareto sets of AbYSS, FastPGA, MOCeLL, NSGAII, and AdjW

a visualization can not be included in a paper we will present only heuristically obvi-
ous conclusion that the quality of approximation of Pareto set by adjustable weights
method is better than that by evolutionary methods. However the latter are less
computing intensive. Therefore development of a hybrid method seems promising.
47 Approximation of Pareto Set in Multi Objective Portfolio Optimization 561

Table 47.2 Performance metrics for two objective problem; the performance measures for AdjW
were GD D 0:0, IGD D 1:48e-4, HV D 0:735; F denotes the number of function evaluations
by the evolutionary methods
Method GD IGD HV
Avg. Std. Avg. Std. Avg. Std.
F D 25; 000
AbYSS 1.44e-3 4.68e-4 2.16e-4 1.16e-4 7.15e-1 6.30e-3
FastPGA 1.42e-3 3.85e-4 2.08e-4 1.90e-5 7.16e-1 2.10e-3
MOCeLL 1.16e-3 3.62e-4 2.12e-4 1.90e-5 7.18e-1 1.30e-3
NSGAII 1.33e-3 3.02e-4 2.10e-4 2.20e-5 7.15e-1 1.90e-3
F D 50,000
AbYSS 1.22e-3 4.04e-4 2.13e-4 1.90e-5 7.16e-1 1.50e-3
FastPGA 1.24e-3 3.59e-4 2.06e-4 1.50e-5 7.18e-1 1.40e-3
MOCeLL 1.14e-3 2.77e-4 2.12e-4 1.80e-5 7.19e-1 1.20e-3
NSGAII 1.65e-3 5.05e-4 2.12e-4 2.20e-5 7.16e-1 1.50e-3

47.7 Conclusions

From the results of the experiments for two criteria portfolio optimization it follows
that MOCeLL is the best of four considered evolutionary methods with respect to
all three performance criteria. The results of the experiments with these methods for
three criteria portfolio optimization reveal that MOCeLL provides the best results
in terms of Hypervolume, and Generational distance, but is slightly outperformed
by FastPga with respect to the Inverted generational distance. The evaluated perfor-
mance criteria of evolutionary methods are only slightly worse than those of method
of adjustable weights who is advantageous in the considered cases. Summarizing
the results it seems promising to develop a hybrid method trying to combine the
advantages of evolutionary methods with those of the method of adjustable weights.

Acknowledgements The second author acknowledges the support by the Lithuanian State Sci-
ence and Studies Foundation.

References

1. Markowitz, H.: Portfolio selection. Journal of Finance, 7, 77–91 (1952).

2. Li, J., Taiwo, S.: Enhancing Financial Decision Making Using Multi-Objective Financial
Genetic Programming. Proceedings of IEEE Congress on Evolutionary Computation 2006,
pp. 2171–2178 (2006).
3. Chang, T.-J., Meade, N., Beasley, J. E., Sharaiha, Y. M.: Heuristics for cardinality constrained
portfolio optimisation. Computers and Operations Research, 27, 1271–1302 (2000).
4. Wang, S.-M., Chen, J.-C., Wee, H. M., Wang, K. J.: Non-linear Stochastic Optimization Using
Genetic Algorithm for Portfolio Selection. International Journal of Operations Research, 3,
No. 1, 16–22 (2006).
5. Ehrgott, M., Klamroth, K., Schwehm, C.: Decision aiding an MCDM approach to portfolio
optimization. European Journal of Operational Research, 155, 752–770 (2004).
562 I. Radziukyniene, A. Zilinskas

6. Xia, Y., Wang, S., Deng, X.: Theory and methodology: a compromise solution to mutual
funds portfolio selection with transaction costs. European Journal of Operation Research, 134,
564–581 (2001).
7. Mukerjee, A., Biswas, R., Deb, K., Mathur, A. P.: Multi-objective evolutionary algorithm for
the risk-return trade-off in bank loan management. International Transactions in Operational
Research, 9, 583–597 (2002).
8. Stummer, C., Sun, M.: New Multiobjective Metaheuristic Solution Procedures for Capital
Investment Planning. Journal of Heuristics, 11, 183–199 (2005).
9. Ehrgott, M., Waters, C., Gasimov, R. N., Ustun, O.: Multiobjective Programming
and Multiattribute Utility Functions in Portfolio Optimization. 2006. Available via
http://www.esc.auckland.ac.nz/research/tech/esc-tr-639.pdf, cited 20 June 2008.
10. Mukerjee, A., Biswas, R., Deb, K., Mathur, A. P.: Multi-objective evolutionary algorithm for
the risk-return trade-off in bank loan management. International Transactions in Operational
Research, 9, 583–597 (2002).
11. Steuer, R. E., Qi, Y., Hirschberger, M.: Portfolio Selection in the Presence of Multiple Criteria.
In Zopounidis, C., Doumpos, M., Pardalos, P. M. (eds.) Handbook of Financial Engineering,
Springer, New York (2008).
12. Clausen, J., Zilinskas, A.: Global Optimization by Means of Branch and Bound with simplex
Based Covering, Computers and Mathematics with Applications, 44, 943–955 (2002).
13. Zilinskas, A., Zilinskas, J.: Global Optimization Based on a Statistical Model and Simplicial
Partitioning, Computers and Mathematics with Applications, 44, 957–967 (2002).
14. Nebro, A. J., Durillo, J. J., Luna, F., Dorronsoro, B., Alba, E.: A Cellular Genetic Algorithm
for Multiobjective Optimization. Proceedings of NICSO 2006, pp. 25–36.
15. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A Fast and Elitist Multiobjective Genetic
Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6, No. 2, 182–197
(2002).
16. Zizler, E., Deb, K., Thiele, L.: Comparison of Multi Objective Evolutionary Algorithms:
Empirical Results, Evolutionary Computation, 8, No. 2, 173–195 (2000).
17. Eskandari, H., Geiger, C. D.: A Fast Pareto Genetic Algorithm Approach for Solving
Expensive Multiobjective Optimization Problems, Journal of Heuristics, 14, No. 3, 203–241
(2008).
18. Nebro, A. J., Luna, F., Alba, E., Beham, A., Dorronsoro, B.: AbYSS Adapting Scatter Search
for Multiobjective Optimization. Technical Report ITI-2006–2, Departamento de Lenguajes y
Ciencias de la Computación, University of Malaga (2006).
Chapter 48
The Scaling Approach for Credit Limit
Management and Its Application

M.Y. Konovalikhhin, D.O. Sergienko, and O.V. Balaeva

Abstract When bank decides to provide the loan to a client, decision has to be made
on the credit amount and the term. Some studies looked into possibility of taking
into consideration the probability of default while choosing the optimal credit pol-
icy by introducing synthetic coefficient or VaR technology. This paper suggests the
scaling methodology for calculation of coefficients, which are required for optimal
credit limit estimation. Scaling rationale is based on the goal-oriented parameters.
The method is built on comparison of two competing forces: (i) potential increase
of the limit caused by client’s positive characteristics such as income, good credit
history, ownership of movable and immovable properties etc. and (ii) potential limit
decrease, which is a result of the probability of default calculated from the client
score. Such model can be related to the quantitative cause-effect model type, which
is focused on the final result without consideration of the process’ dynamics.

Keywords Credit Limit Management  Scaling Approach  goal-oriented parame-

ters  cause-effect model

48.1 Introduction

One of the consequences of impressive growth of Russian retail market is a very

strong competition between the banks. This makes Basel Committee on Banking
Supervision (Basel II) [1] requirement related to the necessity of development of the
modern flexible bank systems for client reliability assessment [2] becomes particu-
larly important. Increased competition and growing pressures for revenue generation
have made financial institutions to look for more effective ways to attract new

M.Y. Konovalikhhin (B)

The head of analytical division of risk analysis department, Ph.D., Bank VTB 24
(Closed joint-stock company), Moscow, Russia,
E-mail: Konovalihin.MY@vtb24.ru

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 563
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
564 M.Y. Konovalikhhin, et al.

creditworthy customers and at the same time to control the losses. Aggressive mar-
keting efforts have resulted in deeper penetration of the risk pool amongst potential
customers. The need to process them rapidly and effectively has initiated growing
automation of the credit and insurance application and adjudication processes.
Banks need rapid credit decision making process in order to maintain high bor-
rowing power, but on the other hand it should not result in deterioration of portfolio
quality. The risk manager is now challenged to provide solutions that not only assess
the creditworthiness, but also keep the per-unit processing cost low, while reducing
turnaround time for the customers. In addition, customer service quality requires
this automated process to be able to minimize the denial of credit to creditworthy
customers, while keeping out as many potentially delinquent customers as possible.
In the recent years a particular attention has been paid to risk scoring, which
along with other predictive models is a tool evaluating the risk level associated with
applicants or customers. While it does not identify “good” (no negative behavior
expected) or “bad” (negative behavior expected) applications on individual basis, it
provides the statistical odds, or probability, that an applicant with any given score
will be either “good” or “bad”. These probabilities or scores along with other busi-
ness considerations, such as expected approval rates, profits and losses are used as
a basis for decision making.
When bank decides to provide the loan to a client, decision has to be made on
the credit amount and the term. Credit limit is the approved level of a loan amount,
which theoretically must represent client’s financial rating. In other words, correct
calculation of the optimal credit limit is a powerful tool for minimizing the credit
losses.
At the same time it should be noted that most of the existing methodologies
for credit limit calculation do not take into consideration the explicit assessment of
the probability of default. Traditional formula for credit limit calculation does not
deal with the credit risk. It is postulated that the limit, calculated by this formula,
provides zero credit risk, or, in other words, the probability of default reduces to
zero [3].
Some studies looked into possibility of taking into consideration the probability
of default while choosing the optimal credit policy by introducing synthetic coef-
ficient [4] or VaR technology [5]. These studies are based on the estimation of a
hypothecation value for inter-bank credits, which is not always applicable to retail
business.
This paper suggests the scaling methodology for calculation of coefficients,
which are required for optimal credit limit estimation. While developing such
scaling methodology, three principles have been considered:
First, scaling rationale is based on the goal-oriented parameters. In this case,
scaling parameters are determined from credit limit quality perspective.
Second, while seeking the simplicity of the scaling equation, it is required
to reduce the (potential) effects of characteristics, involved in calculations, on
component phenomena.
Third, equations and correlations that were used in development of the scaling
methodology are transparent and can be separately validated.
48 The Scaling Approach for Credit Limit Management and Its Application 565

The method is built on comparison of two competing forces: (i) potential increase
of the limit caused by client’s positive characteristics such as income, good credit
history, ownership of movable and immovable properties etc. and (ii) potential limit
decrease, which is a result of the probability of default calculated from the client
score. Such model can be related to the quantitative cause-effect model type, which
is focused on the final result without consideration of the process’ dynamics.

48.2 Analysis

48.2.1 Model Description

Existing credit limit calculation methodology, which is currently used at some

banks is presented below. Here, the total client income .D/ is defined as a sum
of documented and undocumented incomes:

D D DCI C B3  UCI; (48.1)

where B3 is a coefficient showing the presence of some attributes (movable and

immovable properties for instance), that could indicate additional income.
Then the D value is corrected according to the client’s score:

BP D D  B2  OL; (48.2)

where B2 is a coefficient, which represents client’s rating, OL is a total amount of

promissory notes. Client’s expected income is influenced by a number of dependen-
cies, which is reflected by B1 coefficient.

SD D BP  B1 (48.3)

The credit limit is calculated as following:

SD  T
LD ; (48.4)
1C RT
12

where T is the loan term and R is the bank interest rate (%)/100.
The weakness of this method is in the fact, that coefficients B1, B2 and B3 are
defined by the expert judgment. Calculation of the limits, which is based on the
expert judgement may result in ungrounded over- or understatement of credit limits.
In the case of understatement of the credit limits the bank cannot use all the credit
resources, that in turn leads to reduced profit. One of the main consequences of
credit limit overstatement is the increase in credit risks, which leads to additional
losses.
566 M.Y. Konovalikhhin, et al.

On the other hand, applying precise methods to the credit limit calculation allows
to reach full bank credit potential and to maximize bank’s profit. Also, one should
take into account client’s rating (scoring) in order to minimize the risk of default.

48.2.2 B3 Coefficient Calculation

B3 coefficient (Eq. (48.1)) can be estimated by examining the dependency between

the ratio DCIUCI
CUCI , which indicates the share of client’s undocumented income
in the total income, and the default probability P .X / D 1
 1
, where X is the
1Ce X
client score. We assume, that this value can grow only with the improvement of
client’s rating or with the score increase and, respectively, with the decrease of the
probability of default:
UCI 1
D f .P .X // D C.X /  ; (48.5)
DCI C UCI P .X /

where C.X / – a proportional coefficient, which may be independent of the score X .

In order to assess this dependency a pool of 37,890 clients (for one product within
certain timeframe) was analyzed. Then a number of clients were selected that had
declared UCI incomes. From this data two groups were selected – 9,681 clients
without delinquency and 1,334 with delinquency over 30 days.
The first problem, which had to be investigated, was to check for a potential
UCI
correlation between DCICUCI and the presence of delinquency. The comparison of
average values in both cases (with and without delinquency) showed that the differ-
ence between these two values is about 6%, which can be explained by statistical
UCI
uncertainty (Figs. 48.1 and 48.2). This fact confirms the random nature of DCICUCI .
With the knowledge of relation (48.5) it is possible to find the level of undoc-
umented income UCI, which statistically corresponds to the documented income
DCI:  
1 1
.UCI/calc D DCI  C.X/  = 1  C.X/ 
P.X/ P.X/
(48.6)
DCI  C.X/
D
P.X/  C.X/
First, it was necessary to estimate parameter C.X/ by using statistical data. For this
purpose two regions of possible risk levels (high and low) were selected. The line
“good clients average ratio” was chosen as upper conservative boundary because
the region, which is located above this line and correspondingly above the line “bad
clients average ratio” was considered a high risk zone. The lower boundary was
estimated by averaging (dashed line “low risk region”, Figs. 48.1 and 48.2) bad
client data, which occurs under the line “good clients average ratio”. The region
under the line “low risk region” was defined as a low risk region, which contains the
values .UCI/calc .
48 The Scaling Approach for Credit Limit Management and Its Application 567

UCI
Fig. 48.1 Dependency of average DCICUCI
value on the score of the “good” clients

UCI
Fig. 48.2 Dependency of average DCICUCI
value on the score of the “bad” clients
568 M.Y. Konovalikhhin, et al.

Fig. 48.3 Comparison of calculated UCI values and original data

Then two alternatives were examined, when Eq. (48.1) applied undocumented
income .UCI/appl is less than .UCI/calc and Eq. (48.2) when it is greater than
.UCI/calc . In the first case the whole UCI amount can be taken into account and be
employed for client income calculation or, in other words, B3 D 1 in Eq. (48.1). In
the second case the UCI value can be calculated as a sum of two items: .UCI/calc
and some part of the value Œ.UCI/appl  .UCI/calc  W

UCI D .UCI/calc C C3  Œ.UCI/appl  .UCI/calc  (48.7)

During the next step the coefficient C3 , which demonstrates the optimal degree
of confidence in the client, should be estimated. Figure 48.3 shows the difference
between real client undocumented income UCI and calculated values .UCI/calc . One
of the possible solutions is to use the cause-effect relations, which have been suc-
cessfully applied to quantitative analysis of various technical problems [6]. Such
equations provide solution to the problem without describing process’ dynamics.
The result in this case is the ratio between positive and negative characteristics,
that are represented by client’s financial condition, such as movable and immovable
properties, etc. in the numerator and the default probability P .X / in the denomi-
nator:
P
Ai Zi
i
C3 D C3,corr  ; (48.8)
P .X /
48 The Scaling Approach for Credit Limit Management and Its Application 569

Fig. 48.4 Distribution of C3 coefficient

where C3, corr is the coefficient, which correlates Eq. (48.8) to statistical data, Zi 
parameters of client financial condition, Ai – regressive coefficients, that are cal-
culated by the Interactive Grouping methodology. In order to estimate C3 the
following parameters were used: cost of movable properties (car, yacht etc.), country
house, apartment and business shareholding. Results of the calculation are presented
in Fig. 48.4.

48.2.3 B2 Coefficient Calculation

Similar approach could be applied for calculation of B2 coefficient in Eq. (48.2).

Here again we compare two competing parameters through their ratio. The upper
parameter, which is responsible for potential limit increase, is represented by a
regressive sum of characteristics such as documented income, occupation, employ-
ment position etc. The lower parameter, which is pushing down the credit limit, is
the probability of default P .X /.

E0
B2 D C2 ; corr  ; (48.9)
P .X /
where X
E0 D Di Ei ; (48.10)
i
570 M.Y. Konovalikhhin, et al.

Fig. 48.5 Calculated B2 coefficient distribution

where Ei are values of the positive characteristics, Di are regressive coefficients,

and C2 ; corr is a correlation coefficient between calculated values and statistical
data range.
In order to estimate B2 coefficient two samples were chosen – 33,664 clients with
good credit histories and 4,226 clients with delinquency over 30 days. The B2 coef-
ficient distribution, which is based on an expert judgment, is shown on the Fig. 48.5
(black points). Then specific weights of “good” and “bad” clients in each score range
were determined. The comparison of these values indicates that they are almost
equal and this fact attests to independency of B2 distribution on the client quality.
Also Fig. 48.5 shows that this expert distribution covers a small part of possible B2
values calculated for clients from the “good” sample. Each score corresponds to a
number of B2 values, which are determined by the personal characteristics of each
client.

48.2.4 B1 Coefficient Calculation

In order to find B1 coefficient in Eq. (48.3) it would be natural to use P .X / value,

which reflects client’s reliability:
BP
B1 D C1 ; corr  ; (48.11)
FM  P .X /
48 The Scaling Approach for Credit Limit Management and Its Application 571

Fig. 48.6 Calculated B1 coefficient distribution

where FM is a number of dependencies, BP is the income, calculated from

Eq. (48.2), C1, corr is the coefficient, which represents the correlation between this
ratio and statistical data. As shown on Figs. 48.7 and 48.8 B1 calculation through
Eq. (48.11) provides mapping of discrete values to the value field (Fig. 48.6),
demonstrating individual approach to each client.

48.3 Analysis Results

Based on the described approach the credit limits for all clients from the samples
were recalculated. The results are summarized in Table 48.1.
From the analysis results the following observations can be made:
Total increase of the sample portfolio was C14:4% while maintaining the initial
risk level. Calculations were performed in relative values.
Number of clients, that had undocumented income (UCI) in the low risk region
..UCI /calc /, was 32.3% of the total number of clients, that had undocumented
income.
Table 48.1 shows that the maximum average increase was obtained in the low
score range. This can be explained by the fact that clients in the range were
undervalued.
It should be noted that all B coefficients can be adjusted by the corresponding
coefficients C , that depend on the product statistical data.
572 M.Y. Konovalikhhin, et al.

Fig. 48.7 The average credit limit dynamic of one product (the results are calculated at the end of
each month)

Fig. 48.8 The relative delinquency (from 1 to 30 days) level dynamic of one product (the results
are calculated at the end of each month)
48 The Scaling Approach for Credit Limit Management and Its Application 573

Table 48.1 The credit limits for all clients from the samples
The score, X X < 0:6 0:6  X < 0:8 0:8  X  1

Average limit change (%) C38:6 C14:34 C7:9

Because of the normal score distribution most clients have scores in the range
0:6  X < 0:8. This explains the fact that the average changes in this range and
in the total portfolio are very close.
The model enabled to substantially increase (Fig. 48.7) the credit portfolio (about
C17%) while maintaining or improvement (Fig. 48.8) its quality.

48.4 Conclusions

This paper presented the scaling model that can be applied to the calculation of the
optimal credit limit. Particular attention has been paid to this problem due to grow-
ing competition between the banks in the retail market, that leads to the increase
in credit risk. Correct credit limit calculation plays one of the key roles in risk
reduction.
The model described in the paper is based on comparison of two opposite client
characteristics, one leading to potential credit limit increase, and the second leading
to decrease of the credit limit. Credit limit calculation involved client ’s personal
data, which allows to approach each client on the individual basis during credit
amount allocation. The scaling method was applied in order to analyze the data
obtained from the real client database. The scaling ratio provides reasonable predic-
tive capability from the risk point of view and therefore has been proposed to serve
as a model for credit limit allocation. The model’s flexibility allows coefficients’
adjustments according to new statistical data. Although the present work is quite
preliminary, it does indicate that presented solution allows to substantially increase
the credit portfolio while maintaining its quality.

References

1. Bank for International Settlements, 2004. International Convergence of capital Measurement

and Capital Standards. A Revised Framework, Basel, 2004.
2. R. Jankowitsch, S. Pichler, W.S.A. Schwaiger, “Modeling of Economic Value of Credit Rating
System”, Journal of Banking & Finance 31, 181–198, 2007.
3. M. Leippold, S. Ebnoether, P. Vanini, “Optimal Credit Limit Management”, National Centre
of Competence in Research Financial Valuation and Risk Management, Working Paper No.
72, September 2003.
4. I.T. Farrahov, “Credit Limit Calculation. Non-traditional approach”, The Analytical Bank
Journal, 04(83), 2002.
574 M.Y. Konovalikhhin, et al.

5. I.V. Voloshin, “Time Factor Consideration in the Credit Limit Calculation with the VaR
Technology”, Thesis report, Bank analyst club, http://www.bankclub.ru/library.htm
6. M.Y. Konovalikhin, T.N. Dinh, R.R. Nourgaliev, B.R. Sehgal, M. Fischer, “The Scaling
Model of Core Melt Spreading: Validation, Refinement and Reactor Applications”, Organi-
zation of Economical Cooperation and Development (OECD) Workshop on Ex-Vessel Debris
Coolability, Karlsruhe, Germany, 15–18 November 1999.
Chapter 49
Expected Tail Loss Efficient Frontiers for CDOS
of Bespoke Portfolios Under One-Factor Copula
Marginal Distributions

Diresh Jewan, Renkuan Guo, and Gareth Witten

Abstract The global structured credit landscape has been irrevocably changed with
the innovation of Collateralized Debt Obligations (abbreviated as CDOs). As of
2006, the volume of synthetic CDO structures outstanding grew to over $1.9 trillion.
Understanding the risk/return trade-off dynamics underlying the bespoke collateral
portfolios is crucial when optimising the utility provided by these instruments. In
this paper, we study the behaviour of the efficient frontier generated for a collat-
eral portfolio under heavy-tailed distribution assumptions. The convex and coherent
credit risk measures, ETL and Copula Marginal ETL (abbreviated as CMETL), are
used as our portfolio optimisation criterion. iTraxx Europe IG S5 index constituents
are used as an illustrative example.

Keywords Collateralized Debt Obligations  Expected Tail Loss  Efficient Fron-

tiers  Bespoke Portfolio  One-Factor Copula Marginal Distribution

49.1 Introduction

The global structured credit landscape has been irrevocably changed with the inno-
vation of Collateralized Debt Obligations (abbreviated as CDOs). As of 2006, the
volume of synthetic CDO structures outstanding grew to over $1.9 trillion, making it
the fastest growing investment vehicle in the financial markets. Bespoke deals made
up 21% of the total volume [1]. Understanding the risk/return trade-off dynamics
underlying the bespoke collateral portfolios is crucial when optimising the utility
provided by these instruments.

R. Guo (B)
Department of Statistical Sciences, University of Cape Town, Private Bag, Rhodes’ Gift, Ronde-
bosch 7701, Cape Town, South Africa
E-mail: Renkuan.Guo@uct.ac.za

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 575
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
576 D. Jewan, et al.

The loss distribution, a key element in credit portfolio management, is generated

through one-factor copula models. This approach has been widely used in credit
portfolio optimisation [2–4].
The main feature of these models is that default events, conditional on some
latent state variable are independent, and simplifies the computation of the aggregate
loss distribution [5]. For a comparative analysis, both the Gaussian and Clayton
copula models shall be used to model the default dependence structure inherent in
credit portfolios.
According to [6], there exist two types of risk measures: relevant and tractable.
Relevant measures capture key properties of a credit loss distribution, while tractable
measures can be optimized using computationally efficient methods. Efficient fron-
tiers are defined by a collection of optimal risk/return portfolios.
Expected Tail Loss (abbreviated as ETL) was initially introduced by [7], in the
portfolio optimisation context. ETL has proved a more consistent risk measure,
since it is sub-additive and convex [8]. Reference [7] has shown that the portfo-
lio optimisation under this risk measure results in a linear programming problem.
ETL is a superior measure to derive empirical efficient frontiers that act as a useful
approximation to the true efficient frontier [9, 10].
In this paper, we study the behaviour of the efficient frontier generated for a
collateral portfolio under heavy-tailed distribution assumptions. The structure of
the paper is as follow: Section 49.2 introduces bespoke CDOs and its underly-
ing mechanism. Section 49.3 introduces the Stable Paretian distribution and the
copulas for heavy-tail default dependence modelling. Section 49.4 then introduces
convex and coherent credit risk measures, particularly, ETL and Copula Marginal
ETL (abbreviated as CMETL), which is used as our portfolio optimisation crite-
rion. Section 49.5 provides the results of the numerical analysis conducted on a
collateral test portfolio. iTraxx Europe IG S5 index constituents are used as an
illustrative example. Concluding remarks on the analysis are presented in the final
section.

49.2 Bespoke CDO Mechanics

A bespoke CDO is a popular second-generation credit product. This standalone

single-tranche transaction is referred to as a bespoke because it allows the investor
to customise the various deal characteristics such as the collateral composition, level
of subordination, tranche thickness, and rating. Other features, such as substitution
rights, may also play an important role [11].
In a typical bespoke CDO transaction, the main decision step for potential
investors is to select the portfolio of credits, to which they want exposure. A short
position can also be held on a subset of names to have overweight views on certain
credits or sectors. The analysis will focus on the portfolio optimisation problem that
exists within this step of the structuring process.
49 Expected Tail Loss Efficient Frontiers for CDOS 577

Fig. 49.1 Placement of a mezzanine tranche in a bespoke CDO transaction

The next step is to determine the level of subordination, and tranche thickness
corresponding to the risk appetite of the investor. These deal parameters determine
the degree of leverage and the required premium payments [11]. This step of the
process adds a further two constraints to the transaction optimisation problem. The
first is related to the tranche thickness, and the second is the credit rating assigned to
the tranche. At this level of the problem, from an investor’s perspective, the return
generated by the CDO tranche is maximised.
A typical placement of a bespoke CDO is outlined in the following Fig. 49.1.
The cash flow payments that occur between the investor and the bank is indicated
by the dashed lines.
The challenge of issuing bespoke CDOs is the ongoing need and expense of the
risk management of the tranche position [12].

49.3 Heavy-Tail Modelling and Copulas

The use of heavy-tailed distributions for the margins of copula models is becoming
an industry standard for credit risk applications.

49.3.1 Stable Paretian Distributions

The Stable Paretian distribution can be expressed by its characteristic function, '.t/,
is given by [13]:
8
< jtj˛ 1  iˇsign.t/tan 2˛ C it; if ˛ ¤ 1
In .t/ D (49.1)
:
jtj 1  iˇsign.t/ 2 int C it; if ˛ D 1
578 D. Jewan, et al.

Maximum likelihood estimation techniques cannot be used directly. The density

function must first be approximated before it can be used in the parameter estimation
process. This is usually done either through the direct numerical integration [13], or
the application of the inverse Fourier transform to the characteristic function [14].

49.3.2 Copulas

Copula functions are useful tools in the modelling of the dependence between
random variables.
Definition 49.1. A copula is defined as a multivariate cumulative distribution func-
tion (abbreviated cdf) with uniform marginal distributions:

C.u1 ; u2 ; : : : ; uk /; ui Œ0; 1 for all sP D 1; 2; : : : ; n;

where,
C.ui / D ui for all sP D 1; 2; : : : ; n
The ideas behind multivariate distribution theory carries through to copula func-
tions. The above definition indicates that copulas represent the joint cdf of uniform
marginal distributions. The probability integral transformation of continuous non-
uniform random variables allows copulas to define the joint multivariate distribution
H with margins F1 ; F2 ; : : : ; ; Fn . By the strict monotocity of the cdf,

H.F1 ; F2 ; : : : ; Fk / D Pr.X1  !1 ; : : : ; Xk  !k /
(49.2)
D C.F1 .!1 /; F2 .!2 /; : : : ; Fk .!k //:

An alternative representation of copulas is used for credit portfolio modelling.

This approach introduced by [15], is analogous to the Capital Asset Pricing Model
(abbreviated as CAPM) in Modern Portfolio Theory.
De Finetti’s theorem for exchangeable sequence of binary random variables
provides a theoretical background for the use of factor models in credit risk appli-
cations [5].
Reference [16] relates the factor and copula approaches. These one-factor copula
models are presented in Section 49.5.

49.4 Credit Risk Measures

Recently, growing literature has been devoted to an axiomatic theory of risk mea-
sures [8], in which authors presented an axiomatic foundation of coherent risk
measures to quantify and compare uncertain future cashflows. [17] extended the
notion of coherent risk measures to convex risk measures.
49 Expected Tail Loss Efficient Frontiers for CDOS 579

49.4.1 Convex and Coherent Risk Measures

Let X denote the set of random variables defined on the probability space (; =; P ).
We define  as a set of finitely many possible scenarios for the portfolio. Financial
risks are represented by a convex cone M .; =; P / of random variables. Any
random variable X in this set, will be interpreted as a possible loss of some credit
portfolio over a given time horizon. The following provides a definition of a convex
cone.

Definition 49.2. M is a convex cone if 8X1 ; X2 2 M then X1 C X2 2 M and

X1 2 M for every  0.

Definition 49.3. Given some convex cone, M of random variables, a measure of risk
‚ with domain M is a mapping: ‚ W M ! R.

From an economic perspective ‚, can be regarded as the capital buffer that

should be set aside for adverse market movements. In order to measure and control
the associated risk [8] introduced an axiomatic framework of coherent risk measures
which were recently ‘generalized’ by [17] to convex risk measures.
Definition 49.4. A mapping ‚ is called a convex risk measure, if and only if it is,
Convex: for every X1 2 M and X2 2 M , we have ‚.X1 C .1  /X2 / 
‚.X1 / C .1  /‚.X2 /, for some  2 R,
Monotone:: for every X1 2 M and X2 2 M with X1  X2 we have ‚.X1 / 
‚.X2 /, and
Translation Invariant: if a is a constant then ‚.X C aI /  ‚.X /  a where I
denotes the unit vector.
Adding to these properties, positive homogeneity one obtains:
Definition 49.5. A risk measure ‚ is called coherent, if in addition it is, Positive
homogeneous: for every  0 we have ‚.X / D ‚.X /.

49.4.2 Expected-Tail-Loss

In the CDO portfolio optimisation problem, we represent the portfolio risk objective
function by the Expected-Tail-Loss measure.
Definition 49.6. Expected-Tail-Loss is defined by the following:

ET Lp .ˇ/ D EŒLp j Lp > ULp .ˇ/;

where Lp denotes the loss for portfolio p and ULp , the unexpected losses at the
confidence level ˇ.
580 D. Jewan, et al.

Portfolio optimisation with ETL as object function will result in a smooth, convex
problem with a unique solution [7].

49.4.3 Portfolio Optimisation Under Expected-Tail-Loss

A Copula Marginal Expected Tail Loss (abbreviated as CMETL) which is similar

to that of Stable (Distribution) Expected Total Loss (abbreviated as SETL) optimal
portfolio discussed in [9], is one that minimizes credit portfolio ETL subject to
a constraint of achieving expected credit portfolio returns at least as large as an
investor defined level, along with other typical constraints on weights, where both
quantities are evaluated in the CMETL framework.
In order to define the above CMETL precisely we use the following quantities:
Rp : the random return of portfolio p,
CMERp : the expected return of portfolio p with respect to the copula marginal
distribution, and
ˇ: tail loss probability.
The following assumptions are imposed by [9] for the CMETL investor:
The universe of assets is Q(the set of admissible credit portfolios)
The investor may borrow or deposit at the risk-free rate rf without restriction
The investor seeks an expected return of at least 

Definition 49.7. The CMETL investor’s optimal portfolio is then given by the
following:
˚ 
X0 . j ˇ/ D arg min CMETLp .ˇ/ ;
q Q

s; t:
X
x D 1;

`  x  u; 8 r;
CMERp : (49.3)

Where X  denotes the resulting portfolio’s weights. The subscript c indicates that
the problem is defined under the copula framework.
The constraints, in the order written above, require that (a) the sum of the port-
folio weights should equal to one; (b) the position weights should lie within the
trading limits l and u to avoid unrealistic long and short positions; (c) the expected
return on the portfolio in the absence of rating transitions should be at least equal to
some investor defined level .
Together, these optimal risk/return trade-offs define the efficient frontier.
49 Expected Tail Loss Efficient Frontiers for CDOS 581

49.5 Copula Marginal ETL Efficient Frontier for the CDO

Collateral

The analysis is performed in a one-period, default-only mode structural frame-

work. Consider a portfolio composed of n credit instruments. The exposure in each
instrument is denoted by Ni . Under the binomial setting, the loss of counterparty i
for scenario k is given by the following:

Lki D Ni  Vik ; f or i D 1; : : : ; n; k D 1; : : : ; s; (49.4)

where Vik is the value of exposure i at the given horizon in scenario k. In particular:
(
Ni ; Ri ; default statc;
Vi D
k
(49.5)
Ni otherwise;

where Ri is the recovery rate for reference entity i .

Definition 49.8. The portfolio loss function in scenario k; Lk W RnC1 ! R, over
the chosen time horizon is given by the following:

X
n
Lk .x/ D Lki xi ; (49.6)
i D1

where x D .x1 ; x2 ;    ; xn /T is the vector of positions held in the portfolio.

In order to generate scenarios for the credits states of each reference entity, we
follow the factor copula approach.
Definition 49.9. The ‘default-like’ stochastic process under the Gaussian copula
assumption is defined by the following:
p
Ai .x1 ; x2 ; : : : ; xn ; i / WD i M C 1  i i ; (49.7)

where M and "i are independent standard normal variants in the case of a Gaussian
copula model, and cov."k ; "l / ¤ 0, for all k ¤ l.
The default dependence comes from the factor M . Unconditionally, the stochas-
tic processes are correlated but conditionally they are independent. The default
probability of an entity i , denoted by Fi , can be observed from market prices of
credit default swaps, and defined by the following:
2 t 3
Z
Fi .t/ D 1  exp 4 hi .u/du5 ; (49.8)
0

where hi .u/ represents the hazard rate for reference entity i . For simplicity, we
assume that the CDS spread term structure is flat, and calibration of the hazard rates
for all reference entities is straightforward.
582 D. Jewan, et al.

The default barrier Bi .t/ is defined as: Bi .t/ D G 1 .Fi .t//, where G defines
the inverse distribution function. In the case of a Gaussian copula, this would be the
inverse cumulative Gaussian distribution function.
A second type of copula model considered comes from the Archimedean family.
In this family we consider the Clayton copula.

Definition 49.10. The ‘default-like’ stochastic process Ai under the one-factor

Clayton copula assumption is defined by the following:

   1
In."i / In."i /
Ai WD ' 1  D  C1  ; (49.9)
M M

where  ./ is the Laplace transform of the Gamma.1=/ distribution, "i is a uniform
random variable and M is a Gamma.1=/ distributed random variable.

Using the credit models presented above, the loss distribution for the test portfo-
lio described in [18] can easily be derived. This is shown in Fig. 49.2. Reference [18]
also provides empirical evidence for the stationarity of credit spreads. Stationarity
is a crucial issue, since credit risk models implicitly assume that spreads follow
stationary processes.
The following Table 49.1 displays the parameters of the two models for few of
the iTraxx Europe IG index constituents.
The coefficients in the regression analysis for all constituents in the bespoke test
portfolio were significant at a 99% confidence level.
The following Table 49.2 summarizes the expected loss, unexpected loss, and
CMETL at the 99.9% confidence level.
Although the Gaussian copula predicts a slightly higher expected loss than the
Clayton copula, there is a 1.758% absolute difference in the CMETL, and a 3%
difference in the maximum loss between the copula models.
We optimise all positions and solve the linear programming problem represented
by Eq. (49.3). Three scenarios are considered:

12%
11%
10%
9%
8%
Probabilty

7%
6%
5%
4%
3% Gaussian Copula
2%
Clayton Copula
1%
0%
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36

Portfolio Loss %

Fig. 49.2 Credit loss distributions for homogenous test portfolio at 5-year horizon under different
default dependence assumptions
49 Expected Tail Loss Efficient Frontiers for CDOS 583

Table 49.1 Calibrated copula parameters

Corporate entity Copula parameter

(Gaussian) (Clayton)
BAE Systems PLC 0.249616 0.043221
Unilever NV 0.239667 0.035097
Continental AG 0.264289 0.062691
Peugeot SA 0.243152 0.041494
Commerzbank AG 0.220350 0.025312

Table 49.2 Risk measure differences for initial homogenous test portfolio under both Gaussian &
Clayton copulas assumptions
Risk measure Gaussian copula Clayton copula
Expected loss 3.7248% 3.6952%
Unexpected loss @ 99.9% 20.349% 20.328%
CMETL @ 99.9% 25.534% 27.292%
Maximum loss 30.447% 33.434%

1. An examination of the efficient frontiers for a well-diversified portfolio under

both Gaussian and Clayton copula assumptions. Only long positions in the under-
lying credit default swaps are firstly allowed. The upper trading limit is set to 5%
in this case.
2. An investigation of the behaviour of the efficient frontiers under the Clayton
copula assumption when the upper trading limit is increased, is then studied.
This consequently increases concentration risk.
3. The efficient frontiers under the Clayton copula assumption when both long and
short positions are permitted, is lastly investigated.
In the first case lower and upper trading limit of 0.5% and 5% are set respectively.
This is to ensure that no reference asset is excluded from the portfolio. This also
maintains the well-diversified features of the iTraxx Europe IG portfolio, in the
bespoke structure. The following Fig. 49.3 presents the efficient frontiers under the
two default dependence assumptions.
The above Fig. 49.3 shows the difference in efficient frontiers between the Gaus-
sian and Clayton copula assumptions. For a given level of risk, a higher return should
be expected if the portfolio composition is based on the Clayton copula assumption
rather then the Gaussian copula. The above Fig. 49.3 also indicates the inefficient
risk/return levels for the original portfolio under both the Gaussian and Clayton cop-
ula assumptions. Under the Clayton copula assumption, the same level of return is
achieved with less than one-fifth of the original risk.
In the second case, the upper trading limit is increased from 10% to 100%.
The effect of concentration risk on the efficient frontiers is then investigated.
The following Figs. 49.4 and 49.5 presents the efficient frontiers under these
circumstances.
584 D. Jewan, et al.

200
180
160
Excess Return (bps)

140
120
100
80
60
40
20
0
0% 5% 10% 15% 20% 25% 30%

CMETL (% Losses in Portfolio)

Clayton Gaussian Original Portfolio (Gaussian)

Orginal Portfolio (Clayton) Poly. (Gaussian) Poly. (Clayton)

Fig. 49.3 Comparison of efficient frontiers under Gaussian and Clayton copula assumptions

200
180
Excess Return (bps)

160
140
120
100
80
60
40
20
0
5% 10% 15% 20% 25% 30%
CMETL (% Losses in Portfolio)
10% Upper Bound 15% Upper Bound 100% Upper Bound
Poly. (100% Upper Bound) Poly. (15% Upper Bound) Poly. (10% Upper Bound)

Fig. 49.4 Behaviour of second order polynomial fit of efficient frontiers, under increasing upper
trading limit

200
180
Excess Return (bps)

160
140
120
100
80
60
40
20
0
0% 5% 10% 15% 20% 25% 30%

CMETL (% Losses in Portfolio)

Long Only Long & Short Poly. (Long & Short) Poly. (Long Only)

Fig. 49.5 Effect of short positions on the efficient frontier

The figure displays the expected variation in efficient frontiers when the upper
trading limit is slowly relaxed. Under these circumstances, concentration risk is
introduced into the portfolio. Investors demand a higher premium for taking on this
risk. For a 20% level of risk, investors demand an extra 25 bps premium for holding
49 Expected Tail Loss Efficient Frontiers for CDOS 585

a more concentrated portfolio. At these levels, the concentrated portfolio only holds
positions in 49 of the 114 names, with the largest position size being 19.5%.
In the final case, the effect on the efficient frontier for allowing short positions is
examined. Under this scenario, only the well-diversified portfolio case is considered.
The lower and upper trading limits are set at 5% and 5% respectively.
The figure displays an important result: allowing short positions in credits, pro-
vides the investor with superior returns to those in the long-only case. At a 20%
level of risk, investors can earn an extra 30 bps premium for taking on overweight
views on certain credits in the portfolio. This indicates why hedge fund strategies
involving bespoke CDO structures have become increasingly popular.
The results for all efficient frontier second order polynomial regressions were
significant at a 95% confidence level. The resulting R2 coefficients were all above
90%.

49.6 Concluding Remarks

In this paper, we propose a new framework, the CMETL model, for the investi-
gations on CDO market patterns and the efficient frontier. The Gaussian copula
asset allocation is proved sub-optimal. Clayton copula efficient frontiers provided a
higher return for a given level of credit risk. A closer examination of the original risk
profile shows that the credit risk can be reduced to one-fifth of the original amount,
under the Clayton asset allocation method.
The permission of short positions in the bespoke CDO portfolio allows investors
to increase returns beyond a specific risk tolerance level. In the case study con-
sidered, a maximum increase of 37.5% in investor-defined return is achieved by
allowing overweight positions in certain credits.

References

1. E. Beinstein, “A review of themes from 2006 and outlook for 2007,” JP Morgan, in Corporate
Quantitative Research, 2007.
2. C. Finger, “Conditional approaches for credit metrics portfolio distributions,” Credit Metrics
Monitor, 2(1), 1999, 14–33.
3. P. Schonbucher, “Taken to the limit: Simple and not so simple loan loss distributions,” Working
paper, Department of Statistics, Bonn University.
4. O. Vasicek, “Probability of loss on a loan portfolio,” Technical report, KMV Corporation,
1987, Available: www.moodys.com.
5. J. P. Laurent, & J. Gregory, “Basket Default Swaps, CDO’s and Factor Copulas,” Working
paper, 2002.
6. H. Mausser, & D. Rosen, “Efficient risk/return frontiers for credit risk,” in Algo Research
Quarterly, 2(4), 1999, 9–22.
7. R. T. Rockafellar, & S. Uryasev, “Optimisation of conditional value-at-risk,” in Journal of
Risk, 3, 2000, 21–41.
586 D. Jewan, et al.

8. P. Artzner, F. Delbaen, J. M. Eber, & D. Heath, “Coherent measures of risk,” in Mathematical

Finance, 9, 1999, 203–28.
9. S. T. Rachev, R. D. Martin, B. Racheva, & S. Stoyanov, “Stable ETL optimal portfolios &
extreme risk management,” FinAnalytica, 2006, Available: www.statistik.uni-karlsruhe.de/
download/Stable Distributions White Paper.pdf
10. M. Kalkbrener, H. Lotter, & L. Overbeck, “Sensible and efficient capital allocation for credit
portfolios,” Deutsche Bank AG, 2003 Available: http://www.gloriamundi.org/detailpopup.
asp?ID = 453057574
11. A. Rajan, G. McDermotte, & R. Roy, The Structured Credit Handbook, Wiley Finance Series,
New Jersey, 2006, pp. 153–155.
12. G. Tejwani, A. Yoganandan, & A. Jha, “How do dealers risk manage synthetic CDOs,”
Lehman Brothers, in Structured Credit Research, 2007.
13. J. P. Nolan, “Numerical Calculation of Stable densities and distribution functions,” in
Communications in Statistics – Stochastic Models, 13, 1997, 759–774.
14. S. Mittnik, T. Doganoglu, & D. Chenyao, “Computing the probability density function of
Stable Paretian distributions,” in Mathematical & Computer Modelling, 29, 1999, 235–240.
15. O. Vasicek, “Probability of loss on a loan portfolio,” Technical report, KMV Corporation,
1987, Available: www.moodys.com
16. F. R. Frey, A. McNiel, & M. Nyfeler, “Copulas and credit models,” in Risk, October, 2001, pp.
111–113.
17. H. Föllmer, & A. Schied, “Robust preferences and convex measures of risk,” In S. Klaus, P.
J. Schonbucher, & D. Sondermann, editors, Advances in Finance and Stochastics, Springer,
New York, pp. 39–56, 2002.
18. D. Jewan, R. Guo, & G. Witten, “Copula marginal expected tail loss efficient frontiers for
CDOs of bespoke portfolios,” in Proceedings of the World Congress in Financial Engineering
Conference, 2, 1124–1129, London, July 1–7, 2008.
Chapter 50
Data Mining for Retail Inventory Management

Pradip Kumar Bala

Abstract As part of the customer relationship management (CRM) strategy, many

researchers have been analyzing ‘why’ customers decide to switch. However, despite
its practical relevance, few studies have investigated how companies can react to
defection prone customers by offering the right set of products. Consumer insight
has been captured in the work, but not used for inventory replenishment. In the
present research paper, a data mining model has been proposed which can be used
for multi-item inventory management in retail sale stores. The model has been
illustrated with an example database.

Keywords Retail Inventory Management  Data Mining  multi-item  Decision tree

 Association rule

50.1 Introduction

Customer relationship management (CRM) aims at stronger loyalty of customers

with greater market share. With competition for shelf space intensifying, there is
a pressing need to provide shoppers with a highly differentiated value proposition
through “right product mix in right amount at right time”. Mining or extracting cus-
tomer insight from structured and unstructured data, call center records, customer
complaints, and other sources will be of tremendous importance for inventory man-
agement in retail stores. Purchase dependency is one type of consumer behavior
which has not been addressed or used properly for managing inventory in retail
sale. Purchase or non-purchase of item or items by a customer may depend on the
purchase or non-purchase of some other item or items made by him or her. This is
referred as purchase dependency in this paper.

P. K. Bala
Xavier Institute of Management, Bhubaneswar, PIN-751013, India
E-mail: p k bala@rediffmail.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 587
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
588 P.K. Bala

In most of the real situations, we do not find a one-item inventory, but multi-
item inventory with different periods of replenishment. These inventories are usually
managed in aggregate because of the complexity of handling each individual item.
Intuition and commonsense rules with the aid of historical data have been useful
to some extent, but they are not often efficient or cost effective. Some examples of
the multi-item inventory are retail sale store, spare parts for maintenance, medicine
store etc.
Out of thousands of items held in an inventory of a typical organization, only a
small percentage of them deserve management’s closest attention and tightest con-
trol. It is not economical to exercise same degree of inventory control on all the
items. Using selective inventory control, varying degree of control are exercised on
different items. Most widely-used technique used to classify items for the purpose
of selective inventory control is ABC classification. Inventory replenishment policy
deals with ‘how-much-to-order’ and ‘when-to-order’ aspects. Purchase pattern is an
important consumer insight. It is obvious that the knowledge of purchase pattern
will be an important input for designing inventory replenishment policy.
Data mining is used to find new, hidden or unexpected patterns from a very large
volume of historical data, typically stored in a data warehouse. Knowledge or insight
discovered using data mining helps in more effective individual and group deci-
sion making. Irrespective of the specific technique, data mining methods may be
classified by the function they perform or by their class of application. Using this
approach, some major categories of data mining tools, techniques and methods can
be identified as given below.

(i) Association Rule: Association rule is a type of data mining that correlates
one set of items or events with another set of items or events. Strength of
association rule is measured in the framework of support, confidence and
lift [1].
(ii) Classification: Classification techniques include mining processes intended to
discover rules that define whether an item or event belongs to a particular pre-
defined subset or class of data. This category of techniques is probably the most
broadly applicable to different types of business problems. Methods based on
‘Decision Tree’ and ‘Neural Network’ are used for classification.
(iii) Clustering: In some cases, it is difficult or impossible to define the parame-
ters of a class of data to be analyzed. When parameters are elusive, clustering
methods can be used to create partitions so that all members of each set are sim-
ilar according to a specified set of metrics. Various algorithms like k-means,
CLARA, CLARANS are used for clustering. Kohonen’s map is also used for
clustering.
(iv) Sequence Mining: Sequencing pattern analysis or time series analysis methods
relate events in time, such as prediction of interest rate fluctuations or stock
performance, based on a series of preceding events. Through this analysis, var-
ious hidden trends, often highly predictive of future events, can be discovered.
GSP algorithm is used to mine sequence rules [2].
(v) Summarization: Summarization describes a set of data in compact form.
50 Data Mining for Retail Inventory Management 589

(vi) Regression: Regression techniques are used to predict a continuous value. The
regression can be linear or non-linear with one predictor variable or more than
one predictor variables, known as ‘multiple regression’.

Numerous techniques are available to assist in mining the data, along with numer-
ous technologies for building the mining models. Data mining is used by business
intelligence organizations to learn consumer behavior for the purpose of customer
relationship management, fraud detection, etc.

50.2 Literature Review

Inventory items have been classified or clustered in groups for the purpose of joint
replenishment policy in [3–6]. Generic inventory stock control policies are derived
using multi-item classification [3]. Clustering of items has also been done in produc-
tion and inventory systems [4]. Multi-criteria inventory classification has been done
by parameter optimization using genetic algorithm [5]. Artificial neural networks
(ANNs) have been used for ABC classification of stock keeping units (SKUs). ANN
has been used in a pharmaceutical company [6].
The interacting items showing the mutual increase in the demand of one com-
modity due to the presence of the other has been considered in the model given
in [7]. Correlation between the demands of two items has been considered. The arti-
cle gives a new approach towards a two-item inventory model for deteriorating items
with a linear stock-dependent demand rate. It has also assumed a linear demand rate,
that is, more is the inventory, more is the demand. The model has made an attempt
to capture demand interdependency to some extent.
Market segmentation has been done using clustering through neural network
and genetic algorithm [8]. Clustering of customers on the basis of their demand
attributes, rather than the static geographic property has been done in [9]. However,
the findings have not been used for the purpose of inventory management.
As part of the customer relationship management (CRM) strategy, many resear-
chers have been analyzing ‘why’ customers decide to switch. However, despite its
practical relevance, few studies have investigated how companies can react to defec-
tion prone customers by offering the right set of products [10]. Consumer insight
has been captured in the work, but not used for inventory replenishment. For cross-
selling consideration, a method to select inventory items from the association rules
has been developed in [11] which gives a methodology to choose a subset of items
which can give the maximal profit with the consideration of cross-selling effect.
However, this does not talk about any inventory replenishment policy.
In the present research paper, a data mining model has been proposed which can
be used for multi-item inventory management in retail sale stores. The model has
been illustrated with an example database.
590 P.K. Bala

50.3 Purchase Dependencies in Retail Sale in the Context

of Inventory Management

In the context of data mining, various aspects of purchase dependencies which may
be useful for the purpose of retail inventory management have been discussed as
given below.
Demand Interdependency: The problem of multi-item inventory is more challeng-
ing when there is association in the demand or usage pattern amongst the items or
item-sets (Item-set refers to a set of one or more items, hence if we say item-set, it
may refer to one item or a number of items.). The correlation in the demand amongst
the items can be one to one, one to many, many to one or many to many. In many
situations, a customer buys an item or item-set only when another item or item-set
is also in stock.
To explain the above situation further, say, in a store, item B is in stock and item
A is out of stock. One customer is interested in purchasing B, provided A is also
available in the store, so that he can purchase both A and B. As the demand for
B depends on demand for A, he will not purchase B, if A is not available. Under
this situation, we can say that stock of B is as good as a stock-out situation for that
customer. Hence, if A is not in stock, there will be no sale of B also in many cases.
This example depicts a case of interdependency in demand. Interdependency in the
demand of two items can be captured in various ways and can be used for inventory
management. Association rules can be used to mine demand interdependencies.
Customer Profile and Demand Pattern: Customer profile is the detailed features of
a customer. The profile may contain income, age, marital status, gender, education,
number of cars, family size etc. The profile may also contain frequency of shopping,
credit rating, loyalty index etc. In retail sale, demand pattern depends a lot on the
customer profile along with the other factors. Hence, customer profile is an impor-
tant parameter which may be used for learning the purchase pattern of a customer
and this may be useful for inventory modeling. Classification and clustering can
be used for learning the impact of customer profile on demand pattern. Customer
profile will be a useful input for forecasting the demand of the items.
Sequence of Purchase: Many a times, a sequence of purchase gets repeated most of
the times with a time gap between two purchases. A sequence may be of with two
or more events of purchases. In each event of purchase, certain itemset is purchased.
Once a repetitive sequence rule is identified, it can be used as an input for inventory
modeling. Sequence rules are mined using GSP algorithm.
Time-Dependent Purchase Pattern: On different days of the week, there may be
different purchase pattern of the customers. Purchase patterns on the weekdays
and the weekends are generally different. Similarly, in different months or seasons,
different patterns may be observed. Time-dependent purchase pattern may also be
observed at different hours of the day. Purchase pattern in the evening hours hap-
pens to be different from the day hours. Segregated data with respect to time can be
used to learn time-dependent purchase patterns. For example, data of weekend and
weekdays can be segregated for this purpose.
50 Data Mining for Retail Inventory Management 591

Location-Dependent Purchase Pattern: There may be an impact of culture, cli-

matic condition and other factors on the purchase pattern. Segregated data with
respect to location or space can be used to learn location-dependent purchase
patterns.
Negative Dependency: Sometimes, purchase of one item or itemset results in non-
purchase of another item or itemset. Similarly, non-purchase of one item or itemset
may result in purchase of another item or itemset. This kind of purchase dependency
may be observed within a transaction made by a customer or in successive transac-
tions made by the customer with a time gap.
As discussed, various purchase patterns can be mined using appropriate data
mining tool.

50.4 Model Development

The model proposed has been described in Fig. 50.1. In the present research, it has
been proposed to discover purchase patterns using data mining. For this purpose,
sale transaction data of the inventories contained in the ‘Materials’ module can be

Application of
Data
Multi-Item Inventory
Mining
Management
Techniques

Data Warehouse
Data Store Replenishment
Transaction Data Policy

User
DSS

Data
Collection
Other
Modules

Materials
CRM Subjective
Considerations

Fig. 50.1 Block diagram of the model

592 P.K. Bala

used for mining association rules describing demand interdependencies. Further,

time-dependent and location or space-dependent association rules can be mined by
proper segregation of the past sale transaction data from the “Materials’ module.
Using the modules of ‘Materials’ and ‘CRM (Customer Relationship Management)’
containing the demographical and other profiles of the customers, ‘classification’
technique of data mining can be applied to learn the impact of customer profiles on
purchase pattern. Clustering of customers can also be done. The model can integrate
various transactions based on the decision and action taken in other domains of
finance, maintenance, human resource management, marketing, supply chain etc.
In such cases, we find association rules with a mixture of real items with ‘virtual
items’. Virtual items are basically events, decisions and attributes etc. representing
fields in a database or a data warehouse.
Transaction data is stored in database which is also known as data store and use-
ful for operational decisions. These can not address the requirements of decision
support system (DSS) for inventory management. Data Mining, facilitated by data
warehousing, addresses the need of strategic decision-making. Knowledge discov-
ered from data mining is used in Decision Support System (DSS). In this model,
various types of consumer insights discovered using data mining can be used in
multi-item inventory modeling. The consumer insights mined act as additional input
in designing the inventory replenishment policies.
The model provides scope for subjective considerations also as an input to DSS
along with the objective findings obtained using data mining. The focus in this
model has been to reduce the subjectivity by objective decision-making.
One example in Section 50.5 has been illustrated where association rules have
been mined and classification has been done by building decision tree.

50.5 Illustration with an Example

Eleven (11) items have been considered in a retail sale store with one thousand
(1,000) past sale transaction data along with the profiles of the customers. Eleven
items are named as ‘a’, ‘b’, ‘c’, ‘d’, ‘e’, ‘f’, ‘g’, ‘h’, ‘i’, ‘j’ and ‘k’. A transaction con-
tains the items purchased by a customer along with a small profile of the customers.
The profile contains ‘customer id’, ‘value’ (as rated by the retailer), ‘pmethod’
(payment method by cheque/cash/card etc.), ‘sex’ (Male/Female expressed in M/F),
‘hometown’ (yes/no), ‘income’ and ‘age’. Customer id is an identification code for
the customer and it finds use for capturing the sequence of purchases made by the
same customer and hence in mining sequence rules. Otherwise, customer id is not
of any use for mining other patterns.
To understand the database, as it is not possible to show 1,000 records of the
database in this paper, five records have been shown in Table 50.1 from which one
can visualize various fields in the database.
50 Data Mining for Retail Inventory Management 593

Table 50.1 Sale transaction data with customer profile showing five transactions (part of the
database with 1,000 rows)
Customer profile Items
Value Payment Sex Hometown Income (in $ Age a b c d e f g h i J k
method (Y/N) thousands)
43 Ch M N 27 46 F T T F F F F F F F T
25 Cs F N 30 28 F T F F F F F F F F T
21 cs M N 13 36 F F F T F T T F F T F
24 cd F N 12 26 F F T F F F F T F F F
19 cd M Y 11 24 F F F F F F F F F F F

Table 50.2 Association Rules amongst d, f and g

Antecedent Consequent Association Support Confidence Rule support Lift
rule (%) (%) (%)

d and g F (d, g) =>(f) 17.8 87.4 15.5 2.7

f and g d (f, g) =>(d) 18.1 85.9 15.5 2.7
d and f g (d, f) =>(g) 18.4 84.4 15.5 2.7

In Table 50.1, ‘T’ (under ‘Items’) implies purchase of the corresponding item in
the corresponding transaction and ‘F’ (under ‘Items’) implies that the corresponding
item has not been purchased in the corresponding transaction. Association rules
have been mined from the database of 1,000 transactions with the threshold values
of support and confidence to be 10% and 80% respectively. Association rules mined
have been shown in Table 50.2.
Only three rules have qualified for the chosen threshold values of support and
confidence. In each of these three rules, we observe that same three items (i.e., d, f
and g) are involved.
With respect to the simultaneous purchase of all three items (i.e., d, f and g) in
the same transaction, classification of the customers has been done based on their
profiles. Using data mining, decision tree has been built up from the database for the
purpose of classification which is given in Fig. 50.2. Considering threshold value of
80% confidence for the decision rules, only one rule (at node 3) qualifies. The rule
(with 84.242% confidence) is – IF “income is less than or equal to 16,950 and sex
is F (Female)”, THEN “the customer purchases all three items, i.e., d, f, and g”. The
rule has been observed in 84.242% cases.
The association rules and the decision rule mined from the database can be used
as input for designing inventory replenishment policy. Inventory managers must
consider these rules for placing the orders for replenishment.
594 P.K. Bala

All three of d, f, g purchased

Node.

Category % n
F 85.4 854
T 14.6 146
Total 100 1000

Income
< or = 16950 >16950
Node 4
Node 1
Category % n
Category % n
F 99.2 657
F 58.3 197
T 0.8 5
T 41.7 141
Total 66.2 662
Total 33.8 338

Sex

M F

Node 2 Node 3

Category % n Category % n
F 98.8 171 F 15.8 26
T 1.2 2 T 84.2 139
Total 17.3 173 Total 16.5 165

Fig. 50.2 Decision tree

50.6 Case Discussion

Based on focussed interview with sales people and customers in various grocery
retail outlets, 45 items were selected for the research. The objective was to select
a small number of frequently purchased items where some purchase dependencies
can be anticipated and the list will also contain items where no purchase dependency
is expected. In this way, for the purpose of research, the small subset of items can
be representative of the superset of items with large number of items in retail store.
Hence, the intention of the interview was to foresee a few purchase dependencies
from the experience of sales people and customers and to include the related items
for the purpose of the research. In this way, 45 items were selected where one may
expect purchase dependency for some items and no purchase dependency for some
other items. The list of 45 items with description is as given in Table 50.3. Names
50 Data Mining for Retail Inventory Management 595

Table 50.3 Names of the grocery tems

Item No. Item-name Description

1 All Out liquid (35 ml) Mosquito repellent

2 Atta – Ashirvad (5 kg) Flour
3 Basmati (1 kg) Special rice
4 Bournvita – Cadbury (500 g) Beverage
5 Cashew nuts (100 g) Cashew nuts
6 Chilli sauce – Kissan (200 g) Chilli sauce
7 Coffee – Nescafe (50 g) Beverage
8 Hair oil – Parachute Jasmine (100 ml) Hair oil
9 Hair oil – Shalimar Coconut (100 g) Hair oil
10 Horlicks (500 g) Beverage
11 Kismis (100 g) Dry grapes
12 Lux – International soap (100 g) Bathing soap
13 Maida – Rishta (1 kg) Refined flour
14 MDH chicken masala Packaged chicken spices
15 MDH meat masala Packaged meat spices
16 Noodles – Maggi (100 g) Snacks item
17 Noodles – Maggi (200 g) Snacks item
18 OK powder (500 g) Detergent powder
19 Parle – Hide & Seek (100 g) Biscuit
20 Parle – Hide & Seek – orange cream (100 g) Biscuit with orange cream
21 Pears soap – Green (100 g) Bathing soap
22 Pepsodent G (80 g) Toothpaste
23 Poha (Chura) (1 kg) Foodstuff
24 Pond’s Dreamflower powder (100 g) Body powder
25 Priya pickles – Mango (200 g) Pickles
26 Refined oil – Safflower – fora (1 l) Refined oil
27 Rin Supreme (100 g) Washing soap
28 Ruchi chilli powder (100 g) Chilli powder
29 Ruchi chilli powder (50 g) Chilli powder
30 Ruchi curry powder (50 g) Curry powder
31 Ruchi haldi powder (100 g) Turmeric powder
32 Salt – Tata (1 kg) Salt
33 Shampoo – Clinic All Clear (50 ml) Shampoo
34 Shampoo – Clinic Plus (50 ml) Shampoo
35 Shaving cream – Old Spice (100 g) Shaving cream
36 Sooji (1 kg) Foodstuff
37 Sugar (1 kg) Sugar
38 Sundrop refined oil (1 l) Refined oil
39 Surf Excel (500 g) Detergent powder
40 Tea – Tata Leaf (100 g) Beverage
41 Tide detergent (200 g) Detergent powder
42 Tide detergent (500 g) Detergent powder
43 Tomato sauce (Lalls) (200 g) Tomato sauce
44 Tomato sauce (Maggi) (200 g) Tomato sauce
45 Vim bar (100 g) Utensil cleaning soap
596 P.K. Bala

Table 50.4 Association rules amongst the grocery Items

Rule # Antecedent Consequent

1 Noodles – Maggi (200 g). and Tomato MDH chicken masala

sauce (Maggi) (200 g)
2 MDH chicken masala Noodles – Maggi (200 g). and Tomato
Sauce
3 MDH chicken masala. and Noodles – Tomato sauce (Maggi) (200 g)
Maggi (200 g)
4 Tomato sauce (Maggi) (200 g). MDH chicken masala. and Noodles –
Maggi (200 g)
5 Kismis (100 g) Basmati (1 kg)
6 Tomato sauce (Maggi) (200 g) MDH chicken masala
7 MDH chicken masala Tomato sauce (Maggi) (200 g)
8 Coffee – Nescafe (50 g) Bournvita – Cadbury (500 g)
9 MDH chicken masala. and Tomato Noodles – Maggi (200 g)
sauce (Maggi) (200 g)
10 Poha (chura) (1 kg). and Sooji (1 kg). Noodles – Maggi (200 g)
11 Poha (chura) (1 kg) Noodles – Maggi (200 g)
12 Sooji (1 kg) Noodles – Maggi (200 g)
13 MDH chicken masala Noodles – Maggi (200 g)
14 Tomato sauce (Maggi) (200 g) Noodles – Maggi (200 g)

of the items have been written as usually told by the Indian shoppers. The list of
45 items contains 29 food items and 16 miscellaneous grocery items. Miscellaneous
grocery items constitute of cosmetics, detergents, toothpaste and mosquito repellent.
In view of identifying ‘purchase dependencies within a group of items’ and ‘inter-
group purchase dependencies’ the variety of items were chosen.
A data volume of 8,418 sales transactions (at least one of the selected 45 items
is present in each transaction) with quantity of the items was collected from various
grocery retail outlets of different sizes (small and large) in two different cities in
India.
For threshold of 20% support and 65% confidence, 14 association rules were
obtained as given in Table 50.4. In this table, it is observed that rules # 1, 2, 3, 4 and 9
are amongst the items, Noodles – Maggi (200 g), Tomato Sauce (Maggi) (200 g) and
MDH Chicken Masala. Rules 6,7,13 and 14 are between any two items of Noodles
– Maggi (200 g), Tomato Sauce (Maggi) (200 g) and MDH Chicken Masala. This
is like the example illustrated in Section 50.5 which involves three items d, f and g.
The rules which involve these three items have been shown separately in Table 50.5.

50.7 Conclusions

Consumer insight extracted using data mining tools has been used for business deci-
sion making in marketing, shelf design, product management etc. There has been
limited use in inventory management. Purchase dependency as a type of consumer
50 Data Mining for Retail Inventory Management 597

Table 50.5 Association rules amongst three selected items

Rule # Antecedent Consequent

1 Noodles – Maggi (200 g). and Tomato MDH chicken masala

sauce (Maggi) (200 g)
2 MDH chicken masala Noodles – Maggi (200 g). and Tomato
sauce (Maggi) (200 g)
3 MDH chicken masala. and Noodles – Tomato sauce (Maggi) (200 g)
Maggi (200 g)
4 Tomato sauce (Maggi) (200 g) MDH chicken masala. and Noodles –
Maggi (200 g)
6 Tomato sauce (Maggi) (200 g) MDH chicken masala.
7 MDH chicken masala. Tomato sauce (Maggi) (200 g)
9 MDH chicken masala. and Tomato Noodles – Maggi (200 g)
sauce (Maggi) (200 g)
13 MDH chicken masala Noodles – Maggi (200 g)
14 Tomato sauce (Maggi) (200 g) Noodles – Maggi (200 g)

insight which can be applied in inventory modeling has been discussed for the first
time in this paper. The analysis and findings in this research throws some light
on various aspects of purchase dependencies which may be useful for inventory
management in retail stores. The model, illustrative example and the case discus-
sion in this research paper will motivate researchers and inventory managers to
develop methodologies for using the findings of data mining in multi-item inventory
management.
There is scope for further work for identifying the guiding profile of the cus-
tomers which causes in simultaneous purchase of three items, Noodles – Maggi
(200 g), Tomato Sauce (Maggi) (200 g) and MDH Chicken Masala. This knowl-
edge will of tremendous importance for designing inventory replenishment policy
in retail sale.

References

1. Han, J., & Kamber, N., “Data Mining: Concepts and Techniques”, Morgan Kaufmann, San
Francisco, CA, 2006, pp. 227–284.
2. Pujari, A.K., “Data Mining Techniques”, Hyderabad, University Press, 2002, pp. 259–260.
3. Cohen, M.A., & Ernst, R., “Multi-item classification and generic inventory stock control
policies”, Production and Inventory Management Journal, 29(3), 1988, pp. 6–8.
4. Ernst, R., & Cohen, M.A., “A clustering procedure for production/inventory systems”, Journal
of Operations Management, 9(4), 1990, pp. 574–598.
5. Guvenir, H.A., & Erel, E., “Multicriteria inventory classification using a genetic algorithm”,
European Journal of Operations Research, 105(1), 1998, pp. 29–37.
6. Partovi, F.Y., & Anandarajan, M., “Classifying inventory using an artificial neural network
approach. Computers & Industrial Engineering”, 41, 2002, pp. 389–404.
7. Bhattacharya, D.K., “Production, manufacturing and logistics on multi-item inventory”,
European Journal of Operational Research, 162(3), 2005, pp. 786–791.
598 P.K. Bala

8. Kuo, R.J., An, A.L., Wang, H.S., & Chung, W.J., “Integration of self-organizing feature maps
neural network and genetic K-means algorithm for market segmentation”, Expert Systems
with Applications, 30, 2006, pp. 313–324.
9. Hu, T. L., & Sheu, J.B., “A fuzzy-based customer classification method for demand-responsive
logistical distribution operations”, Fuzzy Sets and Systems, 139, 2003, pp. 431–450.
10. Larivie‘re, B., & Piel, D.V.D., “Investigating the role of product features in preventing cus-
tomer churn, by using survival analysis and choice modeling: The case of financial services”,
Expert Systems with Applications, 27, 2004, pp. 277–285.
11. Wong, W., Fu, A.W., & Wang, K., “Data mining for inventory item selection with cross-selling
considerations”, Data Mining and Knowledge Discovery, 11(1), 2005, pp. 81–112.
Chapter 51
Economic Process Capability Index for Product
Design and Process Planning Economic Process
Capability Index

Angus Jeang

Abstract The process capability index (PCI) is a value which reflects real-time
quality status. The PCI acts as the reference for real-time monitoring that enables
process controllers to acquire a better grasp of the quality of their on site processes.
The PCI value is typically defined as the ability to carry out a task or achieve a
goal. However, simply increasing the PCI value can easily create additional and
unnecessary production costs that result from extra efforts and expensive devices
for ensuring tolerance control. Hence, there is a need to balance customer demands
for quality and production costs. In this regard, the off-line PCI value is intro-
duced, in consideration of quality loss and production cost, simultaneously in this
research. The quality loss is expressed by quality loss function, and the production
cost is represented by tolerance cost function. Then, this new PCI expression can be
used as linkage for concurrent product design and process planning, prior to actual
production.

Keywords Process capability index  real-time quality status  production cost 

off-line  tolerance cost function, concurrent product design

51.1 Introduction

In recent years, as the concept of concurrent engineering has become widely

accepted, design engineers hope to achieve simultaneous product design and pro-
cess planning, side by side, at an early stage of product development [1]. The goals
are: to shorten the time span required for introducing the new product onto the mar-
ket and to attain the lowest production cost and premium product quality. Hence,

A. Jeang
Department of Industrial Engineering and Systems Management, Feng Chia University, P.O. Box
25-150, Taichung, Taiwan, R.O.C.
E-mail: akjeang@fcu.edu.tw

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 599
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
600 A. Jeang

what is needed is a way to measure the degree of the producer’s process capabil-
ity, in satisfying the customer’s quality requirement. More importantly, a growing
number of producers include this measurement value in their purchase contracts
with customers, as a documentation requirement [2]. One such measurement is the
process capability index (PCI).
The process capability index (PCI) is a value which reflects real-time quality
status. The PCI acts as the reference for real-time monitoring that enables process
controllers to acquire a better grasp of the quality of their on site processes [3, 4].
Although the PCI is considered as one of the quality measurements employed during
on-line quality management, several authors have pointed out that the PCI should
be addressed at the beginning of the design stage rather than at the production stage,
where process capability analysis is typically done [Shina, 1995]. For the sake of
convenience, let us call the PCI for the former one, off-line PCI, and the latter one,
on-line PCI. The on-line PCI has realized process mean and process variance that
are obtained from the existing process. Conversely, the off-line PCI has the process
mean and process variance as two unknown variables, which the product designer
and process planner would have to determine. When cost is not considered as a
factor for off-line PCI analysis; normally the process planners would do their best
to set the process mean close to the design target, and minimize the process variance
to the process limit. Because the additional cost incurred for tightening the variance
is not considered, obviously, the establishment of mean and variance values will
result in a high PCI scale [6]. Thus, a PCI expression which contains cost factors
for an Off-line application is developed.
The PCI value is typically defined as the ability to carry out a task or achieve a
goal. The controllable factors are the process mean and process variance [7]. The
deviation between process mean and design target can be reduced by locating the
process mean close to the design target without additional cost being incurred.
The process variance can be lowered by tightening the process tolerance, with extra
cost incurred. In case the conventional on-line PCI is used for process capability
analysis during the product and process designs, designer engineers naturally intend
to raise the PCI value by locating the process mean near the target value, and by
reducing the tolerance value to ensure a better product quality. However, simply
increasing the PCI value can easily create additional and unnecessary production
costs that result from extra efforts and expensive devices for ensuring tolerance
control. Hence, there is a need to balance customer demands for quality and pro-
duction costs. In this regard, the off-line PCI value is introduced, in consideration
of quality loss and production cost, simultaneously in this research. The quality
loss is expressed by quality loss function, and the production cost is represented
by tolerance cost function. Then, this new PCI expression can be used as linkage
for concurrent product design and process planning, prior to actual production. The
rationale will be discussed in the latter sections.
51 Economic Process Capability Index for Product Design and Process Planning 601

51.2 Process Capability Indices (PCI)

The frequently seen PCI includes Cp , Cpk , and Cpm expressions. Cp can be defined
as follows [3, 4, 8–10]:
USL  LSL
Cp D (51.1)
6
The expression (USL  LSL) refers to the difference between the upper and lower
limits which are specified by the customer’s quality requirement;  is the standard
deviation which is actually incurred in the production process. However, during the
production process, the process means U can be located at positions other than
design target. If the process variance  2 did not change, the above Cp value would
also remain unchanged; this was the major defect owing to the facts that only the
spread of the process is reflected, and the deviation of process mean can not be
reflected in the measurement. These are the main reasons why Cpk was developed;
the Cpk expression is defined as below:
 
USL  U U  LSL
Cpk D Min ; (51.2)
3 3

There is still a deficiency for Cpk expression: the same Cpk values may be consti-
tuted with different process means and variances. This situation has created a great
deal of confusion, and uncertainty as to which would be the best process capability
among the alternatives. To cope with the above arguments, another form of PCI,
Cpm , was developed. Cpm is defined as follows:

USL  LSL
Cpm D q (51.3)
6  2 C .U  T /2

When the process mean is equal to design target, the Cpm can be simplified as
Cp . For the purpose of comparison, three processes: A, B, and C are depicted in
Fig. 51.1. The Cp , Cpk , and Cpm values from processes A, B, and C are shown in
Table 51.1. Because process C has the greatest Cp value, it is might be mistakenly
concluded that process C had the best process capability among processes A, B,
and C, when Cp is considered as a reference value. However, this erroneous conclu-
sion originates from the fact that the Cp value is solely based on the magnitude of
variance, and disregards the negative impact from the process deviation. Similarly,
when Cpk was used in representing the levels of process capability, the process Cpk
values for processes A, B, and C, are all just equal to one. Thus again, quite obvi-
ously, there is difficulty in ordering the superiority of process capability of the three
processes. To overcome the defects appearing in Cp and Cpk expressions, another
PCI expression, Cpm is introduced. Unlike the previous two expressions, Cpm can
simultaneously reflect the impact from process deviation and process variance. This
feature is particularly important because process mean and process variance are
602 A. Jeang

LSL T USL

35 50 65

Fig. 51.1 The Distribution for process A, B and C

Table 51.1 PCI values for processes A, B, and C

Process Cp Cpk Cpm Cpmc
A 1 1 1 0.051298
B 1.25 1 1 0.047673
C 2 1 0.632456 0.031782

generally changed at the same time in most production process. Unfortunately,

with Cpm , processes A and B are the best two choices. The non-different outcomes
between processes A and B result from the fact that the contribution of Cpm value,
from process mean deviation and process variance magnitude, is identical. Hence,
there must be a way of measurement being provided to make mean deviation and
process variance magnitude distinguishable in PCI expression.
Note: Process A: UA D 50:0, A D 5:00, CM.t/ D $2;000,
Process B: UB D 53:0, B D 3:83, CM.t/ D $3;500,
Process C: UC D 57:5, C D 2:50, CM.t/ D $6;000
The conventional PCI values are mainly applied to measure the on-line spread
resulting from actual process deviation and process variance. The reality of this
spread is a consequence of the production process, in which there are no more
controllable elements. Hence, the conventional PCI expression is valid only in the
on-line application. In this regard, there is an attempt to explore the process capabil-
ity analysis, before production process, to enable designers to integrate the aspects
of product design and process planning at an early time. According to the preced-
ing discussion, unlike the other two PCI expressions, Cpm is able to simultaneously
reflect the influences of process deviation and process variance. However, this is
51 Economic Process Capability Index for Product Design and Process Planning 603

only legitimate after production process when realized U and t are no more con-
trollable variables for a design. However, when Cpm is used as a measurement scale
before production process, U andtbecome controllable variables. Then, it is possi-
ble that various combinations of U and t will result in the same Cpm value. Thus,
it is difficult to make a distinction among alternatives, in order to make a correct
choice among them. Additionally, the designers would most likely establish the pro-
cess mean U as close as possible to the design target T , within the process feasible
range, and attempt to decrease the process variance as much as possible within the
process capability limits in order to attain a higher Cpm value. In other words, with
the exclusive use of the process mean and process tolerance as the determinants of
conventional Cpm expression, regardless of the cost impact on customer and pro-
duction, there is a tendency for designers to position the process mean as close to
the target value as possible, and solely cut down the process tolerance to lower capa-
bility limit in order to increase the Cpm value. Apparently, the found PCI value is
erroneous.
The degree of proximity reflects the different quality loss according to the cus-
tomer’s view. Reducing the process variance is normally completed by tightening
the tolerance value through tolerance design which usually involves additional cost.
Therefore, in addition to the constraints from feasible ranges and capability lim-
its, the influence exerted by the relevant costs representing the selected process
mean and process tolerance, should be considered as well. This brings us to the
next section, a discussion on the requirement that costs related to process mean and
process tolerance must be contained in PCI expression, when referred to as off-line
process capability analysis, during product design and process planning.

51.3 Proposed Process Capability Index

51.3.1 Single Quality Characteristic

Because various combinations of U and t will result in the same Cpm value, this
unhelpful facet will prevent the conventional Cpm from being a suitable index for
differentiating possible alternatives during product design or process planning. To
overcome the above weakness, the lack of consideration of the cost influence from
various U and t values, should be resolved. As is known, all costs incurred within
a product life cycle, include the material and production costs which are incurred
before the product reaches the consumer, and quality loss, which occurs after a
sale. In these regards, let the denominator of Cpm be replaced with total cost, TC,
which is the sum of quality and production related cost, which includes quality
loss KΠ2 C .U  T /2  and tolerance cost CM .ti /. To evaluate the tolerance cost,
this paper adopts the tolerance cost function as developed in the literature [11].
CM .t/ D a C b exp .ct/, where a, b, and c are the coefficients for the tolerance
cost function, and t is the process tolerance.
604 A. Jeang

Table 51.2 Various CPMC when CPM is 1.2

(U,t) (50.00, (50.05, (51.00, (51.05, (52.00, (52.05,
12.50) 12.49) 12.13) 12.09) 10.96) 10.88)
Cpm 1.2000000 1.2000000 1.2000000 1.2000000 1.2000000 1.2000000
Cpmc 0.0345994 0.0346019 0.0356385 0.0357504 0.0394258 0.039727

Have both U and t be controllable variables so that a maximum Cpmc can

be achieved. This developed cost effectiveness PCI expression Cpmc is shown as
follows.
USL  LSL
Cpmc D q (51.4)
6 KΠ2 C .U  T /2  C CM .t/
tL  t  tU (51.5)
UL  U  UU (51.6)

The values of tL , tU , UU , UL , S , and T are known in advance.  is t/P, P

is given [12]. There are infinite combinations of U and t which have the same
Cpm value. Table 51.2 shows that different Cpmc values can be obtained under vari-
ous combinations of U and t, when Cpm is fixed as 1.2. Apparently, the conventional
Cpm is not capable of representing all possible alternatives for off-line application
during design and planning stage. Table 51.1 shows that processes A and B are
identical choices based on Cpm ; however, process A is the only selection based on
Cpmc value. Different selections are being made because Cpm lacks consideration
of the combined influence from quality loss for customers and tolerance costs for
production.
Example 1 – Let design target value T D 30 mm, design tolerance S D 0:05 mm,
quality loss coefficient K D 1;200, coefficients of tolerance function a D 50:11345,
b D 119:3737, c D 31:5877, P D 3, tL D 0:024 mm, tU D 0:086 mm. Substitute these
values into Eqs. (51.3–51.6), to proceed the following discussion.

The optimal Cpmc is 0.0019. The optimal process mean U  is 30.0000 and the

process tolerance t is 0.05 mm. When the process mean is located at the target with
fixed tolerance value, the maximum Cpm and Cpmc can be reached. The explanation
is discernable by looking into the common expression, .U  T /2 , in the denom-
inator of Eqs. (51.3 and 51.4). On the other hand, with fixed process mean at the
target value, the maximum value Cpm , which is infinite, is reached when t is near
to zero and the maximum Cpmc , which is finite, arrives when t is 0.04. The fact
behind different optimal t values being found when the PCI is at its maximum, is
comprehensible because the variance in Cpmc is cost related, while the variance in
Cpm is cost unrelated. The t value in Cpm can be any small value regardless of the
cost impact resulting from tolerance cost.
51 Economic Process Capability Index for Product Design and Process Planning 605

51.3.2 Multiple Quality Characteristics

As discussed in the preceding section, a lower quality loss (better quality) implies
a higher tolerance cost, and a higher quality loss (poorer quality) indicates lower
tolerance cost. Hence, the design parameters must be adjusted to reach an eco-
nomic balance between reduction in quality loss and tolerance cost, so that the cost
effectiveness PCI expression, Cpmc , is maximized. The model developed in the fol-
lowing illustration attempts to achieve this objective in the case of multiple quality
characteristics.
Before model development, a functional relationship between the dependent vari-
able Y and independent variable X should be identified and thoroughly understood.
Based on this relationship, the resultant overall quality characteristic such as Y
and UY can be estimated from the set of individual quality characteristics in both
product and process. The proposed process capability index, Cpmc; for the multiple
quality characteristics is:
USL  LSL
Cpmc D s (51.7)
P0
M
6 KŒ.UY  TY /2 C Y2  C CM .ti /
i D1

where M0 is the total number of quality characteristics in a product or process.

Example 2 – A DC circuit has four resistances: R1 , R2 , R3 , and R4 , as shown
in Fig. 51.2 [13]. The design function representing the functional relationship
between

dependent variable Y and independent variables R1 , R2 , R3 , and R4 , is
R
E1  1C R2
R
1
 E2R1R2 , where E1 is 1.0 V and E2 is 1.0 V. M0 is 4, so i is 1, 2, 3, 4.
1C R3
4

R1 R2
E2

-
V-
V0
R3
V+
E1 +

R4

Fig. 51.2 A DC differential amplifier

606 A. Jeang

Table 51.3 The numerical results of example 2

C/K I II
Cpmc Cpm Cpmc Cpm
U1 = 11 U1 = 11 U1 = 11 U1 = 11
U2 = 105.47 U2 = 105 U2 = 105.47 U2 = 105
U3 = 8 U3 = 8 U3 = 8 U3 = 8
U4 = 115 U4 = 115 U4 = 115 U4 = 115
t1 = 0.6 t1 = 0.6 t1 = 0.6 t1 = 0.6
t2 = 1.2 t2 = 1.2 t2 = 1.6769 t2 = 1.2
I t3 = 0.8 t3 = 0.5 t3 = 0.8 t3 = 0.5
t4 = 6 t4 = 4.9477 t4 = 6 t4 = 4.9477
UY = 19.494 UY = 18.87 UY = 19.494 UY = 18.87
tY = 1.0363 tY = 0.5704 tY = 1.0363 tY = 0.5704
TC = 5.0377 TC = 26.895 TC = 4.2606 TC = 17.562
C=pm = 1.1467 Cpm = 3.2544 C=pm = 1.1355 Cpm = 3.2544
Cpmc = 0.1782 Cpmc = — Cpmc = 0.1937 Cpmc = —
U1 = 11 U1 = 11 U1 = 11 U1 = 11
U2 = 105.47 U2 = 105 U2 = 105.47 U2 = 105
U3 = 8 U3 = 8 U3 = 8 U3 = 8
U4 = 115 U4 = 115 U4 = 115 U4 = 115
t1 = 0.6 t1 = 0.6 t1 = 0.6 t1 = 0.6
t2 = 1.2 t2 = 1.2 t2 = 1.2 t2 = 1.2
II t3 = 0.8 t3 = 0.5 t3 = 0.8 t3 = 0.5
t4 = 6 t4 = 4.9477 t4 = 6 t4 = 4.9477
UY = 19.494 UY = 18.87 UY = 19.494 UY = 18.87
tY = 1.0363 tY = 0.5704 tY = 1.0363 tY = 0.5704
TC = 7.4205 TC = 31.215 TC = 6.6713 TC = 20.188
C=pm = 1.1721 Cpm = 3.2544 C=pm = 1.1577 Cpm = 3.2544
Cpmc = 0.1468 Cpmc = — Cpmc = 0.1548 Cpmc = —

The design target value TY is 20. Design specification SY is 2. Namely, USL 

LSL is 4.0. The quality coefficient K is 6,000. Assume P is 3. M0 is 4. The coeffi-
cients of tolerance cost function, a, b, and c, are given in Table 51.3. The ci value is
further divided into coefficients I, II for the purpose of sensitivity analysis done in
Table 51.2. The intention is to determine the process mean U and process tolerance
t, so that the present PCI expression, Cpmc , is maximized. The formulation is given
in the following expression.
First, let the expressions UY , Y , tY be found by expansion design function
through Taylor’s series.
2 !   !
U U
2U2 2 1C U2 U32 2 1C U2 U3
2
C 2U2 t42 1
 1
6 t1 U13 U  3  2
U2 6 U13 1C U3 U
1C U3 U44
U
1C U3 U43
C 0:5 6
4
UY D 6 C 4 4

U1 4 P2 P2

 3
2t32 1 C U U2
7 1C
2
U1 U1
C  3 5C U3
(51.8)
2 U3
P 1 C U4 U4 2 1C U4
51 Economic Process Capability Index for Product Design and Process Planning 607
!2 !2
t22 1
U1
C  1
U t12  UU22   U2
U
U1 1C U3 1 U12 1C U3
4 4
Y2 D C
P2 P2
 2  2
t42 1 C U
U1
2
U32 t32 1 C U
U1
2

C  4
C  4
(51.9)
P2 1 C UU4
3 4
U4 P2 1 C U 3
U4
U42

0 12 0 12
1 1 A C t12 @ U2  U2
tY2 D t22 @ C  2
 A
U1 U1 1 C U 3 U1 U12 1 C U3
U4 U4
 2  2
t42 1 C U
U1
2
U32 t32 1 C U 2
U1
C  4
C  4
(51.10)
1C U 3
U4 U 4
4 1C U 3
U4 U42

Objective function:
USL  LSL
Max Cpmc D r P (51.11)
6 KŒ.UY  TY /2 C Y2  C CM .ti /
i D1

Subjective to:

j TY  U Y j <
D SY  tY (51.12)
tLi  ti  tU i (51.13)
ULi  Ui  UUi (51.14)

decision variables are ti and Ui , where i is 1, 2, 3, and 4.

UY ; Y2 ; and tY2 are obtained from Eqs. (51.8)–(51.10).
Then, the result is summarized in Table 51.2. For the purpose of comparison,
let the above objective function Cpmc , Eq. (51.11), be replaced by Cpm , Eq. (51.3),
with the same constraints. Then, relevant solutions are also listed in Table 51.2.
The solutions from objective functions Cpmc and Cpm are arranged pair wisely for
easy comparison. Clearly, the solutions found with Cpm as an objective function are
insensitive to the changes of quality loss and tolerance cost; instead, the solutions
based on Cpmc fully reveal the impact from various combinations of quality loss
and tolerance cost. Let the optimal solutions, U  and t  , found from Cpmc as an
=
objective function, be substituted into the Cpm expression to have Cpm , the values
=
Cpm still fall within acceptable range for general application. The reason for opti-
=
mal value Cpm being greater than Cpm is due to the irrational combination of U and
t, when Cpm is used as an objective function. This irrational combination, resulting
from disregarding the effect of quality loss and tolerance cost in Cpm expression,
608 A. Jeang

drives the Cpm value to become unusually high. Similar reasons can also explain
why the total cost, TC, is always higher when Cpm is used as objective function than
is the case for Cpmc . In other words, Cpmc is an appropriate expression for appli-
cation in the process capability analysis during off-line application; particularly, for
product design and process planning at the stage of blueprint.

51.4 Summary

Conventionally, production related issues are usually dealt with via process analysis
after the product design or process planning has been completed. These approaches
will likely result in poor product quality and high production cost, as a consequence
of a lack of consideration concerning quality and production related costs. These
are some of the reasons that concurrent engineers suggesting that possible issues
occurring in the production stages should first be considered at the time that the new
product is developed. That can reduce the time span for introducing a new product
onto market, and increase the chance for obtaining a superior edge among competi-
tors. Thus, the present research introduces a PCI measurement, Cpmc , for the process
capability analysis, to ensure that lower production cost and high product quality
can be achieved at the earlier time of the blue print stage. For an application, prior
to production, process engineers can establish process mean and process tolerance
based on the optimized mean and tolerance values via the above three approaches.
As the expectation, the produced quality values after production process must be
distributed with statistic values as the process mean and process tolerance estab-
lished before production process. As a result, an effective PCI for product life cycle
becomes actualized.

References

1. Carter, D. E. and Baker, B. S., 1992, Concurrent Engineering: The Product Development
Environment for the 1990s, Addison-Wesley, New York.
2. Schneider, H. and Pruett, J., 1995, Use of Process Capability Indices in the Supplier
Certification Process, Qual. Eng., 8: 225–235.
3. Kotz, S. and Johnson, N. L., 1993, Process Capability Indices, Chapman & Hall, London.
4. Kotz, S. and Lovelace, C. R., 1998, Process Capability Indices in Theory and Practice, Arnold,
a Member of the Hodder Headline Group, New York.
5. Shina, S. G. and Saigal, A., 2000, Using Cpk as a Design Tool for New System Development,
Qual. Eng., 12: 551–560.
6. Spiring, F. A., 2000, “Assessing Process Capability with Indices” in Statistical Process Mon-
itoring and Optimization, edited by S. H. Park and G. G. Vining, Marcel-Dekker, New
York.
7. Jeang, A., 2001, Combined Parameter and Tolerance Design Optimization with Quality and
Cost Reduction, Int. J. Prod. Res., 39(5): 923–952.
8. Boyle, R. A., 1991, The Taguchi Capability Index, J. Qual. Technol., 23: 17–26.
51 Economic Process Capability Index for Product Design and Process Planning 609

9. Chan, L. K., Cheng, S. W., and Spring, F. A., 1989, A New Measurement of Process
Capability: Cpm , J. Qual. Technol., 20: 162–175.
10. Kane, V. E., 1986, Process Capability Index, J. Qual. Technol., 18: 41–52.11.
11. Chase, K. W., Greenwood, W. H., Loosli, B. G., and Haugland, L. F., 1990, Least Cost Toler-
ance Allocation for Mechanical Assemblies with Automated Process Selection, Manuf. Rev.,
3(1): 49–59.
12. Jeang, A., Liang, F., and Chung, C. P., 2008, Robust Product Development for Multiple Quality
Characteristics Using Computer Experiments and an Optimization Technique, Int. J. Prod.
Res., 46(12): 3415–3439.
13. Boyd, R. R., 1999, Tolerance Analysis of Electronic Circuits Using Matlab, CRC Press, New
York.
Chapter 52
Comparing Different Approaches for Design
of Experiments (DoE)

Martı́n Tanco, Elisabeth Viles, and Lourdes Pozueta

Abstract Design of Experiments (DoE) is a methodology for systematically apply-

ing statistics to experimentation. Since experimentation is a frequent activity at
industries, most engineers (and scientists) end up using statistics to analyse their
experiments, regardless of their background. OFAT (one-factor-at-a-time) is an old-
fashioned strategy, usually taught at universities and still widely practiced by com-
panies. The statistical approaches to DoE (Classical, Shainin and Taguchi) are far
superior to OFAT. The aforementioned approaches have their proponents and oppo-
nents, and the debate between them is known to become heated at times. Therefore,
the aim of this paper is to present each approach along with its limitations.

Keywords Design of Experiments  Classical  Shainin  Taguchi  statistical app-

roach

52.1 Introduction

Lye [1] defined the Design of Experiments (DoE) as a methodology for systemat-
ically applying statistics to experimentation. More precisely, it can be defined as a
series of tests in which purposeful changes are made to the input variables of a pro-
cess or system so that one may observe and identify the reasons for these changes
in the output response(s) [2].
Since experimentation is a frequent activity at industries, most engineers (and
scientists) end up using statistics to analyse their experiments, regardless of their
background [3]. Not all engineers are exposed to statistics at the undergraduate level,
and this leads to problems when tools are required in practice; either they don’t know

M. Tanco (B)
Department of Industrial Management Engineering at TECNUN (University of Navarra),
Paseo Manuel Lardizabal 13, 20018 San Sebastian, Spain
E-mail: mtanco@tecnun.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 611
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
612 M. Tanco et al.

what type of experimental strategy is required for their problem and they select
something inappropriate, or they select the correct strategy and apply it incorrectly,
or they select the wrong strategy in which case it probably doesn’t matter whether
they use it correctly or not [4].
OFAT (one-factor-at-a-time) is an old-fashioned strategy, usually taught at uni-
versities and still widely practiced by companies. It consists of varying one variable
at a time, with all other variables held constant. On the other hand, DoE is an
efficient technique for experimentation which provides a quick and cost-effective
method for solving complex problems with many variables. Numerous case studies
in different areas of application prove the advantages and potential of DoE [5].
The statistical approach to Design of Experiments and the Analysis of Variance
(ANOVA) technique was developed by R.A. Fisher in 1920. Since then, many have
contributed to the development and expansion of this technique. Most techniques,
defined throughout this article as “Classical”, have adapted Fisher’s ideas to various
industries, including agriculture. However, engineers Shainin and Taguchi are espe-
cially influential due to their contribution of two new approaches to DoE. Both new
approaches offer more than just Design of Experiments, as they can be considered
quality improvement strategies [6].
All three approaches to DoE (Classical, Shainin and Taguchi) are far superior
to OFAT. The aforementioned approaches have their proponents and opponents, and
the debate between them is known to become heated at times. Dr. Deming once said,
“Any technique is useful as long as the user understands its limitations.” Therefore,
the aim of this paper is to present each approach along with its limitations.
Most engineers have surely at least heard of Design of Experiments (DoE),
Taguchi Methods or the Shainin SystemTM . Yet how many of them can really
say that they understand the differences among them? Or that they are capable of
correctly deciding when to use which technique?
The answer to these questions is heavily influenced by the knowledge and expe-
rience one has of each approach to Design of Experiments. Since the majority of
practitioners only have experience with a single approach, this article aims to open
their minds and compare available approaches.
Each approach to Design of Experiments will be briefly described in chronolog-
ical order in the following section. In Section 52.3, the main criticisms published in
literature about each approach are highlighted. Finally, the conclusion and recom-
mendations for engineers and managers working in the manufacturing industry are
presented in Section 52.4.

52.2 Approaches to Design of Experiments

52.2.1 The Classical Approach

Although books on Design of Experiments did not begin to appear until the twen-
tieth century, experimentation is certainly about as old as mankind itself [7]. The
one-factor-at-a-time strategy (OFAT) was, and continues to be, used for many years.
52 Comparing Different Approaches for Design of Experiments (DoE) 613

However, these experimentation strategies became outdated in the early 1920s when
Ronald Fisher discovered much more efficient methods of experimentation based
on factorial designs [1]. Those designs study every possible combination of fac-
tor settings, and are especially useful when experimentation is cheap or when
the number of factors under study is small (less than five). Fisher first applied
factorial designs to solve an agricultural problem, where the effect of multiple vari-
ables was simultaneously (rain, water, fertilizer, etc.) studied to produce the best
crop of potatoes. His experiences were published in 1935 in his book “Design of
Experiments” [8].
Fractional Factorial designs were proposed in the 1930s and 1940s in response
to the overwhelming number of experiments that are involved with full factorial
designs. This design consists of a carefully selected fraction of the full factorial
experimental design. They provide a cost-effective way of studying many factors
in one experiment, at the expense of ignoring some high-order interactions. This is
considered to be low risk, as high order interactions are usually insignificant and
difficult to interpret anyway.
According to Montgomery [2], the second stage in the era of the classical
approach to DoE began in the 1950s when Box & Wilson [9] developed what was
later called Response Surface Methodology (RSM). Their methodology allowed
DoE to be applied in the chemical industry and afterwards in other industries as
well. They touted the advantages of industrial experiments compared to agricultural
experiments in the two following areas: (a) Immediacy: response can be obtained
quicker than with agricultural experiments, when results can sometimes take up to a
year to be obtained; (b) Sequentially: the experimenter is able to carry out few exper-
iments, analyse them and again plan new experiments based on findings obtained
from the previous experiments. In this era Central Composite designs (CCD) and
Box-Behnken designs (BBD) were created.
The third era of the classical approach started with the appearance of the Taguchi
and Shainin approach in the US in the 1980s as a simple and efficient method
of experimentation. Eventually, statisticians and academics began to acknowledge
the value of certain engineering ideas of Taguchi and Shainin. This led to positive
changes, adopting many ideas of the new approaches (for example, the reduction
of variance became an important research area within classical design), and giving
importance to constructing methodologies and guidelines to ease application.
In this last era, the democratization of statistics, thanks in part to software pack-
ages and the spread of Six Sigma thinking throughout industries [10], helped spread
Design of Experiments to all types of industries. Moreover, an increasing interest
in literature was advocated to Design of Experiments [11]. Furthermore, software
packages have made the construction of graphs and calculus easier, further facilitat-
ing the application of DoE. Recent books by Funkenbusch (2005) [12] and Robinson
(2000) [13] show how this approach can be understood by engineers.
Many scientists and statisticians have contributed to DoE development, making
the classical approach a valid and robust methodology for Design of Experiments.
Usual reference books are Box et al. [14] and Montgomery [2].
614 M. Tanco et al.

52.2.2 The Taguchi Approach

As a researcher at the Electronic Control Laboratory in Japan, engineer Genechi

Taguchi carried out significant research on DoE techniques in the late 1940s.
Although he published his first book in Japanese in the 1950s, the standardized
version of DoE, popularly known as the Taguchi Method, wasn’t introduced in the
US until the early 1980s. His most influential books were “Introduction to Quality
Engineering” [15] and “System of Experimental Design” [16]. The following decade
was rife with heated debate mounted by two distinct camps of professionals, one
unfailingly extolling the new-found virtues and power of Taguchi methods, while
the other persistently exposed the flaws and limitations inherent to them [17].
Taguchi used and promoted statistical techniques for quality from an engineer-
ing perspective rather than from an statistical perspective [18]. Although Taguchi
has played an important role in popularising DoE, it would be wrong to consider
Taguchi Methods as just another way to perform DoE.
He developed a complete problem solving strategy [6], which he dubbed “Quality
Engineering”. However, there is general confusion in industry literature when deal-
ing with systems aiming to reduce variation, as the terms robust design, Taguchi
Methods, quality engineering and parameter design are used as synonyms [19].
The basic elements of Taguchi’s quality philosophy can be summarized as
follows [15, 18, 20]:

A quality product is a product that causes a minimal loss to society during

its entire life. The relation between this loss and the technical characteristics
is expressed by the loss function, which is proportional to the square of the
deviations of the performance characteristics from its target value.
Taguchi breaks down his quality engineering strategies into three phases: System
design, Parameter design and Tolerance design. System design deals with inno-
vative research, looking for what factors and levels should be. Parameter design
is what is commonly known as Taguchi Methods and is covered in this paper.
This technique is intended to improve the performance of processes/products
by adjusting levels of factors. Finally, Tolerance Design aims to determine the
control characteristics for each factor level identified in earlier studies.
Change experimentation objectives from “achieving conformance to specifica-
tions” to “reaching the target and minimising variability”.
Since the core of Taguchi’s parameter design is based on experimental methods [19],
he went to great lengths to make DoE more user-friendly (easy to apply). Basically,
Taguchi simplified the use of DoE by incorporating the following: a standard set
of experimental design matrices (Orthogonal arrays), a graphical aid to assign the
factors to the experimental matrix (linear graphs), clear guidelines for the interpreta-
tion of results (cookbook), special data transformation to achieve reduced variation
(S/N Ratios) and a formal study of uncontrollable factors using the robust design
technique, among others [20]. Finally, he simplified Tolerance Analysis through the
use of DoE [21].
52 Comparing Different Approaches for Design of Experiments (DoE) 615

Taguchi’s main contribution to experimental design was a strong emphasis on

variation reduction. Quality is something that cannot be characterised solely by
means of a defined quality characteristic such as yield. The variability of those char-
acteristics must also be considered [22]. Therefore, he proposed a novel design,
where factors (included in experimentation) are classified into two main groups:
Control factors and Noise Factors. The first one includes parameters that can be
easily controlled or manipulated, whereas noise factors are difficult or expensive
to control. Therefore, the basic idea in parameter design is to identify, through
exploiting interactions between control parameters and noise variables, the appro-
priate setting of control parameters at which the system’s performance is capable of
withstanding uncontrollable variation among noise factors [23]. Since the goal is to
make the system resistant to variation of noise variables, the approach has also been
called “Robust design”.
A recent bibliography on Taguchi’s approach to DoE may be found in Roy [20]
and Taguchi et al.’s [24] Quality Engineering Handbook.

52.2.3 The Shainin Approach

The Shainin SystemTM is the name given to a problem solving system developed by
Dorian Shainin. In 1975, he established his own consulting practice: Shainin LLC.
His sons Peter and Richard later joined the family business. Shainin described his
colourful method as the American approach to problem solving, with the same goals
of the Taguchi approach [25].
Shainin viewed his ideas as private intellectual property, which he was known to
sell to clients to help them gain a competitive advantage [26]. As Shainin SystemsTM
are legally protected trademarks and some of its methods are rarely discussed in
literature, it is difficult to obtain a complete overview of the approach [27].
Keki R. Bhote was authorised to publish information in the first and only book
about these methods. His company, Motorola, won the Malcolm Baldrige National
Quality Award, which stipulates that its methods be shared with other US Com-
panies [28]. Interest in Dorian Shainin’s problem solving techniques rose with the
1991 publication of this book (a second edition was published in 2000 [28]).
Dorian Shainin included several techniques – both known and newly invented –
in a coherent step-by-step strategy for process improvement in manufacturing envi-
ronments [27]. Among those powerful tools, he considered Design of Experiments
as the centrepiece. Moreover, he didn’t believe that DoE was limited to the exclu-
sive province of professionals, but could rather be extended so that the whole factory
could be turned loose on problem-solving [28].
The foundation of Shainin’s DoE strategy rests on:
The Pareto Principle: Among many, even hundreds of candidate factors, a single
one will be the root cause of variation of the response y. That root cause is called
the Red X R
and may be a single variable or the interaction of two more separate
616 M. Tanco et al.

variables [25]. There may be then a second or a third significant cause, called the
Pink X R
and Pale Pink X R
, respectively.
Shainin strongly objected to the use of the Fractional Factorial technique. He
proposed instead to identify and diagnostically reduce most of the sources of
variation down to a manageable number (three or four), at which time he allowed
the use of full factorials [29].
“Talk to the parts, they are smarter than engineers”. First, talk to the parts. Then,
talk to the workers on the firing line. Last, the least productive methods are to
talk to the engineers [28].
The Shainin System presents many tools in a sequence of progressive problem solv-
ing. It can be divided into three main groups: Clue Generation tools, Formal DoE
and Transition to SPC. The Shainin DoE technique considers as many variables as
can be identified [30]. The first groups try to generate clues (like Sherlock Homes)
with a number of tools (Multi Vary, Components SearchTM, Paired ComparisonTM,
Process SearchTM and Concentration ChartTM) to reduce the number of variables
involved in the problem through on-line experimentation. In the second stage, the
Variable SearchTM technique is used to reduce the number of variables by sequen-
tially (not randomly) experimenting off-line, based on engineering judgement with
binary search. Once a few factors are obtained, full factorials are used to analyse
their effects and interactions. Afterwards, other techniques (B vs. CTM , Response
Surface, and ScaterPlots) are used to confirm the results and optimise them when
necessary. Finally, in the last step, PositrolTM , Process Certification and Pre-Control
are recommended to guarantee that results will be obtained in the future.

52.3 Limitations of the Different Approaches

52.3.1 The Classical Approach

Up to the time when “Taguchi Methods” were propagated in the U.S., Design of
Experiments (classical) and associated techniques were treated as mathematical
tools, more like an adjunct to an engineer’s technical resources for the study of
product and process characteristics [17].
Taguchi and Shainin were the biggest critics of the Classical Approach. They
believed that managers, engineers and workers found the use of DoE to be a
complicated, ineffective and frustrating experience [28].
However, it is worth mentioning that after a decade of strong opposition to
their new ideas, the “classical” proponents began to acknowledge the importance
of some of the new ideas proposed by Taguchi and Shainin. For example, there
have been many attempts to integrate Taguchi’s parameter design principle with
well-established statistical techniques [18]. As a consequence, the exploitation of
response surface methodology for variance reduction has become a major area of
research [31, 32]. Moreover, the simplicity demanded by critics was reached (or at
52 Comparing Different Approaches for Design of Experiments (DoE) 617

least greatly improved) by the aid of software capable of designing and analysing
experimental design. Furthermore, great emphasis was placed on presenting guide-
lines and including graphical tools to clarify every step in the planning stage. For
example, the Pareto Charts of effects or the use of Multi-Vary charts to correctly
define the problem were included.

52.3.2 Taguchi

There was great debate about these methods during the first decade after their
appearance in the U.S. [21–23, 33, 34]. Taguchi’s approach was criticised for being
inefficient and often ineffective. Moreover, Shainin was highly critical of Taguchi,
challenging the myth of the “secret super Japanese weapon” [25].
Nair [23] identified three general criticisms of Taguchi’s work: (a) The use of
the SNR as a measure of the basis of analysis, (b) his analysis methods and (c) his
choice of experimental designs.
Many have criticized the use of SNR as a performance measurement. Better
approaches to the parameter design problem have been addressed in recent years,
such as Box’s performance measurement [35] or the Response Surface approach
[31].
There has also been much criticism of the new analysis techniques proposed
by Taguchi. Most of them are more difficult and less effective than previous ones.
An example is the use of accumulation analysis [21, 36].
Finally, and most importantly, is the criticism of experimental designs. Firstly,
orthogonal arrays were criticised for underestimating interactions. In response to
this criticism, Taguchi stated [29], “A man who does not believe in the existence of
non-linear effects is a man removed from reality”. He believes, however, in the abil-
ity of engineers to decide on the levels of factors (called sliding levels), in order to
make some interactions for that particular experiment insignificant. Unfortunately,
there is evidence that interaction should be avoided. This is accepted as a matter of
faith among Taguchi’s followers [17].
On the other hand, the designs proposed by Taguchi to simultaneously study
both the mean and the variance (crossed arrays) were also criticized, since they
require multiple runs and generally don’t allow to study control factor interactions.
Therefore, Welch [37], among others, proposed using a combined array to reduce
the number of runs. There has been much debate in recent years on this topic and it
is not yet clear what the best approach is. Pozueta et al. [38] and Kunert et al. [39]
have demonstrated, for example, how classical designs are sometimes worse than
Taguchi’s designs.
It is worth mentioning that whereas some classical statisticians and academics
have acknowledged the value of certain engineering ideas of Taguchi, the Taguchi
camp has shown few signs of compromise, steadfastly vouching for the unchallenge-
able effectiveness of the Taguchi package of procedures in its original form [17].
618 M. Tanco et al.

For more details on the technical debate of Taguchi’s approach, refer to Nair [23],
Pignatello [21], Box [33] and Robinson [22], among others.

52.3.3 Shainin

As the Shainin SystemTM is legally protected, one of the only ways to learn about his
method is to go to his classes. The only other alternative is to read Bhote’s book [28].
Unfortunately, this book is full of hyperbolic, over optimistic, extremely biased and
sometimes even intellectually dishonest claims in its urge to prove that Shainin’s
techniques are superior to all other methods. For example, Bhote claims, [28] “We
are so convinced of the power of the Shainin DoE that we throw out a challenge. Any
problem that can be solved by classical or Taguchi Methods, we can solve better,
faster, and less expensively with the Shainin methods”. Moreover, he presents DoE
as the panacea to all kinds of problems and criticises other alternatives such as Six
Sigma and TQM without academic basis behind his claims.
Although there is little written in industry literature about Shainin, the existing
material is enough to fuel criticism of his methods. The most criticised tech-
niques are the Variable SearchTM [29, 30, 40, 41] and Pre-Control [26, 41]. Variable
SearchTM has much in common with the costly and unreliable “one-factor-at-a-
time” method [29]. Shainin’s technique relies heavily on using engineering judge-
ment. Its weakness lies in the skill and knowledge required to carry out two tasks:
firstly, to correctly identify the variables and secondly, to allocate those variables to
the experiment [30].
On the other hand, although Pre-Control is presented as an alternative to sta-
tistical process control (SPC), it is not an adequate substitution. In particular, it is
not well-suited for poorly performing processes where its use will likely lead to
unnecessary tampering [26].
A recent review and discussion by Steiner et al. [41, 42] provides an in-depth
look at the Shainin SystemTM and its criticisms.

52.4 Conclusions and Recommendations

Three different approaches to DoE have been presented throughout the last two
sections. These approaches, compared with OFAT strategies, are more successful.
However, this does not prove that each technique is necessarily the best. Therefore,
we will give some recommendations and conclusions on each approach based on
our experience and research.
Firstly, we must commended Shainin for stressing the importance of statisti-
cally designed experiments [29]. His methods are easy to learn and can be applied
to ongoing processing during full production. Despite their obvious simplicity
(which stems from the fact that they are, most of the time, simple versions of
one-factor-at-a-time methods), they do not seem to offer a serious alternative to
52 Comparing Different Approaches for Design of Experiments (DoE) 619

other well-established statistical methodologies. Classical and Taguchi approaches

are still much more powerful, more statistically valid and more robust [29]. Shainin
methods could perhaps have some applicability in medium-to-high volume pro-
cesses for which a high level of quality has already been achieved [29]. They can
also be used when dealing with binary response (for example, a machine works or
not) and the reason behind the response may be due to a huge amount of variables.
On the other hand, the most important contribution of the Taguchi approach is in
the area of quality philosophy and engineering methodology, which includes the loss
function and robust design [18]. Taguchi Methods have the potential to bring about
first-cut improvements in industrial applications. However, owing to their theoretical
imperfections, success cannot be assured in every instance [17]. As a general rule,
we don’t recommend using Taguchi Methods unless you have to deal with two types
of problems: Tolerance analysis and robustness to noise factors in products and
processes. However, in those cases, the classical approach may also be suitable.
The two new approaches were a reaction to the existing complexity of the tech-
niques used in industry. Both approaches are presented as an easy and effective
technique for experimentation. However, since the appearance of both initiatives,
the Classical approach has further developed the technique, including the engineer-
ing ideas of the new approaches. The introduction of software and graphical aids
has made this approach much more user friendly than in the 1980s. Therefore, this
approach is the most established and has the most statistically valid methods.
However, although the classical approach has outgrown its “competitors”, there
is still a need to shape an easy experimental design methodology to develop the
information necessary for good planning, design and analysis [18,43,44]. Therefore,
unless classical DoE is developed into easily understood strategies for solving spe-
cific problems and made widely available through teaching materials and software
packages, we cannot expect it to become widely used on the factory floor.
Another strategy is to combine, with statistical rigor, the best DoE approaches
in an easy, simple and integrated methodology from an engineering point of view.
Instead of following one approach or another, we should use the most powerful tools
available to gain the necessary knowledge of a process [34]. For example, the stress
that Shainin placed on carrying out several on-line experiments in a diagnostic stage
to analyse and characterise the problem before experimentation must be considered.
Moreover, the awareness of the importance of variation in industrial experimenta-
tion carried out by Taguchi must force Quality Loss function and Crossed arrays
to be integrated in this methodology. Finally, the importance of confirmatory runs,
proposed by both of them, should also be stressed.

References

1. Lye, L.M., Tools and toys for teaching design of experiments methodology. In 33rd Annual
General Conference of the Canadian Society for Civil Engineering. 2005 Toronto, Ontario,
Canada.
2. Montgomery, D.C., Design and Analysis of Experiments. 2005, New York: Wiley.
620 M. Tanco et al.

3. Gunter, B.H., Improved Statistical Training for Engineers – Prerequisite to quality. Quality
Progress, 1985. 18(11): pp. 37–40.
4. Montgomery, D., Applications of Design of Experiments in Engineering. Quality and
Reliability Engineering International, 2008. 24(5): pp. 501–502.
5. Ilzarbe, L. et al., Practical Applications of Design of Experiments in the Field of Engineering.
A Bibliographical Review. Quality and Reliability Engineering International, 2008. 24(4):
pp. 417–428.
6. De Mast, J., A Methodological Comparison of Three Strategies for Quality Improvement.
International Journal of Quality and Reliability Management, 2004. 21(2): pp. 198–212.
7. Ryan, T.P., Modern Experimental Design. 2007, Chichester: Wiley.
8. Fisher, R.A., The Design of Experiments. 1935, New York: Wiley.
9. Box, G.E.P. and K.B. Wilson, On the Experimental Attainment of Optimum Conditions.
Journal of the Royal Statistical Society, 1951. Series B(13): pp. 1–45.
10. Montgomery, D.C., Changing Roles for the Industrial Statisticians. Quality and Reliability
Engineering International, 2002. 18(5): pp. 3.
11. Booker, B.W. and D.M. Lyth, Quality Engineering from 1988 Through 2005: Lessons from
the Past and Trends for the Future. Quality Engineering, 2006. 18(1): pp. 1–4.
12. Funkenbusch, P.D., Practical Guide to Designed Experiments. A Unified Modular Approach.
2005, New York: Marcel Dekker.
13. Robinson, G.K., Practical Strategies for Experimentation. 2000, Chichester: Wiley.
14. Box, G.E.P., J.S. Hunter, and W.G. Hunter, Statistics for Experimenters – Design, Innovation
and Discovery. Second Edition. Wiley Series in Probability and Statistics, ed. 2005, New York:
Wiley.
15. Taguchi, G., Introduction to Quality Engineering. 1986, White Plains, NY: UNIPUB/Kraus
International.
16. Taguchi, G., System of Experimental Design: Engineering Methods to Optimize Quality and
Minimize Cost. 1987, White Plains, NY: UNIPUB/Kraus International.
17. Goh, T.N., Taguchi Methods: Some Technical, Cultural and Pedagogical Perspectives. Quality
and Reliability Engineering International, 1993. 9(3): pp. 185–202.
18. Tay, K.-M. and C. Butler, Methodologies for Experimental Design: A Survey, Comparison
and Future Predictions. Quality Engineering, 1999. 11(3): pp. 343–356.
19. Arvidsson, M. and I. Gremyr, Principles of Robust Design Methodology. Quality and
Reliability Engineering International, 2008. 24(1): pp. 23–35.
20. Roy, R.K., Design of Experiments Using the Taguchi Approach: 16 steps to Product and
Process Improvement. 2001, New York: Wiley.
21. Pignatello, J. and J. Ramberg, Top Ten Triumphs and Tragedies of Genechi Taguchi. Quality
Engineering, 1991. 4(2): pp. 211–225.
22. Robinson, T.J., C.M. Borror, and R.H. Myers, Robust Parameter Design: A Review. Quality
and Reliability Engineering International, 2004. 20(1): pp. 81–101.
23. Nair, V.N., Taguchi’s Parameter Design: A Panel Discussion. Technometrics, 1992. 31(2):
pp. 127–161.
24. Taguchi, G., S. Chowdhury, and Y. Wu, Taguchi’s Quality Engineering Handbook. First
edition. 2004, New York: Wiley Interscience.
25. Shainin, D. and P. Shainin, Better than Taguchi Orthogonal Tables. Quality and Reliability
Engineering International, 1988. 4(2): pp. 143–149.
26. Ledolter, J. and A. Swersey, An Evaluation of Pre-Control. Journal of Quality Technology,
1997. 29(2): pp. 163–171.
27. De Mast, J. et al., Steps and Strategies in Process Improvement. Quality and Reliability
Engineering International, 2000. 16(4): pp. 301–311.
28. Bhote, K.R. and A.K. Bhote, Word Class Quality. Using Design of Experiments to Make it
Happen. Second edition. 2000, New York: Amacom.
29. Logothetis, N., A perspective on Shainin’s Approach to Experimental Design for Quality
Improvement. Quality and Reliability Engineering International, 1990. 6(3): pp. 195–202.
52 Comparing Different Approaches for Design of Experiments (DoE) 621

30. Thomas, A.J. and J. Antony, A Comparative Analysis of the Taguchi and Shainin DoE Tech-
niques in an Aerospace Enviroment. International Journal of Productivity and Performance
Management, 2005. 54(8): pp. 658–678.
31. Vining, G.G. and R.H. Myers, Combining Taguchi and Response Surface Philosophies:
A Dual Response Approach. Journal of Quality Technology, 1990. 22(1): pp. 38–45.
32. Quesada, G.M. and E. Del Castillo, A Dual Response Approach to the Multivariate Robust
Parameter Design Problem. Technometrics, 2004. 46(2): pp. 176–187.
33. Box, G.E.P., S. Bisgaard, and C. Fung, An Explanation and Critique of Taguchi’s Contribution
to Quality Engineering. International Journal of Quality and Reliability Management, 1988.
4(2): pp. 123–131.
34. Schmidt, S.R. and R.G. Lausnby, Understanding Industrial Designed Experiments. Fourth
Edition. 2005, Colorado Springs, CO: Air Academy Press.
35. Box, G.E.P., Signal to Noise Ratios, Performance Criteria, and Transformations. Technomet-
rics, 1988. 30(1): pp. 1–17.
36. Box, G.E.P. and S. Jones, An Investigation of the Method of Accumulation Analysis. Total
Quality Management & Business Excellence, 1990. 1(1): pp. 101–113.
37. Welch, W.J. et al., Computer Experiments for Quality Control by Parameter Design. Journal
of Quality Technology, 1990. 22(1): pp. 15–22.
38. Pozueta, L., X. Tort-Martorell, and L. Marco, Identifying Dispersion Effects in Robust Design
Experiments - Issues and Improvements. Journal of Applied Statistics, 2007. 34(6): pp. 683–
701.
39. Kunert, J. et al., An Experiment to Compare Taguchi’s Product Array and the Combined Array.
Journal of Quality Technology, 2007. 39(1): pp. 17–34.
40. Ledolter, J. and A. Swersey, Dorian Shainin’s Variables Search Procedure: A Critical
Assessment. Journal of Quality Technology, 1997. 29(3): pp. 237–247.
41. De Mast, J. et al., Discussion: An Overview of the Shainin SystemTM for Quality Improve-
ment. Quality Engineering, 2008. 20(1): pp. 20–45.
42. Steiner, S.H., J. MacKay, and J. Ramberg, An Overview of the Shainin SystemTM for Quality
Improvement. Quality Engineering, 2008. 20(1): pp. 6–19.
43. Tanco, M. et al., Is Design of Experiments Really Used? A Survey of Basque Industries.
Journal of Engineering Design, 2008. 19(5): pp. 447–460.
44. Viles, E. et al., Planning Experiments, the First Real Task in Reaching a Goal. Quality
Engineering, 2009. 21(1): pp. 44–51.
Chapter 53
Prevention of Workpice Form Deviations in
CNC Turning Based on Tolerance Specifications

Gonzalo Valiño Riestra, David Blanco Fernandez, Braulio Jose Álvarez,

Sabino Mateos Diaz, and Natalia Beltrán Delgado

Abstract The simultaneous action of cutting forces and gradients of tempera-

ture due to different effects in turning, end result in deflections of the machine-
workpiece-tool system and in a structural thermal drift. In turn, these factors lead to
a deviation of the tool-point away from the theoretical path and consequently, to final
dimensional and form errors on the workpiece manufactured. Traditionally, these
non-desired deviations have been compensated by on-line corrections of the tool
position, even although they were not strictly necessary according to the tolerance
specifications on the workpiece drawings. These type of compensations require the
use of machine-tools equipped with multiple sensors, already available in modern
and precise machine-tools but not always in common ones, which are more reason-
able from an economic point of view. On the other hand, on-line compensations
modify the predicted working conditions and in turn, chatter or other undesirable
effects can appear such as moving away from optimal cutting conditions. Taking
this into account, a new approach is proposed in this work for prediction and com-
pensation of deviations in turning by considering them at the planning stages prior to
machining operations, mainly focused on development of CAPP systems for CNC
turning. Tolerance specifications are considered as limiting values for the admissi-
ble workpiece deviations which are predicted by using a developed model based on
the strain energy theory.

Keywords CNC turning  form deviation  machining error  tolerance  CAPP

G.V. Riestra (B)

Department of Manufacturing Engineering, University of Oviedo, Campus of Gijón, Spain
E-mail: gvr@uniovi.es

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 623
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
624 G.V. Riestra et al.

53.1 Introduction

The action of cutting forces in turning results in a deflection of the machine-

workpiece-tool system. Moreover, thermal effects such as the increment of tempera-
ture at the cutting point, the temperature related to mechanical friction in movements
and the environmental temperature also have influence on whole system deflection.
In addition, dynamics of the cutting process also affects the accuracy of the manu-
factured workpiece. As a result, the actual workpiece obtained under certain cutting
conditions is different from that expected initially. In this way, both dimensional
and form deviations take place since the tool-point does not accurately follow the
programmed path. A lack of flatness, cylindricity, coaxiality, perpendicularity, par-
allelism among other geometrical errors may occur as result, as well as distance and
diameter deviations.
Most efforts for dealing with this scenery have focused on the prediction and
compensation of workpiece dimensional deviations by means of different tech-
niques. In this way, some authors [1] have implemented functions to modify the
programmed depth of cut in the machine-tool numerical control. Others [2, 3]
have made predictions of workpiece deflection based on FEM analysis. Only in
a few cases [4–7] the influence of the whole machine-part-tool system stiffness
has been considered. Taking apart the effect of cutting forces, some authors [8–10]
have proposed models for compensation of deviations related to the machine-tool
thermal drift.
Although numerous contributions have been made for error prediction and com-
pensation, none of them analyzes whether error compensation enables to meet
workpiece tolerances or even if it results really necessary to do any compensation
when tight tolerances are not required.
Taking this into account, a new approach to deviations in CNC turning will
be proposed in the next sections. The most common errors in turning are initially
described and a model for the calculation of deviations is developed based on the
strain energy theory. Workpieces with both external and internal features and vari-
able section along the axis of rotation are supported by the model as well as the
main clamping methods in turning (on a chuck, between-center and chuck-tailstock)
are considered. Quantitative relationships between the different types of machining
errors and the different types of tolerances are determined and proposed as con-
straints for setting up the optimal cutting conditions at the planning stages prior to
machining.

53.2 Form Errors in Turning

Once finished machining operations, the deflection due to cutting forces disappears
and the elastic recovery of material causes the workpiece axis to return to its original
position, but this entails a defect of form in the workpiece surfaces equivalent to the
deflection that the axis had undergone. Actually, turning of cylindrical or conical
53 Prevention of Workpice Form Deviations in CNC Turning 625

Fig. 53.1 Form errors in turning and facing due to deflections

surfaces leads to a third-order polynomial surface profile [4] (Fig. 53.1a). Similarly,
facing of front surfaces leads to conical surfaces with a lack of perpendicularity with
regard to the symmetry axis (Fig. 53.1b). The conicity of real front faces is directly
related to the turn .z/ of the section and to the total elastic deflection ıT .z/. For
the section in which the cutting tool is applied .z D a/ this relation will be:

d
tan .a/ D ıT .a/ (53.1)
dz

53.3 Calculation of Radial Deviations

The total deviation in the radial direction ıT .z/ of a turned component at a distance z
from the chuck is composed by simultaneous factors related to deformations caused
by cutting forces, such as the spindle-chuck system ısc .z/, the toolpost ıtp .z/,
the workpiece deflection ıp .z/ and the thermal effects [5, 11] ıt h .z/. This can be
expressed as:

ıT .z/ D ısc .z/ C ıtp .z/ C ıp .z/ C ıt h .z/ (53.2)

Deviations can be obtained for the tangential (Z) and radial (X ) directions
(Fig. 53.2). The contribution of the spindle-chuck system can be expressed for these
directions [5] as:
   
Fr Fr Fa D D  2 Fr
ısc r .z/ D C z2 C  z
k2r k3r 2 k3r k2r
 
Fr Fa D Lc Fr Lc 2
C  C (53.3)
k1r 2  k2r k2r
626 G.V. Riestra et al.

Fig. 53.2 Cutting force and deflection components

   
Ft Ft Fa D D  2 Ft
ısct .z/ D C z2 C  z
k2t k3t 2 k3t k2t
 
Ft Fa D Lc Ft Lc 2
C  C (53.4)
k1t 2 k2t k2t

where sub-indices r and t refer to radial and tangential directions respectively;

Fr , Ft and Fa are the cutting force components in the radial, tangential and feed
directions (Fig. 53.2); D is the machining diameter; k1 , k2 and k3 are the stiffness
constants for the spring-model of the system [5, 6].
The contribution of the toolpost is mainly due to a shortening of the tool and the
tool-holder in the radial direction. Therefore, it can be expressed as:
 
lt lh
ıtp .z/ D Fr  C (53.5)
At  Et Ah  Eh

where Fr is the radial component of the cutting force, lt , At and Et the cantilever
length, cross section and elastic modulus for the tool respectively, and lh , Ah and
Eh the length, cross section and elastic modulus for the tool-holder (Fig. 53.2).
Based on strain-energy, the contribution of the workpiece deflection for the chuck
clamping method when a cutting force is at z D a (distance between force and
chuck) will be:

X Z Zi C1
4 F .a  z/2
ıp .a/ D   dz
i
E Zi ri ext  ri4int
4

Z Zi C1 !
F 1
C   dz (53.6)
G Zi ri2ext  ri2int
53 Prevention of Workpice Form Deviations in CNC Turning 627

where i represents each zone with constant cross section, E is the elastic modulus,
G the shear modulus, I the moment of inertia and is the shear factor. For the
between-centre clamping method, the expression is:
Z Zi C1 !
4 F .L  a/2 X z2
ıp .a/ D dz
L2 E
i Zi ri4ext  ri4int
Z Zi C1 ! (53.7)
F .L  a/2 X 1
C dz
L2 G
i Zi ri2ext  ri2int

And in the case of the chuck-tailstock clamping, the expression becomes:

h i2
Z a  F .L  a/2 3L2  .L  a/2
ıp .a/ D dz
0 4L6 A G
Z   2
a F .L  a/2 2L2  a .a  2L/ z C L  a  .a  2L/
C dz
0 4L6 E I
(53.8)

Finally, the contribution of the thermal drift takes place mainly in the radial direction
and depends on factors such as spindle speed, feed rate, machine-tool operation
time and ambient temperature. Some authors use these parameters as input data to a
neural network, which output is the radial thermal drift. [10, 11]

53.4 Relationship Between Deviations and Tolerances

Although dimensional and geometrical specifications of a part are expressed on

drawings by means of dimensional and geometrical tolerances, no relationships
were found in literature between form deviations derived from machining process
and value of tolerances. Nevertheless, it is essential to know these relationships
for setting up the appropriate machining conditions. Therefore, in the next sections
each type of tolerance will be analyzed according to ISO standard and geometrically
related to the feasible deviation types in turned parts.

53.4.1 Length Tolerance

Due to the form error, the real distance between two front faces depends on the
zone in which the measurement is taken. Actual measurements will be between the
values L0min and L0max (Fig. 53.3a) and they will be admissible when the following
expression is satisfied:

L0max  Ln C dLu and L0min Ln C dLl (53.9)

628 G.V. Riestra et al.

Fig. 53.3 Accepting conditions for deflections according to tolerance values

where Ln is the nominal distance and dLu and dLl are the upper and lower limits
of longitudinal tolerance, respectively. Considering distances ei and ej in Fig. 53.3a,
the following relation can be established:

L0max  L0min D ei C ej (53.10)

Taking into account Eqs. (53.9)–(53.10) and that the tolerance zone t is the dif-
ference between the upper and lower limits, the following is the condition for the
measurement to meet the tolerance:

ei C ej  dLu  dLl D t (53.11)

53 Prevention of Workpice Form Deviations in CNC Turning 629

where distances ei and ej can be expressed in terms of the turn of sections i

and j :

DM i  Dmi DMj  Dmj

ei D  tan i and ej D  tan j (53.12)
2 2

53.4.2 Diametrical Tolerance

Due to the form error of the cylindrical surfaces, the diameter of the workpiece
depends on the location of the measurement point, varying between a maximum
0 0
d and Dmd ) (Fig. 53.3b). The measurement is admissible
and a minimum value (DM
when the following expression is satisfied:
0 0
DM d  Dn C dDu and Dmd Dn C dDl (53.13)

where Dn is the nominal diameter and dDu and dDl are the upper and lower limits
of the diametrical tolerance, respectively.
Once machined the cylindrical surface, the maximum error is given by the differ-
ence between the maximum ıTMd and the minimum ıTmd deflections. The following
relation can be deduced from geometry (Fig. 53.3b):
0 0
DM d  Dmd D 2 jıTMd  ıT md j (53.14)

By considering Eqs. (53.13)–(53.14) and that the tolerance zone t is the difference
between the upper and lower deviations, the following is the condition to meet the
tolerance:
2 jıTMd  ıT md j  dDu  dDl D t (53.15)

53.4.3 Geometrical Tolerance of Flatness

Definition of each type of geometrical tolerance has been considered based on ISO
standard (ISO 1101:1983). According to that, the flatness tolerance establishes a
zone of acceptance t with respect to the nominal orientation of the controlled face,
within which the real machined surface must stand. In order to satisfy this condition,
the form error must be lower than the width of the tolerance zone (Fig. 53.3c).
Considering this and being distance ei calculated as in Eq. (53.12), the condition
will be:
ei  t (53.16)
630 G.V. Riestra et al.

53.4.4 Geometrical Tolerance of Cylindricity

The cylindricity tolerance establishes a volume between two coaxial cylinders

whose difference of radii is the value of the tolerance zone t. The machined surface
must lie between these two cylinders.
The maximum error obtained when machining the cylindrical surface is given by
the difference between the maximum ıTMd and minimum ıTmd deviations. Being
0 0
DM d
and Dmd the maximum and minimum surface diameters, the following
relations can be deduced from geometry (Fig. 53.3d):
0 0 0 0
DM d  Dmd  2  t and DM d  Dmd D 2  jıTMd  ıTmd j (53.17)

Therefore, it is obtained the condition:

jıTMd  ıT md j  t (53.18)

53.4.5 Geometrical Tolerance of Profile of a Surface

This tolerance specifies a zone between two surrounding spherical surfaces whose
diameter difference is the value of the tolerance zone t and whose centres are located
onto the theoretical surface. According to norm ISO 3040:1990, for the case of a
conical surface, the tolerance zone becomes the space between two cones of the
same angle than the datum and equidistant to that cone the half of the tolerance
value.
All diameters of the real machined surface must lie within the tolerance zone,
0 0
including the diameters in which deviation is maximum DM d
and minimum Dmd ,
which will satisfy:
ˇ 0 ˇ ˇ 0 ˇ
ˇD ˇ ˇD  Dnmd ˇ
M d  DnM d t md t
 and  (53.19)
2 2  cos ˛ 2 2  cos ˛
where DnM d and Dnmd are the nominal diameters of the cone measured in the
positions of maximum and minimum diametrical deviation (Fig. 53.3e).
On the other hand, workpiece deflections in the zones of maximum and minimum
deviations (ıTMd and ıT md ) can be expressed as:
ˇ 0 ˇ ˇ 0 ˇ
ˇD  DnM d ˇ ˇD  Dnmd ˇ
ıTMd D Md
and ıT md D md
(53.20)
2 2
The final condition is derived from Eqs. (53.19–53.20):
t
ıTMd C ıT md  (53.21)
cos ˛
53 Prevention of Workpice Form Deviations in CNC Turning 631

53.4.6 Geometrical Tolerance of Parallelism

Parallelism is applied between front faces in turned parts. Let be i the datum and
j the related face. According to the definition, the related surface must lie within
an space between two planes parallel to the datum and distanced one another the
value of the tolerance t (Fig. 53.3f). Although even the datum has a form error,
according to norm ISO 5459:1981, this surface will be considered perpendicular
to the part axis and, consequently, also the planes which define the tolerance zone.
Considering this, and calculating ej as in Eq. (53.12), the geometrical condition to
meet the tolerance will be:
ej  t (53.22)

53.4.7 Geometrical Tolerance of Perpendicularity

Perpendicularity is applied between a front face and a feature axis in turned parts.
Two different situations must be considered depending on which of both elements
is the datum and which the controlled one.
When the axis is the datum and a face is controlled, the tolerance zone is the
space between two planes perpendicular to the datum axis and distanced one another
the value of the tolerance t (Fig. 53.3g). Considering this, and calculating ej as in
Eq. (53.12), the condition will be same than in Eq. (53.22).
On the other hand, when a face is the datum and the axis is controlled, the toler-
ance zone is the space within a cylinder of diameter t, and axis perpendicular to the
datum face. The elastic recovery of the part axis after machining implies that any
deviation of this element does not depend directly on the cutting action but on other
technological aspects such as a bad alignment of the workpiece in the lathe. There-
fore, no relation can be established between this error and the deviation caused by
cutting forces.

53.4.8 Geometrical Tolerance of Coaxiality

This tolerance limits the relative deviation between two zones of the part axis. As in
the previous case, the elastic recovery of the workpiece axis after machining implies
that the possible error is not due to deviations caused by cutting forces but to other
causes.

53.4.9 Geometrical Tolerance of Circular Runout

For circular-radial runout, the tolerance zone is the area between two concentric
circles located into a plane perpendicular to the axis, whose radii difference is the
tolerance t and whose centre is located at datum axis. Since form errors derived
632 G.V. Riestra et al.

from cutting forces are completely symmetrical with respect to the workpiece rota-
tion axis, the error evaluated throughout this tolerance does not depend on these
forces, but on the workpiece deflection caused by the clamping force or by a lack of
workpiece alignment.
A similar situation takes place regarding circular-axial runout.

53.4.10 Geometrical Tolerance of Total Runout

The total-radial runout is used to control cumulative variations of circularity and

cylindricity of surfaces constructed around a datum axis. As in the case of circular-
radial runout, circularity does not depend on deviations from cutting forces but
cylindricity does. Therefore, the condition to be satisfied in this case is the same
than in Eq. (53.18) (Fig. 53.3h).
The total-axial runout controls cumulative variations of perpendicularity and
flatness of front faces at a right angle to the datum axis. Therefore, the relation
of this error with deviations coincides with expressions obtained for flatness and
perpendicularity respectively (Fig. 53.3i).

53.5 Deviations as Optimization Conditions

The optimization of cutting conditions in turning is the final stage of process plan-
ning in which not only mathematical considerations about the objective function
have to be done (e.g., time, cost or benefit) but also there are several constraints that
restrict the best solution.
Common constraints are related to ranges of cutting parameters (cutting speed,
feed-rate and depth of cut), ranges of tool-life and other operating limits such as sur-
face finish, maximum power consumption and maximum force allowed. The cutting
force constraint is imposed to limit the deflection of the workpiece or cutting tool,
which result in dimensional error, and to prevent chatter. [12] Traditionally, the value
of this constraint has not been clearly determined. Nevertheless, the relationships
between workpiece deviations and tolerances described in the previous sections can
also be considered as relationships between cutting forces and maximum deviations
allowed and, therefore, all together can be used as optimization constraints.

53.6 Conclusions

This work provides a mathematical model based on strain energies for prediction
of deviations in the turning process of workpieces with complex external and/or
internal geometry and which takes into account the main clamping procedures in
turning.
53 Prevention of Workpice Form Deviations in CNC Turning 633

Likewise, an analysis of maximum deviations is developed according to each

tolerance specification and an accepting criterion is proposed in each case so that
compensation of the calculated errors must be carried out only when tolerances are
not met. This becomes a different approach with regard to other works developed
up to date which propose to make error compensations in all cases, even when they
are not necessary according to tolerances. Moreover, the results of this analysis are
utilized by a turning CAPP system as constraints for the optimization of cutting
conditions and it also can be useful for deciding the most suitable clamping method
for setting up the workpiece on a lathe.

Acknowledgements This work is part of a research project supported by the Spanish Education
and Science Ministry (DPI2007-60430) and FEDER.

References

1. S. Yang, J. Yuan, and J. Ni, Real-time cutting force induced error compensation on a turning
center, Int. J. Mach. Tools Manuf. 37(11), 1597–1610 (1997).
2. A.-V. Phan, L. Baron, J.R. Mayer, and G. Cloutier, Finite element and experimental studies of
diametrical errors in cantilever bar turning, J. App. Math. Model. 27(3), 221–232 (2003).
3. L.Z. Qiang, Finite difference calculations of the deformations of multi-diameter workpieces
during turning, J. Mat. Proc. Tech. 98(3), 310–316 (2000).
4. L. Carrino, G. Giorleo, W. Polini, and U. Prisco, Dimensional errors in longitudinal turning
based on the unified generalized mechanics of cutting approach. Part II: Machining pro-
cess analysis and dimensional error estimate, Int. J. Mach. Tools Manuf. 42(14), 1517–1525
(2002).
5. G. Valiño, S. Mateos, B.J. Álvarez, and D. Blanco, Relationship between deflections caused
by cutting forces and tolerances in turned parts, Proceedings.of the 1st Manufacturing
Engineering Society International Conference (MESIC), Spain (2005).
6. D. Mladenov, Assessment and compensation of errors in CNC turning, Ph.D. thesis, UMIST,
UK (2002).
7. S. Hinduja, D. Mladenov, and M. Burdekin, Assessment of force-induced errors in CNC
turning, Annals of the CIRP 52(1), 329–332 (2003).
8. J. Yang, J. Yuan, and J. Ni, Thermal error mode analysis and robust modelling for error
compensation on a CNC turning center, Int. J. Mach. Tools Manuf. 39(9), 1367–1381 (1999).
9. V.A. Ostafiev and A. Djordjevich, Machining precision augmented by sensors, Int. J. Prod.
Res. 37(1), 1999, pp. 91–98.
10. X. Li, P.K. Venuvinod, A. Djorjevich, and Z. Liu, Predicting machining errors in turning using
hybrid learning, Int. J. Adv. Manuf. Tech. 18, 863–872 (2001).
11. X. Li and R. Du, Analysis and compensation of workpiece errors in turning, Int. J. of Prod.
Res. 40(7), 1647–1667 (2002).
12. Y.C. Shin and Y.S. Joo, Optimization of machining conditions with practical constraints,
Int. J. Prod. Res. 30(12), 2907–2919 (1992).
Chapter 54
Protein–Protein Interaction Prediction
Using Homology and Inter-domain Linker
Region Information

Nazar Zaki

Abstract One of the central problems in modern biology is to identify the complete
set of interactions among proteins in a cell. The structural interaction of proteins and
their domains in networks is one of the most basic molecular mechanisms for bio-
logical cells. Structural evidence indicates that, interacting pairs of close homologs
usually interact in the same way. In this chapter, we make use of both homology and
inter-domain linker region knowledge to predict interaction between protein pairs
solely by amino acid sequence information. High quality core set of 150 yeast pro-
teins obtained from the Database of Interacting Proteins (DIP) was considered to
test the accuracy of the proposed method. The strongest prediction of the method
reached over 70% accuracy. These results show great potential for the proposed
method.

Keywords Protein–Protein Interaction  Prediction  Homology  Inter-domain

Linker Region  yeast protein

54.1 Introduction

54.1.1 The Importance of Protein–Protein Interaction

The term protein–protein interaction (PPI) refers to the association of protein

molecules and the study of these associations from the perspective of biochem-
istry, signal transduction and networks. PPIs occur at almost every level of cell
function, in the structure of sub-cellular organelles, the transport machinery across

N. Zaki
Assistant Professor with the College of Information Technology, UAE University.
Al-Ain 17555, UAE,
E-mail: nzaki@uaeu.ac.ae

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 635
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
636 N. Zaki

the various biological membranes, the packaging of chromatin, the network of sub-
membrane filaments, muscle contraction, signal transduction, and regulation of gene
expression, to name a few [1]. Abnormal PPIs have implications in a number of neu-
rological disorders; include Creutzfeld-Jacob and Alzheimer’s diseases. Because of
the importance of PPIs in cell development and disease, the topic has been studied
extensively for many years. A large number of approaches to detect PPIs have been
developed. Each of these approaches has strengths and weaknesses, especially with
regard to the sensitivity and specificity of the method.

54.1.2 Current Methods to Predict PPI

One of the major goals in functional genomics is to determine protein interac-

tion networks for whole organisms, and many of the experimental methods have
been applied to study this problem. Co-immunoprecipitation is considered to be
the gold standard assay for PPIs, especially when it is performed with endogenous
proteins [2]. Yeast two-hybrid screen investigates the interaction between artifi-
cial fusion proteins inside the nucleus of yeast [3]. Tandem Affinity Purification
(TAP) detects interactions within the correct cellular environment [4]. Quantita-
tive immunoprecipitation combined with knock-down (QUICK) detects interactions
among endogenous non-tagged proteins [5]. High-throughput methods have also
contributed tremendously in the creation of databases containing large sets of pro-
tein interactions, such as Database of Interacting Proteins (DIP) [6], MIPS [7] and
Human Protein Reference Database (HPRD) [8].
In addition, several in silico methods have been developed to predict PPI based
on features such as gene context [9]. These include gene fusion [10], gene neigh-
borhood [11] and phylogenetic profile [12]. However, most of the in silico methods
seek to predict functional association, which often implies but is not restricted to
physical binding.
Despite the availability of the mentioned methods of predicting PPI, the accu-
racy and coverage of these techniques have proven to be limited. Computational
approaches remain essential both to assist in the design and validation of the exper-
imental studies and for the prediction of interaction partners and detailed structures
of protein complexes [13].

54.1.3 Computational Approaches to Predict PPI

Some of the earliest techniques predict interacting proteins through the similarity
of expression profiles [14], description of similarity of phylogenetic profiles [12]
or phylogenetic trees [15], and studying the patterns of domain fusion [16]. How-
ever, it has been noted that these methods predict PPI in a general sense, meaning
54 Protein–Protein Interaction Prediction 637

joint involvement in a certain biological process, and not necessarily actual physical
interaction [17].
Most of the recent works focus on employing protein domain knowledge [18–
22]. The motivation for this choice is that molecular interactions are typically medi-
ated by a great variety of interacting domains [23]. It is thus logical to assume that
the patterns of domain occurrence in interacting proteins provide useful information
for training PPI prediction methods [24].
An emerging new approach in the protein interactions field is to take advantage of
structural information to predict physical binding [25, 26]. Although the total num-
ber of complexes of known structure is relatively small, it is possible to expand this
set by considering evolutionary relationships between proteins. It has been shown
that in most cases close homologs (>30% sequence identity) physically interact in
the same way with each other. However, conservation of a particular interaction
depends on the conservation of the interface between interacting partners [27].
In this chapter, we propose to predict PPI using only sequence information. The
proposed method combines homology and structural relationships. Homology rela-
tionships will be incorporated by measuring the similarity between protein pair
using Pairwise Alignment. Structural relationships will be incorporated in terms
of protein domain and inter-domain linker region information. We are encouraged
by the fact that compositions of contacting residues in protein sequence are unique,
and that incorporating evolutionary and predicted structural information improves
the prediction of PPI [28].

54.2 Method

In this work, we present a simple yet effective method to predict PPI solely by amino
acid sequence information. The overview of the proposed method is illustrated in
Fig. 54.1. It consists of three main steps: (a) extract the homology relationship by
measuring regions of similarity that may reflect functional, structural or evolution-
ary relationships between protein sequences (b) downsize the protein sequences by
predicting and eliminating inter-domain linker regions (c) scan and detect domain
matches in all the protein sequences of interest.

54.2.1 Similarity Measures Between Protein Sequences

The proposed method starts by measuring the PPI sequence similarity, which reflects
the evolutionary and homology relationships. Two protein sequences may interact
by the mean of the amino acid similarities they contain [24]. This work is moti-
vated by the observation that an algorithm such as Smith-Waterman (SW) [29],
which measures the similarity score between two sequences by a local gapped
alignment, provides relevant measure of similarity between protein sequences. This
638 N. Zaki

Input Dataset of Protein Sequences in

FASTA Format

Measuring Homology Between Protein

Sequences through Pairwise Alignment
(SW Algorithm)

A Matrix
Summarizes
Eliminating Inter-Domain Linker Regions
the Protein–
Before Using the Pairwise Alignment
Protein
Interaction

Detecting Domain Matches which

Reflect the Associated Structural
Relationships Between Proteins

Fig. 54.1 Overview of the proposed method

similarity incorporates biological knowledge about protein evolutionary structural

relationships [30].
The Smith-Waterman similarity score S W .x1 ; x2 / between two protein seque-
nces x1 and x2 is the score of the best local alignment with gaps between the two
protein sequences computed by the Smith-Waterman dynamic programming algo-
rithm. Let us denote by  a possible local alignment between x1 and x2 , defined by
a number n of aligned residues, and by the indices 1  i1 < ::: < in  jx1 j and
1  j1 < ::: < jQ n  jx2 j of the aligned residues in x1 and x2 respectively. Let
us also denote by .x1 ; x2 / the set of all possible local alignments
Q between x1 and
x2 , and by p.x1 ; x2 ; / the score of the local alignment  2 .x1 ; x2 / between x1
and x2 , the Smith-Waterman score S W .x1 ; x2 / between sequences x1 and x2 can
be written as:
S W .x1 ; x2 / D Qmax p.x1 ; x2 ; / (54.1)
2 .x1 ;x2 /

The similarity matrix M can be calculated as follow:

2 3
S W .x1 ; x1 / S W .x1 ; x2 / ::: S W .x1 ; xm /
6 S W .x2 ; x1 / S W .x2 ; x2 / ::: S W .x2 ; xm / 7
M D6 4
7
5 (54.2)
W W W W
S W .xm ; x1 / S W .xm ; x2 / ::: S W .xm ; xm /

where m is the number of the protein sequences. For example, suppose we have
the following randomly selected PPI dataset: YDR190C, YPL235W, YDR441C,
YML022W, YLL059C, YML011C, YGR281W and YPR021C represented by x1 ,
x2 , x3 , x4 , x5 , x6 , x7 and x8 respectively. The interaction between these eight
proteins is shown in Fig. 54.2.
54 Protein–Protein Interaction Prediction 639

YDR190C YPL235W X1 X2

YDR441C YML022W X3 X4
OR

YLL059C YML011C X5 X6

YGR281W YPR021C X7 X8

Fig. 54.2 The interaction between the randomly selected proteins

Then the SW similarity score matrix M will be calculated as:

x1 x2 x3 x4 x5 x6 x7 x8
x1 X 465 28 30 25 30 34 29
x2 465 X 30 24 32 33 50 47
x3 28 30 X 553 29 27 32 29
x4 30 24 553 X 29 20 25 40
x5 25 32 29 29 X 24 28 49
x6 30 33 27 20 24 X 25 26
x7 34 50 32 25 28 25 X 36
x8 29 47 29 40 49 26 36 X

From M , higher score may reflect interaction between two proteins. S W .x1 ; x2 /
and S W .x2 ; x1 / scores are equal to 465; S W .x3 ; x4 / and S W .x4 ; x3 / scores are
equal to 553, which confirm the interaction possibilities. However, S W .x5 ; x6 / and
S W .x6 ; x5 / scores are equal to 24; S W .x7 ; x8 / and S W .x8 ; x7 / scores are equal
to 36, which are not the highest scores. To correct these errors more biological
information is needed, which lead us to the second part of the proposed method.

54.2.2 Identify and Eliminate Inter-domain Linker Regions

The results could be further enhanced by incorporating inter-domain linker regions

knowledge. The next step of our algorithm is to predict inter-domain linker regions
solely by amino acid sequence information. Our intention here is to identify and
eliminate all the inter-domain linker regions from the protein sequences of interest.
By doing this step, we are actually downsizing the protein sequence to shorter ones
with only domains information to yield better alignment scores. In this case, the
prediction is made by using linker index deduced from a data set of domain/linker
segments from SWISS-PROT database [31]. DomCut developed by Suyama et al.
[32] is employed to predict linker regions among functional domains based on the
difference in amino acid composition between domain and linker regions. Following
640 N. Zaki

Fig. 54.3 An example of linker preference profile generated using Domcut. In this case, linker
regions greater than the threshold value 0.093 are eliminated from the protein sequence

Suyama’s method [32], we defined the linker index Si for amino acid residue i and
it is calculated as follows:
 Linker 
fi
Si D  ln (54.3)
fiDomain

Where fiLinker is the frequency of amino acid residue i in the linker region and
fiDomain is the frequency of amino acid residuei in the domain region. The negative
value of Si means that the amino acid preferably exists in a linker region. A thresh-
old value is needed to separate linker regions as shown in Fig. 54.3. Amino acids
with linker score greater than the set threshold value will be eliminated from the
protein sequence of interest.
When applying the second part of the method, the matrix M will be calculated
as follows:

x1 x2 x3 x4 x5 x6 x7 x8
x1 X 504 30 30 25 32 34 27
x2 504 X 30 21 32 32 50 36
x3 30 30 X 775 29 24 38 29
x4 30 21 775 X 19 21 53 37
x5 25 32 29 19 X 28 28 24
x6 32 32 24 21 28 X 23 27
x7 34 50 38 53 28 23 X 339
x8 27 36 29 37 24 27 339 X
FromM , it’s clearly noted that, more evidence is shown to confirm the interac-
tion possibility between proteins x7 and x8 , and therefore, the result is furthermore
enhanced. In the following part of the method, protein domain knowledge will be
incorporated in M for better accuracy.
54 Protein–Protein Interaction Prediction 641

54.2.3 Detecting Domain Matches and Associated Structural

Relationships in Proteins

In this part of the method, protein domains knowledge will be incorporated in M .

Protein domains are highly informative for predicting PPI as they reflect the poten-
tial structural relationships between them. In this implementation, we employed
ps scan [33] to scan one or several patterns, rules and profiles from PROSITE
against our protein sequences of interest. Running ps scan through the eight proteins
identifies the following domains:
DR441C (x3 ) ! PS00103
YML022W (x4 ) ! PS00103
YGR281W (x7 ) ! PS00211, PS50893 and PS50929
YPR021C (x8 ) ! PS50929
Which reveals structural relationships between proteins x3 and x4 ; and proteins
x7 and x8 . Based on this relationship, S W .x3 ; x4 / and S W .x7 ; x8 / scores will be
incremented by specified value (We have randomly selected a value equal to 300).
Unfortunately, these results have not added more accuracy in this case, however,
it confirmed the interacting possibilities between proteins x3 and x4 ; x7 and x8 .

54.3 Experimental Work

To test the method, we obtained the PPI data from the Database of Interacting
Proteins (DIP). The DIP database catalogs experimentally determined interactions
between proteins. It combines information from a variety of sources to create a
single, consistent set of PPIs in Saccharomyces cerevisiae. Knowledge about PPI
networks is extracted from the most reliable, core subset of the DIP data [34].
The DIP version we used contains 4,749 proteins involved in 15,675 interactions
for which there is domain information [6]. However, only high quality core set
of 2,609 yeast proteins was considered in this experimental work. This core set
is involved in 6,355 interactions, which have been determined by at least one small-
scale experiment or two independent experiments [35]. Furthermore, we selected
proteins interacts with only one protein and not involved in any other interactions.
This process results in a dataset of 150 proteins with 75 positive interactions as
shown in Fig. 54.4. The intention is to design a method capable of predicting pro-
tein interaction partner, which facilitate a way to construct PPI using only protein
sequences information.
We started our experimental work by measuring the protein–protein sequence
interaction similarity using SW algorithm as implemented in FASTA [36]. The
default parameters are used: gap opening penalty and extension penalties of 13
and 1, respectively, and a substitution matrix BLOSUM62 matrix. The choice of
which substitution matrix to use is not trivial because there is no correct scoring
scheme for all circumstances. The BLOSUM matrix is another very common used
642 N. Zaki

Fig. 54.4 Dataset of core interaction proteins used in the experimental work

amino acid substitution matrix that depends on data from actual substitutions. This
procedure produces the matrix M150X150 . This matrix was then enhanced by incor-
porating inter-domain linker regions information. In this case, only well defined
domains with sequence length ranging from 50 to 500 residues were considered.
We skipped all the frequently matching (unspecific) domains. A threshold value of
0.093 is used to separate the linker regions. Any residue generates an index greater
than the threshold value results in eliminating it. This procedure downsized the pro-
tein sequences without losing the biological information. In fact, running the SW
algorithm on a sequence having pure domains, results in better accuracy. A linker
preference profile is generated using the linker index values along an amino acid
sequence using a siding window. A window of size w D15 was used which gives
the best performance (different window sizes were tested).
Further more, protein domains knowledge will be incorporated in M150X150 .
In this implementation, ps scan [33] is used to scan one or several patterns from
PROSITE against the 150 protein sequences. All frequently matching (unspecific)
patterns and profiles are skipped. The M150X150 is then used to predict the PPI
network.

54.4 Results and Discussion

The performance of the proposed method is measured by how well it can predict
the PPI network. Prediction accuracy, whose value is the ratio of the number of
correctly predicted interactions between protein pairs to the total number of interac-
tions and non-interactions possibilities in network, is the best index for evaluating
the performance of a predictor. However, approximately 20% of the data are truly
interacting proteins, which leads to a rather unbalanced distribution of interacting
and non-interacting cases. Suppose we have six proteins, then the interacting pairs
are 1R 2 3
, R 4 and 5 R 6
, which result in three interactions cases out of a total of 15
interaction possibilities (12 non-interactions).
To assess our method objectively, another two indices are introduced in this
paper, namely specificity and sensitivity commonly used in the evaluation of
54 Protein–Protein Interaction Prediction 643
Specificity Sensitivity
0.703 0.61
Accuracy (100%)

Accuracy (100%)
0.7025 0.6 0.6067
0.702 0.7026 0.7026 0.59
0.7015 0.58
0.701
0.57
0.7005 0.5733
0.7 0.56
0.55 0.5467
0.6995 0.6992
0.699 0.54
Using only SW Eliminating inter- Adding protein Using only SW Eliminating inter- Adding protein
Algorithm domain linker domains Algorithm domain linker domains
regions evidence regions evidence

The 3 steps of the method The 3 steps of the method

Fig. 54.5 Sensitivity and specificity results

70.3
Accuracy (100%)

70.2
70.1 70.17 70.19
70
69.9
69.8
69.7 69.82
69.6
Using only SW Eliminating inter- Adding protein
Algorithm domain linker domains
regions evidence

The 3 steps of the method

Fig. 54.6 Overall accuracy

Table 54.1 Overall performance evaluation

tp fp tn fn Sens. Spec. Accuracy
Similarity using 82 68 15,523 6,677 0.5467 0.6992 60.82
SW algorithm
Inter-domain 86 64 15,597 6,603 0.5733 0.7026 70.17
linker regions
Structural domain 91 59 15,597 6,603 0.6067 0.7026 70.19
evidence

information retrieval. A high sensitivity means that many of the interactions that
occur in reality are detected by the method. A high specificity indicates that most
of the interactions detected by the screen are also occurring in reality. Sensitivity
and specificity are combined measures of true positive (tp), true negative (tn), false
positive (fp) and false negative (fn) and can be expressed as:
Based on these performance measures, the method was able to achieve encour-
aging results. In Figs. 54.5 and 54.6, we summarized the sensitivity and specificity
results based on the three stages of the method. The figures clearly show improve-
ment in sensitivity but not much in specificity and that’s because of the big number
of non-interacting possibilities.
The overall performance evaluation results are summarized in Table 54.1.
644 N. Zaki

54.5 Conclusion

In this research work we make use of both homology and structural similarities
among domains of known interacting proteins to predict putative protein interaction
pairs. When tested on a sample data obtained from the DIP, the proposed method
shows great potential and a new vision to predict PPI. It proves that the combination
of methods predicts domain boundaries or linker regions from different aspects and
the evolutionary relationships would improve accuracy and reliability of the pre-
diction as a whole. However, it is difficult to directly compare the accuracy of our
proposed method because all of the other existing methods use different criteria for
assessing the predictive power. Moreover, these existing methods use completely
different characteristics in the prediction. One of the immediate future works is to
consider the entire PPI network and not to restrict our work on binary interaction.
Other future work will focus on employing more powerful domain linker region
identifier such as profile domain linker index (PDLI) [37].

References

1. I. Donaldson, J. Martin, B. Bruijn, C. Wolting, V. Lay, B. Tuekam, S. Zhang, B. Baskin, G.

D. Bader, K. Michalickova, T. Pawson, and C.W. Hogue, PreBIND and Textomy – mining the
biomedical literature for protein-protein interactions using a support vector machine, BMC
Bioinformatics, 4(11) (2003).
2. E. Gharakhanian, J. Takahashi, J. Clever, and H. Kasamatsu, In vitro assay for protein–protein
interaction: carboxyl-terminal 40 residues of simian virus 40 structural protein VP3 contain a
determinant for interaction with VP1, PNAS, 85(18), 6607–6611 (1998).
3. P. L. Bartel and S. Fields, The yeast two-hybrid system. In Advances in Molecular Biology,
Oxford University Press, New York, 1997.
4. G. Rigaut, A. Shevchenko, B. Rutz, M. Wilm, M. Mann, and B. Seraphin, A generic protein
purification method for protein complex characterization and proteome exploration, Nature
Biotechnology, 17, 1030–1032 (1999).
5. M. Selbach and M. Mann, Protein interaction screening by quantitative immunoprecipitation
combined with knockdown (QUICK), Nature Methods, 3, 981–983 (2006).
6. L. Salwinski, C. S. Miller, A. J. Smith, F. K. Pettit, J. U. Bowie, and D. Eisenberg, The
Database of Interacting Proteins: 2004 update, Nucleic Acids Research, 1(32), 449–51 (2004).
7. H. W. Mewes, MIPS: analysis and annotation of proteins from whole genomes, Nucleic Acids
Research, 32, 41–44 (2004).
8. S. Peri, Human protein reference database as a discovery resource for proteomics, Nucleic
Acids Research, 32, 497–501 (2004).
9. J. Espadaler, Detecting remotely related proteins by their interactions and sequence similarity,
Proceedings of the National Academy of Sciences USA, 102, 7151–7156 (2005).
10. E. Marcotte, Detecting protein function and protein–protein interactions from genome
sequences, Science, 285, 751–753 (1999).
11. T. Dandekar, Conservation of gene order: a fingerprint of proteins that physically interact,
Trends in Biochemical Sciences, 23, 324–328 (1998).
12. M. Pellegrini, E. M. Marcotte, M. J. Thompson, D. Eisenberg, and T. O. Yeates, Assigning
protein functions by comparative genome analysis: protein phylogenetic profiles, Proceedings
of National Academy of Sciences USA, 96, 4285–4288 (1999).
13. A. Szilàgyi, V. Grimm, A. K. Arakaki, and J. Sholnick, Prediction of physical protein-protein
interactions, Physical Biology, 2, 1–16 (2005).
54 Protein–Protein Interaction Prediction 645

14. E. M. Marcotte, M. Pellegrini, M. J. Thompson, T. O. Yeates, and D. Eisenberg, A combined

algorithm for genome-wide prediction of protein function, Nature, 402, 83–86 (1999).
15. F. Pazos and A. Valencia, Similarity of phylogenetic trees as indicator of protein-protein
interaction, Protein Engineering, 14, 609–614 (2001).
16. J. Enright, I. N. Ilipoulos, C. Kyrpides, and C. A. Ouzounis, Protein interaction maps for
complete genomes based on gene fusion events, Nature, 402, 86–90 (1999).
17. D. Eisenberg, E. M. Marcotte, I. Xenarios, and T. O. Yeates, Protein function in the post-
genomic era, Nature, 405, 823–826 (2000).
18. J. Wojcik and V. Schachter, Protein-Protein interaction map inference using interacting domain
profile pairs, Bioinformatics, 17, 296–305 (2001).
19. W. K. Kim, J. Park, and J. K. Suh, Large scale statistical prediction of protein-protein
interaction by potentially interacting domain (PID) pair, Genome Informatics, 13, 42–50
(2002).
20. S. K. Ng, Z. Zhang, and S. H. Tan, integrative approach for computationally inferring protein
domain interactions, Bioinformatics, 19, 923–929 (2002).
21. S. M. Gomez, W. S. Noble, and A. Rzhetsky, Learning to predict protein-protein interactions
from protein sequences, Bioinformatics, 19, 1875–1881 (2003).
22. C. Huang, F. Morcos, S. P. Kanaan, S. Wuchty, A. Z. Chen, and J. A. Izaguirre, Predict-
ing protein-protein interactions from protein domains using a set cover approach, IEEE/ACM
Transactions on Computational Biology and Bioinformatics, 4(1), 78–87 (2007).
23. T. Pawson and P. Nash, Assembly of cell regulatory systems through protein interaction
domains, Science, 300, 445–452 (2003).
24. N. M. Zaki, S. Deris, and H. Alashwal, Protein–protein interaction detection based on
substring sensitivity measure, International Journal of Biomedical Sciences, 1, 148–154
(2006).
25. P. Aloy and R. B. Russell, InterPreTS: protein interaction prediction through tertiary structure,
Bioinformatics, 19, 161–162 (2003).
26. L. Lu, Multiprospector: an algorithm for the prediction of protein–protein interactions by
multimeric threading, Proteins, 49, 350–364 (2002).
27. J. Espadaler, O. Romero-Isart, R. M. Jackson, and B. Oliva, Prediction of protein–protein
interactions using distant conservation of sequence patterns and structure relationships,
Bioinformatics, 21, 3360–3368 (2005).
28. O. Keskin, A new, structurally nonredundant, diverse data set of protein–protein interfaces and
its implications, Protein Sciences, 13, 1043–1055 (2004).
29. T. Smith and M. Waterman, “Identification of common molecular subsequences”, Journal of
Molecular Biology, 147, 195–197 (1981).
30. H. Saigo, J. Vert, N. Ueda, and T. Akutsu, Protein homology detection using string alignment
kernels, Bioinformatics, 20(11), 1682–1689 (2004).
31. A. Bairoch and R. Apweiler, The SWISS-PROT protein sequence database and its supplement
TrEMBL in 2000, Nucleic Acids Research, 28, 45–48 (2000).
32. M. Suyama and O. Ohara, DomCut: prediction of inter-domain linker regions in amino acid
sequences, Bioinformatics, 19, 673–674 (2003).
33. A. Gattiker, E. Gasteiger, and A. Bairoch, ScanProsite: a reference implementation of a
PROSITE scanning tool, Applied Bioinformatics, 1, 107–108 (2002).
34. I. Xenarios, L. Salwı́nski, X. J. Duan, P. Higney, S. Kim, and D. Eisenberg, DIP, the Database
of Interacting Proteins: a research tool for studying cellular networks of protein interactions,
Nucleic Acids Research, Oxford University Press, 30, 303–305 (2002).
35. C. M. Deane, L. Salwinski, I. Xenarios, and D. Eisenberg, Protein interactions: two meth-
ods for assessment of the reliability of high throughput observations, Molecular & Cellular
Proteomics, 1, 349–56 (2002).
36. W. R. Pearson, Rapid and sensitive sequence comparisons with FASTAP and FASTA method,
Methods in Enzymology, 183, 63–93 (1985).
37. Q. Dong, X. Wang, L. Lin, and Z. Xu, Domain boundary prediction based on profile domain
linker propensity index, Computational Biology and Chemistry, 30, 127–133 (2006).
Chapter 55
Cellular Computational Model of Biology

Alex H. Leo and Gang Sun

Abstract One basic way that gene expression can be controlled by regulating tran-
scription. In prokaryotes, our basic understanding of how transcriptional control of
gene expression began with Jacob and Monod’s 1961 model of the Lac operon.
Recently, attempts have been made to model operon systems in order to understand
the dynamics of these feedback systems. In the present study, it was independently
attempted to model the Lac operon, based on the molecules and their activities on
the lac operon of an E. coli. By various ways and means in engineering, a flow chart
on the metabolic pathway of the lac operon activities of the E. coli was established.

Keywords Cellular Computational model  Lac operon  E. coli  metabolic pathway

55.1 Introduction

The control of gene expression has been a major issue in biology. One basic way
that gene expression can be controlled by regulating transcription. In prokaryotes,
our basic understanding of how transcriptional control of gene expression began
with Jacob and Monod’s 1961 model of the Lac operon (see Fig. 55.1), in which
a regulator protein controls transcription by binding to a target region called the
operator (see Fig. 55.2).
In the absence of this binding, RNA transcription is initiated when an RNA
polymerase is able to bind to a region, called the promoter, just upstream from
the operator. The RNA polymerase travels down stream and transcribes a series
of structural genes coding for enzymes in a particular metabolic pathway (see
Fig. 55.3). The operon system acts as a molecular switch involving positive or
negative control [1].

G. Sun (B)
Mechanics & Engineering Science Department, Fudan University, Shanghai 200433, China
E-mail: gang sun@fudan.edu.cn

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 647
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
648 A.H. Leo and G. Sun

Fig. 55.1 Organization of the lac genes of E. coli and the associated regulatory elements,
the operator, promoter, and regulatory gene [3]

Fig. 55.2 Functional state of the lac operon in E. coli growing in the absence of lactose [3]

Fig. 55.3 Functional state of the lac operon in E. coli growing in the presence of lactose as the sole
carbon source [3]
55 Cellular Computational Model of Biology 649

Recently, attempts have been made to model operon systems in order to under-
stand the dynamics of these feedback systems. In 2003, Yidirim modeled the Lac
operon as a series of coupled differential equations in terms of chemical kinetics [2].
Yidrim and Mackay note that their formulation assumes that they are dealing with
a large number of cells and that the dynamics of small numbers of individual cells
might be quite different. Meanwhile, it is hardly to obtain analytic solutions by
using concentrations, average-values, as the variables of dynamics in differential
equations, thus Laplace transform is necessary to solve the equations. Whereas, in
the present study, it was independently attempted to model the Lac operon, based
on the molecules and their activities on the lac operon of an E. coli. By various
ways and means in engineering, a flow chart (Fig. 55.4) on the metabolic pathway
of the lac operon activities of the E. coli was established. In terms of the amount of
each relevant molecule present in the cell, a set of difference equations, mathematic
model, comes on the basis of the flow chart. Since in form of difference equation,
the numerical solution of molecular state-variable are available, by recurring the
state-variables upon given a set of initial state or condition intuitionally.

Lactose

switch
E. coli membrance

RNA Polymerase
transcribing Lac gene
L

Lactose ~1 0<a<<1 Lactose

breaking down ~1 0<b<<1 isomerizing mRNA

tran scribin
y g
Allolactose A
Glucose?

n inducer/effector
G
translation
Galactose = = >
glucose R
¦ Âgalactosidase
enzyme

Glucose metabolizing
Transacetylase
enzyme
1

energy + Carbon N+5 Î (NO3)-1

Permease
enzyme
G

E. coli Constitutive Gene

growing Enzymes
Transcribing

G = 3000 in the presence of lactose

G=3 in the absence of lactose
repressor= 4 in the absence of lactose

Fig. 55.4 Flow chart of the activity pathways of an E. coli

650 A.H. Leo and G. Sun

55.2 List of Variables

mR(n) amount of mRNA polymerases at step n

R half-life of mRNA polymer
rTRSCRB transcriptional rate of mRNA in the presence of allolactose
G.n/ amount of ˇ-galactosidase molecules at step n
G half-life for ˇ-galactosidase
rTRSLTN translational rate of ˇ-galactosidase, permease and transacetylase
molecules per mRNA polymer
L.n/ amount of lactose molecules transported by the permease at step n in the
cell
Lext .n/ amount of lactose molecules outside of the cell at step n; Lext .0/ is the
amount of lactose molecules supplied at the beginning
MaxrTRSPT the maximum rate of lactose transported in by each permease
molecule
A.n/ amount of ATP molecules, produced from the lactose at the step n
A half-life of ATP
rmeta rate of a ˇ-galactosidase molecule breaking down the lactose into the
glucose, and thus into the ATP
 set step length, e.g.  D 1(min), with regard to the half-life and the creating-
rate for the substance presented by the variable

55.3 Determining mRNA Polymerases, mR(n + 1)

An E. coli cell contains about 3,000 copies of holoenzyme, a form of RNA poly-
merase [3]. The DNA-length of lacZYA is [1],

3;510 C 780 C 825 D 5;115 bp

It was found that mRNA half-lives were globally similar in nutrient-rich media,
determined from a least-squares linear fit. Approximately, 80% of half-lives ranged
between 3 and 8 min in M9 C glucose medium, and 99% of the half-lives measured
were between 1 and 18 min. In the LB, 99% of half-lives were between 1 and 15 min.
The mean half-life was 5.7 min in M9 C glucose and 5.2 min in the LB [4]. The
transcription of a polymerase approaches at a rate of average 30–50 nucleotides
per second [3]. But in some defined media, the general time was approximately
tripled [4]. The lac mRNA is extremely unstable, and decays with a half-life of only
3 min [3]. Another article indicates the average half-life of an mRNA in E. coli is
about 1.8 min. As lactose is no longer present, the repressor will be activated, and
thus bind to the operator, making the transcription action stopped.
Considering the above conditions, mR.n C 1/; rTRSCRB, and R are able to be
determined in the presence of lactose. The variable, mR (nC1), equates, elapsing the
period of , the amount remained of mR (n) through the halving attenuation, plus
the amount produced by transcription triggered in the presence of the allolactose
from the lactose.
55 Cellular Computational Model of Biology 651

5000

4000

3000

2000

1000

0
0 20 40 60 80 min

Fig. 55.5 Simulated amount of the mRNA molecules of the cell

Therefore, a polymerase needs 1.705–2.842 min for transcribing an mRNA, or

0.5865–0.3519 copies of mRNA come per minute per polymerase. So to speak,
1,759–1,055 copies of mRNA are maximally created per minute, with regard to
3,000 copies of the polymerases within an E. coli. Herein, rTRSCRB D 1;759 – 586
mRNA= min, R D 5 min. That is, the resulting formula is as follows:
0:69
mR.n C 1/ D e R  mR.n/ C rTRSCRB   (55.1)

A simulated curve of mR(n) is shown as in Fig. 55.5.

55.4 Determining ˇ-Galactosidases, G.n C 1/

Given supplying E. coli cells lactose, ˇ-galactosidase enzyme will appears in min-
utes, so do lac permease in membrane and transacetylase, a third protein. The level
of ˇ-galactosidase enzyme can accumulate to 10% of cytoplasmic protein [5]. Once
a ribosome approaches away of the initiative site on an mRNA, another one will
locate at the initiative site. Thus, many ribosomes may well simultaneously be trans-
lating each mRNA. An average mRNA has a cluster of 8–10 ribosomes, named as
polysome, for synthesizing protein [3].
Where cells of E. coli are grown in the absence of lactose, there is no need of
ˇ-galactosidase. An E. coli cell contains 3–5 molecules of the enzyme [3]. That
is, a dozen of the repressors bind and unbind rather than just bind and stay on the
mRNA. In a fraction of a second, after one repressor unbinds and before another
binds, an mRNA polymerase could initiate a transcription of the operon, even in
the absence of lactose. [3]. In 2–3 min after adding lactose, there soon are 3,000–
5,000 molecules of ˇ-galactosidase per E. coli cell [1, 3]. The ˇ-galactosidase in
the E. coli is more stable than the mRNA, whose half-life is more than 20 h. The
half-life of the permease is more than 16 h [1]. So that, the ˇ-galactosidase activity
remains at the induced level for longer. To add or change codons at the 50 end of the
652 A.H. Leo and G. Sun

gene to encode N-terminal amino acid within the DNA-lacZYA will provide greater
resistance to ˇ-galactosidase production [6]. As a porter, the permease compound
can be to concentrate lactose against the gradient across the cellular membrane in
ad hoc manner for E. coli. The concentration of the substrate against a gradient can
be achieved up to 103 to 106 -fold. The existing ˇ-galactosidase molecules will be
diluted out by cell division [5]. Whereas, replicating the DNA of an E. coli needs
84 min [3], the half-life is 100 min for simplicity, or 30 min for conservation. This
point of view could be used to derive a model in colony of the bacteria. The DNA-
length of lac ZYA [1]:

3;510 C 780 C 825 D 5;115 bpI

The mRNA-length of lac ZYA:

5;115  3  1 D 1;705  1 D 1;704 amino acids:

Protein synthetic rate is 350–400 amino acids (1/4–1/5 ˇ-galactosidase) per

minute per ribosome [5]. There are approximate 8 ribosomes attached to an mRNA
That is, for a ribosome to make a ˇ-galactosidase needs 4–5 min (say 4.5 min) by
running all the way of the mRNA, and there are 1.8 copies of the ˇ-galactosidase
translated per minute per mRNA (Fig. 55.6).
As a function of mR(n), G.n C 1/ is the amount of ˇ-galactosidase, same as
permease in count. G.nC1/, then rTRSLTN and G are able to be determined. G.nC1/
equates, elapsing the period of , the ˇ-galactosidase amount remained of G.n/
through the halving attenuation, plus the amount produced by translation of mR(n),
herein G D 1;000 min; rTRSLTN D 8  4:5 D 1:8; G.0/ D 3. That is, the resulting
formula is as follows:
0:69
G.n C 1/ D e G  G.n/ C rTRSLTN    mR.n/ (55.2)

x 105
2

1.5

0.5

0
0 20 40 60 80 min

Fig. 55.6 The simulated curve of G(n), the permease or ˇ -galactosidase of the cell
55 Cellular Computational Model of Biology 653

55.5 Determining Lactoses with in and out the Cell, Lext(n C 1)

and L(n C 1)

Lac permease is a highly lipophilic protein with an apparent subunit molecular

weight of about 30,000 [7]. The lactose transport-in rate of permease is 120 nmol
per mg per min. That is, 2.4 copies of lactose are transported into the cell by a
permease per minute.
rmeta , the rate of lactose breaking down into glucose by G.n/; 2 min is for to
convert a lactose into 72 ATPs [3]. Therefore, rmeta is 0.5, and max rTRSPT is 2.4.
rL is the rate of transporting lactose into the cell by the permease, presented
by G(n). L.n C 1/ is the amount of the lactose at step n C 1 in the E. coli. For
the period of , L.n C 1/ equates to the amount of the lactose brought-in by the
permease, presented by G.n/, but subtracting the amount metabolized into glucose.
Each glucose molecule, in turn, is immediately transformed into 72 ATPs (Figs. 55.7
and 55.8).
That is, the resulting formula is as follows:

Lext .n C 1/ D Lext .n/  max rTRSPT    G.n/

f .Lext .n/  L.n// (55.3)
L.n C 1/ D L.n/ C max rTRSPT    G.n/  f .Lext .n/
 L.n//  rmeta    L.n/ (55.4)

( x
2e  when x 0
f .x/ D x (55.5)
e else

in which,  is a constant, whose magnitude close to that of the lactose supplied.

x 106
5

0
0 20 40 60 80 min

Fig. 55.7 The simulated curves of the Lext (n) and L(n), the external and internal lactose molecules
of the cell
654 A.H. Leo and G. Sun

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2
-2 0 2 4 6
x 106

Fig. 55.8 The simulated function curve of the f (x)

55.6 Determining ATP Molecules, A(n)

ATP half-life,A, is about 15 min in water at room temperature, estimating 2 min for
to convert a copy of lactose into 72 ATPs. The energy resource for E. coli growth
is from the lactose in form of ATP at 72 in scale. The number of ATP for the
transportation-in of a lactose by a permease in energy equates about one. There-
fore, the transcription of an mRNA needs 5,115 ATPs (in forms of ATP, UTP, CTP,
or GTP), and the translation of an mRNA needs 1,700 ATPs.
Given the consumption of energy, comprising transcribing, translating, as well
as transporting lactose into the cell, count 80% of total energy, as is the primary life
activity, then a coefficient d for energy distribution set at 0.8. The remaining energy
is for the other metabolic activities.
With regard to the rmeta , G.n/, , steps n C 1 and n, the energy generated by the
glucose is distributed among producing mRNA and ˇ-Galactosidase, and bringing
the lactose into E. coli cell (Fig. 55.9–55.11).
An energy-distributional equation is as follows:
0:69
A.n C 1/ D e A A.n/ C 72  rmeta  G.n/  Œ5115rTRSCRB
C 1700rTRSLTN  mR.n/ C max rTRSPT  f .Lext .n/  L.n//  G.n/
(55.6)

In the above equation, first term is accumulated energy before step n; second term
is the energy created from the lactose, and third term is the consumed for the three
activities. If the third is greater or equal to the second, there will not be energy
55 Cellular Computational Model of Biology 655

x 107
2.5

1.5

0.5

0
0 20 40 60 80 min

Fig. 55.9 Simulated curve of the ATP molecules consumed within the cell

x 107
3

2.5

1.5

0.5

0
0 20 40 60 80 min

Fig. 55.10 Simulated function curve of the ATP molecules created within the cell

Fig. 55.11 A classic experimental data of E. coli

656 A.H. Leo and G. Sun

accumulation. An comprehensive energy distributive coefficient equation, c(n), can

be determined as following,
(
e.n/ when e.n/  1
c.n/ D (55.7)
1 else

e.n/ D 72  rmeta  G.n/  d  Œ5115rTRSCRB C 1700rTRSLTN  mR.n/

CrTRSPT  f .Lext .n/  L.n//  G.n/
(55.8)
The domain of c.n/ is within [0, 1].
If the energy is inadequate, the energy-distributional coefficient equation comes,
and thus the Eqs. (55.1)–(55.4) are modified into energy-constrained Eqs. (55.1)’–
(55.4).

0:69
mR.n C 1/ D e R mR.n/ C rTRSCRB   c.n/; (55.1)
0:69
G.n C 1/ D e G G.n/ C rTRSLTN    m  R.n/  c.n/; (55.2)
Lext .n C 1/ D Lext .n/  max rTRSPT    G.n/;
f .Lext .n/  L.n//  c.n/; (55.3)
L.n C 1/ D L.n/ C max rTRSPT    G.n/  f .Lext .n/  L.n//;
c.n/  rmeta    L.n/: (55.4)

55.7 The System Structure

Eqs. (55.1)–(55.8), or Eqs. (55.1)’–(55.4)’ in place if necessary, are integrated into

a systematic function Eq. (55.9), with respect to the initial condition presented as
Eq. (55.10).

X.n C 1/ D T.X.n// C U (55.9)

X.n/ D ŒR.n/; G.n/; Lext .n/; L.n/; A.n/ T
(55.10)

They are solved numerically by MATLAB through setting the following initial
condition:

X.0/ D Œ0 3 supply 0 0T

U D ŒrTRSCRB  0 0 0 0T
55 Cellular Computational Model of Biology 657

55.8 In Comparison to Traditional Biological Experiment

A traditional experiment culturing E. coli in tubes offered the following graph [6].
In comparison of the outcomes of the presented equations and the traditional
experimental shown as Fig. 55.2, the results are consistent in a mutual time scale;
and so does mRNA in Fig. 55.1. In addition, more information on the cell, such as
external and internal lactose, ATP consumption, etc. in a cell of E. coli are available
by the model, but not in the other methods.

55.9 Conclusion

By computational simulation, the Figs. 55.2–55.7 were obtained. Supplied with

mere lactose as nutrition, an E. coli cell of the Operon activities can be described
by the eight difference equations, or a mathematical model. The Relevant outcomes
of the model are highly consistent with traditional experimental results. The math-
ematic model can be developed into a general method to govern the growth of an
E. coli colony by gene expression. The model is justified through the traditional
biological experiment.

55.10 Discussion

This paper was based on the authors’ background of computer science, Electri-
cal Engineering, and Genetics. This paper discusses the growth of an E. coli cell
supplied with mere lactose. As a core governing the state-variables of the E. coli
cell’s molecules on the lac operon, the model could be developed into another set
of formula for colony, with regard to cellular divisions in compliance with normal
distribution in mathematics. The set of formula may well model a colony of the bac-
teria by halving the values of state-variable in various periods, simulating cellular
division. Finally, the formula could be transformed into corresponding differential
equations. In addition, effects of parameters  and d in the Eqs. (55.5) and (55.8)
should be studied further respectively. The rate or behavior pattern of the permeases
transporting lactose across the membrane of an E. coli cell is rather approximate.
Finally, it is noticeable that the simulated numerical solutions of the model vary
slightly with  value. The phenomenon probably is of the theoretical approximate
hypothesis in the Section 55.2.
Since on individual cellular level, a colony model could be developed rather to
describe and simulate more complex activities and regular patterns of the represen-
tative substances of the colony, by considering individual cell division, and could
further be developed into a general quantitative analysis for cellular biology in high
precise.
658 A.H. Leo and G. Sun

References

1. Lewis, Benjamin, 2004. Genes VIII. Prentice-Hall, Upper Saddle River, NJ

2. Yidirim, Necmettin et al., 2003. Feedback Regulation in the Lactose Operon: A Mathematical
Modeling Study and Comparison with Experimental Data, Biophys. J. 84: 2841–2851.
3. Russell, Peter J., 2002. iGenetics, Benjamin Cumming, New York.
4. Bernstein, Jonathan, A. et al., 2002 July 23. Global Analysis of mRNA Decay and
Abundance in Escherichia coli at Single-Gene Resolution Using Two-Color Fluorescent
DNA Microarrays. Proc. Natl. Acad. Sci. USA 99(15): 9697–9702; http://www.pnas.org/
cgi/content/full/99/15/9697
5. Terry, Thomas, M., 2002. Regulation of Metabolism. Microbiology, Chapter 12, pp. 237–
250, 4th Edition; http://www.biologie.unihamburg.de/bonline/library/micro229/terry/229sp00/
lectures/regulation.html.
6. Southern Illinois University, 2003. BIOTECHNOLOG.Y. MICR 421, chapter 6; http://www.
science.siu.edu/microbiology/micr421/chapter6.html
7. Ullmann, Agnes, 2001. Escherichia coli Lactose Operon, Encyclopedia of Lifesciences;
http://216.239.41.104/search?q = cache:mzCMeayarnMJ:www.cellcycle.bme.hu/oktatas/
mikrofiz/extra/lac%2520operon.pdf + how + weight + lac + permease + is + in + e. + coli&hl =
en&ie = UTF-8.
8. Weinglass, Adam, B. et al., 2001. Manipulating Conformational Equilibria in the Lactose
Permease of Escherichia coli, Encyclopedia of Life sciences; http://darwin.bio.uci.edu/
bardwell/weinglass02jmb.pdf.
Chapter 56
Patient Monitoring: Wearable Device
for Patient Monitoring

Robert G. Lupu, Andrei Stan, and Florina Ungureanu

Abstract The increasing request of patients, suffering of chronic diseases, who wish
to stay at home rather then in a hospital and also the increasing need of homecare
monitoring for elderly people, have lead to a high demand of wearable medical
devices. Also, extended patient monitoring during normal activity has become a very
important target. Low power consumption is essential in continuously monitoring of
vital-signs and can be achieved combining very high storage capacity, wireless com-
munication, and ultra-low power circuits together with firmware management of
power consumption. This approach allows the patient to move unconstraint around
an area, city or country. In this paper the design of ultra low power wearable
monitoring devices based on ultra low power circuits, high storage memory flash,
bluetooth communication and the firmware for the management of the monitoring
device are presented.

Keywords Patient Monitoring  Wearable Device  wearable medical Device 

monitoring device

56.1 Introduction

Phenomena of ageing population observed in most developed countries [1] and

prevalence of chronic diseases have increased the need for chronic and geriatric
care at home [2]. The task of patient monitoring may be achieved by telemedicine
(enabling medical information-exchange as the support to distant-decision-making)
and telemonitoring (enabling simultaneous distant-monitoring of a patient and his
vital functions) both having many advantages over traditional practice. Doctors can

R.G. Lupu (B)

Gh. Asachi” Technical University of Iasi, Faculty of Automatic Control and Computer
Engineering, 700050 Romania
E-mail: robert@cs.tuiasi.ro

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 659
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
660 R.G. Lupu et al.

receive information that has a longer time span than a patient’s normal stay in
a hospital and this information has great long-term effects on home health care,
including reduced expenses for healthcare.
Despite the increased interest in this area, a significant gap remains between
existing sensor network designs and the requirements of medical monitoring. Most
telemonitoring networks are intended for deployments of stationary devices that
transmit acquired data at low data rates and rather high power consumption. By
outfitting patients with wireless, wearable vital sign devices, collecting detailed real-
time data on physiological status can be greatly simplified. But, the most important
effect is the widely social integration of peoples with disabilities or health problems
that need discreet and permanent monitoring.
It is well known that Bluetooth represents a reliable and easy solution for sig-
nal transmission from a portable, standalone unit to a nearby computer or PDA [3].
When a real time rudimentary analysis program detects pathological abnormality
in the recorded data an alert is sent to the telemonitoring centre via internet or
GSM/GPRS using a PDA or computer.
The needs of the society and market trend suggest that healthcare-related applica-
tions will be developed significantly. Particularly, PDA phones interfaced with wear-
able sensors have an enormous market as they will enlarge the goal of healthcare and
mobile devices [4].
The development of ultra low power wearable monitoring device unit is pro-
pounded. The main functions of the wearable device consist in signal acquisition,
rudimentary processing, local data storage and transmission to the remote care
centre. In Fig. 56.1, the general architecture of the telemonitoring network is
presented.
This paper presents the initial results and our experiences for a prototype medi-
cal wearable device for patient monitoring taking into account some hardware and
software aspects regarding vital signs measurement, robustness, reliability, power
consumption, data rate, security and integration in a telemonitoring network.

Fig. 56.1 The general structure of telemonitoring network

56 Patient Monitoring: Wearable Device for Patient Monitoring 661

Respiratory
Acceleromatters
rate
SD/MMC
ECG ADC
ECG Amplifier CARD
3 Leads
Ultra low power
Digital Port Keyboard
microcontroller
LED Drive Circuit
SaO2
Input Front End DAC Bluetooth
Heart rate
Circuit and LED
Temperature Digital Port Digital
Temperature sensor
board

Fig. 56.2 Block diagram of the monitoring device

56.2 Wearable Monitoring Device

The monitoring device is build using custom developed hardware and application
software. The block diagram is presented [5] in Fig. 56.2. Low power amplifiers
and sensor are connected to the device [6], for vital parameters acquisition. The
health parameters acquired are: heart rate, heart rhythm regularity, respiratory rate,
oxygen saturation and body temperature. The monitoring device includes also a
custom made three-lead ECG amplifier.
The digital board includes an ultra low power microcontroller and interface cir-
cuits for keyboard, SD/MMC card and Bluetooth. The low power sensors for the
above mentioned parameters are connected to the digital board using digital port or
specific converters.
The wireless communication via Bluetooth with a computer or PDA and the flash
memory for raw data information storage are two of the most important facilities of
the proposed wearable device.
The software application running on computer and PDA assures the communica-
tion via internet respectively GSM/GPRS [3, 7].
The monitoring device is built using an ultra low power microcontroller (MSP430
from Texas Instruments) that has a 16 bit RISC core with clock rates up to 8 MHz.
An important advantage of this family of microcontrollers is the high integration of
some circuits that offer the possibility to design devices with open architecture: dig-
ital ports, analog to digital converter (ADC) and digital to analog converter (DAC).
MSP430 microcontroller also has a built in hardware multiplier which is a valuable
resource for digital filter implementation. The custom made electronic boards used
for data acquisition are all designed with low power circuits. The system is modular:
every board is detachable.
The on board storage device (SD/MMC card) with FAT32 file system is used
for raw data recording together with signals processing results. The radio module
Bluetooth [8] designed for short-range communication is used for data transmis-
sion between monitoring device and PC or PDA. This module has low power
consumption being used only when data transfer is performed.
662 R.G. Lupu et al.

56.3 Functionality

The ultra low power wearable monitoring device is able to acquire simultane-
ously the physiological parameters mentioned in the previous section and also
to perform rudimentary digital signal processing on board. The signals are con-
tinuously recorded in separate files on flash memory for feature analysis. Once
pathological abnormality is detected, the monitoring device requests a transmission
trough PC or PDA to the remote care centre.
For the respiratory rate it is used a low power three axis low-g accelerometer
MMA7260QT [9]. This is a low cost capacitive micromachined accelerometer that
features signal conditioning, a one-pole low pass filter, temperature compensation
and g-Select which allows for the selection among four sensitivities.
The accelerometer outputs are read and processed once at 10 ms (100 times per
second).
Because this is a low power system, the accelerometer is disabled between two
conversion cycles. Each conversion cycle has a 2 ms startup period for accelerome-
ters to recover from sleep mode.
Due to the sensitivity of the accelerometer and the high resolution of analog to
digital converter, the useful signal (the output voltages of accelerometer) has a large
amount of superposed noise with possible sources in body moving or even heart
beating. In Fig. 56.3, the acquired signal affected by noise is presented. There can
be seen the initial rise of voltage corresponding to the start of chest movement. This
is explained by the fact that the movable central mass of g cell moves in opposite
direction of the movement. As the chest movement velocity becomes constant, the

Fig. 56.3 The accelerometer time response

56 Patient Monitoring: Wearable Device for Patient Monitoring 663

acceleration decreases and movable central mass comes back to initial position. This
settling process implies acceleration opposite to the direction of movement.
Noise filtering is achieved by using a 128 point rolling average filter. Math-
lab simulation was used with real datasets for filter calibration. Without filtering,
the signal is almost unusable for further processing. After filtering, the noise is
substantially reduced.
To easily distinguish between positive and negative accelerations (e.g. to deter-
mine the sense of movement) the accelerometer has an offset of half of supply volt-
age for 0 g acceleration. So, for positive or negative accelerations the corresponding
output voltage is positive in all cases.
To make the difference between positive and negative accelerations, it is neces-
sary to fix the reference value corresponding for 0 g. This is done at the application
startup, when an average with 512 points is made.
In Fig. 56.4, the results of the algorithm are plotted. The first plot is the raw
unfiltered data from accelerometers outputs in millivolts. This plot is a 2,500 points
record of accelerometer output for a time period of 5 s. The second plot represents
the filtered input acceleration.
The custom made ECG module uses three leads. The ECG signals are acquired
using micro power instrumentation amplifiers and micro power operational ampli-
fiers. Both are single power supply. A very important feature of the instrumen-
tation amplifier is that it can be shutdown with a quiescent current of less than
1 μA (Fig. 56.5). Returning to normal operations within microseconds, the shut-
down feature makes is optimal for low-power battery or multiplexing applications
[10]. The propounded functioning is: on power on the instrumentation ampli-
fier is shutdown. With 10 μs before a conversion cycle starts the instrumentation
amplifier is returned from shutdown mode and back again when the conversion cycle
finishes.

Fig. 56.4 The processed signals

664 R.G. Lupu et al.

Fig. 56.5 The instrumentation amplifier output

The temperature is monitored every minute using an ultra low power temperature
sensor.
Meanwhile, the sensor is shutdown saving maximum power by shutting down
all device circuitry other then the serial interface, reducing current consumption to
typically less than 0:5 μA [11].
Because the read/write operation with flash card can reach a current consumption
of max 60 mA, these are done with blocks of data [12]. The file system on flash card
is FAT32 witch gives the possibility of reading those files with any PC with a card
reader.
The average current consumption of Bluetooth circuit is around of 35 mA. For
this reason, the Bluetooth communication is used in two ways: always connected
transmitting live or stored data (signal processing and data storing are not per-
formed) and not connected, the circuit is in sleep mode (signal processing and data
storing are performed). If a critical situation occurs the circuit is awaken, the com-
munication starts, and in the same time data storage is performed. When files are
downloaded every monitoring operation is stopped.
In the case of blood oxygenation measurement, for driving and controlling the
LEDs a solution proposed by Texas Instruments was used [9]. It is based on two
LEDs, one for the visible red wavelength and another for the infrared wavelength.
The photo-diode generates a current from the received light. This current signal is
amplified by a trans-impedance amplifier. OA0 (microcontroller built in), one of the
three built in op-amps, is used to amplify this signal. Since the current signal is very
small, it is important for this amplifier to have a low drift current [9].
56 Patient Monitoring: Wearable Device for Patient Monitoring 665

Because of the high level of analog circuits’ integration the external components
involved in hardware development are very few. Furthermore, by keeping the LEDs
ON for a short time and power cycling the two light sources, the power consumption
is reduced [9].

56.4 Firmware Issue

The developed firmware consists of several tasks and each of them manages a par-
ticular resource as presented in Fig. 56.6. The communication between tasks is
implemented with semaphores and waiting queues allowing a high level of paral-
lelism between processes. Each process may be individually enabled or disabled.
This feature is very important in increasing the flexibility of the application: if real
time monitoring is desired, then SD Card Process may be disabled and Bluetooth
Process is enabled, if only long term monitoring is desired then SD Card Process
is enabled and Bluetooth Process may be disabled. This has a positive impact on
power consumption because only the resources that are needed are enabled for use.
A logical cycle of the device operation is presented in Fig. 56.6. In the first step
(thread), a buffer of data is acquired. The ADC module works in sequence mode and
acquires data from several channels at a time. The ADC channels are associated with
temperature, accelerometer, ECG signals and SaO2 inputs of electronic interface
modules. One sequence of ADC conversion is made by one sample for each of
these inputs. The data is formatted and then stored in a buffer. Only when the buffer
is full the following thread may begin. Depending on application requirements the
storage, transmission or analyze of the data may be performed.
Encryption of personal data is provided. Some files that store personal data
related to the patient are encrypted and can be accessed only using the proper
decryption algorithm. Only a part of the files are encrypted because there must be
kept a balance between power consumption and computing power requirements. For
space saving, a light compression algorithm may be activated as an extra feature of
the device. The activation of this feature has a negative impact on the overall power
consumption.
The communication process implements an application layer protocol for data
exchange between the device and other Bluetooth enabled devices. The communi-
cation requirements must be limited to a minimum rate in order to save power. The
Bluetooth module is powered on when a message has to be sent. The message trans-
mission rate is kept low by using internal buffering and burst communication. The
Bluetooth module can be also waked up by an incoming message that may embed
commands to the device.
The analysis process implements some rudimentary signal processing tasks
in order to detect anomalies in physiological activities. Only few vital parame-
ters anomalies are locally detected (by the wearable device) and trigger an alarm
event. More complex signal processing and dangerous condition detection are
666 R.G. Lupu et al.

Fig. 56.6 The tasks cycle

implemented on a remote computer which has the required computing power. The
remote computer receives raw data from the device and does the required analysis.
The physical processes that are monitored have a slow changing rate leading to
a low sampling rate requirement. This allows the controller to be in sleep mode for
an important percentage of operating time.
56 Patient Monitoring: Wearable Device for Patient Monitoring 667

The acquisition of one full data buffer takes the most time of the operation cycle.
The time needed for the execution of each remaining thread is shorter compared to
the ADC process.
A double buffering scheme is used: meanwhile a buffer is filled up with new data
samples, the second buffer may be processed (stored on SD card, transmitted via
Bluetooth or locally analyzed in order to detect conditions that should trigger alarm
events). When the new data buffer is acquired the buffers change (simple pointer
assignment) occurs: the previous processed buffer becomes empty, the acquired data
will be placed in it and the actual filled buffer will be processed.
The device must be reliable. For this purpose a power monitoring function is
designed. To implement this feature the ADC module is used. One of its inputs
monitors the voltage level across the batteries. The alarm events are designed to
preserve as much power as possible. If a voltage threshold is reached then an alarm
event is triggered and a resource could be disabled or its usage restricted only if
configured so. For example, the messages content and its transmission period could
be modified in order to save power but also to ensure a safe operation.
The computer software (see Fig. 56.7) is a MDI application (Multiple Document
Interfaces). A connection can be established with more than one device (each one
has a unique id). The software is always listening for connection with registered
Bluetooth monitoring device. This means that before data transmission, the device
id has to be known by the software. This is done by entering the code manually,
or by answering yes, when the application prompts for a communication request.
The ID is stored for further use. The communication session is established in two
ways: by the device when a critical situation occurs or by the computer when a
real time monitoring or file download is requested. This is done by sending a break
condition to the device to wake it from the deep sleep state [8]. The software assures
the connection to a server via Internet and also data transmission.

Fig. 56.7 Software user interface

668 R.G. Lupu et al.

56.5 Conclusion

The work of this paper focuses on design and implementation of an ultra low
power wearable device able to acquire patient vital parameters, causing minimal
discomfort and allowing high mobility. The proposed system could be used as a
warning system for monitoring during normal activity or physical exercises. The
active collaboration with The Faculty of Biomedical Engineering from University
of Medicine and Pharmacy from Iasi (Romania) and with several hospitals offered
the opportunity to test the prototype. Preliminary results were so far satisfactory.
This wearable device will be integrated into SIMPA healthcare systems to provide
real-time vital parameters monitoring and alarms.
In conclusion, this paper presents some challenges of hardware and software
design for medical wearable device based on low-power medical sensors and
microcontroller with a recent tremendous impact in many medical applications.
Obviously, the results demonstrate that there is still significant work to be done if
the wearable device is effectively integrated in a large network of medical sensors.

References

1. United Nations Population Division Department of Economic and Social Affairs.

http://www.un.org/esa/population/publications/aging99/a99pwld.htm
2. United Nations Department of Economic and Social Affairs, Report on the World Social
Situation 2007 http://www.un.org/esa/socdev/rwss/docs/rwss07 fullreport.pdf
3. K. Hung, Wearable Medical Devices for Tele-Home Healthcare Proceedings of the 26th
Annual International Conference of the IEEE EMBS, CA, USA, September 1–5, 2004.
4. K. J. Liszka, Keeping a beat on the heart IEEE Pervasive Computing, vol. 3, no. 4, October–
December 2004 p. 42.
5. Custom build using low power Texas Instruments technology http://www.ti.com
6. C. Rotariu, R. G. Lupu, An ultra low power monitoring device for telemedicine, 1st
National Symposium on e-Health and Bioengineering, EHB 2007, November 16, 2007, IASI,
ROMANIA.
7. R. G. Lupu, Solution for home care monitoring via GSM network, Master degree thesis,
University of Naples “Federico II”, Department of Electronic Engineering and Telecom-
munucations, Biomedical Engineering Unit, 2003.
8. F2M03GLA is a Low power embedded BluetoothTM v2.0 + EDR module with built-
in high output antenna. Available: http://www.free2move.net/index.php?type = view&pk =
24&pk language = 1&PHPSESSID = dc94542acd247be5916d98a1f5d088b9
9. A Single-Chip Pulsoximeter Design Using the MSP430. Available: http://focus.ti.com/lit/an/
slaa274/slaa274.pdf
10. The INA322 family is a series of low cost, rail-to-rail output, micropower CMOS instru-
mentation amplifiers that offer wide range, single-supply, as well as bipolar-supply operation.
Available: http://focus.ti.com/lit/ds/symlink/ina322.pdf
11. The TMP102 is a two-wire, serial output temperature sensor available in a tiny SOT563
package. http://focus.ti.com/lit/ds/symlink/tmp102.pdf
12. ATP MMC plus Specification. Available: http://flash.atpinc.com/articles/images/1031/MMC
Plus spec.pdf
Chapter 57
Face Recognition and Expression Classification

V. Praseeda Lekshmi, Dr. M. Sasikumar, and Divya S Vidyadharan

Abstract In this chapter, faces are detected and facial features are located from
video and still images. ‘NRC-IIT Facial video database’ is used as image sequences
and ‘face 94 color image database’ is used for still images. Skin pixels and non-skin
pixels are separated and skin region identification is done by RGB color space. From
the extracted skin region, skin pixels are grouped to some meaningful groups to
identify the face region. From the face region, facial features are located using seg-
mentation technique. Orientation correction is done by using eyes. Parameters like
inter eye distance, nose length, mouth position, and DCT coefficients are computed
which is used for a RBF based neural network.

Keywords Face Recognition  Expression Classification  neural network  facial

features

57.1 Introduction

The technological advancement in the area of digital processing and imaging has
led to the development of different algorithms for various applications such as auto-
mated access control, surveillance, etc. For automated access control, most common
and accepted method is based on face detection and recognition. Face recognition
is one of the active research areas with wide range of applications. The problem is
to identify facial image/region from a picture/image. Generally pattern recognition
problems rely upon the features inherent in the pattern for efficient solution. Though
face exhibits distinct features which can be recognized almost instantly by human
eyes, it is very difficult to extract and use these features by a computer. Human can
identify faces even from a caricature. The challenges associated with face detection

V.P. Lekshmi (B)

College of Engineering, Kidangoor, Kottayam, Kerala, India
E-mail: vplekshmi@yahoo.com

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 669
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
670 V.P. Lekshmi et al.

and recognition are pose, occlusion, skin color, expression, presence or absence of
structural components, effects of light, orientation, scale, imaging conditions etc.
Most of the currently proposed methods use parameters extracted from facial
images. For access control application, the objective is to authenticate a person
based on the presence of a recognized face in the database.
Here skin and non-skin pixels are separated and the pixels in the identified skin
region are grouped to obtain face region. From the detected face area, the facial fea-
tures such as eyes, nose and mouth are located. Geometrical parameters of the facial
features and the DCT coefficients are given to a neural network for face recogni-
tion. The expressions with in the face regions are analyzed using DCT and classified
using a neural network.

57.2 Background and Related Work

A lot of research has been going on in the area of human face detection and
recognition [1]. Most face detection and recognition methods fall into two cate-
gories: Feature based and Holistic. In feature-based method, face recognition relies
on localization and detection of facial features such as eyes, nose, mouth and
their geometrical relationships. In holistic approach, entire facial image is encoded
into a point on high dimensional space. Principal Component Analysis (PCA) and
Active Appearance Model (AAM) [2] for recognizing faces are based on holistic
approaches. In another approach, fast and accurate face detection is performed by
skin color learning by neural network and segmentation technique [3]. Indepen-
dent Component Analysis (ICA) was performed on face images under two different
conditions [4]. In one condition, image is treated as a random variable and pixels
are treated as outcomes and in the second condition pixels are treated as random
variables and image as outcome. Facial expressions are extracted from the detailed
analysis of eye region images is given in [5]. Large range of human facial behavior
is handled by recognizing facial muscle actions that produce expressions is given
in [6]. Video based face recognition is explained in [7]. Another method of classi-
fication of facial expression using Linear discriminant analysis (LDA) is explained
in [8] in which the Gabor features extracted using Gabor filter banks are compressed
by two stage PCA method.

57.3 Discrete Cosine Transform

Discrete Cosine Transform of an N  N cosine transform matrix C D fc.k; n/g is

defined [6] as
57 Face Recognition and Expression Classification 671
p
C.k; n/ D 1= N; k D 0; n D 0: : : N  1 (57.1)
p
D 2=N cos.….2 n C 1/ k/=2 N;
k D 1: : :N  1
N D 0: : : N  1

The cosine transform is orthogonal, that is

C D C D> C  1 D CT (57.2)

On applying the DCT, the input signal will get decomposed into a set of basis
images. For highly correlated data, cosine transforms show excellent energy com-
paction. Most of the energy will be represented by a few transform coefficients.

57.4 Radial Basis Function

Radial functions are a special class of function. Their characteristic feature is that
their response decreases (or increases) monotonically with distance from a central
point. The center, the distance scale, and the precise shape of the radial function
are parameters of the model, all fixed if it is linear. A typical radial function is the
Gaussian which, in the case of a scalar input, is

h.x/ D exp..x  c/2 =r2 /: (57.3)

Its parameters are its centre c and its radius r. Universal approximation theorems
show that a feed forward network with a single hidden layer with non linear units can
approximate any arbitrary function [Y]. No learning is involved in RBF networks.
For pattern classification problems, the numbers of input nodes are equal to the
number of elements in the feature vector and the numbers of output nodes are equal
to the number of different clusters.A Radial Basis Function network is shown in
Fig. 57.1.

57.5 Method

The method explains detection of faces from video frames and still images. This is
followed by extraction of facial features, recognition of faces and analyzing facial
expressions.
672 V.P. Lekshmi et al.

f(x)

ω1 ωj ωm

h1(x) hj(x) hm(x)

x1 xj xn

Fig. 57.1 Radial basis function network

57.5.1 Facial Region Identification

The first step in face detection problem is to extract facial area from the background.
In our approach, both still and video images are used for face detection. Image
frames from video are extracted first. The input images contain regions other than
face such as hair, hat etc. Hence it is required to identify the face. Each pixel in
the image is classified as skin pixel or non-skin pixel. Different skin regions are
detected from the image. Face regions are identified from the detected skin region as
in [9] which addressed the problem of face detection in still images. Some randomly
chosen frames with different head pose, far away from the camera, and expressions
from video face images are extracted as shown in Fig. 57.2. The difference image at
various time instances is shown in Fig. 57.3.
The portion of the image that is moving is assumed as head. The face detection
algorithm used RGB color space for the detection of skin pixels. The pixels corre-
sponding to skin color of the input image is classified according to certain heuristic
rules. The skin color is determined from RGB color space as explained in [10, 11].
A pixel is classified as skin pixel if it satisfies the following conditions.

R > 95 AND G > 40 AND B > 20 AND max fR; G; Bg  min fR; G; Bg
(57.4)
> 15 AND jR  Gj > 15 ANDR > GANDR > B

OR
57 Face Recognition and Expression Classification 673

Fig. 57.2 Some frames from face video database

Fig. 57.3 Difference image

Fig. 57.4 Edge detection and skin region identification from video frames

R > 220ANDG > 210ANDB > 170ANDjR  Gj  15ANDRBANDG > B

(57.5)
Edge detection is performed in each frame. Edge detection and skin regions identi-
fied from the color images of video frames and still images are shown in Figs. 57.4
and 57.5.
674 V.P. Lekshmi et al.

Fig. 57.5 Edge detection and

skin region identification from
still images

From these skin regions, it is possible to identify whether a pixel belongs to skin
region or not. To find the face regions, it is necessary to categorize the skin pixels
in to different groups so that it will represent some meaningful groups such as face,
hand etc. Connected component labeling is performed to classify the pixels. In the
connected component labeling operation, pixels are connected together geometri-
cally. In this, 8-connected component labeling used so that each pixel is connected
to its eight immediate neighbors. At this stage, different regions are identified and
each region has to be classified as a face or not. This is done by finding the skin
area of each region. If the height
p to width ratio of those skin region falls with in the
range of golden ratio (.1 C 5/=2˙ tolerance), then that region is considered as a
face region.

57.5.2 Face Recognition

57.5.2.1 Segmentation

Image segmentation is a long standing problem in computer vision. There are differ-
ent segmentation techniques which divides spatial area with in an image to different
meaningful components. Segmentation of images is based on the discontinuity and
similarity properties of intensity values. Cluster analysis is a method of grouping
objects of similar kind in to respective categories. It is an exploratory data analy-
sis tool which aims at sorting different objects in to groups in a way that degree
of association between two objects is maximal if they belong to same group and
minimal otherwise. K-Means is one of the unsupervised learning algorithms that
solve the clustering problems. The procedure follows a simple and easy way to clas-
sify a given data set through a certain number of clusters (assume K clusters) fixed a
57 Face Recognition and Expression Classification 675

priori. The idea is to define K centroids, one for each cluster. These centroids should
be placed in a cunning way because of different location causes different results. So
the better choice is to place them as much as possible far away from each other. The
next step is to take each point belonging to a given dataset and associate it to the
nearest centroid. When no point is pending, the first step is completed and an early
groupage is done. At this point K new centroids need to be re-calculated as barycen-
tres of the clusters resulting from previous step. After getting these K new centroids,
a new binding has to be done between the same dataset points and the nearest new
centroid. A loop has been generated. As a result of this loop it may be noticed that
the K centroids change their location step by step until no more changes are done.
If we know the number of meaningful groups/classes based on the range of pixel
intensity, weighted K-means clustering can be used to cluster the spatial intensity
values.
In facial images the skin color and useful components can be generally classified
as two different classes. But if two class based clustering is used, it may result
in components that may be still difficult to identify. So three classes are used and
are able to cluster the data in useful manner. Initial cluster centers are calculated
using histogram. Then K-means clustering algorithm computes distance between
the different pixels and cluster centers and selects a minimum distance cluster for
each pixel. This process continues until all pixels are classified properly.
The results of clustering algorithm are shown in Fig. 57.6. Class-I is selected
since it is possible to separate the components properly compared to other classes.
Then connectivity algorithm is applied to all the components in the clustered face.

57.5.2.2 Feature Extraction

Eye regions are located in the upper half of skin region and can be extracted using
the area information of all the connected components. Using eye centers, orientation
is corrected by rotation transformation. After calculating the inter eye distance, nose
and mouth are identified as they generally appear along the middle of two eyes in
the lower half. The inter eye distance, nose length and mouth area are computed.
To keep track of the overall information content in the face area, DCT is used. On
applying DCT, most of the energy values can be represented by a few coefficients.
First 64  64 coefficients are taken as part of the feature set. The face area calcu-
lated from the skin region is taken as another parameter. Parameters like inter eye

Fig. 57.6 Clustering

676 V.P. Lekshmi et al.

distance, nose length, mouth position, face area and DCT coefficients are computed
and given to a RBF based neural network.

57.5.3 Facial Expression Analysis

For expression analysis, face images are considered from JAFFE face database [12].
Figure 57.7 shows some images in the database. There were ‘K’ images with ‘N’
expressions for each face so that ‘K  N’ face images are used as the database. Nor-
malization is done to make the images with uniform scale. The facial expressions
are analyzed by facial feature extraction. Facial expressions are dominated in eye
and mouth regions. These regions are located first and then extracted by cropping
the image. The face image with its features cropped is shown in Fig. 57.8.
DCT is applied to these cropped images. The DCT coefficients are given to
the RBF based neural network which classifies the expressions. Four expressions
namely ‘Happy’, ‘Normal’, ‘Surprise’ and ‘Angry’ are considered for the analysis.

Fig. 57.7 Various facial

expressions

Fig. 57.8 Original image and cropped image

57 Face Recognition and Expression Classification 677

57.6 Results

In this experiment 200 frames of video images with considerable variations in head
poses, expressions, camera viewing angle were used. The ‘face 94’ color image
database was used for still images. One hundred face images were selected with
considerable expression changes and minor variation in head turn, tilt and slant.
The performance ratios were 90% for video image sequence and 92% for still
images. The results of face detection for both still and video images are shown in
Figs. 57.9 and 57.10. Figure 57.11 shows distinct connected components. Rotation
transformation is done by eyes. Figure 57.12 shows the located features. Geomet-
ric parameters and DCT coefficients were given to a classification network for
recognition.
Facial expressions are analyzed using JAFFE face database. Three sets of datasets
from JAFFE database were used. Four expressions namely ‘Happy’, ‘Normal’, ‘Sur-
prise’ and ‘Anger’ were analyzed. The DCT coefficients obtained from the cropped
regions of faces which were very sensitive to facial expressions were given to the
neural network which classifies the expressions with in the face. The average per-
formance ratio is 89.11%. The efficiency plots for three sets of images are shown in
Fig. 57.13.

Fig. 57.9 Face detection from

still images

Fig. 57.10 Face detection from video images

Fig. 57.11 Distinct connected

components
678 V.P. Lekshmi et al.

Fig. 57.12 Located features

Fig. 57.13 Efficiency plot of facial expressions

57.7 Conclusion

In this paper, faces are detected and facial features are located from video and still
images. ‘NRC-IIT Facial video database’ is used as image sequences and ‘face
94 color image database’ is used for still images. Skin pixels and non-skin pixels
are separated and skin region identification is done by RGB color space. From the
extracted skin region, skin pixels are grouped to some meaningful groups to identify
the face region. From the face region, facial features are located using segmentation
technique. Orientation correction is done by using eyes. Parameters like inter eye
distance, nose length, mouth position, and DCT coefficients are computed which is
used for a RBF based neural network. In this experiment only one image sequence
is used for detection of faces.
Facial expressions namely ‘Happy’, ‘Neutral’, ‘Surprise’ and ‘Anger’ are ana-
lyzed using JAFFE face database. Facial features such as eyes and mouth regions
are cropped and these areas are subjected to discrete cosine transformation. The
DCT coefficients are given to a RBF neural network which classifies the facial
expressions.
57 Face Recognition and Expression Classification 679

References

1. Rama Chellappa, Charles L. Wilson and Saad Sirohey, “Human and Machine Recognition of
Faces: A Survey” In Proceedings of the IEEE, Vol. 83, No. 5, 1995, 705–740.
2. Nathan Faggian, Andrew Paplinski, and Tat-Jun Chin, “Face Recognition from Video
Using Active Appearance Model Segmentation”, 18th International Conference on Pattern
Recognition, ICPR 2006, Hong Kong, pp. 287–290.
3. Hichem Sahbi, Nozha, Boueimma, Tistarelli, J. Bigun, A.K. Jain (Eds), and “Biometric
Authentication” LNCS2539, Springer, Berlin/Heidelberg.
4. Marian Stewrt Bartlett, Javier R, Movellan and Terrence J Seinowski, “Face Recognition by
Independent Component Analysis”, IEEE Transactions on Neural Networks, Vol. 113, No. 6,
November 2002.
5. Tsuyoshi Moriyama, Takeo Kanade, Jing Xiao and Jeffrey F. Cohn “Meticulously Detailed
Eye region Model and Its Application to Analysis of Facial Images”, IEEE Transactions in
Pattern Analysis and Machine Intelligence, Vol. 28, No. 5, May 2006.
6. Yan Tong, Weinhui Lio and Qiang Ji, “Facial Action Unit Recognition by Exploiting Their
Dynamic and Semantic Relationships”, IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 29, No. 10, October 2007.
7. Dmitry O. Gorodnichy “Video-based Framework for Face Recognition in Video”, Second
Workshop on Face Processing in Video. (FPiV’05) in Proceedings of Second Canadian Con-
ference on Computer and Robot Vision (CRV’05), Victoria, BC, Canada, 9–11 May, 2005, pp.
330–338.
8. Hong-Bo Deng, Lian-Wen Jin, Li-Xin Zhen and Ian-Cheng Huang, “A New Facial Expression
Recognition Method Based on Local Gabor Filter Bank and PCA plus LDA”, International
Journal of Information Technology, Vol. 11, No. 11, 2005 pp. 86–96.
9. K. Sandeep and A.N. Rajagopalan, “Human Face Detection in Cluttered Color Images Using
Skin Color and Edge Information”. Proc. Indian Conference on Computer Vision, Graphics
and Image Processing, Dec. 2002.
10. P. Peer and F. Solina, “An Automatic Human Face Detection Method”, Proceedings of 4th
Computer Vision Winter Workshop (CVWW’99). Rastenfeld, Australia, 1999, pp. 122–130.
11. Franc Solina, Peter Peer, Borut Batagelj and Samo Juvan,“15 Seconds of Frame-An Interac-
tive Computer Vision Based Art Installation”, Proceedings of 7th International Conference on
Control, Automation, Robotics and Vision (ICARCV 2002), Singapore, 2002, pp. 198–204.
12. Michael J. Lyons, Shigeru Akamatsu, Miyuki Kamachi and Jiro Gyoba,“Coding Facial
Expressions with Gabor Wavelets”, Proceedings, Third IEEE International Conference on
Automatic Face and Gesture Recognition, April 14–16, 1998, Nara Japan, IEEE Computer
Society, pp. 200–205.
Chapter 58
Spiking Neurons and Synaptic Stimuli: Neural
Response Comparison Using Coincidence-Factor

Mayur Sarangdhar and Chandrasekhar Kambhampati

Abstract In this chapter, neural responses are generated by changing the Inter-
Spike-Interval (ISI) of the stimulus. These responses are subsequently compared
and a coincidence factor is obtained. Coincidence-factor, a measure of similarity, is
expected to generate a high value for higher similarity and a low value for dissim-
ilarity. It is observed that these coincidence-factors do not have a consistent trend
over a simulation time window. Also, the lower-bound limit for faithful behaviour
of coincidence factor shifts towards the right with the increase in the reference ISI
of the stimulus. In principle, if two responses have a very high similarity, then their
respective stimuli should be very similar and could possibly be considered the same.
However, as results show, two spike trains generated by highly-varying stim-
uli have a high coincidence-factor. This is due to limitations imposed by the
one-dimensional comparison of coincidence-factor.

Keywords Spiking Neuron  Synaptic Stimuli  Neural Response  Coincidence-

Factor  Inter-Spike-Interval

58.1 Introduction

The responses of a neuron to various types of stimuli have been studied extensively
over the past years [1–9]. Stimulus-dependent behaviour of neurons has already
been pursued to understand the spiking responses and it is thought that either the
firing rate or firing time of individual spikes carries specific information of the neu-
ronal response [3, 10–16]. The response of the neurons studied above has a constant
magnitude whose variance is very low. In this paper, the neural responses fluctuate

M. Sarangdhar (B)
Department of Computer Science, University of Hull, Cottingham Road, Hull,
East-Yorkshire HU6 7RX
E-mail: M.Sarangdhar@2006.hull.ac.uk

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 681
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
682 M. Sarangdhar, C. Kambhampati

and a one-dimensional analysis based on firing times is shown to be insufficient for

comparison.
A supra-threshold static current stimulus is sufficient to induce spiking behaviour
in a neuron. The magnitude of these action potentials is considered to be almost the
same and their variance is thus ignored. Such responses have been studied and mod-
els to depict their spiking behaviour have been proposed and implemented [17, 28].
On the other hand, a synaptic current is used to stimulate the same neuron [3]. This
synaptic stimulus comprises of a static and a pulse component and is of particular
interest as it induces fluctuations in the membrane voltage. These responses can be
compared by their firing times [18,20,23–26] using a measure of comparison known
as coincidence-factor. Here, the generality of this approach is investigated for a
Hodgkin-Huxley (H-H) neuron [29] for which a synaptic current induces membrane
fluctuations.
In this chapter, neural responses are generated by changing the Inter-Spike-
Interval (ISI) of the stimulus. These responses are subsequently compared and
a coincidence factor is obtained. Coincidence-factor, a measure of similarity, is
expected to generate a high value for higher similarity and a low value for dissim-
ilarity. It is observed that these coincidence-factors do not have a consistent trend
over a simulation time window. Also, the lower-bound limit for faithful behaviour
of coincidence factor shifts towards the right with the increase in the reference ISI
of the stimulus. In principle, if two responses have a very high similarity, then their
respective stimuli should be very similar and could possibly be considered the same.
However, as results show, two spike trains generated by highly-varying stimuli have
a high coincidence-factor. This is due to limitations imposed by the one-dimensional
comparison of coincidence-factor. Elsewhere, [30,31] have worked on temporal pat-
terns of neural responses but do not specifically address this issue. Thus, in order
to differentiate spike trains with fluctuating membrane voltages, a two dimensional
analysis is necessary taking both firing time and magnitude of the action potentials.

58.2 Neuronal Model and Synapse

58.2.1 The Neuron Model

The computational model and stimulus for an H-H neuron is replicated from [3].
The differential equations (Eqs. (58.1–58.3)) of the model are the result of non-
linear interactions between the membrane voltage V and the gating variables m, h
and n. for NaC and KC .

dt D gNa m h.V  VNa /  gK n .V  VK /

C dv 3 4
(58.1)
gL .V  VL / C Ii
58 Spiking Neurons and Synaptic Stimuli 683
9
dt D .˛m C ˇm /m C ˛m >
dm
>
>
>
=
dh
D .˛h C ˇh /h C ˛h (58.2)
dt >
>
>
>
;
dn
dt D .˛ n C ˇ n /n C ˛ n

9
˛m D 0:1.V C 40/=Œ1  e .V C40/=10  >
>
>
>
>
>
˛h D 0:07e .V C65/=20 >
>
>
>
>
>
>
>
.V C55/=10 >
˛n D 0:01.V C 55/=Œ1  e >
=
(58.3)
>
>
ˇm D 4e .V C65/=18 >
>
>
>
>
>
>
>
ˇh D 1=Œ1 C e .V C35/=10  >
>
>
>
>
>
;
ˇn D 0:125e .V C65/=80
The variable V is the resting potential where as VNa , VK , and VL are the reversal
potentials of the NaC , KC channels and leakage. VNa D 50 mV, VK D 77 mV
and VL D 54:5 mV. The conductance for the channels are gNa D 120 mS=cm2 ,
gK D 36 mS=cm2 and gL D 0:3 mS=cm2 . The capacitance of the membrane is
C D 1 μ F=cm2 .

58.2.2 The Synaptic Current

An input spike train described in (Eq. (58.4)) is used to generate the pulse compo-
nent of the external current.
X
Ui .t/ D Va ı.t  tf / (58.4)
n

where, tf is the firing time and is defined as

tf.nC1/ D tf.n/ C T (58.5)

tf.1/ D 0 (58.6)
T represents the ISI of the input spike train and can be varied to generate a different
pulse component. The spike train is injected through a synapse to give the pulse
current IP . X
IP D gsyn ˛.t  tf /.Va  Vsyn / (58.7)
n
684 M. Sarangdhar, C. Kambhampati

gsyn , Vsyn are the conductance and reversal potential of the synapse. The α-function
is defined in [32] as
˛.t/ D .t=/e t = ‚.t/; (58.8)
where,  is the time constant of the synapse and ‚.t/ is the Heaviside step function.
V D 30 mV,  D 2 ms, gsyn D 0:5 mS=cm2 and Vsyn D 50 mV.

58.2.3 The Total External Current

The total external current applied to the neuron is a combination of static and pulse
component
I i D IS C IP C " (58.9)
where, IS is the static and IP is the pulse current, " is the random Gaussian noise
with zero mean and standard deviation  D 0:025. [3] has ignored the noise in the
external current and the current consists of only two terms. However, the presence
of noise is necessary in the simulation of a biological activity and hence considered.

58.3 Comparison of Two Spike Trains

58.3.1 Responses of the Neuron

The static component IS of the external current is set at 25 μ A. The H-H neuron is
stimulated with a current Ii D IS C IP and its response is recorded. The fluctuations
in the membrane are due to the specific nature of synaptic stimulus. The amplitude
of the action potential in Fig. 58.1 is not constant and the standard deviation is
Amp D 3:0978. Hence, the amplitude of the response is not ignored. This is one
major difference between [3, 30, 31] and this work. The synaptic time constant of
2 ms defines the shape of the pulse current. As the refractory period of an H-H
neuron is 1–2 ms, we choose a 2 ms bound for coincidence detection. The simulation
activity is divided into three sets of ISIs. Each set has a corresponding reference
ISI (Tref /. The first set compares responses generated using stimulus ISI between
14–16 ms while the second set compares responses of ISIs between 13–15 ms. The
third set compares responses for ISIs varied between 15–17 ms. The responses for
each set are compared with a fixed response known as the reference response. The
reference response for each set is unique and is generated by varying the stimulus
ISI. Reference ISIs for the sets are 15, 14 and 16 ms respectively. Neural responses
are recorded for various ISIs within a set and compared with the reference response
for that set. For set 1, the reference spike train is generated with T D 15 ms (Tref /
and compared with responses generated with T D 14–16 ms. Coincidence factors
are calculated to estimate the similarity between these responses.
58 Spiking Neurons and Synaptic Stimuli 685

Fig. 58.1 Response of the H-H neuron to with T D 15 ms causing fluctuations in membrane
voltage. (a) The synaptic spike train input that induces a pulse current. (b)The pulse current gener-
ated. (c) The total external current applied to the neuron. Note that there is a static offset. (d) The
neuronal response to the current

25
Spike points

20

15
10
50 100 150 200 250 300
( a ) T ime T=15 ms
50 T=16 ms
Voltage

-50

-100
50 100 150 200 250 300
( b ) T ime

Fig. 58.2 Comparison of responses. (a) The corresponding magnitude of spikes for the responses
at T D 16 ms and T D 15 ms. (b) The two spike trains not only differ in firing times but also in
magnitudes

58.3.2 Comparison of Responses

The response of the neuron is specific to an input stimulus. In order to generate

different stimuli, we varied the ISI of the synaptic input from T D 14–16 ms with
T D 15 ms as the reference ISI. Figures 58.2 and 58.3 show that the response of the
neuron differs with respect to both firing time and magnitude. The figures indicate
that the variation in the input ISI causes the membrane voltage to fluctuate. They also
686 M. Sarangdhar, C. Kambhampati

25

Spike points 20

15

10
50 100 150 200 250 300
( a ) T ime T=15 ms
50 T=14 ms
Voltage

-50

-100
50 100 150 200 250 300
( b ) T ime

Fig. 58.3 Comparison of responses. (a) The corresponding magnitude of spikes for the responses
at T D 14 ms and T D 15 ms. (b) The two spike trains not only differ in firing times but also in
magnitudes

show the difference in responses generated with T D 14 ms & Tref and T D 16 ms

& Tref :

58.3.3 Coincidence-Factor

The coincidence-factor, as described by [18, 20] is 1 only if the two spike trains
are exactly the same and 0 if they are very dissimilar. Coincidence for an individ-
ual spike is established if its firing time is within 2 ms of the firing time of the
corresponding spike in the reference spike train (in this case T D 15 ms).

Ncoinc  hNcoinc i 1
D (58.10)
1=2.N1 C N2 / N
where, N1 is the number of spikes in the reference train, N2 is the number of spikes
in the train to be compared, Ncoi nc is the number of coincidences with a precision
ı D 2ms between the spike trains.hNcoi nc i D 2ıN1 is the expected number of
coincidences generated by a homogeneous Poisson process with the same rate as
the spike train to be compared. N D 1  2ı is the normalising factor. For set
1, N1 is the number of spikes in the reference spike train (Tref D 15 ms) and N2
is the number of spikes in the train to be compared (T D 14–16 ms). Figure 58.4
shows that the coincidence-factors for responses generated using T D 14–16 ms do
not follow a fixed pattern. The coincidence-factor () is expectedly 1 when spike
train generated with T D 15 ms is compared with the reference spike train Tref (T D
15 ms). However, the coincidence factor for spike trains generated at T D 16 ms and
Tref is 1. This indicates that the two highly varying currents have an exactly similar
58 Spiking Neurons and Synaptic Stimuli 687

1.2 Reference, T=15ms

Concidence factor =1
1
Coincidence factor

0.8

Expected
0.6 decrease in the
coincidence
factor with
0.4 increase in ISI
from the
Inconsistent result reference spike Inconsistent result
train
0.2

0
14 14.2 14.4 14.6 14.8 15 15.2 15.4 15.6 15.8 16
Inter-Spike-Interval

Fig. 58.4 Coincidence-factor versus ISI. The coincidence-factor decreases expectedly between
T D 15–14:65 ms and T D 15–15:25 ms. At other times the result is inconsistent and does not
have a fixed pattern

response or conversely as the responses are same; the two input stimuli are similar,
which is an incorrect inference. The coincidence factor for the spike trains generated
at T D 14 ms and Tref is 0.1207 indicating very low similarity. From a mathematical
and signal transmission standpoint, the coincidence-factor should decrease as the
input stimulus increasingly varies from Tref . However, this can only be observed
between T D 14:65–15:25 ms (30% of the 2 ms time window). The coincidence-
factor  increases from T D 14–14:5 ms but then drops till T D 14:65 ms. 
steadily increases to 1 when T D 15 ms and drops for 0.25 ms. There is an upward
rise from T D 15:25–15:5 ms, a sharp drop from T D 15:5–15:75 ms followed by a
steep increase to  D 1 at T D 16 ms. Traversing from the reference the expected
trajectory of the coincidence-factor breaks at T D 14:65 ms and T D 15:25 ms.
These are therefore taken as limits for faithful behaviour of the coincidence-
factor approach. However, for set 2 reference spike train is chosen as Tref D 14 ms,
limits of faithful behaviour change (Fig. 58.5, left). The coincidence factor steadily
rises to unity, stays there for 0.5 ms and drops gradually. Ideally, the coincidence-
factor should be not 1 for T D 13:5, 13.65 and 13.75. While in set 3, Fig. 58.5, right,
reference spike train chosen is at Tref D 16 ms. The limits of faithful behaviour
change with a change in the stimulus. There is a sharp rise in the coincidence factor
from 15.75 to 16 ms where it reaches unity. From 16 to 17 ms the coincidence-
factor executes a perfect curve as expected. From Figs. 58.4 to 58.6 it is conclusive
that the lower-bound of faithful behaviour increases with the increase in the input
reference ISI. The difference between the reference ISI (Tref / and the lower-bound
limit decreases with the increase in the reference ISI. It is also important to note
that within each set of simulation, there are some false coincidences. The term false
coincidence is used to identify comparisons whose coincidence factor is 1 – when
it should not be. In Fig. 58.4, there is a false coincidence when ISI D 16 ms is
688 M. Sarangdhar, C. Kambhampati

Reference Reference
T=14ms T=16ms
1 1

0.8 0.8

Coincidence Factor
Coincidence Factor

0.6 0.6

0.4 0.4

0.2 0.2
Faithful behaviour
Faithful behaviour
0 0

13 13.2 13.4 13.6 13.8 14 14.2 14.4 14.6 14.8 15 15 15.2 15.4 15.6 15.8 16 16.2 16.4 16.6 16.8 17

Inter-Spike-Interval Inter-Spike-Interval

Fig. 58.5 Coincidence-factor versus ISI. Left – The coincidence-factor has a faithful behaviour
between T D 13:15 ms and T D 14:65 ms. Right – The coincidence-factor has a faithful
behaviour between T D 15:75–17 ms. It executes a perfect curve after 16 ms

compared with Tref D 15 ms. In Fig. 58.5, left, false coincidences can be seen when
ISI varied between 13.5–13.75 ms is compared with Tref D 14 ms while in Fig. 58.5,
right, false coincidences can be observed for ISI varied between 15 and 15.15 ms and
compared with Tref D 16 ms.

58.3.4 Two-Dimensional Analysis

The coincidence-factors over the 2 ms time window show an inconsistent trend. A

one-dimensional approach of the coincidence-factor determination is thought to be
the cause of this inconsistency. The coincidence-factor is highly accurate for spike
trains with a constant amplitude response however; the coincidence-factor does not
give a proper estimate of similarity between two spike trains with varying ampli-
tudes. As a result, two visually distinct spike trains would still generate a high
coincidence-factor (Figs. 58.2 and 58.3). A two-dimensional analysis of spike trains
with fluctuating magnitudes can resolve this inconsistency. To support this, a simple
binary clustering algorithm is used. It shows that the clustering solution for each
response is unique to itself and therefore helps to eliminate any ambiguity.

58.3.5 Binary Clustering

The peak of each spike in a spike train is considered as an object. The number of
objects for each spike train is equal to the number of spikes.

Obj D ŒFiringtime; Amplitude (58.11)

58 Spiking Neurons and Synaptic Stimuli 689

2
drs D .Nr  Ns /.Nr  Ns /0 (58.12)
The Euclidean distance between object-pairs is calculated using (Eq. (58.12)) where,
Nr , NS are the objects in the spike train. Once the distance between each pair
of objects is determined, the objects are clustered based on the nearest neighbour
approach using
d.r; s/ D min.d i st.Nri  Nsj //
(58.13)
i 2 .1; :::; nr /; j 2 .1; :::; ns /
where nr ; ns is the total number of objects in the respective clusters. The binary clus-
ters are plotted to form a hierarchical tree whose vertical links indicate the distance
between two objects linked to form a cluster. A number is assigned to each cluster
as soon as it is formed. Numbering starts from (m C 1), where m = initial number
of objects, till no more clusters can be formed. We investigated the case described
in Section 58.3.3 for the response generated at Tref D 15 ms and T D 16 ms
(false coincidence). The coincidence-factor for these responses is 1 (Fig. 58.4)
and indicates an exact match. The clustering tree shows that these responses are
actually different from each other by a margin not captured by coincidence-factor
(Fig. 58.6a, b). The clustered objects are shown on the X-axis and the distance
between them is shown on the Y-axis. A comparison of the clustering solutions
shows that the shape, form, height as well as linkages are different for the two spike
trains. In Fig. 58.6a, objects 12 and 13 are clustered together at a height of 11.5
while in Fig. 58.6b, objects 11 and 12 are clustered at a height of 13.5 – shown
in green circles. Also, objects 4 and 5 are clustered in Fig. 58.6a while objects
3 and 4 are clustered in Fig. 58.6b – shown in red circles. This means that the
spike trains are inherently different from each other. The results hence prove that
the two spike trains are not an exact match. We therefore believe that though deter-
mining coincidence-factor is important, a two-dimensional analysis is necessary for
responses with fluctuating membrane voltages.

15
14
14

13
13
Distance between objects

Distance between objects

12
12

11
11

10
10

9 9

17 18 19 20 21 14 15 16 22 23 24 12 13 1 2 3 4 5 6 7 8 9 10 11 23 24 22 19 20 21 16 17 18 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12
Objects being clustered Objects being clustered
(a) ISI = 15 ms (b) ISI = 16 ms

Fig. 58.6 (a) Clustering solution for T D 15 ms indicating objects being clustered. (b) Clustering
solution for T D 16 ms indicating objects being clustered. The shape, form, height as well as
linkages for each spike train are different
690 M. Sarangdhar, C. Kambhampati

58.4 Conclusions

A synaptic stimulus known to induce fluctuations in the membrane voltage is used

to stimulate an H-H neuron [3] to verify if firing time alone is enough to differ-
entiate between these responses. The time constant of the pulse component of the
external current is 2 ms and due to refractoriness of the neuron, coincidence-bound
is also chosen as 2 ms. The coincidence-factors are calculated for time windows
t1 D 14–16 ms, t2 D 13–15 ms and t3 D 15–17 ms with reference spike trains at
T D 15, 14 and 16 ms respectively. In all three sets of results, there is no consis-
tent trend exhibited by the coincidence-factor. Also, the limits of faithful behaviour
change and the percentage of acceptable results varies. The percentage of faithful
behaviour for the three time windows is 30, 75 and 62.5 respectively. It is observed
that: (a) the limits of faithful behaviour change with a change in the reference ISI. (b)
The lower-bound limit of faithful behaviour increases with the increase of the ref-
erence ISI. (c) The difference between the reference ISI and the lower-bound limit
of faithful behaviour decreases with the increase in the reference ISI. (d) In order
to differentiate between these responses accurately, a two-dimensional analysis is
required. A simple clustering algorithm easily differentiates two visually-distinct
responses as opposed to the coincidence-factor. It effectively demonstrates the
necessity of a two-dimensional approach to comparison of neural responses [33,34].
Clustering is primarily used to demonstrate the requirement of a two-dimensional
analysis for comparing spike trains with fluctuating membrane voltages. We take
this as a supporting claim for our future work.

References

1. Lundström I (1974). Mechanical wave propagation on nerve axons. Journal of Theoretical

Biology, 45: 487–499.
2. Abbott L F, Kepler T B (1990). Model neurons: From Hodgkin Huxley to Hopfield. Statistical
Mechanics of Neural Networks, Edited by Garrido L, pp. 5–18.
3. Hasegawa H (2000). Responses of a Hodgkin-Huxley neuron to various types of spike-train
inputs. Physical Review E, 61(1): 718.
4. Kepecs A, Lisman J (2003). Information encoding and computation with spikes and bursts.
Network: Computation in Neural Systems, 14: 103–118.
5. Fourcaud-Trocmé N, Hansel D, van Vreeswijk C, Brunel N (2003). How spike genera-
tion mechanisms determine the neuronal response to fluctuating inputs. The Journal of
Neuroscience, 23(37): 11628–11640.
6. Bokil H S, Pesaran B, Andersen R A, Mitra P P (2006). A method for detection and clas-
sification of events in neural activity. IEEE Transactions on Biomedical Engineering, 53(8):
1678–1687.
7. Davies R M, Gerstein G L, Baker S N (2006). Measurement of time-dependent changes in the
irregularity of neural spiking. Journal of Neurophysiology, 96: 906–918.
8. Diba K, Koch C, Segev I (2006). Spike propagation in dendrites with stochastic ion channels.
Journal of Computational Neuroscience, 20: 77–84.
9. Dimitrov A G, Gedeon T (2006). Effects of stimulus transformations on estimates of sensory
neuron selectivity. Journal of Computational Neuroscience, 20: 265–283.
58 Spiking Neurons and Synaptic Stimuli 691

10. Rinzel J (1985). Excitation dynamics: Insights from simplified membrane models. Theoretical
Trends in Neuroscience Federal Proceedings, 44(15): 2944–2946.
11. Gabbiani F, Metzner W (1999). Encoding and processing of sensory information in neuronal
spike trains. The Journal of Biology, 202: 1267–1279.
12. Panzeri S, Schultz S R, Treves A, Rolls E T (1999). Correlations and the encoding of infor-
mation in the nervous system. Proceedings of the Royal Society of London, B 266(1423):
1001–1012.
13. Agüera y Arcas B, Fairhall A L (2003). What causes a neuron to spike? Neural Computation,
15: 1789–1807.
14. Agüera y Arcas B, Fairhall A L, Bialek W (2003). Computation in a single neuron: Hodgkin
and Huxley revisited. Neural Computation, 15: 1715–1749.
15. Izhikevich E M (2006). Polychronization: Computation with spikes. Neural Computation, 18:
245–282.
16. Li X, Ascoli G A (2006). Computational simulation of the input-output relationship in
hippocampal pyramidal cells. Journal of Computational. Neuroscience, 21: 191–209.
17. Kepler T B, Abbott L F, Marder E (1992). Reduction of conductance-based neuron models.
Biological Cybernetics, 66: 381–387.
18. Joeken S, Schwegler H (1995). Predicting spike train responses of neuron models; in
M.Verleysen (ed.), Proceedings of the 3rd European Symposium on Artificial Neural Net-
works, pp. 93–98.
19. Wang X J, Buzsáki G (1996). Gamma oscillation by synaptic inhibition in a hippocampal
interneuronal network model. The Journal of Neuroscience, 16(20): 6402–6413.
20. Kistler W M, Gerstner W, Leo van Hemmen J (1997). Reduction of the Hodgkin-Huxley
equations to a single-variable threshold model. Neural Computation, 9: 1015–1045.
21. Izhikevich E M (2003). Simple model of spiking neurons. IEEE Transactions on Neural
Networks, 14(6): 1569–1572.
22. Shriki O, Hansel D, Sompolinsky H (2003). Rate models for conductance-based cortical
neuronal networks. Neural Computation, 15: 1809–1841.
23. Jolivet R, Gerstner W (2004). Predicting spike times of a detailed conductance-based neuron
model driven by stochastic spike arrival. Journal of Physiology – Paris, 98: 442–451.
24. Jolivet R, Lewis T J, Gerstner W (2004). Generalized integrate-and-fire models of neuronal
activity approximate spike trains of a detailed model to a high degree of accuracy. Journal of
Neurophysiology, 92: 959–976.
25. Jolivet R, Rauch A, Lüscher H-R, Gerstner W (2006). Integrate-and-fire models with adap-
tation are good enough: Predicting spike times under random current injection. Advances in
Neural Information Processing Systems, 18: 595–602.
26. Jolivet R, Rauch A, Lüscher H-R, Gerstner W (2006). Predicting spike timing of neocortical
pyramidal neurons by simple threshold models. Journal of Computational Neuroscience, 21:
35–49.
27. Clopath C, Jolivet R, Rauch A, Lüscher H-R, Gerstner W (2007). Predicting neuronal activity
with simple models of the threshold type: Adaptive exponential integrate-and-fire model with
two compartments. Neurocomputing, 70: 1668–1673.
28. Djabella K, Sorine M (2007). Reduction of a cardiac pacemaker cell model using singular
perturbation theory. Proceedings of the European Control Conference 2007, Kos, Greece, pp.
3740–3746.
29. Hodgkin A, Huxley A (1952). A quantitative description of membrane current and its
application to conduction and excitation in nerve. Journal of Physiology, 117:500–544.
30. Maršálek, P (2000). Coincidence detection in the Hodgkin–Huxley equations. Biosystems,
58(1–3).
31. Victor J D, Purpura K P (1997). Metric-space analysis of spike trains: Theory, algorithms and
application. Network: Computation in Neural Systems, 8: 127–164.
32. Park M H, Kim S (1996). Analysis of phase models for two coupled Hodgkin-Huxley neurons.
Journal of the Korean Physical Society, 29(1): 9–16.
692 M. Sarangdhar, C. Kambhampati

33. Sarangdhar M, Kambhampati C (2008). Spiking neurons: Is coincidence-factor enough to

compare responses with fluctuating membrane voltage? In World Congress on Engineering
2008: The 2008 International Conference of Systems Biology and Bioengineering, London,
UK, 2–4 July 2008, Vol. 2, pp. 1640–1645.
34. Sarangdhar M, Kambhampati C (2008). Spiking neurons and synaptic stimuli: Determining
the fidelity of coincidence-factor in neural response comparison. Special Issue of IAENG
Journal (in print).
Chapter 59
Overcoming Neuro-Muscular Arm Impairment
by Means of Passive Devices

Federico Casolo, Simone Cinquemani, and Matteo Cocetta

Abstract The present research was originated by the request of an organization of

disables affected by muscular distrophy (UILDM) for the development of a passive
system. Most of UILDM affiliated in fact, up to now, prefer to avoid any motor
driven device except for the electrical wheelchair on which they are forced to live.
Therefore, aim of the first part of the research is the development of a passive
device as simple as possible, capable to enhance the upper limb mobility by acting
against the gravity action. The device must be mounted on the frame of the subjects’
wheelchair. Until now we designed and used only passive systems for preliminary
analyses of patients response. These results are the basis for the development of
our next passive device that will be optionally provided of an active module for the
weaker subjects.

Keywords Neuro-Muscular Arm Impairment  Passive Device  muscular

dystrophy

59.1 Introduction

Some degenerative neuromuscular diseases, such as dystrophy, affect muscles

strength and force the subjects, few years after the appearance of the symptoms,
to use electric wheelchairs. In fact, not only the legs muscles become too weak to
sustain the upper body weight during walking, but also the upper limbs become
soon inadequate to power an ordinary wheelchair. At a certain stage of the evolution
of the pathology, due to the lack of muscular strength, the upper limbs are not able
to counterbalance the gravity action, thus the subjects can only act, at the most, by

F. Casolo (B)
Politecnico di Milano, Dipartimento di Meccanica – Campus Bovisa Sud, via La Masa 34,
20156 Milano, Italy
E-mail: federico.casolo@polimi.it

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 693
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
694 F. Casolo et al.

moving the arm horizontally when it is supported by a plane. The forced immobil-
ity produces a faster degeneration of the muscular structure. The recovery of some
active mobility of the arm is then important both to carry out some autonomous
daily activity and to execute exercises as self physiotherapy. Assistive systems may
help the subject in this task: they can be active, that means in general motor driven,
or passive. Some active systems require to be driven by the controlateral limb –
e.g. by replicating its movement or by means of a joystick – producing unnatural
movements [1, 2]. The present research was originated by the request of an organi-
zation of disables affected by muscular distrophy (UILDM) for the development of
a passive system. Most of UILDM affiliated in fact, up to now, prefer to avoid any
motor driven device except for the electrical wheelchair on which they are forced to
live. Therefore, aim of the first part of the research is the development of a passive
device as simple as possible, capable to enhance the upper limb mobility by acting
against the gravity action. The device must be mounted on the frame of the subjects’
wheelchair. Until now we designed and used only passive systems for preliminary
analyses of patients response. These results are the basis for the development of
our next passive device that will be optionally provided of an active module for the
weaker subjects.

59.2 Background

Presently some mechatronic devices are designed to increase subject physical abil-
ities [3–7], but only few of them are suitable for subjects affected by muscular
dystrophy.
In [8], the aim of the research is to increase the autonomy of people affected
by muscular dystrophy and spinal muscular atrophy through a passive device. The
system has four degrees of freedom: flexion extension of the shoulder, abduction
adduction of the shoulder, flexion extension of the elbow and prono-supination of
the forearm. The torque value at the joint is a non linear function of position, how-
ever gravity balance is achieved by means of linear springs (Fig. 59.1). To obtain

Fig. 59.1 (a) Rahman device – (b) schema of the static balance of a one d.o.f. structure with linear
spring
59 Overcoming Neuro-Muscular Arm Impairment by Means of Passive Devices 695

Fig. 59.2 Arm model

Fig. 59.3 DAS balancing

system

the system balance regardless the value of , the spring stiffness value is set to:
k D mgl=.ab/.
Armon orthosis [9, 10] is designed for people affected by spinal muscular atro-
phy. It assumes that arm weight can be sustained by natural shoulder and elbow
(Fig. 59.2). Static balance is obtained by springs for about 75% of the arm weight,
the interface between the device and the arm is located in the forearm, near the
elbow. The system has two degrees of freedom in the vertical plane; the spring
action is transmitted to the frame by a system of pulleys.
In [11], the Dynamic Arm Support (DAS) is designed to be mounted on the
wheelchair of people affected by muscular dystrophy, muscular atrophy and amy-
otrophic lateral sclerosis. It’s an hybrid device with an active operator-controlled
electrical system to adjust weight compensation. Patient’s forearm is balanced by a
gravity compensation system, made by a planar linkage (Fig. 59.3) through a linear
spring. The balance is guaranteed by equation: ra ak D mgl.

59.3 Experimental Test Apparatus Set Up

Starting from a kinematic model of the human arm [12] with seven degrees of free-
dom – five d.o.f. for shoulder and two d.o.f. for elbow – a linkage system was
designed to be coupled in parallel with the arm, avoiding to add constraints to the
scapula-clavicle complex (Fig. 59.4).
696 F. Casolo et al.

Fig. 59.4 First prototype mounted on a wheelchair

Fig. 59.5 Simplified scheme of loading actions on the arm

Aim of the system is to evaluate experimentally the gain of performance in arm

motion due to the help of an anti gravity device. With this approach the device
directly acts on the elbow (force F2 on Fig. 59.5), providing the counterweight force
and couple for balance, leaving subject’s shoulder and trunk free to move. Thus the
residual force becomes adequate to move the weight-compensated arm in its natural
working volume, without any limitation of the motion due to the external device.
The linkage includes two sliding pairs (for horizontal and vertical translations) and
one revolute pair. Weight compensation at the elbow is achieved by a spiral spring
at the revolute pair for the forearm flexo-extension and by a counterbalance moving
in vertical to compensate the whole arm weight. In order to balance the system,
forces and couples have been preliminary evaluated versus the arm position in space
neglecting the inertia actions [13], which are small for the range of motion typical
of this application.
The counterbalancing force F2 and the counterbalancing torque MC can be
simply obtained as follow:
59 Overcoming Neuro-Muscular Arm Impairment by Means of Passive Devices 697

Fig. 59.6 Torque at the elbow

X  
AC  AB
FV D 0 ! F2 D 1  F1 C F3 C F4 (59.1)
AC
X
MO D 0 ! MC D F3 CD sin ./ C F4 CF sin ./ (59.2)

The forearm counterbalancing torque (Fig. 59.6 – line A) is only function of the
elbow angle g, otherwise counterbalance weight is independent from the arm
position and both of them are functions of the subject weight.
For the prototype a spiral spring has been designed to follow approximately the
torque diagram and the vertical guide has been equipped with masses counterbalanc-
ing, through a cable and a pulley, the subject arm and the sliding structure weight.
Line C in Fig. 59.6 shows the torque exerted by the spring and C1 represents the
residual torques that the subject must exert to move the forearm. Lines B and B1 are
related to a preloaded spring.
While counterweights can be easily adjusted on patients’ weight and skills,
spring action can only be tuned by varying the preload thanks to its adjustable sup-
port. Major variations of the torque can only be obtained by changing the spring.
Preliminary tests allow to adjust the spring preload for each specific subject.

59.4 Experimental Tests

The prototype has been tested with some patients to evaluate its effectiveness for
helping them to move the arm and to handle simple objects for daily tasks (e.g.
drinking). The analyzed parameters are the changes produced by the external device
to the joints excursions and to the trajectories followed by the hand to complete a
task. The test protocol gives a description of the exercises performed by patients
without any initial training. Every task is executed three times, starting from sim-
ple movements of a single joint, to more complex exercises (as to take a glass of
water from a table and bring it to the mouth). Markers on upper body (arm and
trunk) are acquired by an infrared motion capture system together with the EMG
signals of the most important muscles. Therefore the system kinematics could be
698 F. Casolo et al.

reconstructed and correlated with the timing of the muscular activity. For the first
series of tests a performance gain index was also extrapolated by comparing the
results with the preliminary tests obtained without using the assistive device. The
four tests performed by all the subjects are listed in Table 59.1.
For example, Fig. 59.7 shows the data processed during the execution of move-
ment n.2 and Fig. 59.8 highlights the increase of the range of motion of the subject’s
elbow produced by the helping device. Figure 59.9 shows how the same device is
also helpful during the drinking task to reduce the lateral flexion of the trunk.
Overall tests results clearly show that an anti-gravity device can increase the
autonomy of people affected by M.D. The subject’s motion skills increase test after
test and, moreover, in most cases the trunk posture improves.

Table 59.1 Test protocol

Target movement Involved muscles
1 Shoulder rotation to bring the forearm to the Deltoid, pectorals et al.
chest starting from a position where the elbow
angle is 90ı
2 Elbow flexo-extension Biceps, triceps
3 Shoulder abdo-adduction Deltoid, pectorals, triceps
4 Drinking: picking a glass it from a table in front Deltoid, pectorals et al.
of the subject

Fig. 59.7 Joints excursion with the helping device (movement 2)

59 Overcoming Neuro-Muscular Arm Impairment by Means of Passive Devices 699

Fig. 59.8 Comparison of elbow max excursion

Fig. 59.9 Comparison of trunk bending (average values of one subject) for drinking task with the
device (1) and without (2); areas show the st. deviat

59.5 Device Evolution

The prototype design is developed in order to obtain a better structure, statically

balanced in every position, with a better interface with the operator and lower
dimensions.
Two main solutions have been analyzed [14]: one with counterbalancing masses
and the other with springs. Both structures support two degrees of freedom and
provide static weight balance of the masses m1 and m2 of the arm and of the forearm
in a vertical plane (Fig. 59.10).
The four-bars linkage provides weight compensation for every position of the
structure (α and ˇ angles). The second solution overcomes the problems due to
counterweights, reducing the total weight of the device and its inertia. Relationships
700 F. Casolo et al.

Fig. 59.10 Structures with counterweights and springs

Fig. 59.11 Arm and forearm segments

Fig. 59.12 Counterbalance

between masses, lengths and springs stiffness guarantee static equilibrium of the
structure in every position.
Arm and forearm segments (respectively OA and AE in Fig. 59.11) are balanced
for each value of α and ˇ. For segment AE:

m2 gAEcosˇ C FBC BA sin ˛ cos ˇ C FBC BA sin ˇ cos ˛ D 0 (59.3)

For the counterbalance segment (Fig. 59.12) equilibrium:

M2 gOD cos ˇ C FBC OC sin ˛ cos ˇ C FBC OC sin ˇ cos ˛ D 0 (59.4)

From Eqs. (59.3) and (59.9):

m2 AE D M2 OD (59.5)

Analogously for balancing OA segment:

m1gOGcos˛  M1 gOF cos ˛ C FyA OAcos˛ C FxA OA sin ˛ D 0 (59.6)

59 Overcoming Neuro-Muscular Arm Impairment by Means of Passive Devices 701

Fig. 59.13 Spring actions

From which:
m1 OG C m2 OA D M1 OF (59.7)
Equations (59.5) and (59.7) describe the static equilibrium of the structure; two
values among M1 , M2 , OF and OD (directly related to structure dimensions and
weight) can be arbitrarily chosen while the two others result from equations.
The equilibrium equations for the structure with springs can be achieved from
analogous considerations.
Following notations on Fig. 59.13:

F1y OD cos ˇ C F1x OD sin ˇ C FBC OC sin ˛ cos ˇ C FBC OC cos ˛ sin ˇ D 0
(59.8)
Where F1x and F1y are the components of elastic force: F1x D k1 ODcosˇ and
F1y D k1 .OM  ODsi nˇ/. From Eq. (59.8), for the equilibrium:

m2 gAE D k1 OMOD (59.9)

P
k2 stiffness can be evaluated from MO D 0:

m1 gOG cos ˛  F2y OF cos ˛ C F2x OF sin ˛ C FyA OA cos ˛ C FxA OA sin ˛ D 0
(59.10)
With F2x D k2 OF cos ˛ and F2y D k2 .OL C OF / sin ˛.
Thus for OA segment:

.m1 OG C m2 OA/ g D k2 OFOL (59.11)

Like the other configuration of the device, there are four unknown parameters
(OM, OL, k1 , k2 /, but only two of them are independent. Therefore, if the system
geometry is constrained, it is enough to choose the spring stiffness and vice versa.

59.5.1 Simulations

To better evaluate and choose between the planned solutions, a multibody model of
the human arm coupled with the device has been realized and torques in shoulder
and elbow joints have been calculated. Inertia actions have been taken into account.
702 F. Casolo et al.

Simulations were carried out for both the new structures to calculate torques on
shoulder and elbow joints. The anthropometric data (arm and forearm weights and
centers of mass) come from literature [15].
Simulations was completed for each of the following configurations:
Non balanced structure (free arm)
Weight balanced structure
Spring balanced structure
The torques supplied by the subjects have been evaluated for the following these
different motion tasks:
1. Static initial position: ˛ D 60ı , ˇ D 0ı , no motion
2. Motion 1: ˛P D 15ı =s, ˇ D 0ı
3. Motion 2: ˛P D 0ı =s, ˇP D 10ı =s
4. Motion 3: Arm flexion ˛P D 30ı=s
Joint torque values with the balancing apparatus (either by weights or by springs) are
lower than ones in unbalanced conditions, as confirmation of the device efficiency.
Since in practice it is very hard to build a structure whose joints exactly remain
overlapped to the arm articulations, the effect of the misalignment has also been
investigated.
To evaluate how the torques exerted by shoulder and elbow can change, pre-
vious simulations were repeated with the new configurations where both human
joints position have an offset with respect to the ideal one. Figure 59.14 shows
the torques required during motion 2, starting for both joints with the same offset
(30 mm horizontally and 70 mm vertically): although the new torques are higher
than the ones evaluated in the ideal configuration (Figs. 59.14 and 59.15), they still
remain acceptable.

59.6 Conclusions

Gravity counterbalancing systems can be profitable for increasing autonomous

mobility of MD subjects. Also the built and tested simple preliminary device demon-
strated to be able to recover the patient ability of performing some activities which
are impossible without assistance. All the tested subjects demonstrated their satis-
faction. The preliminary results are encouraging for the evolution of the passive
counterbalancing systems optimized on subject anatomy and skills. The passive
devices can help the subjects with enough residual muscular force, while for the late
phases of MD an active system is required. The next step of the research requires a
deeper analysis of the action requested to the muscles for the selected task, in order
to exploit the muscles less weak because of MD, by means of specialized models.
A new prototype is under development and in future it will include active modules.
Active systems design will take advantage by the studies for the evolution of passive
ones.
59 Overcoming Neuro-Muscular Arm Impairment by Means of Passive Devices 703

Fig. 59.14 Motion 2 – shoulder (up) and elbow (down) torques unbalanced and with weight and
spring structure

Fig. 59.15 Arm placed with joint off-set – influence on joints torques
704 F. Casolo et al.

References

1. Perry, J.C., Rosen, J., and Burns, S., “Upper-Limb Powered Exoskeleton Design”, Mechatron-
ics, IEEE/ASME Transactions on, vol. 12, no. 4, pp. 408–417, Aug. 2007.
2. Kiguchi, K., Tanaka, T., Watanabe, K., and Fukuda, T., “Exoskeleton for Human Upper-
Limb Motion Support”, Robotics and Automation, 2003, Proceedings. ICRA ‘03, IEEE
International Conference on, vol. 2, pp. 2206–2211, 14–19 Sept. 2003.
3. Kazerooni, H., “Exoskeletons for Human Power Augmentation”, Proceedings of the
IEEE/RSJ. International Conference on Intelligent Robots and Systems, pp. 3120–3125.
4. Kazuo K., Takakazu T., Keigo W., and Toshio F., “Exoskeleton for Human Upper-Limb
Motion Support”, Proceedings of the 2003 IEEE International Conference on Roboticsn&
Automation Taipei, 2003.
5. Agrawal, S.K., Banala, S.K., Fattah, A., Sangwan, V., Krishnamoorthy, V., Scholz, J.P., and
Hsu, W.L.,“Assessment of Motion of a Swing Leg and Gait Rehabilitation With a Gravity Bal-
ancing Exoskeleton”, IEEE Transactions on Neural Systems and Rehabilitation Engineering,
2007.
6. Ball, S.J., Brown, I.E., and Scott, S.H., “A Planar 3DOF Robotic Exoskeleton for Rehabilita-
tion and Assessment”, Proceedings of the 29th Annual International Conference of the IEEE
EMBS, Lyon, France, pp. 4024–4027, 23–26 Aug. 2007.
7. Kiguki, K. and Tanaka, T. and Watanabe, K. and Fukuda, T., “Exoskeleton for Human Upper-
Limb Motion Support”, Proceedings of the 2003 IEEE 10th International Conference on
Robotics n& Automation, Taipei, Taiwan, pp. 2206–2211, 14–19 Sept.
8. Rahman T., Ramanathan R., Stroud S., Sample W., Seliktar R., Harwin W., Alexander M.,
and Scavina M., “Towards the Control of a Powered Orthosis for People with Muscular
Dystrophy”, IMechE Journal of Engineering in Medicine, vol. 215, Part H, pp. 267–274, 2001.
9. Herder, L., “Development of a Statically Balanced Arm Support: ARMON”, Proceedings of
the 2005 IEEE 9th International Conference on Rehabilitation Robotics, Chicago, IL, USA,
pp. 281–286, June 28–July 1, 2005.
10. Mastenbroek, B., De Haan, E., Van Den Berg, M., and Herder, J.L., “Development of a Mobile
Arm Support (Armon): Design Evolution and Preliminary User Experience”, Proceedings of
the 2007 IEEE 10th International Conference on Rehabilitation Robotics, Noordwijk, The
Netherlands, pp. 1114–1120, 12–15 June.
11. Kramer, G., Romer, G., and Stuyt, H., “Design of a Dynamic Arm Support (DAS) for
Gravity Compensation”, Proceedings of the 2007 IEEE 10th International Conference on
Rehabilitation Robotics, Noordwijk, The Netherlands, pp. 1042–1048, 12–15 June.
12. Keith K. and Barbara W., Anthropometry and Biomechanics, Man-Systems Integration
Standards Revision B, NASA, July 1995.
13. Chandler, R.F., Clauser, C.E., McConville, J.T., Reynolds, H.M., and Young, J.W., Investiga-
tion of Inertial Properties of the Human Body. Technical Report DOT HS-801 430, Aerospace
Medical Research Laboratory, Wright-Patterson Air Force Base, OH, March 1975.
14. Legnani, G., “Robotica Industriale”, Milano, Casa Editrice Ambrosiana, 2003 ISBN 88-408-
1262-8.
15. Chandler, R.F., Clauser, C.E., McConville, J.T., Reynolds, H.M., and Young, J.W. “Investi-
gation of Inertial Properties of the Human Body”. Wright Patterson Air Force Base, Ohio
(AMRL-TR-75-137), 1975.
Chapter 60
EEG Classification of Mild and Severe
Alzheimer’s Disease Using Parallel Factor
Analysis Method
PARAFAC Decomposition of Spectral-Spatial
Characteristics of EEG Time Series

Charles-Francois Vincent Latchoumane, Francois-Benois Vialatte, Jaeseung

Jeong, and Andrzej Cichocki

Abstract Electroencephalograms (EEG) recordings are now widely used more and
more as a method to assess the susceptibility to Alzheimer’s disease. In this study,
we aimed at classifying control subjects from subjects with mild cognitive impair-
ment (MCI) and from Alzheimer’s disease (AD). For each subject, we computed
the relative Fourier power of five frequency bands. Then for each frequency band,
we estimated the mean power of five brain regions: frontal, left temporal, central,
right temporal and posterior. There were an equivalent number of electrodes in each
of the five regions. This grouping is very useful in normalizing the regional repar-
tition of the information. We can form a three-way tensor, which is the Fourier
power by frequency band and by brain region for each subject. From this tensor,
we extracted characteristic filters for the classification of subjects using linear and
nonlinear classifiers.

Keywords EEG Classification  Alzheimer’s Disease  Parallel Factor Analysis

Method  Spectral-Spatial Characteristics  Fourier power

60.1 Introduction

60.1.1 Diagnosis of Alzheimer’s Disease and EEGs

Alzheimer’s disease is the most prevalent neuropathology form leading to demen-

tia; it affects approximately 25 million people worldwide and is expected to have
fast recrudescence in near future [1]. The numerous clinical methods that are now

J. Jeong (B)
Brain Dynamics Lab., Bio and Brain Engineering Dept., KAIST, Yuseong-gu, Guseong-dong,
Daejeon, South Korea, 305-701
E-mail: jsjeong@kaist.ac.kr

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 705
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
706 C.-F. V. Latchoumane et al.

available to detect this disease including imaging [2, 3], genetic methods [4], and
other physiological markers [5], however, do not allow a mass screening of the pop-
ulation. Whereas psychological tests such as Mini Mental State Evaluation (MMSE)
in combination with an electrophysiological analysis (e.g. electroencephalograms
or EEG) would be very efficient and inexpensive screening approach to detect the
patients damaged by the disease.
EEG recordings are now widely used more and more as a method to assess the
susceptibility to Alzheimer’s disease, but are often obtained during steady states
where temporal information does not easily reveal the relevant features for sub-
ject differentiation; interestingly, in those conditions, previous studies could obtain
excellent classification results ranging from 84% to 100%, depending on the con-
ditions (i.e. training or validation conditions for the classifier tests) and on the
confronted groups (i.e. control subject, mild cognitive impairment (MCI), and
different stages of Alzheimer’s disease) demonstrating the promising use of resting
EEGs for diagnosis of Alzheimer’s disease [6–11].
The spectral spatial information estimated during resting states might contain
valuable clues for the detection of demented subjects; however, the inter-subject
variability, also influenced by the differences in the progression of the disease, might
render the study difficult when undertaken subject-by-subject. In that case, a multi-
way analysis would allow the extraction of information that is contained across
subjects simultaneously considering the spectral spatial information. This method-
ology has been applied to epilepsy detection and has successfully characterized the
epilepsy foci in a temporal-frequency-regional manner [12,13]. Classification based
on multi-way modeling has been performed on continuous EEGs [14] showing the
power and versatility of multi-way analyses.

60.1.2 Multi-way Array Decomposition

Previous two-way analyses combining PCA-like techniques [6] (i.e. principal com-
ponent analysis (PCA) and independent component analysis (ICA)) have shown
very high performance in the classification of subjects and have assisted in early
detection. These methods require data in the form of matrices, and then they limit
the order (i.e. number of dimensions or modes) or mix several types of variables
(e.g. using the unfolding method, also known as matricization). Either way, the nat-
urally high-order of EEGs and interactions between the ways (or modes) is lost or
destroyed, limiting further the understanding of underlying processes in the brain.
The multi-way array decomposition is a modeling tool that conserves the original
high dimensional nature of the data. This method uses the common features and
interactions between modes present in the data to create a model fit of the same
data that can be decomposed into components. The decomposition into components
provides the information of interactions between modes in the form of weight, which
is relatively easier to interpret. The application of such methods to diagnosis of
60 EEG Classification of Mild and Severe 707

Alzheimer’s disease would allow characterization of the disease based on simple

markers or weights.
Thus far, no application of multi-way array decomposition has been made in
this type of database, dealing with the classification of Alzheimer’s disease subjects
based on EEG characteristics.

60.2 Subjects and EEG Recordings

The subjects analyzed in this study were obtained from a previously studied database
[6,9,11] and consisted of eyes-open, steady state EEG recordings (20 s in duration),
with over 21 leads disposed according to the 10–20 international system and digital-
ized at 200 Hz. The database contains 38 control (mean age: 71:7 ˙8:3) subjects, 22
mild cognitive impairment (MCI) subjects (mean age: 71:9 ˙ 10:2) who later con-
tracted Alzheimer’s disease, and 23 Alzheimer’s disease patients (AD; mean age:
72:9 ˙ 7:5). The control subjects had no complaints or history of memory prob-
lems, and scored over 28 (mean score: 28:5 ˙ 1:6) on the mini mental state exam
(MMSE). The MCI subjects had complaints about memory problems and scored
over 24 at the MMSE (mean score: 26 ˙ 1:8). The inclusion criterion was set at 24
as suggested in [11], therefore encompassing MCI subjects with various cognitive
deficits, but in the early stage of Alzheimer’s disease. The AD patients scored below
20 on the MMSE and had had full clinical assessment. Thirty-three moderately or
severely demented probable AD patients (mean MMSE score: 15:3 ˙ 6:4; range: 0–
23) were recruited from the same clinic. After obtaining written informed consent
from the patients and controls, all subjects underwent EEG and SPECT examination
within 1 month of entering the study. All subjects were free of acute exacerbations
of AD related co-morbidities and were not taking medication. The medical ethical
committee of the Japanese National Center of Neurology and Psychiatry approved
this study.

60.3 Method

60.3.1 Three-Way Tensor for Analysis

In this study, we aimed at classifying control subjects from subjects with MCI
and from AD. For each subject, we computed the relative Fourier power of five
frequency bands ı (1–4 Hz),  (4–8 Hz), ˛1 (8–10 Hz), ˛2 (10–12 Hz) and ˇ (12–
25 Hz); then for each frequency band, we estimated the mean power of five brain
regions: frontal, left temporal, central, right temporal and posterior. There were an
equivalent number of electrodes in each of the five regions. This grouping is very
useful in normalizing the regional repartition of the information (Fig. 60.1).
708 C.-F. V. Latchoumane et al.

Fig. 60.1 Topological grouping of electrodes in five regions; the numbers 1, 2, 3, 4 and 5 denote
the frontal, left temporal, central, right temporal and posterior regions, respectively

We can form a three-way tensor, Subjects x Frequency Band Power x Brain

Region, which is the Fourier power by frequency band and by brain region for each
subject. From this tensor, we extracted characteristic filters for the classification of
subjects using linear and nonlinear classifiers. The detail of the method is described
in the next sections.

60.3.2 Classification in Divide and Conquer Scheme

For simplicity and interpretation, we opted for two-step classification of the subjects
(divide and conquer scheme): (1) we compared healthy subjects with the patient
group (regrouping MCI and AD); (2) we compared MCI group with AD group. The
schematic diagram of our approach is presented in Fig. 60.2.
For each step of classification, we extracted the features based on the Parallel
Factor Analysis (PARAFAC) decomposition (unsupervised) and the reference fil-
ters (supervised). We estimated the accuracy in classification for subjects using a
k-fold cross-validation with k D 1 (i.e. leave-one-out cross validation) on both
60 EEG Classification of Mild and Severe 709

Fig. 60.2 Divide and conquer method for the classification of control subjects vs. MCI vs. AD
subjects – cascade of classifiers

a quadratic naı̈ve Bayes classifier (one of the simplest statistical classifier with
quadratic boundary) and an artificial neural network (ANN; with n D 1 hidden
neurons in one hidden layer, i.e. nonlinear classifier).

60.3.3 The PARAllel FACtor Analysis (PARAFAC)

The important idea underlying multi-way analysis is the extraction of multiple inter-
actions within the structure of the data. Interactions and common patterns between
modes of the data are often neglected cause studied separately (e.g. subject by
subject, region by region) or folded on the same dimension (i.e. mixed, hence
destroying interactions) as for method using two dimensional PCA or ICA. In this
study, we constructed a three-way tensor in the form Subjects x Frequency Band
Power x Brain Region to preserve relations between spectral and regional character-
istics of each subject; we used a common modeling method for N-way analyses to
extract common interaction in the tensor with relation to the subjects: the parallel
factor analysis (PARAFAC) [15].
The PARAFAC is often referred to as a multilinear version of the bilinear factor
models. From a given tensor, X 2 RI J K , this model is able to extract linear
decompositions of Rank-1 tensors.

X
R
XD ar ı br ı cr C E; (60.1)
rD1

where ar , br , and c r are the r-th column of the component matrices A 2 RI R ,

B 2 RJ R , and C 2 RKR , respectively, and E 2 RI J K is the residual error
matrix (Fig. 60.3).
The operator ı designates the outer product of two vectors. The PARAFAC model
under optimal fitting conditions (e.g. core consistency analysis) is capable of pro-
viding a model with the assumption of trilinearity relations between the dimensions
(also called modes), thus providing a unique fit for the given data.
The fitting of the model depends on the number of components, R, chosen by
the users, and for this approach we opted for validation of the model (i.e. number of
components) based on the core consistency [16]. The optimal number R is chosen
710 C.-F. V. Latchoumane et al.

Fig. 60.3 PARAFAC modeling of a three-way tensor; each component (R D 2) is the outer
product of Rank-1 tensor a, b, and c, and E is a residual tensor

for a core consistency value that is above 90% and with the core consistency value
of R C 1 under 90%.

60.3.4 Filter Estimation and Feature Extraction

60.3.4.1 PARAFAC Filters

The PARAFAC model with a sufficient number of components and an appropriate

fitting (i.e. core consistency [16] and component variability) is capable of fitting a
model that decompose data in component of common characteristics with trilinear
interaction between modes. We are interested in features that differentiate the sub-
jects according to their clinical group. We assume that the common characteristic
contain in the data could significantly discriminate the group based on the difference
between the Frequency Band Power x Brain Region profile and normal control and
demented subjects, i.e. shift in the frequency profile of subjects with AD or “slowing
of frequencies” [17, 18].
The PARAFAC decomposition returns for each component and each mode a
Rank-1 vector which can be considered as a weight at each component and each
mode, respectively. The mode subject contains the discriminative features or weight
of each subject to be compared with the combination of the mode Frequency Band
Power and the mode Brain Region. More precisely, the component-wise combi-
nation (i.e. for each component) of “Frequency Band Power” and “Brain Region”
could be interpreted as a characteristic filter. Similar to the Fourier Power Distribu-
tion which associates a power to each frequency, the decomposition obtained from
PARAFAC can represent a distribution of filters (number of filters R D 3) asso-
ciated with a weight (i.e. subject mode). We calculated the characteristic filters for
each component, Fr , as described in Eq. (60.2):

Fr D br  crT ; (60.2)

where br and cr are the Rank-1 vector of the frequency and region mode in the
PARAFAC model, respectively. Following Eq. (60.2), we obtain a number of filters
R D 3.
60 EEG Classification of Mild and Severe 711

60.3.4.2 Reference Filters

The reference filter are supervised filters in contrast to the filters obtained using
PARAFAC (unsupervised) because they integrate prior knowledge of the member-
ship of the subjects. The reference filters of the control group and demented group
were calculated as the average Frequency Band Power x Brain Region over the nor-
mal subjects and the demented subjects (regrouping MCI and AD), respectively.
The reference filters of the MCI group and AD group were calculated as the aver-
age Frequency Band Power x Brain Region over MCI subjects and AD subjects,
respectively.

60.3.5 From Filter to Feature

The reference filter and the PARAFAC filters contain common characteristics of the
group under this study (i.e. control vs. demented or MCI vs. AD). We use a distance
measure based on the matrix scalar product also known as normalized Frobenius
inner product to compare the estimated filters (i.e. reference filters or PARAFAC
filters) with the Frequency Band Power x Brain Region profile of each subject. The
details on the distance measure are as follows:
Trace.F1T F2 /
Di s.F1 ; F2 / D ; (60.3)
jjF1 jjjjF2 jj
where F1 and F2 are two Frequency Band Power x Brain Region matrices,
T
denotes the conjugate transposition of a matrix and the Trace function returns
of a matrix. The function jj  jj indicates the Frobenius norm defined as
the trace p
jjF jj D Trace.F T F /. For each subject, we obtained R features comparing the
R PARAFAC filter to the subject’s Frequency Band Power x Brain Region profile.
Similarly, we obtained two features comparing the references filters to the subject’s
Frequency Band Power x Brain Region profile.

60.4 Results

The best classification results using the artificial neural network (ANN) and quadratic
naı̈ve Bayes classifier are displayed in Table 60.1. There was no noticeable differ-
ence between the performance of the two classifiers (both nonlinear) using either the
reference filters or the PARAFAC filters; however, while comparing MCI with AD
subjects, the ANN showed better performance using the PARAFAC filters (75.6%)
than using the reference filters (68.9%).
712 C.-F. V. Latchoumane et al.

Table 60.1 Classification results obtained using an ANN and quadratic naı̈ve Bayes classifier
with the leave-one-out cross-validation method; NC denotes normal controls and Patients denotes
patient group (regrouping MCI group and AD group)
NC vs. patients NC vs. patients AD vs. MCI AD vs. MCI
(PARAFAC) (reference) (PARAFAC) (reference)
(%) (%) (%) (%)
ANN 74.7 61.4 75.6 64.4
Bayes 74.7 61.5 68.9 64.4

Fig. 60.4 ROC curve of classification accuracy of (a) control vs. demented subjects and (b)
AD vs. MCI subjects using an ANN and leave-one-out cross-validation method; classification
results obtained using the reference filters (stars) and using the filtered data extracted with a
three-component PARAFAC (R D 3; triangle)

Generally, the PARAFAC filters showed an improvement in the accuracy of clas-

sification as compared with the reference filters and the two classification methods
used.
As shown in Fig. 60.4, the performance in the classification based on the
ANN exhibited higher accuracy using the information from the PARAFAC fil-
ters than when using the reference filters. The best performance was found to be
74.7% (75.6% sensitivity, 73.7% specificity) for the Ctr vs. demented classifica-
tion (Fig. 60.4a) and 75.6% (95.6% sensitivity, 59.1% specificity) for MCI vs. AD
classification (Fig. 60.4b).
Using PARAFAC on each separate group, a clustering of frequency bands (i.e.
frequency band mode) can also be obtained [19] in order to study the interactions
between the frequency ranges (Fig. 60.5). We observed the evolutions in the inter-
relations between the  range and high frequency (α2 and ˇ) ranges: seemingly,
for Control subjects and AD patients, the  range activity is clustered with high
frequencies (distances of 1.5 and 1, respectively); whereas for the MCI subjects, the
 range activity is much less (a distance of 2.5).
60 EEG Classification of Mild and Severe 713

Fig. 60.5 Clustering of frequency bands; dendrograms extracted from the clustering of two-
component PARAFAC models for (a) control subject, (b) MCI patients, and (c) Alzheimer
patients

60.5 Discussion

In this study, we presented a method applied to the classification of subjects based

on a PARAFAC decomposition of their EEG features. This type of application of
multi-way analysis to EEG features has not yet been implemented for AD diagnosis.
We also showed possible interpretation of the fitted model not only based on the
spatial-frequency filters, but also based on the unimodal clustering for each group’s
model. The weight obtained using the PARAFAC decomposition could serve as a
marker for differentiation between control and patient group but also between MCI
and AD subjects.
Previous studies on Alzheimer’s classification and compromising with different
degrees of severity of the disease obtained classification accuracies ranging from
84% to 100% [6–11]. The classification results presented in this study (overall good
classification of about 74%) are considerably good compared with the classification
results using the same database in another study [6] (80%); however, this study
proposes a three-class classification using a two step divide and conquer approach.
The capability of the method to make a stepwise differentiation of subjects
illustrates the applicability of this study to the clinical domain. The method also
demonstrates possibility of mass-screening of subjects for the diagnosis of AD as
well as the differentiation of stages in the progression of the disease (MCI developed
to turn out to be mild and severe).
In this study, we determined to use a three-way tensor confronting subjects, fre-
quency band power and brain regions allowing simplicity of computation and inter-
pretation. The method using multi-way array decomposition allows the integration
714 C.-F. V. Latchoumane et al.

of a large number of variables, organized on higher orders (dimensions of the tensor

equal to D, typically with D higher than 3). The addition of more variables neces-
sitate that those variables would commonly interact and characterize the data or, in
our case, the subjects in the data, otherwise the extracted information might lose
its property of separability. More importantly, increasing the number of variables
or order would not seriously increase the number of features for classification (i.e.
number of components of the PARAFAC model, R) which would avoid artificially
increasing the separability of the groups and reducing the bias due to overfitting for
the classifier.
Moreover, apart from its crucial uniqueness and the resulting easy interpretabil-
ity [20], the PARAFEC model used might not be the best model to use in this
analysis, as it imposes trilinearity conditions. Future investigation include the com-
parison of the accuracy of other models such as PARAFAC2 [21], Tucker3 [22], or
the nonnegative tensor factorization method [20, 23] and its constrained model on
sparsity [24].

Acknowledgements The first author would like to thank the Minister of Information and Technol-
ogy of South Korea, Institute for Information and Technology Advancement (IITA) for his financial
support. The authors would like to thank Dr. T. Musha for his contribution in subjects’ analysis and
the EEG recordings and Dr. M. Maurice and S. Choe for their help in the editing of the content.

References

1. Ferri, C.P. et al. Global prevalence of dementia: a Delphi consensus study. The Lancet 366,
2112–2117 (2006).
2. Alexander, G.E. Longitudinal PET evaluation of cerebral metabolic decline in dementia: a
potential outcome measure in Alzheimer’s disease treatment studies. American Journal of
Psychiatry 159, 738–745 (2002).
3. Deweer, B. et al. Memory disorders in probable Alzheimer’s disease: the role of hippocampal
atrophy as shown with MRI. British Medical Journal 58, 590 (1995).
4. Tanzi, R.E. & Bertram, L. New frontiers in Alzheimer’s disease genetics. Neuron 32, 181–184
(2001).
5. Andreasen, N. et al. Evaluation of CSF-tau and CSF-Aß42 as Diagnostic Markers for
Alzheimer Disease in Clinical Practice, Archives of Neurology, 58, pp. 373–379 (2001).
6. Cichocki, A. et al. EEG filtering based on blind source separation (BSS) for early detection of
Alzheimer’s disease. Clinical Neurophysiology 116, 729–737 (2005).
7. Buscema, M., Rossini, P., Babiloni, C. & Grossi, E. The IFAST model, a novel parallel
nonlinear EEG analysis technique, distinguishes mild cognitive impairment and Alzheimer’s
disease patients with high degree of accuracy. Artificial Intelligence in Medicine 40, 127–141
(2007).
8. Huang, C. et al. Discrimination of Alzheimer’s disease and mild cognitive impairment by
equivalent EEG sources: a cross-sectional and longitudinal study. Clinical Neurophysiology
111, 1961–1967 (2000).
9. Musha, T. et al. A new EEG method for estimating cortical neuronal impairment that
is sensitive to early stage Alzheimer’s disease. Clinical Neurophysiology 113, 1052–1058
(2002).
60 EEG Classification of Mild and Severe 715

10. Pritchard, W.S. et al. EEG-based, neural-net predictive classification of Alzheimer’s disease
versus control subjects is augmented by non-linear EEG measures. Electroencephalography
and Clinical Neurophysiology 91, 118–30 (1994).
11. Woon, W.L., Cichocki, A., Vialatte, F. & Musha, T. Techniques for early detection of
Alzheimer’s disease using spontaneous EEG recordings. Physiological Measurement 28,
335–347 (2007).
12. Acar, E., Aykut-Bingol, C., Bingol, H., Bro, R. & Yener, B. Multiway analysis of epilepsy
tensors. Bioinformatics 23, i10–i18 (2007).
13. Acar, E., Bing, C.A., Bing, H. & Yener, B. in Proceedings of the 24th IASTED International
Conference on Biomedical Engineering 317–322 (2006).
14. Lee, H., Kim, Y.D., Cichocki, A. & Choi, S. Nonnegative tensor factorization for continuous
EEG classification. International Journal of Neural Systems 17, 305 (2007).
15. Andersson, C.A. & Bro, R. The N-way toolbox for MATLAB. Chemometrics and Intelligent
Laboratory Systems 52, 1–4 (2000).
16. Bro, R. & Kiers, H.A.L. A new efficient method for determining the number of components
in PARAFAC models. Contract 1999, 10377 (1984).
17. Coben, L.A., Danziger, W.L. & Berg, L. Frequency analysis of the resting awake EEG in mild
senile dementia of Alzheimer type. Electroencephalography and Clinical Neurophysiology
55, 372–380 (1983).
18. Sloan, E.P., Fenton, G.W., Kennedy, N.S.J. & MacLennan, J.M. Electroencephalography and
single photon emission computed tomography in dementia: a comparative study. Psychologi-
cal Medicine 25, 631 (1995).
19. Atkinson, A.C. & Riani, M. Exploratory tools for clustering multivariate data. Computational
Statistics and Data Analysis 52, 272–285 (2007).
20. Bro, R. PARAFAC. Tutorial and applications. Chemometrics and Intelligent Laboratory
Systems 38, 149–171 (1997).
21. Al Kiers, H., Ten Berge, J. & Bro, R. PARAFAC2: PART I. a direct fitting algorithm for the
PARAFAC2 model. Journal of Chemometrics 13, 275–294 (1999).
22. Kim, Y.D., Cichocki, A. & Choi, S. in IEEE International Conference on Acoustics, Speech,
and Signal Processing (ICASSP-2008) (IEEE, Las Vegas, Nevada, 2008).
23. Cichocki, A., Zdunek, R. & Amari, S. Nonnegative matrix and tensor factorization. Signal
Processing Magazine, IEEE 25, 142–145 (2008).
24. Cichocki, A., Zdunek, R., Plemmons, R. & Amari, S. in ICANNGA-2007 (ed. Science,
L.N.i.C.) 271–280 (Springer, Warsaw, Poland, 2007).
Chapter 61
Feature Selection of Gene Expression Data
Based on Fuzzy Attribute Clustering

Elham Chitsaz, Mohammad Taheri, and Seraj D. Katebi

Abstract In this chapter, a novel approach which uses fuzzy version of k-mode has
been proposed for grouping interdependent genes. The fuzzy approach considers
uncertainty, achieves more stability and consequently improves the classification
accuracy. In addition, other modifications have implemented to fuzzify the selection
of the best features from each cluster. A new method to initialize cluster centers has
also been applied in this work. Moreover, a novel discretization method based on
Fisher criterion is also proposed.

Keywords Feature Selection  Gene Expression  Fuzzy Attribute Clustering 

discretization method

61.1 Introduction

By now, many applications have been introduced in which, feature selection is

utilized as a preprocessing stage for classification. This process speeds up both
the training and reasoning stages, reduces memory space, and improves classifi-
cation accuracy. Reducing the cost of gathering data is another advantage of feature
selection. Biological data (e.g. micro-arrays) are usually wide and shallow.
Small number of samples narrows the acquirable knowledge. Hence it reduces
the probability of correct reasoning whether a specified feature effects on the class
label or not. Moreover, a classifier can generate the classification rules more easily
with small number of features. But increasing the number of features may lead to
ambiguity in training so that it would not even converge.
Finding proper features has been the subject of various approaches in the lit-
erature. (e.g. filter, wrapper, and embedded approaches [1], greedy [2], statistical

E. Chitsaz (B)
Computer Science & Engineering Department of Shiraz University, Shiraz, Iran
E-mail: chitsaz@cse.shirazu.ac.ir

S.I. Ao and L. Gelman (Eds.), Advances in Electrical Engineering and Computational 717
Science, Lecture Notes in Electrical Engineering.
c Springer Science + Business Media B.V. 2009
718 E. Chitsaz et al.

approaches such as the typical principle component analysis (PCA) method [3] and
the linear discriminant analysis (LDA) method [4], GA [5, 6], Neural Network [7],
Fuzzy Systems [8, 9], mutual information-based feature selection [10, 11]). Selec-
tion of the most influential genes, for classification of samples in micro-arrays, is
one of the well known topics in which many investigations have been carried out. In
the context of microarrays and LDA, wrapper approaches have been proposed [12].
Recently, biclustering algorithms [13, 14] have been proposed to cluster both genes
and samples simultaneously. Wai-Ho Au et al., in 2005 [15], proposed a feature
selection technique based on clustering the correlated features by a semi-k-means
method named k-modes.
In this paper a novel approach which uses fuzzy version of k-mode has been pro-
posed for grouping interdependent genes. The fuzzy approach considers uncertainty,
achieves more stability and consequently improves the classification accuracy. In
addition, other modifications have implemented to fuzzify the selection of the best
features from each cluster. A new method to initialize cluster centers has also been
applied in this work. Moreover, a novel discretization method based on Fisher
criterion [4] is also proposed.
C4.5 [16] classifier has been applied in the final stage to asses the feature selec-
tion process. Leukemia dataset [17], a micro-array data containing 73 samples and
7,129 genes, is used in all experimental results.
In next section, the previous work of feature selection by clustering on Micro-
arrays [15] is explained following by its fuzzy approach in Section 61.3. Exper-
imental results are presented and explained in Section 61.4. Finally we draw our
conclusion in Section 61.5.

61.2 Related Work

The attribute clustering algorithm (ACA) has been applied by Wai-Ho Au et al., in
2005 [15], for grouping, selection, and classification of gene expression data which
consists of a large number of genes (features) but a small number of samples.
This approach finds c disjoint clusters and assigns each feature to one of the
clusters. The genes in each cluster should have high a correlation with each other
while they are low correlated to genes in other clusters. This method uses the
interdependence redundancy measure as the similarity measure.
To cluster features, the k-modes algorithm is utilized which is similar to the well
known clustering method, k-means [18]. Mode of each cluster is defined as one of its
features which has the largest multiple interdependence redundancy measure among
other features in that cluster. The multiple interdependence redundancy measure is
calculated for each feature by Eq.(61.1).
X  
MR .Ai / D R Ai W Aj (61.1)
Aj 2 C luster.i /;
j ¤i
61 Feature Selection of Gene Expression Data Based on Fuzzy Attribute Clustering 719

Where, Cluster(i) is the set of features which are in the same cluster with Ai and is
the interdependence measure between the two features, Ai and Aj , which is defined
by Eq. (61.2).  
  I Ai W Aj
R Ai W Aj D   (61.2)
H Ai W Aj
Where, I.Ai W Aj / is the mutual information between Ai and Aj as computed in
Eq. (61.3).
mi mj
X X  
I.Ai W Aj / D Pr Ai D i k ^ Aj D jl
kD1 lD1
 
Pr Ai D i k ^ Aj D jl (61.3)
log  
Pr .Ai D i k / Pr Aj D jl

H.Ai W Aj / is joint entropy of Ai and Aj which is given by Eq.(61.4).

mi mj
X X  
H.Ai W Aj / D  Pr Ai D i k ^ Aj D jl
kD1 lD1
  (61.4)
log Pr Ai D i k ^ Aj D jl

H.Ai W Aj / is used to normalize I.Ai W Aj / in Eq. (61.2). The larger value for
I.Ai W Aj /, the higher interdependency of two features, Ai and Aj . Hence, there
should be some pairs of values for these features which are simultaneously vis-
ited with high frequency and other pairs are less probable. Therefore, having one
of the values, other one may be approximated considering value pairs with high
probability.
K-modes is different from k-means in two points. First, mode of each cluster is
selected as the cluster center instead of the mean. Second, use of Euclidean distance
as the dissimilarity measure is substituted by the interdependency between attributes
as a similarity measure.
In ACA method, genes are grouped into different clusters. A cluster is a set of fea-
tures which are more correlated in comparison with features in other clusters. Hence,
if there is a missed value for one of these features, it may be approximated consid-
ering other co-cluster features one by one. Therefore, a few features in each cluster
may be sufficient to present properties of samples. But the selected features should
be overall correlated with other features in the same cluster. This is the motivation
of selecting features with highest multiple interdependence redundancy measure, as
defined by Eq. (61.1), to represent the cluster.
Computing interdependency is defined on just discrete data types. Therefore, to
determine the interdependence measure between two features, the range of all con-
tinuous features should be first discretized into a finite number of intervals. This is
done by the Optimal Class-Dependent Discretization algorithm (OCDD).
720 E. Chitsaz et al.

61.3 The Proposed Fuzzy Approach

A novel method is proposed which combines the effectiveness of ACA with fuzzy
k-means algorithm [19]. In this method each feature is assigned to different clusters
with different degrees. This comes from the idea that each gene may not belong
to just one cluster and it is much better to consider the correlation of each gene
to features in entire clusters. Hence, during the selection of the best features, more
accurate relations between genes are available. The main point here is that in select-
ing each feature, it is considered among the entire clusters not just one. Hence, in
this case a feature, which is not correlated enough with members of one cluster
but its correlation among entire clusters is high, gains more chance to be selected
in comparison with crisp ACA. In this method better features might be selected
for classification stage over crisp method. The stability of fuzzy clustering is also
much higher in comparison with k-modes according to experimental results. In addi-
tion, Fuzzy Attribute Clustering Algorithm (FACA) converges smoothly whereas
k-modes in ACA often oscillates between final states.
In the proposed method matrix Ukm represents the membership degree of each
gene in each cluster where k is the number of clusters that is fixed and determined
at first and m is the number of features. Matrix U is computed by Eq. (61.5).

1
uri D  (61.5)
P
k
R.Ai ;c /
2
m1
R.Ai ;r /
cD1

Where, k is number of clusters. uri is membership degree of ith feature in rth cluster
and m is a weighting exponent. Afterwards, uri is normalized by Eq. (61.6).

unew
unew
ri D ri
(61.6)
P
k
uold
li
lD1

According to this membership matrix, fuzzy multiple interdependence redundancy

measure is defined by Eq. (61.7) which is a modified version of Eq. (61.1).

unew
unew
ri D ri
(61.7)
P
k
uold
li
lD1

Where, p is the total number of features and r is the cluster number in which the
multiple interdependence redundancy measure of feature Ai is calculated. Hence,
in calculating MRr .Ai / the entire features are considered.
Indeed, fuzzy multiple interdependence redundancy measure should be com-
puted for each cluster separately since each feature is not belonged to just one
cluster. In this approach, mode of a cluster is updated to the feature with the
highest fuzzy multiple interdependence redundancy in that cluster. Fuzzy multiple
61 Feature Selection of Gene Expression Data Based on Fuzzy Attribute Clustering 721

interdependence redundancy of each feature should be calculated regardless of its

own membership degree in associated cluster. Considering this membership degree,
mode of each cluster will never change, since it has high membership degree in that
cluster.
The objective function in the proposed fuzzy method is computed as Eq. (61.8).

X
k X
p
J D ri R .Ai W
um r/ (61.8)
rD1 i D1

Where, k and p are the number of clusters and features respectively and r is mode
of rth cluster which represents center of that cluster.
Selection of the best features is based on the rank of each feature which is
calculated as Eq. (61.9).

X
k
rank .Ai / D um
ri MRr .Ai / (61.9)
rD1

The final stage is the classification of data points according to selected features. The
following items encouraged us to use C4.5 as the final classifier for assessment of
the proposed feature selection.
Both C4.5 and proposed feature selection method can be applied just on discrete
features.
C4.5 is a well known decision tree technique with high performance which is
used as a benchmark in many articles in the state of the art.
To determine the priority of features, it uses Information gain and Entropy which
are similar to mutual information (I) and joint entropy (H), respectively.
The flowchart of the proposed algorithm is depicted in Fig. 61.1. The stop condition
is satisfied when the value of the cost function defined in Eq. (61.8) is the same
as previous iteration with a predefined precision. A maximum number of iterations
have also been defined.
Initialization of cluster centers (modes) is the only random event in ACA. But
uncertainty concept in FACA and taking membership degree of each feature in each
cluster into consideration leads to more stability in the final clustering result. Never-
theless, a simple initialization procedure is proposed here which seems to improve
the final results as explained in the next section.
A method which is inspired by an initialization method proposed in [20] and [21],
is utilized. All patterns are considered as just one cluster and the feature with the
highest multiple interdependence redundancy measure is chosen as the first initial
cluster center . 1 /. Other centers are selected from features by Eq. (61.10).
2 3
X  
r D arg min 4 R Ai W Aj 5 2  r  k (61.10)
Ai …Sr1 Aj 2Sr1
722 E. Chitsaz et al.

Fig. 61.1 Flowchart of the proposed method

Where, r is the center of rth cluster, Ai is the ith feature and k is the desired
number of clusters. Sr represents the set f 1 ; 2 ; : : :; r g. Indeed, after selecting
the first cluster center, others are selected from unselected features one by one. In
any step, the feature, which is the least interdependent feature with selected cluster
centers by now, is selected as a new cluster center. It improves the final objective
function as explained in the next section.

61.4 Experimental Results

In this paper, Leukemia dataset [17] is used as a well known biological dataset. It
is a gene expression micro-array dataset with 73 samples and 7,129 genes (fea-
tures). Due to memory constraints and dynamic implementations, only the first
1,000 features are considered.

61.4.1 Stability

As mentioned in last section, FACA seems to be more stable than ACA. Curve of
objective function Eq. (61.8) (achieved in the final stage of clustering) vs. number
61 Feature Selection of Gene Expression Data Based on Fuzzy Attribute Clustering 723

Fig. 61.2 Objective function vs. the number of clusters. Best, worst, average curves in 20 runs for
random initialization and associated curve with proposed initialization technique. (a) Crisp. (b)
Fuzzy

of clusters, is depicted Figs. 61.2a and b, for crisp and fuzzy approaches, respec-
tively. In these figures, with a fixed number of clusters and specifying the initial
cluster modes randomly, 20 separate clustering runs have been plotted. Best, worst
and average values over these runs are depicted in separate curves. Less difference
between best and worst values for fuzzy approach, in comparison with crisp version,
may show less dependency of the clustering to initial values and consequently more
stability is gained. Also considering the same curve, with the proposed initialization
technique indicates completely better results than the average values for both crisp
and fuzzy approaches, as shown in Fig. 61.2.
Also less slope of these curves for fuzzy approaches shows that the number of
clusters is not influential on the objective function for fuzzy approach as much as
crisp version. Although, objective function should be maximized but, higher objec-
tive value in crisp version is never an evidence of better performance in comparison
with fuzzy version. Indeed, the objective function in the crisp version is summation
of interdependence measure values between high correlated features; whereas, in
fuzzy version, high interdependencies are reduced by membership degree which is
less than one, and this reduction is tried to be counted with low interdependence
measure value and low membership degrees of other feature pairs.
With a predefined number of clusters, objective function is expected to change
after each iteration in fuzzy k-modes. As shown in Fig. 61.3, objective function
oscillates in crisp version whereas, fuzzy objective function converges smoothly.
This smoothness speeds up and guarantees convergence.
724 E. Chitsaz et al.

Fig. 61.3 Objective function vs. iterations, for 4, 5, 6 and 7 clusters. (a) Crisp. (b) Fuzzy

Fig. 61.4 Classification rate resulted by C4.5 on Leukemia dataset for two features per cluster

61.4.2 Classification Accuracy

As another experiment, C4.5 decision tree has been used to classify Leukemia
dataset based on selected features. The resulted classification rate can be encoun-
tered as a measure of assessing the feature selection method. Hence, ACA and Fuzzy
ACA are compared by the resulted classification rate on different number of clus-
ters, as depicted in Figs. 61.4–61.6 for 2, 3 and 4 features per cluster, respectively.
Wai-Ho Au et al. [15] proposed selection of a predefined number of features from
each cluster. But in fuzzy approach, any feature is belonged to all clusters. There-
fore, each feature is first belonged to the cluster with maximum membership degree
before any feature selection based on their ranks. Leaved-One-Out method has been
used here to assess the generalization of classification system. Increasing the num-
ber of features per cluster, improves the classification accuracy for both crisp and
fuzzy approaches.
61 Feature Selection of Gene Expression Data Based on Fuzzy Attribute Clustering 725

Fig. 61.5 Classification rate resulted by C4.5 on Leukemia dataset for three features per cluster

Fig. 61.6 Classification rate resulted by C4.5 on Leukemia dataset for four features per cluster

61.5 Conclusion

In this paper, a fuzzy approach is suggested to group features into clusters in order to
select best ones for classification stage. It proposes a new clustering method which
combines k-modes clustering used in ACA with fuzzy k-means clustering algorithm.
Hence, it leads to more stability, faster convergence and greater classification rate,
in comparison with the previous method.
Both ACA and FACA are based on the interdependence redundancy measure
which is applicable only on discrete data types. Hence, a new method for dis-
cretization of continuous data has also been proposed. Moreover, the initialization
of cluster centers is not random, but a new method has been applied for initialization
in order to increase the probability of better clusters formation.
Further work will extend the suggested discretization method and improve the
similarity measure. Discovering the influential genes on diseases regarding the
selected genes is also worth working on.
726 E. Chitsaz et al.

References

1. I. Guyon and A. Elisseeff, ‘An introduction to variable and feature selection’, Journal of
Machine Learning Research, vol. 3, pp. 1157–1182, 2003.
2. R. Caruana and D. Fratage, ‘Greedy attribute selection’, in Machine Learning: Proceedings of
11th International Conference, San Francisco, CA, pp. 283–288, 1994.
3. T. Joliffe, ‘Principal Component Analysis’, New York: Springer-Verlag, 1986.
4. K. Fukunaga, ‘Introduction to Statistical Pattern Recognition’, New York: Academic, 1972.
5. F. Z. Bril, D. E. Brown, and N. W. Worthy, ‘Fast genetic selection of features for neural
network classifiers’, IEEE Transactions on Neural Networks, vol. 3, pp. 324–328, March 1992.
6. M. L. Raymer, W. F. Punch, E. D. Goodman, L. A. Kuhn, and L. C. Jain, ‘Dimensionality
reduction using genetic algorithms’, IEEE Transactions on Evolutionary Computation, vol. 4,
no. 2, pp. 164–171, 2000.
7. R. Setiono and H. Liu, ‘Neural-network feature selector’, IEEE Transactions on Neural
Networks, vol. 8, pp. 654–662, May 1997.
8. E. C. C. Tsang, D. S. Yeung, and X. Z. Wang, ‘OFFSS: Optimal Fuzzy-Valued Feature Subset
Selection’, IEEE Transactions on Fuzzy Systems, vol. 11, no. 2, pp. 202–213, 2003.
9. M. R. Rezaee, B. Goedhart, B. P. F. Lelieveldt, and J. H. C. Reiber, ‘Fuzzy feature selection’.
PR (32), no. 12, pp. 2011–2019, December 1999.
10. R. Battiti, ‘Using mutual information for selecting features in supervised neural net learning’,
IEEE Transactions on Neural Networks, vol. 5, pp. 537–550, July 1994.
11. N. Kwak and C.-H. Choi, ‘Input feature selection for classification problems’, IEEE Transac-
tions on Neural Networks, vol. 13, pp. 143–159, January 2002.
12. M. Xiong, W. Li, J. Zhao, L. Jin, and E. Boerwinkle, ‘Feature (gene) selection in gene
expression-based tumor classification’, Molecular Genetics and Metabolism, vol. 73, no. 3,
pp. 239–247, 2001.
13. S. C. Madeira and A. L. Oliveira, ‘Biclustering algorithms for biological data analysis: A
survey’, IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 1, no.
1, pp. 24–45, January-March 2004.
14. Y. Cheng and G. M. Church, ‘Biclustering of expression data’, Proceedings of Eighth
International Conference Intelligent Systems for Molecular Biology, pp. 93–103, 2000.
15. W.-H. Au, K. C. C. Chan, A. K. C. Wong, Y. Wang, ‘Attribute clustering for grouping, selec-
tion, and classification of gene expression data’, IEEE/ACM Transactions on Computational
Biology and Bioinformatics (TCBB), vol. 2, no. 2, pp. 83–101, 2005.
16. J. R. Quinlan, ‘C4.5: Programs for Machine Learning’, Morgan Kaufmann, San Francisco,
CA, 1993.
17. T. R. Golub, D. K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J. P. Mesirov, H. Coller,
M. L. Loh, J. R. Downing, M. A. Caligiuri, C. D. Bloomfield, and E. S. Lander, ‘Molecular
classification of cancer: class discovery and class prediction by gene expression monitoring’,
vol. 286, no. 5439, pp. 531–537, 1999.
18. T. Kanungo, D. Mount, N. Netanyahu, C. Piatko, R. Silverman, and A. Wu, ‘A local search
approximation algorithm for k-means clustering’, Computational Geometry, vol. 28, pp. 89–
112, 2004.
19. J. C. Bezdek, ‘Fuzzy mathematics in pattern classification’, Ph.D. thesis, Applied Mathematics
Center, Cornell University, Ithaca, 1973.
20. M. L. Micó, J. Oncina, and E. Vidal, ‘A new version of the nearest-neighbour approximat-
ing and eliminating search algorithm (AESA) with linear preprocessing time and memory
requirements’, Pattern Recognition, vol. 15, pp. 9–17, 1994.
21. M. Taheri and R. Boostani, ‘Novel auxiliary techniques in clustering’, International Confer-
ence on computer science and engineering, 2007.