INTEGRATED LOCATION, ROUTING AND

SCHEDULING PROBLEMS: MODELS AND

ALGORITHMS

by

Zeliha Akça

Presented to the Graduate and Research Committee

of Lehigh University

in Candidacy for the Degree of

Doctor of Philosophy

in

Industrial and Systems Engineering

Lehigh University

January 2010
°
c Copyright by Zeliha Akça 2010
All Rights Reserved

ii
 Approved and recommended for acceptance as a dissertation in partial fulfillment of the re-

quirements for the degree of Doctor of Philosophy.

Date

Dr. Theodore K. Ralphs

Dissertation Director

Accepted Date

Committee:

Dr. Theodore K. Ralphs, Chairman

Dr. Rosemary T. Berger

Dr. Jeffrey T. Linderoth

Dr. Karen Smilowitz

iii
Acknowledgments

I would like to thank a number of people who have supported me in any way during my Ph.D.

study. First of all, I am grateful to my parents and my brother for their unconditional love and

support in my life. I would like to thank to other members of my family for their support to my

parents while I have been abroad for the PhD study.

I would like to thank to my previous dissertation advisor Dr. Rosemary Berger for helping me

to start and develop my research and for her support and directions during my graduate study. I

would like to thank to my dissertation advisor, Assoc. Dr. Ted Ralphs, for accepting being my

advisor, helping me patiently to construct and finalize my Ph.D. research, assisting me to provide

better outputs, challenging me with different views and for his continuous support. I have learned

a great deal from them, I feel very much indebted. I would also like to thank to my other members

of my dissertation committee for their evaluations, directions and insightful comments.

I am grateful to the P.C. Rossin School of Engineering and the Department of Industrial and

Systems Engineering at Lehigh University for providing me financial support during my educa-

tion. Many thanks also go to those who arranged the financial support. Special thanks to Dr. Larry

Snyder for hiring me as a research assistant in his project and helping me to get industrial expe-

rience. I would like to thank to the faculty in Industrial and Systems Engineering Department for

their invaluable assistance in my training. I should thank to Dr. Jeffrey Linderoth for being very

supportive throughout my graduate study. I am truly happy to know him. I also want to thank him

for his help to solve problems we had with MINTO.

iv
 I would like to thank to my friends, Mehmet Oguz Atan, Tolga Seyhan, Umit Ozkan, Wasu

Glankwamdee, Kamonkan Laksana and many others who I could not list all here for their friend-

ship and unconditional assistance when needed. I would like to thank to my friends Tuba Alim

Kilic, Selin Bilgin Ozpeynirci, Ozgul Baysar, Nilgun Esin Aksoy, and Zehra Bal who have sup-

ported me despite the distance between us. I would like to thank to Ashutosh Mahajan for his

friendship, professional collaboration and his help in solving software and hardware problems I

had in my studies. Special thanks to my fellow friends Banu Gemici Ozkan, Menal Guzelsoy,

Pinar Guvelioglu, Zumbul Bulut, Ruken Duzgun and Berrin Aytac for being with me at Lehigh

and sharing my key moments. Their friendship means a lot to me. I also want to thank them for

their professional collaboration during my education. I am grateful to Menal Guzelsoy for being

available for 7 days and 24 hours in order to give me technical support in hardware and software

related problems.

Finally, I would like to thank to Rita Frey and Kathy Rambo, coordinators in Industrial and

Systems Engineering Department, for their assistance since I got acceptance to the PhD program.

v
Contents

Acknowledgments iv

Contents vi

List of Tables x

List of Figures xiii

Abstract 1

1 Introduction 3

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.2 Polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.3 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.2.4 Mixed Integer Linear Programming . . . . . . . . . . . . . . . . . . . . 8

1.2.5 Cutting-Plane Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.2.6 Branch-and-Bound Algorithms . . . . . . . . . . . . . . . . . . . . . . . 14

1.2.7 Branch-and-Cut Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 15

1.2.8 Column Generation Algorithms . . . . . . . . . . . . . . . . . . . . . . 17

vi
 1.2.9 Dantzig-Wolfe Decomposition . . . . . . . . . . . . . . . . . . . . . . . 20

1.2.10 Branch-and-Price Algorithms . . . . . . . . . . . . . . . . . . . . . . . 24

1.2.11 Branch-Cut-and-Price Algorithms . . . . . . . . . . . . . . . . . . . . . 25

1.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.3.1 Facility Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.3.2 Vehicle Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.3.3 Integrated Facility Location and Routing . . . . . . . . . . . . . . . . . 35

1.3.4 Vehicle Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.3.5 Vehicle Routing and Scheduling . . . . . . . . . . . . . . . . . . . . . . 39

1.3.6 Integrated Location, Routing and Scheduling . . . . . . . . . . . . . . . 41

1.4 Contributions of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

1.5 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2 Models and Complexity 45

2.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.2 Problem Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.2.1 Graph-Based Formulation of LRSP . . . . . . . . . . . . . . . . . . . . 48

2.2.2 Set Partitioning-Based Formulation of LRSP . . . . . . . . . . . . . . . 51

2.3 Simple Valid Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

2.4 Relationship Between Graph- and Set Partitioning-based Formulations . . . . . . 58

2.4.1 Dantzig-Wolfe Decomposition of the Graph-based Formulation . . . . . 58

2.4.2 Comparing Alternative Formulations . . . . . . . . . . . . . . . . . . . 69

2.5 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3 Branch-and-Price Algorithm 73

3.1 Notation and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

3.2 Pricing Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3.2.1 Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

vii
 3.2.2 Exact Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

3.2.3 Heuristic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

3.3 Branching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

3.3.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

3.3.2 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

3.4 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.4.1 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.4.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

3.4.3 Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

4 Valid Inequalities 111

4.1 Notation and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4.2 Valid Inequalities For Set Partitioning-Based Formulations . . . . . . . . . . . . 115

4.2.1 Relaxations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

4.2.2 Reformulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

4.3 Valid Inequalities Based on Graph-Based Formulations . . . . . . . . . . . . . . 128

4.4 Disjunctive Valid Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

5 Computational Experiments 149

5.1 Description of the Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

5.2 Comparing the LP Relaxations of Alternative Formulations . . . . . . . . . . . . 153

5.3 Assessing the Strength of Simple Valid Inequalities . . . . . . . . . . . . . . . . 157

5.4 Computational Experiments With Branch and Price . . . . . . . . . . . . . . . . 159

5.4.1 Performance of Exact Pricing Algorithms . . . . . . . . . . . . . . . . . 160

5.4.2 Effect of Using Heuristic Pricing Algorithms . . . . . . . . . . . . . . . 162

5.4.3 Performance of Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . 172

5.4.4 Comparing One- and Two-Stage Branch And Price . . . . . . . . . . . . 176

5.4.5 Performance on Generated Instances . . . . . . . . . . . . . . . . . . . . 179

viii
 5.4.6 Performance on Instances from the Literature . . . . . . . . . . . . . . . 186

5.5 Computational Experiments for Cut Generation . . . . . . . . . . . . . . . . . . 186

5.5.1 Lifted Cover and GAP Inequalities . . . . . . . . . . . . . . . . . . . . . 187

5.5.2 Disjunctive Valid Inequalities . . . . . . . . . . . . . . . . . . . . . . . 189

5.6 Value of Integrating Scheduling Decision . . . . . . . . . . . . . . . . . . . . . 195

5.7 Comparing Integrated and Sequential Optimization . . . . . . . . . . . . . . . . 204

6 Location and Routing Problems 209

6.1 Problem Definition and Formulations . . . . . . . . . . . . . . . . . . . . . . . 210

6.1.1 Set Partitioning Formulation . . . . . . . . . . . . . . . . . . . . . . . . 210

6.1.2 Vehicle Flow Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 212

6.2 Solution Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

6.2.1 Strengthening the Set Partitioning-Based Formulation . . . . . . . . . . 214

6.2.2 Branch-and-Price Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 215

6.3 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.3.1 Generated Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

6.3.2 Instances From the Literature . . . . . . . . . . . . . . . . . . . . . . . 223

7 Conclusions and Future Directions 229

A Tables of Computational Results 248

Vita 272

ix
List of Tables

5.1 Location & Demand Data for LRSP instances generated from Cordeau et al. (1997) 151

5.2 Facility Locations for LRSP instances . . . . . . . . . . . . . . . . . . . . . . . 151

5.3 Summary of Capacity and Cost Parameters . . . . . . . . . . . . . . . . . . . . 153

5.4 LP Relaxation Bounds of (G-LRSP) and (SPP-LRSP): 25-Customer Instances . . 155

5.5 LP Relaxation Bounds of (G-LRSP) and (SPP-LRSP): 40-Customer Instances . . 156

5.6 Results for instances with 25 customers and 5 facilities . . . . . . . . . . . . . . 181

5.7 Results for instances with 40 customers and 5 facilities . . . . . . . . . . . . . . 183

5.8 More Solution Details for 25-customer instances . . . . . . . . . . . . . . . . . 184

5.9 Solution details for instances with 40 customers and 5 facilities . . . . . . . . . . 185

5.10 More Solution Details for 40-customer Instances . . . . . . . . . . . . . . . . . 185

5.11 Comparison of Lin et al. (2002) results and branch-and-price results . . . . . . . 186

5.12 Vehicle capacity values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

5.13 Solution details for LRPDC + BPP . . . . . . . . . . . . . . . . . . . . . . . . . 199

5.14 Solution details for LRSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

5.15 Average number of vehicles for each vehicle capacity value . . . . . . . . . . . . 202

5.16 Solution details for Sequential Optimization . . . . . . . . . . . . . . . . . . . . 207

5.17 Solution details for LRSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208

6.1 Details for the instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

6.2 Performance of Exact Branch-and-Price for 10-customer instances . . . . . . . . 221

x
6.3 Performance of Exact Branch-and-Price for 15-customer instances . . . . . . . . 222

6.4 Performance of Exact Branch-and-Price for 20-customer instances . . . . . . . . 223

6.5 Instances with N F ≥ 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224

6.6 Instances with Capacitated Facilities . . . . . . . . . . . . . . . . . . . . . . . . 225

6.7 Instances with facility fixed cost . . . . . . . . . . . . . . . . . . . . . . . . . . 226

6.8 Characteristics of the Instances and the Optimal Solutions . . . . . . . . . . . . . 226

6.9 Performance of the 2S-EBP for Instances with 30 and 40 Customers, Capacitated

Facilities and Facility Fixed Cost . . . . . . . . . . . . . . . . . . . . . . . . . . 227

6.10 Performance of Heuristic Branch-and-Price . . . . . . . . . . . . . . . . . . . . 227

6.11 Performance of Two-Stage Exact Branch-and-Price . . . . . . . . . . . . . . . . 228

6.12 Performance of Non-Elementary Exact Branch-and-Price . . . . . . . . . . . . . 228

A.1 LP relaxation bounds with simple valid inequalities: 20-cust. Inst. . . . . . . . . 249

A.2 LP relaxation bounds with simple valid inequalities: 25-cust. Inst. . . . . . . . . 250

A.3 LP relaxation bounds with simple valid inequalities: 40-cust. Inst. . . . . . . . . 251

A.4 Performance of 1p-ESPPRC, 1p-ESPPRC-NwD, 2p-ESSPRC: 25-customer in-

stances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

A.5 Performance of 1p-ESPPRC, 1p-ESPPRC-NwD, 2p-ESSPRC: 40-customer in-

stances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

A.6 Performance of 2p-ESSPRC and 2p-ESSPRC40 : 25-customer instances . . . . . 254

A.7 Performance of 2p-ESSPRC and 2p-ESSPRC40 : 40-customer instances . . . . . 255

A.8 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 25-customer instances . . . . 256

A.9 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 25-customer instances, Cont. 257

A.10 Performance with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 40-customer instances . 258

A.11 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 40-customer instances, Cont. 259

A.12 HBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS . . . . . . . . . . . . . . . . . . 260

A.13 2S-EBP with different designs of HBP . . . . . . . . . . . . . . . . . . . . . . . 261

xi
A.14 Performance of I-Heur, BB, HBP: 25 customer instances . . . . . . . . . . . . . 262

A.15 Performance of I-Heur, BB, HBP: 40-customer instances . . . . . . . . . . . . . 263

A.16 I-Heur, BB, HBPs Upper Bounds vs. LP Relaxations & Optimal (Best) IPs: 25-

customer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264

A.17 I-Heur, BB, HBPs Upper Bounds vs. LP Relaxations & Optimal (Best) IPs: 40-

customer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265

A.18 Performance of BB at Root Node of 1S-EBP: 25-customer . . . . . . . . . . . . 266

A.19 Performance of BB at Root Node of 1S-EBP: 40-customer . . . . . . . . . . . . 267

A.20 1S-EBP vs 2S-EBP: Effect of HBP, 25-Customer Instances . . . . . . . . . . . . 268

A.21 1S-EBP vs 2S-EBP: Effect of HBP, 40-customer instances . . . . . . . . . . . . 269

A.22 Cover and GAP Inequalities for 25-customer Instances . . . . . . . . . . . . . . 270

A.23 Cover and GAP Inequalities for 25-customer Instances: Cont. . . . . . . . . . . . 271

xii
List of Figures

1.1 Outline of the Cutting Plane Algorithm . . . . . . . . . . . . . . . . . . . . . . . 10

1.2 Outline of the Branch-and-Bound Algorithm . . . . . . . . . . . . . . . . . . . . 16

1.3 Outline of the Branch-and-Cut Algorithm . . . . . . . . . . . . . . . . . . . . . 17

1.4 Outline of the Column Generation Algorithm . . . . . . . . . . . . . . . . . . . 19

1.5 Outline of the Branch-and-Price Algorithm . . . . . . . . . . . . . . . . . . . . 26

1.6 Outline of the Branch-Cut-and-Price Algorithm . . . . . . . . . . . . . . . . . . 28

2.1 Example of pairings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.1 Example of a network and a path from source to sink (dashed line) . . . . . . . . 80

3.2 Demonstration of Phase 1 and Phase 2 Networks on a Small Example . . . . . . 89

3.3 Outline of the Initial Heuristic I-Heur . . . . . . . . . . . . . . . . . . . . . . . 105

3.4 Column Generation Algorithm for 1S-EBP and 2S-EBP . . . . . . . . . . . . . . 107

3.5 Column Generation Algorithm for HBP . . . . . . . . . . . . . . . . . . . . . . 109

3.6 Variants of the Branch-and-Price Algorithm . . . . . . . . . . . . . . . . . . . . 110

4.1 Heuristic Separation Algorithm for GAP inequalities [de Farias and Nemhauser

(2001)] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.2 Outline of the Modified Heuristic Separation Algorithm for GAP inequalities for

the LRSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4.3 Outline of the Disjunctive Cut Generation Procedure for Branch-Cut-and-Price . 140

xiii
5.1 Performance profiles for LP relaxation gap of (SPP-LRSP), (SPP-LRSP1 ), (SPP-

LRSP2 ), (SPP-LRSP3 ) and (ASPP-LRSP) . . . . . . . . . . . . . . . . . . . . . 158

5.2 Performance profile for the root node solution time using 1p-ESPPRC, 1p-

ESPPRC-NwD and 2p-ESPPRC . . . . . . . . . . . . . . . . . . . . . . . . . . 161

5.3 Performance profile for the root node solution time using 2p-ESPPRC and 2p-

ESPPRC with multiple columns (20, 40 &100 columns per iteration) . . . . . . . 163

5.4 Performance profile for the total time of 1S-EBP employing only 2p-ESPPRC or

performing one or two of 2p-ESPPRC-LL and 2p-ESPPRC-CS before 2p-ESPPRC166

5.5 Performance profile for total time of HBP employing only 2p-ESPPRC-LL, and

both 2p-ESPPRC-LL and 2p-ESPPRC-CS with U = 12 or 20 . . . . . . . . . . 168

5.6 Performance profile for the gap between zU B & best known IP . . . . . . . . . . 169

5.7 Performance profile for execution time of 2S-EBP . . . . . . . . . . . . . . . . . 171

5.8 Performance profile for the optimality gap in 2S-EBP, log2 . . . . . . . . . . . . 171

5.9 Performance profile for the execution time in HBP with and without BB . . . . . 173

5.10 Performance profile for the gap between zU B , and the LP bound and the optimal

solution (IP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

5.11 Performance profile for the execution time of 1S-EBP with and without BB at the

root node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

5.12 Performance profile for the optimality gap in 1S-EBP with and without BB for

40-customer instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

5.13 Performance profile for the total solution time of 1S-EBP and 2S-EBP . . . . . . 178

5.14 Performance profile for the optimality gap of 1S-EBP and 2S-EBP, log2 scale . . 178

5.15 Performance profile comparing 1S-EBP without any cuts, with CUTGENK , CUTGENG

and CUTGENK+G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

5.16 Performance profile comparing 1S-EBP without any cuts, with DISJCUTGENI1 ,

DISJCUTGENI2 , DISJCUTGENI3 , DISJCUTGENI4 . . . . . . . . . . . . . . 191

xiv
5.17 Performance profile for the cut generation time with DISJCUTGENI1 , DISJCUTGENI2 ,

DISJCUTGENI3 , DISJCUTGENI4 . . . . . . . . . . . . . . . . . . . . . . . . 192

5.18 Performance profile comparing the execution time of 1S-EBP for LRSP and LRPDC198

5.19 % Improvement in the total cost with LRSP compared to LRPDC + BPP . . . . . 201

5.20 # of vehicles with LRPDC, LRPD + BPP and LRSP . . . . . . . . . . . . . . . 203

5.21 Performance profile comparing the execution time of LRSP and Sequential opti-

mization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

5.22 % Improvement in the total cost with LRSP compared to sequential optimization 206

5.23 LRSP versus sequential optimization . . . . . . . . . . . . . . . . . . . . . . . . 206

6.1 The main steps of R-COLGEN(m) . . . . . . . . . . . . . . . . . . . . . . . . . 217

xv
Abstract

In this research, we investigate location, routing and scheduling problems, a class of problems

in which the decisions of facility location, vehicle routing, and route assignment are optimized

simultaneously. The problem considers the interdependency between facility location and vehicle

routing decisions and also allows the assignment of multiple routes to a single vehicle, resulting

in a solution that requires fewer vehicles. We refer to this integrated problem, which is in the

complexity class NP-hard, as the location routing and scheduling problem (LRSP).

For a version of this problem with capacitated facilities and time- and capacity-limited ve-

hicles, we present two formulations, one graph-based and one set partitioning-based. We derive

the theoretical relationship between these two formulations by applying Dantzig-Wolfe decom-

position to the graph-based formulation. We show the equivalence of the set partitioning-based

model and the reformulation of the graph-based formulation obtained by Dantzig-Wolfe decom-

position. In addition, we computationally compare the bounds yielded by the LP relaxations of

the different formulations of the LRSP and demonstrate the strength of the LP relaxation of the

set partitioning-based model and some simple valid inequalities introduced for this model.

By adopting a column generation approach derived from the set partitioning-based formulation

of the problem, we provide exact and heuristic solution algorithms for the problem. We describe

two versions of a branch-and-price algorithm, a standard one-stage version and a two-stage ver-

sion. The two-stage version is an extended variant of the one-stage algorithm based on the idea

of ”restart.” It has an initialization phase that provides a good upper bound and a pool of effective

1
columns. We evaluate and computationally demonstrate the performance of the algorithms we

design using instances up to 40 customers and various parameter settings. We extend our compu-

tational study in order to assess the benefit of integrating optimization of facility location, vehicle

routing and vehicle scheduling decisions by comparing the LRSP with some variations of the se-

quential optimization of these decisions. We obtain results confirming our interest in solution of

the LRSP.

We extend the branch-and-price algorithm to include methods for dynamically generating

valid inequalities. We investigate classes of inequalities valid for polytopes related to the LRSP

polytope; discuss their validity for the problem and their integration into the branch-and-price

algorithm. In addition, we explore possible ways of adapting the approach of generating valid

inequalities using the disjunctive procedure. We provide a methodology to derive disjunctive in-

equalities in a branch-and-price algorithm and describe some implementations of the methodology

for our algorithm. We computationally compare these valid inequalities and discuss their effect to

the performance of the branch-and-price algorithm.

Finally, we discuss how we can modify the presented solution approaches for a location and

routing problem (LRP), which is a special case of the LRSP. We present computational results for

instances up to 40 customers and for instances of certain special cases previously considered in

the literature to demonstrate the effectiveness of the solution approach empirically.

2
Chapter 1

Introduction

Logistics consists of the planning and controlling of the transportation and storage activities in

a system that produces and distributes goods and/or services. It is one of the most important

functions of many companies. The design of a distribution system begins with the questions of

where to locate the facilities and how to allocate customers to the selected facilities. Furthermore,

determining how to distribute goods to customers from these facilities and how to use vehicles

and/or distribution resources efficiently are important decisions that arise in the design of logistics

systems.

In this research, we focus on a class of problems that integrates the decisions of facility lo-

cation, vehicle routing, and route assignment and seeks to minimize the total cost. We refer to

such problems as location, routing and scheduling problems (LRSPs). The LRSP generalizes and

subsumes several well-studied problem classes, such as the location and routing problem (LRP),

the vehicle routing problem (VRP), the multi-depot VRP (MDVRP), and the VRP with multiple

trips (VRPMT).

Given a set of candidate facility locations and a set of customer locations, the objective of the

LRSP is to select a subset of the facilities, construct a set of delivery routes, and assign routes to

vehicles in such a way as to minimize total cost, subject to satisfaction of the system’s constraints.

The total cost may include both fixed costs and operating costs of facilities and vehicles.

3
1.1. MOTIVATION

The LRSP arises in the context of many delivery problems, such as gas, oil, food, beverage,

bill and newspaper delivery problems. In practical applications of the LRSP, various constraints

need to be considered. For example, facilities and vehicles may have capacities, the length of

routes may be restricted with time limits, service of customers may include both deliveries and

pick-ups, service to customers might be restricted to occur within time windows, different types of

vehicles might be used, customers might have priorities, and some data such as customer demands

or travel times might be uncertain. In this research, we consider an LRSP with capacity constraints

on the facilities and on the vehicles, as well as time constraints on the vehicles. We assume that

all data is deterministic, and the problem is either purely a pick-up problem or purely a delivery

problem.

1.1 Motivation

The motivation behind the integration of the facility location, vehicle routing and vehicle schedul-

ing decisions is to improve the efficiency of the system and yield more accurate and cost-effective

solutions. In the literature, it was realized early that the location decision and the routing decision

are interdependent when customers demand less-than-truckload quantities (Webb (1968) and Salhi

and Rand (1989)). Therefore, if it is possible to serve customers through routes making multiple

stops, the assumption of single-customer routes in location models will not accurately capture the

transportation cost. Gas and oil delivery systems, distribution of goods to retailers, and newspaper

and bill delivery systems are examples of such systems that require routes with multiple stops.

Location and routing models seek to minimize total cost by simultaneously selecting a set

of facilities and constructing a set of delivery routes that satisfy the specified system constraints.

LRPs, however, implicitly assume that each vehicle covers exactly one route. This may potentially

overestimate the number of vehicles required and the associated distribution cost. In many cases,

it is possible to serve multiple routes with a single vehicle, in which case the decision of assigning

routes to vehicles becomes interdependent with the location and routing decisions. Eilon (1977)

4
1.1. MOTIVATION

reports that fixed costs are equal to the eighty percent of total vehicle costs. Therefore, mod-

els integrating the scheduling decision and considering the fixed costs of the vehicles may yield

more cost efficient solutions. In Section 5.6, we empirically demonstrate the value of integrat-

ing scheduling decision with the location and routing decisions by comparing the LRSP solution

with the LRP solution and the solution to the optimization of scheduling decision based on the

resulting LRP solution. Furthermore, some systems require different types of vehicles, which may

have different costs than the multi-purpose vehicles. In such cases, the number of customized

vehicles is an important decision for the company. Similarly, if the service to the customer re-

quires a trained worker who is more expensive than a regular driver, then the company will want

to consider training costs, which are also fixed costs.

Why is capturing the costs more accurately important? Major components of logistics cost are

warehousing costs including inventory carrying costs and transportation costs. Based on estimates

for the U.S. in 2002, transportation costs were $577 billion, and inventory carrying costs and

warehousing costs were $298 billion. The total logistics costs were $910 billion, which was

equivalent to 8.7% of the U.S. gross domestic product in 2002 (MacroSy Research and Technology

(2005)). Undeniably, logistics costs are very important for the U.S. economy as well as for a

company. Since the contribution of logistics activities to the GDP is large, improvements in the

efficiency of the logistics activities will contribute to the growth of the U.S. GDP. Even a small

amount of improvement in distribution systems can result in a substantial amount of savings.

Another motivation to consider scheduling decision together with the location and routing

decisions is to be able include more system constraints to the model. In distribution systems in

which the company has a constant fleet size, the number of vehicles should be considered in the

model. To be able to cover all customers, it might be necessary to use a vehicle for more than one

route. Therefore, the decision of scheduling of vehicles must be considered with the decisions of

location and routing. In some systems, there are time constraints for the customers or the vehicles.

Delivery of some items such as perishable foods, newspapers and bills are time sensitive. In

addition, the overtime for the workers might be expensive, hence the company is not willing to

5
1.2. BACKGROUND

pay overtime driver hours. Therefore, while designing the model, time limit for the vehicle routes

must be considered.

1.2 Background

In this section, we provide some definitions, briefly explain some terminology and the associated

notation that is used in this thesis. The following background information was prepared based on

several sources, mostly from Nemhauser and Wolsey (1989), Linderoth (2005), Thomas (2005)

and Perregaard and Balas (2001).

1.2.1 Notation

We denote the sets of all (nonnegative) real numbers, integers and rational numbers with symbols

R (R+ ), Z (Z+ ) and Q (Q+ ), respectively. The set of all (nonnegative) real vectors with dimension

n for some positive scalar n is represented by Rn (Rn+ ). Similar notation is used to define sets of

all integer and rational vectors.

We use uppercase mathematical (italic) letters such as A, B, I, J to denote matrices and sets.

When we introduce such notation, we indicate what it represents. We use lowercase mathematical

(italic) letters such as d, n, s, x, for scalars and vectors. The notation c> is used to represent the

transpose of vector c, and the notations bdc and dde are used to represent the floor and the ceiling

of the scalar d. Uppercase calligraphic characters such as P, K, are used to denote polyhedra and

polyhedral sets. The notation conv(C) represents the convex combination of all points in set C

(i.e., the convex hull of C). We use notation (α, β) where α ∈ Rn and β ∈ R to represent an

inequality in the form of αx ≥ β for some vector x ∈ Rn .

1.2.2 Polyhedra

A set R ⊆ Rn is called convex if for every x, y ∈ R and for all λ ∈ [0, 1], λx + (1 − λ)y ∈ R. Let
P
x1 , x2 , ..., xk ∈ Rn . The vector ki=0 λi xi is called a convex combination of vectors x1 , x2 , ..., xk

6
1.2. BACKGROUND

Pk
for some λ ∈ Rk+ satisfying i=0 λi = 1. The set of all convex combinations of members of set

R ⊂ Rn is called convex hull of R, i.e., conv(R).

A set P ⊆ Rn of the form {x ∈ Rn |Ax ≥ d} for a matrix A ∈ Rm×n and a vector d ∈ Rm

is called a polyhedron. All polyhedra are convex sets. A polyhedron P is said to be bounded if

|x| ≤ k is true for some constant k and for all x ∈ P. A bounded polyhedron is called a polytope.

A valid inequality for P = {x ∈ Rn |Ax ≥ d} is a pair (α, β), where α ∈ Rn is the coefficient

vector and β ∈ R is the right-hand side such that αx ≥ β for all x ∈ P. If (α, β) is a valid

inequality for P, F = {x ∈ P|αx = β} is called a face of P. If F is nonempty and F 6= P, then

F is called a proper face of P. Maximal proper faces are called facets. A proper face is a facet

if its dimension is one less than the dimension of the polyhedron. The dimension of a polyhedron

is equal to the maximum number of affinely independent points contained in the polyhedron. The

points x1 , x2 , ..., xk ∈ Rn are called affinely independent if λ = 0 is the unique solution to

Pk i
Pk k
i=0 λi x = 0 and i=0 λi = 0 for λ ∈ R .

1.2.3 Linear Programming

A linear program (LP) is the problem of minimizing (or maximizing) a linear objective func-

tion subject to a set of linear constraints. To be more precise, the general form of an LP is

zL =min{c> x|Ax ≥ d}, where A ∈ Qm×n is the constraint matrix, c ∈ Qn is the objective

function vector and d ∈ Qm is the right hand side vector. The vector x ∈ Rn is called the vector

of decision variables. Any vector x̂ ∈ PL is called a feasible solution to the LP, where PL is the

underlying polyhedron PL = {x ∈ Rn |Ax ≥ d} (also called the feasible region).

The LP is said to be infeasible if PL is empty. In that case, we set zL = ∞. The LP is said

to be unbounded if zL = −∞. If it is not infeasible or unbounded, the LP has a finite optimal.

zL is called the optimal objective (function) value and xL ∈ PL is called an optimal solution if

zL = c> xL ≤ c> x for all x ∈ PL .

The LP zD = max{y > d|y > A = c, y ≥ 0} is called the dual of min{c> x|Ax ≥ d} and

y ∈ Rm is called the dual vector. Let PD = {y ∈ Rm >

+ |y A = c}. For ŷ ∈ PD , vector c − ŷA

7
1.2. BACKGROUND

is called the reduced cost vector corresponding to variable vector x ∈ PL and ŷ > d ≤ c> x for

all x ∈ PL . This property of LPs is referred to as the weak duality property in the literature.

Furthermore, if an LP has a finite optimal value, then so does the associated dual LP and these

optimal values are equal (referred to as strong duality property). If zL = −∞, then zD = −∞

(infeasible). If zD = ∞ (unbounded), then zL = ∞. Both the primal and dual LPs can also be

infeasible.

1.2.4 Mixed Integer Linear Programming

A mixed integer linear program (MILP) is the problem of minimizing or maximizing a linear

objective function subject to a set of linear constraints and integrality restrictions on some or all of

the variables. More specifically, an MILP model (which we refer to as (P) in the rest of the thesis)

can be formulated as follows

zIP = Minimize c> x

(P) subject to Ax ≥ d,
n−r
x ∈ Zr+ × R+ ,

where A ∈ Qm×n is the constraint matrix, c ∈ Qn , d ∈ Qm are rational vectors and x ∈

Zr+ × Rn−r
+ is a vector of decision variables. Without loss of generality, variables that are indexed

from 1 to r are called integer variables, and those indexed from n−r+1 to n are called continuous

variables. If P = {x ∈ Rn+ |Ax ≥ d} is the polyhedron defined by the linear constraints of (P),

the set S = {x ∈ Zr × Rn−r |x ∈ P} is called the feasible solution set of (P) and x ∈ S is called

a feasible solution to (P).

If S = ∅, then (P) is said to be infeasible, and we set zIP = ∞. Otherwise (P) is said to

be feasible. If zIP = −∞, (P) is said to be unbounded. If (P) is feasible and zIP is finite, than

(P) has a finite optimal. zIP is the optimal objective value and xIP ∈ S is an optimal solution if

zIP = c> xIP ≤ c> x for all x ∈ S.

8
1.2. BACKGROUND

If r = n, then (P) is called a (pure) integer linear program (ILP). The LP that is obtained by

relaxing all integrality restrictions on the variables in (P) is called the LP relaxation of (P). We

refer to the value of the LP relaxation of (P) as zLP (i.e., zLP = minx∈P c> x) and the solution

vector as xLP ∈ P.

1.2.5 Cutting-Plane Procedure

The cutting-plane procedure is a methodology used to obtain a lower bound for the optimal solu-

tion value of an MILP by generating approximations of the feasible solution set associated with

the MILP. Consider the MILP (P). The algorithm iteratively provides approximations of the fea-

sible solution set S as the intersection of half-spaces associated with inequalities valid for S. A

cutting plane is a valid inequality that separates a given solution, say x̄ ∈ P, such that x̄ ∈
/ S from

S. In other words, if (α, β) is such a cutting plane, then αx ≥ β for all x ∈ S and αx̄ < β. The

problem of determining whether x̄ ∈ S or finding a valid inequality (α, β) such that αx̄ < β is

called the separation problem. In the standard procedure, P is the initial approximation of S and

zLP is the corresponding lower bound for zIP . The algorithm starts by generating a valid inequal-

ity (α, β) that is violated by xLP , the solution to the LP minx∈P c> x. The polyhedron obtained by

intersecting the half-space associated with (α, β) and P is a better approximation of S and may

provide a better bound for zIP . This process continues iteratively until no violated valid inequality

is obtained or a solution of S is found.

In order to present an outline of the algorithm, we define some notation. Let zLB represent

the lower bound on the optimal objective function value of the MILP that is being handled in the

algorithm. P i denotes the underlying polyhedron at iteration i of the cutting-plane algorithm; Pi

denotes the MILP defined with feasible region P i and the integrality restrictions for the variables;

xiLP and zLP

i denote the solution and the objective function value of the LP relaxation of Pi ,

respectively; and (αi , β i ) is the valid inequality obtained by solving the separation problem for

xiLP . An outline of a cutting-plane algorithm is presented in Figure 1.1. Next, we briefly describe

some commonly used classes of valid inequalities that are employed in our study.

9
1.2. BACKGROUND

CUTTING-PLANE
Input: An instance (P).
Output: A lower bound zLB for zIP .

0. Set i ← 0 and zLB ← −∞. Let P 0 ← P.

1. Solve the LP relaxation of Pi with feasible region P i . Set zLB ← zLP

i .

i
• If zLP = ∞, then Pi is infeasible, STOP.
i
• If zLP = −∞, then Pi is unbounded, STOP.
• If xiLP ∈ S, then xiLP is the optimal solution and zLP
i is the optimal solution
value. STOP.
• Otherwise, go to next step.

2. Solve the separation problem for xiLP .

• If a violated inequality (αi , β i ) is obtained, set P i+1 ← P i ∩{x ∈ Rn |αi x ≥ β i }

and i ← i + 1. Return to step 1.
• Otherwise, STOP.

Figure 1.1: Outline of the Cutting Plane Algorithm

Chvátal-Gomory (CG) Inequalities. Chvátal-Gomory (CG) valid inequalities (described by

Gomory (1963) and Chvátal (1973)) are a class of valid inequalities that are obtained by taking

linear combinations of the inequalities in the current model and applying a rounding procedure.

Using the integrality of the variables, the rounding procedure leads to integer right and left hand

sides in any inequality. Consider the MILP (P) with r = n. A CG inequality is a valid inequality

for S which is of the form dλ> Aex ≥ dλ> de, where λ ∈ Rn+ . The CG rank of any original

inequality and any inequality dominated by a nonnegative linear combination of the original in-

equalities, e.g., λ> Ax ≥ λ> d for λ ∈ Rn+ , is said to be 0. An inequality that can be generated as

a CG inequality based on rank 0 inequalities but is not rank 0 itself has CG rank 1. Inequalities

with higher rank can be obtained recursively from the lower rank inequalities.

Disjunctive Valid Inequalities. A disjunction is a logical operator that evaluates TRUE when-

ever one or more of its operands evaluates TRUE. In the context of MILP, the disjunctions most

10
1.2. BACKGROUND

commonly arising are those whose operands are systems of linear equalities or inequalities. More

specifically, such a disjunction is defined by s systems over set S:

B 1 x ≥ b1 ∨ B 2 x ≥ b2 ∨ ... ∨ B s x ≥ bs ∀x ∈ S, (1.1)

k ×n k
where B k ∈ Qm is a constraint matrix, bk ∈ Qm is a right hand side vector for some integer

mk > 0, ∀k = 1, .., s. The associated disjunction described by (1.1) results in true if and only if

B k x̂ ≥ bk is true for some s ≥ k ≥ 1 and x̂ ∈ S.

In many solution approaches for general MILPs such as branch-and-cut, the concept of valid

disjunction is fundamental. Consider the MILP (P). The disjunction described by (1.1) is called

valid for (P), if

S ⊆ ∪sk=1 {x ∈ Rn+ |Ax ≥ d, B k x ≥ bk }.

The generation of cutting planes based on disjunctions and the idea of disjunctive programming

were introduced in Balas (1974) and Balas (1979). Generally speaking, disjunctive programming

is finding the optimal value of an objective function over a finite union of polyhedra constructed

based on a combination of conjunctive statements (connected with logical operator ”AND”) and

disjunctive statements. Here, we focus on disjunctive statements. We explain the methodology

based on (P) using the valid disjunction for (P) of the form (1.1).

Using the s systems of inequalities in the valid disjunction, a relaxation of (P) (denoted by

(PDR )) can be defined in the form of a disjunctive program:

Minimize c> x

(PDR ) subject to Ax ≥ d,

∨sk=1 B k x ≥ bk ,

x ≥ 0.

11
1.2. BACKGROUND

We interpret the above program as follows. With term k of the disjunction, we associate a polyhe-

dron.

P k = {x ∈ Rn+ |Ax ≥ d, B k x ≥ bk } ∀k = 1..s.

(PDR ) can then be interpreted as an optimization problem over the union of s polyhedra.

zDR =Minimize c> x

subject to x ∈ PDR ,

where PDR = ∪sk=1 P k . Since (PDR ) is a relaxation of (P), any inequality valid for PDR is also

valid for S. The aim of the disjunctive cut generation methodology is to derive an inequality,

(α, β) that is valid for PDR and is violated by some feasible solutions of the LP relaxation of (P).

Balas (1979) showed that (α, β) is valid for PDR if and only if there exist multipliers uk ∈ Qm
k
and v k ∈ Qm such that

uk A + v k B k ≤ α ∀k = 1, .., s,

uk b + v k bk = β ∀k = 1, .., s,

uk , v k ≥ 0 ∀k = 1, .., s.

Hence, to generate an inequality (α, β) that is maximally violated by xLP and valid for PDR , the

12
1.2. BACKGROUND

following optimization model, called the cut generating LP (CGLP) is solved.

Minimize α> xLP − β (1.2)

(CGLP) subject to uk A + v k B k ≤ α ∀k = 1, .., s, (1.3)

uk b + v k bk = β ∀k = 1, .., s, (1.4)
s X
X m mk
s X
X
uki + vik = 1, (1.5)
k=1 i=1 k=1 i=1
k k
u ,v ≥ 0 ∀k = 1, .., s. (1.6)

The objective function (1.2) seeks to maximize the violation for xLP . If the optimal objective

function value is negative, then xLP violates the inequality (α∗ , β ∗ ) which is the optimal solution.

Constraints (1.3) and (1.4) ensure that any feasible solution (α, β) of this problem is a valid in-

equality for PDR . Constraint (1.5) is added to normalize the cut to be generated. The purpose of

the normalization constraint is to truncate the cone defined by constraints (1.3) and (1.4) to a poly-

tope. The normalization removes the unboundedness of the model and transforms extreme rays

to extreme points. Here, we use one of the common normalization constraints, different versions

of normalization constraints are described in the literature. Some examples are α−normalization
P
(i.e., i=1,..,n |αi | = 1) and β−normalization (i.e.,|β| = 1) (Balas et al. (1996)).

As noted in Caprara and Letchford (2003) (based on the results in Balas et al. (1993) and

Balas et al. (1996))), for a fixed disjunction, separation of disjunctive cuts can be accomplished

in polynomial time by solving the above (CGLP). In general, it is common to use disjunctions

including two terms consisting of linear inequalities. One of the most common and successful

procedures based on this principle is referred to as lift-and-project (see Balas et al. (1993) and

Balas et al. (1996)). In lift-and-project, valid inequalities are derived from disjunctions involving

bounds on single binary variables of the form (xi ≤ 0) ∨ (xi ≥ 1) for some index i of a binary

variable. Lift-and-project cuts are one of the widely-used general-purpose cutting planes for 0-1

mixed integer programming models.

13
1.2. BACKGROUND

If the disjunction used to derive a given cut is of the more general form (πx ≤ π0 ) ∨ (πx ≥

π0 + 1) for some integer vector (π, π0 ) ∈ Zn+1 such that πi = 0 for i = r + 1, .., n (continuous

variables), the cut is called a split cut (Cook et al. (1990) and Balas and Saxena (2008)). Caprara

and Letchford (2003) proved that the separation problem for split cuts is strongly NP-Complete.

Many general-purpose classes of valid inequalities for MILPs are in the broad class of split cuts.

The well-known valid inequalities Gomory (Gomory (1963)) and the mixed integer rounding in-

equalities (Nemhauser and Wolsey (1990)), for example, have been shown to be in the class of

split cuts (Cornuejols and Li (2001) and Cornuejols (2006)).

In addition to these studies that derive valid inequalities from valid disjunctions, Letchford

(2001) showed that many facet-inducing inequalities for well-known combinatorial optimization

problems are actually disjunctive valid inequalities derived from certain simple disjunctions. They

used these results to prove that the separation problems for these facet-inducing inequalities are

solvable in polynomial time. They provided results for some facets of the clique partitioning

problem, the maximum cut problem, the acyclic sub digraph problem, the linear ordering problem,

the asymmetric traveling salesman problem and the set covering problem.

1.2.6 Branch-and-Bound Algorithms

The most widely used class of algorithms for solving MILPs is the branch-and-bound algorithms,

which employ a divide-and-conquer approach. The algorithm divides the set of feasible solutions

into subsets to implicitly enumerate all feasible solutions of the problem. Optimizing the objective

function for each subset results in subproblems of the original problem. Consider the MILP (P).

The basic idea is to partition S into k subsets S1 , ..., Sk and recursively optimize over each subset:

minx∈S c> x = mini=1..k {minx∈Si c> x}.

Partitioning a feasible set into smaller subsets is called branching. To avoid the complete enumer-

ation of all feasible solutions, the algorithm iteratively calculates upper and lower bounds on the

14
1.2. BACKGROUND

optimal objective function value for each subproblem. Any feasible solution x̂ ∈ S to the problem

provides an upper bound. The most common lower-bounding procedure used in the algorithm is

to solve an LP relaxation of the subproblem. In such an LP-based branch-and-bound algorithm,

the LP relaxation of the original problem is first solved to obtain xLP . If xLP ∈ S, then it is

also optimal for (P) and the algorithm terminates. If the LP relaxation is unbounded or infeasible,

then (P) is also unbounded or infeasible. Otherwise, we need to branch and solve the resulting

subproblems, recursively.

The development of the algorithm can be visualized by constructing an associated tree (which

is referred to as branch-and-bound tree) in which each node represents a subproblem. The root

node represents the MILP that is going to be solved and successors represent subproblems obtained

by branching. In order to describe the algorithm and present an outline, we define some notation.

We will use the same notation in the rest of the thesis. Pi denotes the subproblem associated

with node i and defined over subset S i obtained by applying terms of the branching disjunctions

associated with node i to set S. P i denotes the polyhedron defined by the linear constraints of the

subproblem Pi . xiLP and zLP

i denote the LP relaxation solution and objective function value for

subproblem Pi , respectively. For consistency, we let P0 denote the subproblem at the root node,

which is (P) and the set S 0 denotes the feasible solution set S. We use zU B and zLB to represent

the global upper and lower bounds on the optimal objective function value for (P). L represents

the candidate list that includes the subproblems to be processed. The subproblems generated by

partitioning the feasible solution set of another subproblem (say subproblem Pi ) are referred to

as the children of Pi , while Pi is referred to as the parent. An outline of a branch-and-bound

algorithm is presented in Figure 1.2.

1.2.7 Branch-and-Cut Algorithms

Another solution methodology applied to MILP is branch and cut, which is also a divide-and-

conquer approach and is a combination of the branch-and-bound algorithm and the cutting-plane

procedure. The basic idea of the algorithm is the same as the branch-and-bound algorithm. The

15
1.2. BACKGROUND

BRANCH-BOUND
Input: An instance (P).
Output: zIP and xIP .

0. Set zU B ← ∞ or to the value of a known feasible solution, zLB ← −∞, and L ← {P0 }.

1. Select a subproblem Pi from L. Set L ← L\{Pi }.

2. Solve the LP relaxation of Pi with feasible region P i and obtain zLP

i .

i
• If zLP ≥ zU B , go to step 5.
• If xiLP ∈ S, then set zU B ← min{zLP
i ,z
U B } and go to step 5.
• Otherwise, go to next step.

3. Branch (partition S i into subsets). Define subproblems associated with subsets. Add
new subproblems to L.
k |Pk is a parent problem for Pi ∈ L}.
4. zLB = min{zLP

5. If zLB ≥ zU B or L = ∅, then set zIP ← zU B (if there is a solution vector that yields
zU B , then it is xIP ), and STOP. Otherwise, return to step 1.

Figure 1.2: Outline of the Branch-and-Bound Algorithm

feasible solution set is partitioned into subsets and each resulting subproblem is optimized, recur-

sively. Recall that in an LP-based branch-and-bound algorithm, a lower bound is determined by

solving the LP relaxation of each subproblem. In a branch-and-cut algorithm, this lower bound is

strengthen by dynamically generating valid inequalities for each subproblem.

One major challenge in a branch-and-cut procedure is the difficulty of the separation problems

to be solved to generate the valid inequalities. In most cases, heuristic approaches are employed

in the solution of the separation problem. Another challenge is to determine a class of valid

inequalities that would improve the LP relaxation bound most. Effectiveness of the algorithm

depends on the classes of the valid inequalities to be generated and the efficiency of the separation

routines required for those.

The steps of a branch-and-cut algorithm are similar to BRANCH-BOUND except for one addi-

tional step in which valid inequalities are generated. We use the same notation used for BRANCH-

BOUND and define some additional notation. We denote the set of valid inequalities generated

16
1.2. BACKGROUND

BRANCH-CUT
Input: An instance (P).
Output: zIP and xIP .

0. Set zU B ← ∞ or to the value of a known feasible solution, zLB ← −∞, and L ← {P0 }.

1. Select a subproblem Pi from L. Set L ← L\{Pi }.

2. Solve the LP relaxation of Pi with feasible region P i and obtain zLP

i .

i
• If zLP ≥ zU B , go to step 6.
• If xiLP ∈ S, then set zU B ← min{zLP
i ,z
U B } and go to step 6.
• Otherwise, go to next step.

3. If desired, search for valid inequalities violated by xiLP . Let F i be the set of inequalities
generated.

• If F i 6= ∅, set P i ← P i ∩ {x ∈ Rn |Di x ≥ di }. Return to step 2.

• Otherwise, proceed.

4. Branch (partition S i into subsets). Define subproblems associated with subsets. Add
new subproblems to L.
k |Pk is a parent node for problem Pi ∈ L}.
5. zLB = min{zLP

6. If zLB ≥ zU B or L = ∅, then set zIP ← zU B (if there is a solution vector that yields
zU B , then it is xIP ), and STOP. Otherwise, return to step 1.

Figure 1.3: Outline of the Branch-and-Cut Algorithm

for subproblem Pi based on xiLP as F i . If F i 6= ∅, we denote the constraint matrix and the right

hand side vector corresponding to the set of valid inequalities in F i using notation [Di , di ] where
i |×n i
Di ∈ Q|F and di ∈ Q|F | . An outline of a branch-and-cut algorithm is presented in Figure

1.3.

1.2.8 Column Generation Algorithms

Column generation is an approach for solving LPs with a large number of variables for which we

want to avoid explicitly enumerating all columns. By solving a sequence of linear programs and

dynamically generating columns eligible to enter the basis, such an algorithm implicitly considers

17
1.2. BACKGROUND

all columns but generates only those that may improve the objective function. The objective of

the algorithm is to determine a subset of the variables that includes those that are nonzero in some

optimal solution. By solving the restricted version of LP including only this subset of columns

rather than all possible columns, we can obtain the optimal LP solution.

Column generation algorithms have been applied to a wide variety of large-scale optimization

problems, such as vehicle routing, crew scheduling, cutting stock, lot sizing and machine schedul-

ing. Lübbecke and Desrosiers (2005) list other problems to which column generation has been

applied and provide an excellent review of the methodology. To describe the algorithm in more

detail, we consider the following LP.

X
Minimize cq θq
q∈Q
X
(CP) subject to Aq θq ≥ d,
q∈Q
θq ≥ 0 ∀q ∈ Q,

where Q is the index set for all columns, A ∈ Qm×|Q| is the constraint matrix, Aq ∈ Qm is

the vector associated with column q and d ∈ Qm is the right hand side vector. Assume that |Q|

is large and thus enumeration of all columns is inefficient. We refer to (CP) as linear master

problem. In each iteration of the column generation algorithm, a restricted version of (CP) which

is referred to as restricted master problem (RMP) that includes only a subset of columns is solved.

Assuming that the RMP is feasible, by solving RMP, we can obtain a dual solution vector π. Next,

the algorithm searches for a column Aq for some q ∈ Q with the smallest reduced cost ĉq , where

ĉq = cq − π > Aq . This search step is called pricing.

In the pricing steps of the algorithm, for a given dual solution π̄, the following optimization

model is solved.

ẑ = Minimize c(x) − π̄ > x

(SP(π)) subject to x ∈ {Aq |q ∈ Q},

18
1.2. BACKGROUND

COLGEN
Input: An instance (CP) with a set Q0 ⊆ Q of initial columns (containing a feasible solu-
tion).
Output: The optimal solution value of (CP).

0. Let Q0 ⊆ Q be the index set for the initial columns. Set k ← 0.

1. Build RMPk with the columns in Qk .

2. Solve RMPk and obtain πk .

3. Call SP(πk ).

a. If ẑk < 0, then set k ← k + 1, Qk ← Qk−1 ∪ {k}. Go to step 1.

b. Otherwise, the solution for RMPk is an optimal solution of (CP). STOP.

Figure 1.4: Outline of the Column Generation Algorithm

where c(x) is a function of x, and results in cq for the corresponding member Aq . (SP(π)) is

referred to as the pricing problem (or subproblem).

The objective of the pricing problem is to either identify new columns to add to the current

RMP or conclude that the RMP solution is optimal for (CP). From the theory of linear program-

ming, we know that if every variable has a non-negative reduced cost, then the solution is optimal.

If ẑ < 0 for a given π̄, then x̂ that results in the objective ẑ can be added to the RMP. Then,

the updated RMP is re-optimized. The algorithm continues iteratively solving the RMP and the

pricing problem until ẑ ≥ 0 for a dual vector obtained in an iteration of the algorithm.

Before we outline the steps of a column generation algorithm, which we refer to as COLGEN,

we describe some notation. Let Qk be the set of indices of the columns at k th iteration of the

algorithm for k ≥ 0. To ease the description, let RMPk be the restricted version of (CP) including

columns in set Qk at iteration k and let πk be the dual variable vector obtained by solving RMPk

for all k ≥ 0. Similarly SP(πk ) represents the pricing routine for a given πk . Let ẑk and x̂k be

the optimal solution value and an optimal solution vector obtained from SP(πk ), respectively. An

outline of a column generation algorithm is presented in Figure 1.4.

19
1.2. BACKGROUND

1.2.9 Dantzig-Wolfe Decomposition

The Dantzig-Wolfe Decomposition (DWD) approach was developed by Dantzig and Wolfe (1960)

to solve large scale linear programs. The methodology aims to obtain a reformulation of the orig-

inal problem that leads to the iterative solution of many smaller problems instead of the original

problem. DWD has become a widely used method in (mixed) integer linear programming to ob-

tain models with stronger LP relaxation bounds. We first explain the steps of the DWD method

for linear programs and then extend the review for mixed integer linear programming. Consider

the LP:

Minimize c> x

subject to A1 x ≥ d1 ,

A2 x ≥ d2 , (1.7)

x ∈ Rn+ ,

where A1 ∈ Qm1 ×n and A2 ∈ Qm2 ×n are constraint matrices, and d1 ∈ Qm1 and d2 ∈ Qm2 are

rational right hand side vectors. Let X = {x ∈ Rn+ : A2 x ≥ d2 } be the polyhedron defined by

the constraint matrix A2 . In this research, we focus on instances in which X is bounded. We can

describe X as the convex combination of its the extreme points:

X X |Q|
X = {x ∈ Rn+ |x = xq θq , θq = 1, θ ∈ R+ }, (1.8)
q∈Q q∈Q

20
1.2. BACKGROUND

where {xq }q∈Q is the set of extreme points of X with finite index set Q. Based on this alternative

definition of X , model (1.7) can be reformulated.

X
Minimize c> ( xq θq )
q∈Q
X
1
subject to A ( xq θq ) ≥ d1 ,
q∈Q
X
θq = 1,
q∈Q
|Q|
θ ∈ R+ .

If we rearrange and let cq = c> xq and Aq = A1 xq ∀q ∈ Q, we obtain the following model:

X
Minimize cq θq
q∈Q
X
subject to Aq θq ≥ d1 ,
q∈Q
X
θq = 1, (1.9)
q∈Q
|Q|
θ ∈ R+ .

As in the previous section, model (1.9) is referred to as linear master problem. To find the solution

of problem (1.9), we have to explore the set of extreme points of X , which is possibly large. The

COLGEN procedure can be employed to solve problem (1.9).

As we noted before, the DWD methodology can be used to obtain a stronger lower bound for

an MILP. Consider the MILP, which is obtained by adding integrality restrictions to model (1.7).

21
1.2. BACKGROUND

In particular,

Minimize c> x

(D) subject to A1 x ≥ d1 , (1.10)

A2 x ≥ d2 , (1.11)
n−r
x ∈ Zr+ × R+ , (1.12)

for some integer n ≥ r > 0. In the context of mixed integer programming, the DWD method aims

to get a possibly better lower bound than the LP relaxation of (D) by optimizing the intersection

of the polyhedron defined by constraints (1.10) with the convex hull of constraints (1.11) and

(1.12). Based on this idea, we are interested in the convex hull of constraints (1.11) and (1.12).

Let X = {Zr+ × Rn−r

+ : A2 x ≥ d2 } be the set of feasible solutions defined by constraints (1.11)

and (1.12), and let conv(X) be the convex hull of X. We can express conv(X) using the convex

combination of the solutions in X:

X X |E|
conv(X) = {x ∈ Rn |x = xq θq , θq = 1, θ ∈ R+ }, (1.13)
q∈E q∈E

where {xq }q∈E is the set of points in X and E is the index set. We can reformulate (D) by

substituting x based on (1.13). The LP relaxation of this reformulation is as follows:

X
Minimize (c> xq )θq
q∈E
X
subject to (A1 xq )θq ≥ d1 ,
q∈E
X
θq = 1, (1.14)
q∈E
|E|
θ ∈ R+ .

The optimal solution value of problem (1.14) is an upper bound for the LP relaxation of (D). If all

22
1.2. BACKGROUND

extreme points of X satisfy integrality restrictions, then the LP relaxation of (D) is equal to the

optimal solution of problem (1.14). If we aim to find feasible solutions to (D) using the reformula-

tion (1.14), we also need to satisfy the integrality restrictions on x. Let θ̄ be a solution of problem
P
(1.14). When we transform this into x space, the corresponding solution, x̄ = q∈E xq θ̄q , may

not satisfy (1.12). In order to satisfy integrality constraints, two modeling approaches are proposed

(see Vanderbeck (1994)). The first one is called discretization in which integrality restrictions are

added for (θ):

X
Minimize (c> xq )θq
q∈E
X
subject to (A1 xq )θq ≥ d1 ,
q∈E
X
θq = 1, (1.15)
q∈E
|E|
θ ∈ Z+ .

The second approach is called convexification in which the integrality conditions are forced for

the variables in the original space (x).

X
Minimize (c> xq )θq
q∈E
X
subject to (A1 xq )θq ≥ d1 ,
q∈E
X
θq = 1,
q∈E
|E|
θ ∈ R+ , (1.16)
X
x = xq θq ,
q∈E
n−r
x ∈ Zr+ × R+ .

Problems (1.15) and (1.16) are the reformulations of (D) using the DWD procedure and in general

23
1.2. BACKGROUND

referred to as integer master problems. In order to solve problems (1.15) and (1.16), the branch-

and-price algorithm is used.

1.2.10 Branch-and-Price Algorithms

Another divide-and-conquer approach that is used to solve MILP models is branch and price,

which is a combination of the branch-and-bound algorithm and the column generation algorithm.

As in the branch-and-bound algorithm, the original feasible solution set is partitioned into subsets

and each subset is recursively treated by applying bounding and branching. Similar to an LP-based

branch-and-bound algorithm, a lower bound is determined by solving the LP relaxation of each

subproblem. However, because of the formulation of the original problem, for each subproblem,

column generation is employed to obtain the LP relaxation bound. After the LP relaxation solution

is obtained, the branching procedure is applied.

The algorithm has been applied to a variety of problems, including various vehicle routing

problems (Fukasawa et al. (2006)), location routing problems (Berger et al. (2007)) and airline

crew scheduling problems (Vance et al. (1997)), all of which are relevant in the context of the

LRSP. Although branch and price has been applied successfully to various problems, there are

many difficulties associated with the algorithm and the success of the algorithm highly depends

on the structure of the problem. Barnhart et al. (1998) discuss the general methodology and the

many challenges in developing a branch-and-price algorithm.

One major challenge in a branch-and-price algorithm is the efficiency of the column genera-

tion algorithm used. Since the column generation algorithm is applied for each subproblem, the

efficiency of the column generation procedure has a major effect on the overall performance of

the algorithm. Another challenge is the determination of the branching disjunctions. To solve the

LP relaxation of each subproblem which is obtained by applying terms of the branching disjunc-

tion associated with the subset, we need to use the column generation algorithm that will provide

columns feasible for the associated subset of the original feasible solution set. In order to do that,

24
1.2. BACKGROUND

the pricing problem in the column generation algorithm may need to be modified. These modifi-

cations may result in more difficult pricing problems. Therefore, in branch-and-price algorithms,

the branching decisions are crucial because of their effect to the pricing problem. They may effect

the performance of the overall algorithm significantly.

Next, we present an outline of a branch-and-price algorithm. We use the same notation that is

used to describe BRANCH-BOUND and define some additional notation. We define set CG global

to the algorithm to represent the set of columns generated so far. We use notations (RMPik ) to

denote the restricted version of the LP relaxation of Pi at the k th iteration of the column generation

algorithm. We let πki represent the dual variable vector obtained by solving RMPik and SP(πki )

represent the pricing routine for a given πki . SP(πki ) results in the set of generated columns that is

denoted by Cki . We let xik denote the solution of RMPik and zik be the solution value corresponding

to xik . If Cki = ∅, then zki = zLP

i . An outline of a branch-and-price algorithm is presented in Figure

1.5.

1.2.11 Branch-Cut-and-Price Algorithms

A branch-cut-and-price algorithm combines the branch-and-price algorithm with a branch-and-cut

algorithm. In a fashion similar to the branch-and-cut procedure, valid inequalities are optionally

generated for each subproblem to strengthen the LP relaxation bound obtained via column gen-

eration algorithm. The cut generation procedure may be done before, after or during the column

generation algorithm. However, at any node, after a set of valid inequalities are added to the re-

stricted version of the subproblem, the column generation algorithm must be called again to solve

or find a lower bound for the LP relaxation of the restricted version of the subproblem extended

with valid inequalities.

As in a branch-and-cut procedure, the decision to employ a given class of valid inequalities

depends on factors such as the efficiency with which the separation problem can be solved and

the degree of improvement on the LP relaxation bound expected by adding these inequalities. In

addition to these factors, in a branch-cut-and-price algorithm, the structure of the valid inequalities

25
1.2. BACKGROUND

BRANCH-PRICE
Input: An instance (P) with initial column set C 0 .
Output: zIP and xIP .

0. Set zU B ← ∞ or to the value of a known feasible solution, zLB ← −∞, CG ← C 0 and

L ← {P0 }.

1. Select a subproblem Pi from L. Set L ← L\{Pi }.

2. Set k ← 0.

3. Build RMPik with the columns in CG and solve. Call SP(πki ). Set CG ← CG ∪ Cki .

(a) If Cki = ∅, set zLP

i ← zki .
i
• If zLP ≥ zU B , go to step 6.
i i ,z
• If xLP ∈ S, then set zU B ← min{zLP U B } and go to step 6.
• Otherwise, go to step 4.
(b) Set k ← k + 1 and return to step 3.

4. Branch (partition S i into subsets). Define subproblems associated with subsets. Add
new subproblems to L.
k |Pk is a parent node for problem Pi ∈ L}.
5. zLB = min{zLP

6. If zLB ≥ zU B or L = ∅, then set zIP ← zU B (if there is a solution vector that yields
zU B , then it is xIP ), and STOP. Otherwise, return to step 1.

Figure 1.5: Outline of the Branch-and-Price Algorithm

may also be important, since, as with the branching decisions, they may affect the structure of the

pricing problem. If generation of a class of valid inequalities significantly affects the efficiency of

the pricing algorithm, employment of these inequalities may not be desirable, as the gain from the

lower bound improvement may not be offset by the overall performance decline because of the

modifications to the pricing algorithm.

In order to outline a branch-cut-and-price algorithm, we use the notation defined before for

BRANCH-BOUND and BRANCH-PRICE, and define some additional notation. We denote the

set of valid inequalities generated for subproblem Pi based on xik as Fki . If Fki 6= ∅, we denote the

constraint matrix and the right hand side vector corresponding to the set of valid inequalities in Fki

26
1.3. RELATED WORK

i i
using notation [Dki , dik ] where Dki ∈ Q|Fk |×n and dik ∈ Q|Fk | . An outline of a branch-cut-and-price

algorithm is presented in Figure 1.6.

1.3 Related Work

The LRSP integrates facility location, vehicle routing and vehicle scheduling decisions. In this

section, we briefly review the literature for each component problem as well as the literature for

related integrated problems.

1.3.1 Facility Location

The first component of the LRSP is the question of where to locate the facilities and how to allocate

the customers to these facilities. Determining where to locate facilities is an important problem

that arises in the design of distribution networks. ReVelle and Eiselt (2005) and Klose and Drexl

(2005) provide recent surveys of the various types of location problems, formulations and solution

algorithms. The facility location literature encompasses a wide variety of problems, which can be

categorized in a variety of ways. Daskin (1995) provides a taxonomy of facility location problems

and describes methods for categorizing location problems. Location problems can be grouped

as planar, network, or discrete problems, based on the allowable location for the facilities and

the customers. Problems in which customers and facilities can be located anywhere on a plane

are planar location problems. Problems in which the customers and the candidate locations are

restricted to a discrete set are discrete location problems. Network location problems are a special

case of discrete location problems in which the candidate facilities and the demand locations are

restricted to the nodes of a network and travel is only possible via the arcs of network.

Most relevant to our study of the LRSP are discrete location problems. Labbé and Louveaux

(1997) provide a survey of the literature related to discrete facility location problems including ap-

plication areas, solution methods and model extensions. Early work on discrete location problems

started with Hakimi (1964, 1965), who introduced the p-median problem on a network, which is

27
1.3. RELATED WORK

BRANCH-CUT-PRICE
Input: An instance (P) with initial column set C 0 .
Output: zIP and xIP .

0. Set zU B ← ∞ or to the value of a known feasible solution, zLB ← −∞, CG ← C 0 ,

and L ← {P0 }.

1. Select a subproblem Pi from L. Set L ← L\{Pi }.

2. Set k ← 0.

3. Build RMPik with the columns in CG and solve. Call SP(πki ). Set CG ← CG ∪ Cki .

(a) If Cki = ∅, set zLP

i ← zki .
i
• If zLP ≥ zU B , go to step 7.
i i ,z
• If xLP ∈ S, then set zU B ← min{zLP U B } and go to step 7.
• Otherwise, go to next step.
(b) If cut generation is desired, go to step 4.
(c) If Cki = ∅ go to step 5. Otherwise, set k ← k + 1 and return to step 3.

4. Generate valid inequalities violated by xik . Let Fki be the set of inequalities generated.

• If Fki 6= ∅, set P i ← P i ∩ {x ∈ Rn |Dki x ≥ dik }. Return to step 2.

• If Cki 6= ∅, return to step 3.
• Otherwise, proceed.

5. Branch (partition S i into subsets). Define subproblems associated with subsets. Add
new subproblems to L.
k |Pk is a parent node for problem Pi ∈ L}.
6. zLB = min{zLP

7. If zLB ≥ zU B or L = ∅, then set zIP ← zU B (if there is a solution vector that yields
zU B , then it is xIP ), and STOP. Otherwise, return to step 1.

Figure 1.6: Outline of the Branch-Cut-and-Price Algorithm

28
1.3. RELATED WORK

known to be NP-hard (Kariv and Hakimi (1979)). The objective is to locate p facilities on a net-

work with demand nodes so as to minimize the total weighted distance between the customers and

the closest facilities to them. ReVelle and Swain (1970) were the first to formulate the p-median

problem as a 0-1 integer program. Since then, researchers have developed both exact methods

(e.g., Galvao (1993), Bilde and Krarup (1977)) and heuristic methods (e.g., Maranzana (1964),

Teitz and Bart (1968), Rosing and ReVelle (1997)) for the p-median problem. Depending on the

difficulty of the instances, optimal or near optimal solutions for p-median problems are reported

for instances up to 3000 customer nodes.

By removing the constraint that p locations must be selected and by adding a fixed cost for

each candidate location, we obtain a simple plant location problem from the p-median problem.

Also called the uncapacitated facility location problem (UFLP), the problem has been well-studied

in the literature (Cornuejols et al. (1990)). The objective of the problem is to select a subset

of the candidate facilities and to assign customers to the selected facilities so as to minimize

the sum of fixed and operating costs. The operating cost in the UFLP is the transportation cost

calculated assuming individual customer routes that originate at a facility, serve one customer

and return to the same facility. The problem is NP-hard (Cornuejols et al. (1990)). A number of

researchers have focused on polyhedral analysis of the UFLP. Although a complete description of

the UFLP polytope is not known, Guignard (1980), Cornuejols and Thizy (1982) and Cho et al.

(1983 a,b) provide partial descriptions. Others have focused on developing algorithms for the

UFLP, for example, branch-and-bound based algorithms by Erlenkotter (1978) and Körkel (1989),

a Lagrangian relaxation algorithm by Cornuejols et al. (1977), and a dual simplex algorithm by

Simao and Thizy (1989). The best results for the problem are reported by Körkel (1989) using

an approach that extends the dual-ascent heuristic procedure developed by Erlenkotter (1978)

(ReVelle and Eiselt (2005)).

By adding capacity constraints on the facilities, we obtain the capacitated facility location

problem (CFLP) from the UFLP. Leung and Magnanti (1989) and Aardal et al. (1995) investigate

the polyhedral structure of the CFLP and introduce facet defining inequalities for the problem,

29
1.3. RELATED WORK

which lead to cutting plane algorithms (Aardal (1998)) for solving the problem. By using an

extended formulation (with other valid inequalities) of the CFLP, Cornuejols and Thizy (1982)

find and compare lower bounds for the CFLP by relaxing different constraints within a Lagrangian

relaxation and decomposition algorithm. Other papers presenting various solution algorithms for

the CFLP are Baker (1986), Roy (1986), Pirkul (1987) and Beasley (1988).

In the CFLP, the demand of a customer can be supplied by more than one facility. However,

in some applications, there may be a single-source requirement for a customer, which means

that the total demand of a customer must be supplied by exactly one facility. When the single

source requirement is added to the CFLP, it is called the capacitated facility location problem with

single sourcing (CFLPSS). The CFLPSS is more difficult than the CFLP since, while the CFLP

is a mixed integer model, the CFLPSS is a pure integer model with same number of variables

and constraints. In addition, for a given set of open facilities, the CFLP solves a transportation

problem to find the assignment of customers to the facilities, but the CFLPSS solves a generalized

assignment problem, which is also an NP-hard problem. In the literature, most of the studies

use heuristics, especially Lagrangian based heuristics, such as Klincewicz and Luss (1986) and

Sridharan (1993). Delmaire et al. (1997) design several metaheuristics and Ahuja et al. (2004)

develop a neighborhood search algorithm for the problem. Holmberg et al. (1999) develop an

exact branch-and-bound algorithm that uses a Lagrangian based heuristic and repeated matching

heuristic that solves successive matching problems. They can solve problems up to 200 customers

and 30 facilities.

1.3.2 Vehicle Routing

The second component of the LRSP is the decision of how to design the vehicle routes. The

classical vehicle routing problem (VRP) finds a set of minimum cost vehicle routes, each of which

starts at the depot, visits some customer locations and returns to the depot. There is a large body of

literature related to the VRP that describes the many problem variations and solution approaches.

The VRP has many variations depending on the characteristics of the vehicles, the facilities and

30
1.3. RELATED WORK

the customers. For example, the vehicles may be either identical or differ with respect to capacity

limits; the problem may be concerned with delivery only, pick-up only, or both delivery and pick-

up; the vehicles might be restricted to serve each customer in a given time window and/or serve

according to precedence relations of the customers. In addition, the problem may involve a single

facility or multiple facilities. Christofides (1985) surveys the different types of formulations used

to model the VRP and the exact and heuristic algorithms designed for the problem prior to 1985.

In a recent study, Toth and Vigo (2002) give a detailed description of the variants of the VRP and

their formulations, and they discuss the solution methods that are applied to these variants.

To formulate the VRP as an integer program, three different approaches are used in the liter-

ature. One approach is to define integer variables for each arc or edge in the model representing

the number of times the arc or edge is traveled. This type of model, which is called a vehicle

flow model, can be divided into two groups: two index formulations and three index formulations

(Laporte and Nobert (1987)). In a two index formulation, a flow variable has two indices repre-

senting the origin and the destination of the arc and in a three index formulation, each variable has

an additional index representing the vehicle traveling the arc. Vehicle flow models, especially two

index formulations, are very common in the VRP literature.

A classical vehicle flow model includes constraints called generalized subtour elimination

constraints, the number of which grows exponentially with the number of nodes in the problem.

There is one subtour elimination constraint for each subset of customers, which forces a lower

bound for the number of vehicles exiting the subset based on the minimum number of vehicles

necessary to serve the customers in the subset. The exponential number of constraints is handled

by considering a relaxation of the model that includes only some of the constraints or by using a

branch-and-cut algorithm. Miller, Tucker, and Zemlin (1960) introduce a polynomial number of

constraints that eliminate subtours, but the LP relaxation of the model with these new constraints

is weaker than the model with generalized subtour elimination constraints. Desrochers and La-

porte (1991) strengthen the Miller-Tucker-Zemlin subtour elimination constraints by lifting them.

They define stronger valid inequalities that can be adapted to variations of the VRP, such as the

31
1.3. RELATED WORK

capacitated VRP (CVRP), the distance constrained VRP (DVRP) and the VRP with time windows

(VRPTW).

A second modeling approach, developed by Garvin et al. (1957), is called a commodity flow

model. In this approach, an additional variable for each arc is added to the variables in the vehicle

flow model. These additional variables represent the amount of demand that flows along the edges

traveled by the vehicles. Letchford and González (2006) analyze three different commodity flow

models: one commodity, two commodity and multi commodity flow models. In two commodity

flow model, in addition to the flow on each arc representing the vehicle load, there is a reverse flow

equal to the vehicle residual capacity. In multi commodity flow model, demand of each customer

represents one commodity and flow of each commodity is denoted by different variables on each

arc. No exact solution algorithms for this approach are reported in the literature. Based on the

computational experiments, Gavish and Graves (1979) report large gaps between the LP solutions

and the optimal integer solutions when the number of customers are greater than 15.

A third approach, proposed by Balinski and Quandt (1964), is a set partitioning formulation, in

which there is one variable corresponding to each feasible vehicle route. The model has possibly

exponentially large number of binary variables representing whether a specific route is selected

or not. Since the feasibility of each route is considered in the process of constructing the set

of routes to be included in the model, various routing constraints, such as vehicle capacity, route

length restrictions and time windows for the customer nodes, can be included in the model without

changing the formulation. Because of the large number of variables, it is usually solved by a

column generation algorithm. The LP relaxation of the model is generally tight (Toth and Vigo

(2002)).

Letchford and González (2006) derive theoretical relationships between these different for-

mulations described in previous paragraphs. Using projection, they compare the LP relaxation

bounds. For capacitated VRP (CVRP), they prove that the three index formulation strengthened

with certain classes of valid inequalities gives the same lower bound with a two index formulation

32
1.3. RELATED WORK

with simpler constraints. Therefore, the LP relaxation bound based on the three index formula-

tion is weaker than the LP relaxation bound based on the two index formulation. They provide

some results for the relationship between a commodity flow model and a two index vehicle flow

model, and they show that some valid inequalities in the two index space are implied by extra flow

variables in commodity flow formulations. In addition, Letchford and González (2006) prove the

strength of the set partitioning formulation by showing that the set partitioning variables imply

some of the valid inequalities in the two index space. Another study investigating relationships

between formulations is Gouveia (1995) which includes some theoretical results that compare

commodity flow models and two index vehicle flow models.

In reviewing the solution algorithms for the VRP, we highlight those for deterministic, capac-

itated VRPs. In a capacitated VRP (CVRP), the total demand of a route must be less than the

capacity of the vehicle. Both exact methods and heuristic methods have been developed to solve

the CVRP. Laporte and Nobert (1987) survey the exact algorithms for solving the VRP. They

give a complete analysis of branch-and-bound algorithms, which are the most common way to

solve the CVRP. They also describe dynamic programming and integer programming approaches

that use different formulations of the problem. In addition to branch-and-bound procedures, La-

porte (1992) describes set partitioning formulations, column generation procedures and heuristic

approaches (e.g. Clark and Wright, tabu search and sweep algorithm) for the problem.

Branch-and-cut algorithms are very successful in finding optimal solutions for large instances

of the symmetric TSP, which is a problem closely related to the CVRP. This success has motivated

the development of branch-and-cut algorithms for vehicle flow models of the CVRP. Toth and

Vigo (2002) describe how to apply a branch-and-cut algorithm to the VRP and the CVRP. Branch-

and-cut algorithms incorporate the results of polyhedral studies of the CVRP polytope, i.e., valid

inequalities (e.g. Laporte and Nobert (1981b) and Laporte et al. (1985)) and facets of the polytope.

The first papers published about the facets of the CVRP polytope were by Araque, Hall, and

Magnanti (1990) and Araque (1990), who assumed unit demand for customers, and by Cornuejols

and Harche (1993), who extended the research to general demand. One of the main difficulties

33
1.3. RELATED WORK

with branch-and-cut algorithms is that the separation procedure for the cuts might be NP-hard. To

overcome this difficulty, researchers seek polynomial time algorithms for exact separation (e.g.,

McCormick et al. (2002) and Blasum and Hochstättler (2000)) and heuristic separation algorithms

(e.g., Augerat et al. (1999), Ralphs et al. (2003)).

Set partitioning-based formulations are used successfully by Cullen et al. (1981) to solve the

VRP heuristically and by Agarwal et al. (1989), Bixby (1999), Hadjiconstantinou et al. (1995) and

Desrochers et al. (1992) to solve the VRP exactly. The first step for solving a set partitioning-based

formulation is to solve the LP relaxation of the model using column generation. The objective

of the column generation subproblem (pricing problem) is to construct feasible vehicle routes

which include each customer at most once (elementary routes) and satisfy some other constraints

required for the feasibility of a route such as vehicle capacity. Since the complexity of the problem

of finding elementary routes is NP-hard, it is common to relax this property and solve the set

partitioning formulation with non elementary routes.

In set partitioning based models, to obtain the LP relaxation bound, people use various ap-

proaches to solve the column generation subproblem. Agarwal et al. (1989) use a branch-and-

bound algorithm, Desrochers et al. (1992) use a branch-and-bound algorithm and dynamic pro-

gramming, Bixby et al. (1997) use a cutting plane algorithm, and Hadjiconstantinou et al. (1995)

use a dynamic programming heuristic and solve the dual of column generation problem. For many

of the problems, the LP relaxation solution is very close to integer solution. However, even if the

LP relaxation is very strong, some branching may be required to obtain an integer solution. Thus,

branch-and-price algorithms (see Desrochers et al. (1992)) are used to obtain integer solutions.

To improve the overall performance, branch-and-price algorithm is extended with cutting planes

(Fukasawa et al. (2006) and Jepsen et al. (2008)).

A generalization of the VRP, which is also a special case of the LRSP, is the Multi-Depot

Vehicle Routing Problem (MDVRP). The MDVRP constructs a set of vehicle routes with mini-

mum cost in which each customer is visited by exactly one vehicle, and each vehicle route starts

and ends at the same depot. Exact solution algorithms are very rare for MDVRP; Laporte et al.

34
1.3. RELATED WORK

(1984) and Laporte et al. (1988) are the only two papers that report exact solution algorithms for

the MDVRP. Both solution methods are based on branch-and-bound algorithms. Many papers

develop heuristic algorithms, some of them are Tillman (1969), Tillman and Cain (1972), Gillett

and Johnson (1976), Chao et al. (1993), Cordeau et al. (1997) and Renaud et al. (1996).

1.3.3 Integrated Facility Location and Routing

In most facility location problems, the implicit assumption is that customers are served on indi-

vidual routes or that customers travel to the facilities. For situations in which customers receive

service from routes making multiple stops, this assumption does not accurately capture the trans-

portation cost. In order to accurately represent the cost of serving customers, we need to know

the order of the customers on the route. Hence, the location and routing problems must be solved

simultaneously. As early as 1968, Webb (1968) compared using a cost function considering multi-

ple stop tours and a simple cost function based on straight line trips to customers to determine the

locations of depots with minimum cost. He concluded that the solutions correspond to minimums

of the cost functions did not coincide, and the solution based on a cost function that considers

multiple-customer trips could be effected with the changes in the trip restrictions. Salhi and Rand

(1989) also investigated the effect of ignoring the routing of vehicles while deciding the locations

of facilities. They showed that solving the location and routing problems independently does not

always result in the best solution in terms of cost. Thus, although solving the combined prob-

lem may be more difficult than solving the individual ones, it may result in a more cost efficient

solution.

Laporte (1988) reviews the work on deterministic location routing problems; he describes the

different formulations of the problem, the solution methods used and the computational results

published up to 1988. More recent review papers by Min et al. (1998) and Nagy and Salhi (2007)

classify both deterministic and stochastic problems in the LRP literature with respect to problem

characteristics and solution methods. They group the LRP papers according to characteristics

35
1.3. RELATED WORK

such as single- or multi-stage, deterministic or stochastic, single or multiple facilities and vehi-

cles, uncapacitated or capacitated vehicles and facilities and solution space (discrete, network or

continuous). Because the LRP combines two NP-hard problems, researchers have developed a

wide variety of both heuristic and exact algorithms for the LRP. Most of the heuristic algorithms

tend to divide the problem into subproblems. Heuristic algorithms handle these subproblems se-

quentially or iteratively. Based on Min (1996), for problems with large facility fixed costs and

capacitated vehicles, sequential approaches perform computationally better. Some examples of

sequential heuristic algorithms are developed by Perl and Daskin (1985), Hansen et al. (1994) and

Barreto et al. (2007).

Perl and Daskin (1985) develop a heuristic for a three index flow formulation of an LRP

with capacitated distribution centers and vehicles. They decompose the problem into three sub-

problems: a multi-depot vehicle dispatch problem, a warehouse location-allocation problem and

a multi-depot routing allocation problem. They develop an algorithm that solves the problems

sequentially but considers the dependency between the problems. They solve the multi-depot

vehicle dispatch problem and the multi-depot routing allocation problem heuristically and the

location-allocation problem exactly. Hansen et al. (1994) introduce flow variables and flow con-

straints to the model in Perl and Daskin (1985) to get an MILP and solve the new model with

a heuristic that is a modification of the heuristic developed by Perl and Daskin (1985). Barreto

et al. (2007) develop a sequential heuristic algorithm for problems with capacitated facilities and

capacitated vehicles. In their heuristic, the first step is to solve routing subproblem. They group

customers using clustering techniques and design TSP solutions. Then, by assuming each group

of customers as a single customer, single source capacitated location routing problem is solved.

Srivastava (1993) presents 3 iterative heuristics called SAV1, SAV2 and CLUST and compares

the results with a classical sequential heuristic in which facility locations are determined first and

vehicle routes are designed based on open facilities. SAV1 heuristic starts with all depots being

open and using a multi depot vehicle routing heuristic calculates approximate routing costs for the

facilities. At each step a facility is closed until the problem becomes infeasible. SAV2 heuristic

36
1.3. RELATED WORK

works similar to SAV1 except initially all depots are closed and at each iteration one depot is

opened. CLUST heuristic generates groups of customers using a minimum spanning tree and

a density search clustering technique. Then a depot is selected for each cluster considering the

distance between cluster centroid and the depot.

Albareda-Sambola et al. (2005) provide a mathematical model that is constructed based on a

network representing an LRP with a single uncapacitated vehicle at each possible depot location.

They construct an initial solution by applying a rounding procedure to the LP relaxation of the

model, and develop a tabu search algorithm that starts with this initial solution. They develop

an additional lower bound by solving an asymmetric traveling salesman problem (ASTP) that in-

cludes all depots and customers. Then, they combine the cost of ASTP and a knapsack problem

that determines a lower bound for the total depot costs and the cost of connecting depots to cus-

tomers. Even though the obtained lower bound is generally better than the LP relaxation bound

of their LRP model, they conclude that the bound is not tight enough, and they report large gaps

between this lower bound and, both the tabu search upper bound and the optimal solution. In

addition, by comparing their tabu search upper bound with an upper bound that is derived via a

sequential heuristic, they restate the strength of integrated approaches.

Prins et al. (2007) develop a heuristic approach that combines Lagrangean Relaxation and

granular tabu search (Toth and Vigo (2003)) for the LRP with capacitated facilities and capacitated

vehicles. They divide the algorithm into two phases and alternate between each phase. In the first

phase, for a given solution, they aggregate the routes into a single customer and solve a location

problem using a Lagrangean relaxation. In the second phase, based on the solution of the location

problem, they construct new routes by solving a MDVRP using the granular tabu search heuristic.

They evaluate their algorithm using randomly generated instances up to 200 customers.

Exact algorithms for the LRP have been developed mostly for two index vehicle flow formula-

tions. Laporte and Nobert (1981a) solve a single depot model with a constraint relaxation method

and a branch-and-bound algorithm. They report solving problems with 50 customer locations.

Laporte et al. (1983) solve a multi-depot model using a constraint relaxation method and Gomory

37
1.3. RELATED WORK

cutting planes to satisfy integrality. They can solve problems with at most 40 customer sites.

Laporte et al. (1986) apply a branch-and-cut algorithm to a multi-depot LRP model with vehicle

capacity constraints. They use subtour elimination constraints and chain barring constraints that

guarantee that each route starts and ends at same facility. The authors report computational results

for problems with 20 customer locations and 8 depots.

Belenguer et al. (2006) provide two formulations of the LRP with capacitated facilities and

capacitated vehicles using two-index variables. They present a set of valid inequalities for the

problem and develop two branch-and-cut algorithms based on the different formulations of the

problem. They report that they solve instances with up to 32 customers to optimality in less than

70 CPU seconds and provide good lower bounds for the rest of the instances, which have up to

134 customers.

Berger (1997) develops a set partitioning-based model for an LRP with route-length con-

straints and uncapacitated vehicles and facilities. Berger et al. (2007) extend the work to develop

an exact branch-and-price algorithm in which they solve the pricing problem as an elementary

shortest path problem with one resource constraint. They report computational results for prob-

lems with 100 customers and various distance constraints.

1.3.4 Vehicle Scheduling

The third component of the LRSP is how to assign routes to vehicles, which is known as the vehicle

scheduling or loading problem. The (multi-depot) vehicle scheduling problem ((MD)VSP) seeks

a least-cost assignment of routes (or trips), each with a starting and ending time, to vehicles such

that each route is assigned to one vehicle, each vehicle starts a route after the starting time and

finishes before the ending time and each vehicle starts from a depot and returns to same depot.

In some variations, there are constraints such as depot capacity and upper and lower bounds on

the number of vehicles at each depot. The MDVSP is a special case of the LRSP; given a set of

selected facilities and a set of vehicle routes connected to these facilities, the LRSP reduces to a

MDVSP.

38
1.3. RELATED WORK

Carraresi and Gallo (1984) describe how to convert a single depot VSP to an assignment

problem or a minimum cost network flow model and therefore they conclude that the single depot

VSP is easy to solve. However, Bertossi et al. (1987) show that if there are more than 2 depots,

the MDVSP is NP-hard.

Both heuristic and exact solution methods are proposed for the MDVSP. Bodin and Golden

(1981) propose two heuristics based on decomposing the problem into single depot VSPs. Bertossi

et al. (1987) model the MDVSP by defining an independent problem for each depot and addi-

tional constraints to cover each trip exactly once. They relax the additional constraints using a

Lagrangian relaxation algorithm and then solve independent VSPs. Dell’Amico et al. (1993) pro-

pose a heuristic method that guarantees a solution with the minimum number of vehicles. They

solve shortest path problems on a graph that includes one node for each route and arcs between

two nodes that can be assigned to the same vehicle. A path on the graph represents a vehicle duty.

Several exact algorithms based on different formulations of the problem are proposed.

Carpaneto et al. (1989) use a branch-and-bound algorithm based on additive lower bounds. Forbes

et al. (1994) develop an algorithm using a dual simplex method to solve their three-index integer

programming model, which is a multi-commodity network flow problem. Ribeiro and Soumis

(1994) and Bianco et al. (1994) use set partitioning-based formulations for which they generate

columns and apply a branch-and-bound algorithm to get an integer solution. Fischetti et al. (2001)

design a branch-and-cut algorithm by defining valid inequalities of the MDVSP.

1.3.5 Vehicle Routing and Scheduling

In the VRP, there is one-to-one relationship between vehicles and routes, which means that one

route is assigned to each vehicle. In a combined vehicle routing and scheduling problem (VRSP),

scheduling constraints are added to the VRP constraints to change this one-to-one relationship

between vehicles and routes. Therefore, the VRSP seek to determine the set of vehicle routes as

well as the assignment of these routes to the available vehicles. In the literature, this version of

the VRSP is sometimes called the VRP with multiple trips (VRPMT). Only a few papers in the

39
1.3. RELATED WORK

literature address the VRPMT and they only describe heuristic approaches.

Taillard et al. (1996) use a three-step algorithm to find good feasible solutions for the VRSP.

In the first step, they generate some number of VRP solutions using a tabu search algorithm. In

the second step, they select a subset of these routes using an enumerative algorithm and, in the

third step, they use a bin packing heuristic to combine the routes in such a way that they will not

violate the time constraint and assign them to vehicles. The authors generate test problems with

50 to 199 customers. Feasible solutions are provided for 82 instances out of 104 instances. They

report that the cost of the feasible solutions are on the average within 1.2% of the VRP solution

obtained by Rochat and Taillard (1995).

Brandao and Mercer (1997) use a tabu search algorithm and the GENI algorithm (developed

by Gendreau et al. (1992) to solve TSP) to solve the VRPMT with time window constraints. They

design a three-phase algorithm in which the solution is improved in each phase. The input for the

first phase is an initial set of routes constructed using VRP constraints. The first phase checks each

route in the initial set for the time window infeasibility and, if there is an infeasible route, the route

is divided into smaller routes that satisfy the time windows. The second phase constructs a solution

with smaller cost by searching the set of routes. Similar to the second phase, the third phase also

seeks a solution with lower cost but the solution space is restricted to the feasible space during the

search. They evaluate their algorithm by comparing their solutions with manual solutions in a case

study. They report on average 20% better solutions. In addition, Brandao and Mercer (1998) solve

the instances created by Taillard et al. (1996). They can find feasible solutions for 89 instances

out of 104 instances.

Petch and Salhi (2004) develop a multi-stage heuristic algorithm in which they create vehicle

routes and then combine the routes using bin packing heuristics. Olivera and Viera (2007) solve the

VRPMT using an adaptive memory algorithm that is an enhancement of tabu search. The basic

idea is to construct new solutions using the components of the previously generated solutions.

They initialize the memory by generating some feasible solutions to the VRP, which they use to

construct feasible assignments of routes to vehicles. Then, they use tabu search to improve the

40
1.3. RELATED WORK

constructed solution. This heuristic differs from the others in that it constructs feasible vehicle

assignments in each iteration, while the others only construct the assignment at the final stage.

They can find feasible solutions for 95 instances out of 104 Taillard et al. (1996)’s instances. The

feasible solutions are on the average within 1.6% of the best known VRP solutions. In infeasible

solutions, in which overtime is used, the solutions provided by Olivera and Viera (2007) are better

than the solutions provided by other authors.

1.3.6 Integrated Location, Routing and Scheduling

Integrating facility location, vehicle routing and vehicle scheduling decisions gives rise to a class

of problems called integrated location routing and scheduling problems (LRSP). As far as we

know, the papers by Lin et al. (2002) and Lin and Kwok (2006) are the only LRSP papers to appear

in the literature. Lin et al. (2002) describe an LRSP, which arises in the delivery of telephone

bills, in which they seek to locate facilities and to route and assign vehicles so as to minimize

cost. To solve the problem, they develop a heuristic that divides the problem into three phases,

facility location, routing and loading, and that iterates through the phases. The location phase

determines the set of facilities via enumeration. For a given set of opened facilities, the routing

phase constructs vehicle routes using the Clarke and Wright algorithm (Clarke and Wright (1964))

followed by an insertion heuristic and more extensive metaheuristics. The loading phase optimally

assigns routes to vehicles considering the time limits. The authors report the development of a

branch-and-bound algorithm to solve the exact problem to evaluate their heuristics but they do

not provide a formulation. They test their algorithm on five instances with 10-12 customers and

4 potential depots, but the branch-and-bound algorithm can solve only three of the five instances

to optimality within the specified time limit. For these instances, they report that their heuristic

algorithm can find optimal or good solutions quickly. In addition, they solve a real life instance

with 27 customers and 4 depots using their heuristic algorithm and compare the results with the

management’s manual solution.

41
1.4. CONTRIBUTIONS OF THE THESIS

Lin and Kwok (2006) extend the study by Lin et al. (2002) to solve a multi-objective LRSP

that considers workload balance in terms of total vehicle time and total vehicle load as well as

total cost. They use the same three-phase heuristic algorithm designed by Lin et al. (2002). Since

they consider multi-objective problems, for each feasible solution in the algorithm, they keep

three performance measures, the total cost, the difference between the maximum and the mini-

mum vehicle time and the difference between maximum and minimum vehicle load. The feasible

solutions are evaluated based on these three performance measures and only non-dominated so-

lutions are kept. In addition, they design a tabu search (TS) algorithm for the routing phase of

the algorithm and compare its performance with the simulated annealing (SA) algorithm applied

in Lin et al. (2002). While the TS algorithm and the SA algorithm have similar performance in

some instances, the SA algorithm is better than the TS algorithm when the number of demand

nodes in a route increases. Furthermore, they compare the performance of the algorithm in two

approaches: assigning vehicles to routes after the routing phase (sequential approach) and assign-

ing vehicles while constructing routes (multiple approach). They conclude that the performance

of the approaches depends on the difficulty of the instance. The multiple approach is better than

the sequential approach when the solution includes routes with more demand nodes. They test

their algorithm with two sets of real data, one with 27 customers and one with 89 customers, and

a set of simulated data including instances with 100 and 200 customers.

1.4 Contributions of the Thesis

We summarize our contributions as follows.

• We are the first to consider the LRSP formally as a class of problems and to provide mathe-

matical programming models.

• We show that the Dantzig-Wolfe reformulation of the graph-based formulation is equivalent

to the set partitioning-based formulation of the problem.

42
1.4. CONTRIBUTIONS OF THE THESIS

• We show that some valid inequalities defined for the set partitioning-based formulation are

not dominated by the original inequalities.

• We are the first to provide an exact solution approach for this class of problems. We develop

an exact branch-and-price algorithm. We describe a two-phase solution approach for the

pricing problem based on the algorithms available in the literature. We define a set of

problem-specific branching disjunctions which guarantees to provide valid partitions of the

original feasible solution set and an optimal solution to the problem.

• We extend the branch-and-price algorithm to a two-stage branch-and-price algorithm using

the idea of restart and a heuristic extension of the algorithm.

• We evaluate the effect of the inequalities valid for polytopes that are relaxations of the LRSP

polytope and its reformulations.

• We show that the chain barring inequalities that are used in the graph-based formulations of

the LRP are satisfied by the LP relaxation of our set-partitioning based formulation for the

LRSP.

• We develop a methodology to derive disjunctive inequalities in a branch-and-price algorithm

and provide examples of various implementations of the methodology for the LRSP.

• We do extensive computational study to demonstrate the effectiveness of the solution ap-

proaches we propose and the value of integrating location, routing and scheduling decisions.

• We contribute to the LRP literature by providing an exact branch-and-price algorithm for

the problem with capacitated facilities and vehicles. We are not aware of any publication

in the literature solving the capacitated version of the problem using a branch-and-price

algorithm.

43
1.5. OUTLINE OF THE THESIS

1.5 Outline of the Thesis

We summarize the contents of the remaining chapters as follows. In Chapter 2, we define the LRSP

considered in this thesis and present two formulations of the problem: a graph-based and a set

partitioning-based. We describe the application of the Dantzig-Wolfe decomposition methodology

to one formulation of the LRSP. We describe some valid inequalities to strengthen one formulation

of the LRSP.

In Chapter 3, we describe our solution methodology for the LRSP, which is an exact branch-

and-price algorithm. We describe the two main components of the algorithm: the pricing problem

and the branching disjunctions. We present some implementations used to improve the overall

performance of the solution algorithm.

In Chapter 4, we investigate classes of inequalities valid for the LRSP and discuss how we can

extend the branch-and-price algorithm to a branch-cut-and-price algorithm. We discuss the adap-

tation of the disjunctive procedure to generate valid inequalities in a branch-and-price algorithm.

In Chapter 5, we present computational experiments for instances up to 40 customers evaluat-

ing the performance of the algorithms described in the previous chapters. In addition, we empiri-

cally compare the LRSP solutions with those of some variations of the sequential optimization of

facility location, vehicle routing and vehicle scheduling decisions.

In Chapter 6, we discuss solution of the location and routing problem (LRP) under capac-

ity restrictions using a modified version of the algorithm described for the LRSP. We report on

our computational experience solving both randomly generated instances and instances from the

literature.

In Chapter 7, we summarize the conclusions of the thesis and conclude the thesis with future

directions.

44
Chapter 2

Models and Complexity

In this chapter, we provide a formal description of the location, routing and scheduling problem

(LRSP) considered in this research and present two formulations of the problem, one using the

commodity flow paradigm and the other using a set partitioning-based approach. We then show

that the two formulations are related in that one can be obtained by applying DWD to the other.

Finally, we discuss the complexity of the LRSP.

2.1 Problem Definition

The LRSP integrates the decision-making process for determining the optimal number and loca-

tions of facilities to be open; an optimal assignment of customers to facilities; an optimal set of

vehicle routes from facilities to customers; and an optimal assignment of routes to vehicles subject

to scheduling constraints. The objective of the problem is to minimize the total fixed and operating

costs associated with facilities and vehicles. We consider an LRSP with capacity constraints on

the facilities and on the vehicles, as well as time constraints on the vehicles. In particular, given a

set of candidate facility locations and a set of customer locations, we seek to determine a subset of

facilities to be open and a routing of vehicles to customers from open facilities in such a way that

a. each customer is visited exactly once,

45
2.1. PROBLEM DEFINITION

b. each customer is assigned to exactly one facility,

c. each route starts and ends at the same facility,

d. the total demand of the customers assigned to a route is at most the vehicle capacity,

e. the total working time of a vehicle is no more than the time limit, and

f. the total demand of the customers assigned to a facility does not exceed the capacity of the

facility.

In the context of this research, a route is a path that starts at a facility and ends at the same facility

after visiting at least one customer. A route may only visit customer nodes with the exception

of the origin and destination stops, which are at a facility. A route may visit a given customer at

most once. Figure 2.1 displays examples of routes. The LRSP is to construct routes originating at

open facilities, and also to determine the schedules for vehicles by assigning groups of constructed

routes to individual vehicles.

Pairings. To define a schedule for a single vehicle, we adapt the pairing concept from the crew

scheduling literature (e.g., Desrosiers et al. (1991), Vance et al. (1997), and Anbil et al. (1998)).

In the crew scheduling literature, a pairing is a sequence of duty periods and represents a possible

schedule for a single crew. In the LRSP context, a pairing is a sequence of routes that represents

a possible (feasible) schedule for a single vehicle. Based on the constraints considered here, a

pairing is feasible if

a. each route included in the pairing starts and ends at the same facility,

b. for each route in the pairing, the total demand of the customers assigned to the route is at

most the vehicle capacity,

c. the total travel time of the pairing is at most the vehicle time limit,

d. each customer included in the pairing is visited only once, i.e., included on only one route.

46
2.1. PROBLEM DEFINITION

Figure 2.1 shows two possible pairing constructions. Pairing 1 includes routes 1, 2 and 3; pairing

2 includes routes 4 and 5. Each pairing specifies the routes covered by a single vehicle and has an

associated cost, which is defined as the sum of the costs of the component routes plus a fixed cost

for the vehicle. Since the cost of a multiple-stop route depends on the sequence of nodes visited,

the pairing information implicitly includes sequencing information for each route. In the absence

of time windows, the cost of a pairing, however, is not affected by the ordering of the routes and

thus is arbitrary.
PAIRING 1

2 5

1 Route 1 = (facility − 1 − 2 − 3 − 4 − facility)

Route 2 = (facility − 5 − 6 − 7 − facility)
Route 3 = (facility − 8 −facility)
6 CUSTOMER Route 4 = (facility − 9 − facility)
Route 5 = (facility − 10 − 11 − facility)
3 7

FACILITY

8
4

9 11

10

PAIRING 2

Figure 2.1: Example of pairings

We develop models for the delivery form of the LRSP and assume that there is no service time

at the customers. The models, however, can be adapted easily for pickup problems and for nonzero

service times. We assume that the demand information and distance matrix are known and fixed.

We assume that each facility has enough vehicles to serve the allocated customers. Furthermore,

we consider one class of vehicles, all having the same characteristics and working under the same

conditions.

47
2.2. PROBLEM FORMULATIONS

2.2 Problem Formulations

At the heart of the LRSP is a network flow structure, deriving from the routing of customer de-

mands, that is constrained through the vehicle and facility capacities. We consider both graph-

based and path-based formulations for the LRSP, as each has been widely applied for related

problems. In Section 2.2.1, we present a three-index commodity flow formulation of the LRSP. In

Section 2.2.2, we present a set-partitioning-based formulation of the problem.

2.2.1 Graph-Based Formulation of LRSP

For problems related to the LRP and the VRP, the literature contains two main types of graph-

based models, vehicle flow and commodity flow. Recall that in both formulations, integer variables

represent the number of times a vehicle traverses a given arc or edge in an underlying graph. For

the LRP and the VRP, two-index vehicle flow formulations are preferred in solution algorithms

based on branch-and-cut, since these formulations include a smaller number of variables and their

LP relaxations yield better bounds. However, it is not possible to formulate the LRSP using

variables with only two indices because of the time constraint for the vehicles. In this context, we

require a three-index commodity flow model. To formulate the model, we introduce the following

notation.

Sets

I = set of customer locations,

J = set of candidate facility locations,

N = I ∪ J = set of all nodes,

A = (J × I) ∪ (I × N ) = set of arcs,

Hj= set of vehicles located at facility j, ∀j ∈ J, such that |Hj | = |I|,

[
H = Hj = set of all vehicles.
j∈J

48
2.2. PROBLEM FORMULATIONS

Parameters

CjF = daily equivalent fixed cost of opening facility j, ∀j ∈ J,

C O = vehicle operating cost per unit travel time,

CV = daily equivalent fixed cost of a vehicle (including the driver cost),

Di = demand of customer i, ∀i ∈ I,

LFj = capacity of facility j, ∀j ∈ J,

LV = capacity of a vehicle,

LT = time limit for a vehicle and the driver, and

Tij = travel time between locations i and j, ∀(i, j) ∈ A.

Decision Variables


 1 if vehicle h travels on arc (i, k), ∀h ∈ H, (i, k) ∈ A,
xikh =

 0 otherwise,

yikh = flow on arc (i, k) carried by vehicle h, ∀(i, k) ∈ A, h ∈ H,



 1 if facility j is selected, ∀j ∈ J,
tj =

 0 otherwise,


 1 if vehicle h is used, ∀h ∈ H, and
vh =

 0 otherwise.

Based on this notation, we develop the commodity flow model for the LRSP, which we refer to as

(G-LRSP) in the rest of the thesis.

49
2.2. PROBLEM FORMULATIONS

(G-LRSP)

X X X X
Min CjF tj + C V vh + C O Tik xikh (2.1)
j∈J h∈H h∈H (i,k)∈A
X X
s.t. xikh − xkih = 0 ∀i ∈ N, h ∈ H, (2.2)
k∈N k∈N
XX
xikh = 1 ∀i ∈ I, (2.3)
h∈H k∈N
X X
yjkh − LFj tj ≤ 0 ∀j ∈ J, (2.4)
h∈Hj k∈I

yikh − LV xikh ≤ 0 ∀(i, k) ∈ A, h ∈ H, (2.5)

X X X
yikh − ykih + Di xikh = 0 ∀i ∈ I, h ∈ H, (2.6)
k∈N k∈N k∈N
X
Tik xikh − LT vh ≤ 0 ∀h ∈ H, (2.7)
(i,k)∈A
xjkh = 0 ∀j ∈ J, k ∈ N, h ∈ Ht , t ∈ J\{j}, (2.8)

xikh ∈ {0, 1} ∀(i, k) ∈ A, h ∈ H, (2.9)

yikh ≥ 0 ∀(i, k) ∈ A, h ∈ H, (2.10)

tj ∈ {0, 1} ∀j ∈ J, and (2.11)

vh ∈ {0, 1} ∀h ∈ H. (2.12)

The objective function (2.1) states that the total cost, which includes the fixed cost of the

selected facilities, the fixed cost of the vehicles (including the driver cost), and the operating

cost of the vehicles, should be minimized. Constraints (2.2) require that a vehicle should enter

and leave a node an equal number of times. Constraints (2.3) specify that exactly one vehicle

must service customer node i. Constraints (2.4) ensure that the total flow to the customer nodes

from each facility does not exceed its capacity. Constraints (2.5) are vehicle capacity constraints

and define the relationship between the binary variable xikh and the flow variable yikh ; each

constraint ensures that the flow between any pair of nodes cannot exceed the capacity of a vehicle.

Constraints (2.6) require conservation of flow at each customer node; these constraints must be

50
2.2. PROBLEM FORMULATIONS

modified for pick-up problems.Constraints (2.5) and (2.6) together ensure that no route violates

the vehicle capacity. The combined sets of constraints (2.2), (2.3), and (2.6) ensure that only valid

routes that include a facility are formed, as in Nambiar et al. (1981). Constraints (2.7) limit the

total time of a vehicle’s schedule to the time limit. In this research, we assume that the set of

vehicles in two different facilities are disjoint and that a vehicle located at facility j can only visit

customer nodes and the facility j during its trip. Therefore, constraints (2.8) restrict travel on the

arcs originating from a facility to vehicles located at that facility. Constraints (2.9), (2.10), (2.11),

and (2.12) are the integrality and non-negativity requirements on the variables.

2.2.2 Set Partitioning-Based Formulation of LRSP

In this section, we introduce a set-partitioning-based formulation of the LRSP in which the pairing

concept defined in Section 2.1 is used to formulate the LRSP. Given a set of candidate facility

locations and the set of feasible pairings that can be constructed for each candidate facility, the

objective of the LRSP is to select a subset of the facilities and an associated subset of pairings in

such a way as to minimize the total cost subject to the following constraints:

a. each customer must be covered by exactly one pairing,

b. the total demand of the customers in the pairings assigned to a facility must not exceed the

capacity of the facility, and

c. vehicles cannot be associated with a facility unless it is open.

To present the formulation, we define additional parameters and decision variables. Associated

with each candidate facility j ∈ J, we let Pj be the set of all feasible pairings for facility j.

51
2.2. PROBLEM FORMULATIONS

Parameters



 1 if customer node i is in pairing p of facility j, ∀i ∈ I, p ∈ Pj , j ∈ J,
aip =

 0 otherwise, and
Cp = cost (vehicle fixed cost plus operating cost) of pairing p of facility j, ∀p ∈ Pj , j ∈ J.

Decision Variables



 1 if facility j is selected, ∀j ∈ J,
tj =

 0 otherwise,


 1 if pairing p is selected for facility j, ∀p ∈ Pj , j ∈ J, and
zp =

 0 otherwise.

Based on the notation, we develop the set-partitioning-based model for the LRSP which we refer

to as (SPP-LRSP) in the rest of the thesis.

(SPP-LRSP)

X XX
Min CjF tj + Cp zp (2.13)
j∈J j∈J p∈Pj
XX
s.t aip zp = 1 ∀i ∈ I, (2.14)
j∈J p∈Pj
XX
aip Di zp − LFj tj ≤ 0 ∀j ∈ J, (2.15)
p∈Pj i∈I
tj ∈ {0, 1} ∀j ∈ J, and (2.16)

zp ∈ {0, 1} ∀p ∈ Pj , j ∈ J. (2.17)

The objective function (2.13) seeks to minimize the total cost including the fixed cost of the se-

lected facilities and the cost of the selected pairings. We note that the cost Cp of a pairing includes

both the vehicle fixed cost and variable costs. Constraints (2.14) guarantee that each customer

52
2.3. SIMPLE VALID INEQUALITIES

location is served by exactly one pairing. Constraints (2.15) ensure that the total demand of the

selected pairings does not exceed the facility capacity. Constraints (2.16) and (2.17) are standard

binary restrictions on the variables. We use notation SSP P to denote the set of feasible solutions

to (2.14) - (2.17).

The structure of the (SPP-LRSP) is very similar to the CFLP with single sourcing (CFLPSS).

In the CFLPSS, only pairings with one customer are considered and assigned to capacitated facil-

ities. However, in the LRSP, pairings with multiple customers must be taken into account.

2.3 Simple Valid Inequalities

In this section, we discuss how we can strengthen the set-partitioning-based formulation (SPP-

LRSP) by adding simple valid inequalities a priori. First, we introduce a lower bound on the total

number of facilities required in any integer feasible solution. Let R denote a lower bound on the

minimum number of required facilities. Then we can add to (SPP-LRSP) a constraint to force the

number of selected facilities to be at least R:

X
tj ≥ R ∀(t, z) ∈ SSP P . (2.18)
j∈J

For any instance, we can calculate a valid value for R from the demands of the customers and the

capacities of the facilities. There are two possible cases:

Case 1. If all of the facilities have the same capacity value, LFj = k ∀j ∈ J, then

»P ¼
i∈I Di
R= .
k

Case 2. If at least one facility capacity differs, then a lower bound on the number of required

facilities can be calculated using a greedy algorithm: compute the total demand and assign it to

the facilities in non-increasing order of their capacities. Let n = |J| and let (j1 , j2 , .., jn ) be an

53
2.3. SIMPLE VALID INEQUALITIES

ordering of the facilities such that LFj1 ≥ LFj2 ≥ ... ≥ LFjn . Then,

à l
!
X X
R = argmin{l=1..n} LFjs ≥ Di .
s=1 i∈I

Constraint (2.18) is not necessary in the formulation of the LRSP. However, as a simple lower

bound on the number of required facilities, it may eliminate some solutions in which a fractional

number of facilities is used, thereby improving the LP relaxation bound. The following proposition

can be stated:
³P ´
i∈I Di
Proposition 2.1 The Chvátal-Gomory rank (CG rank) of constraint (2.18) is 1, if k is
µP ¶
i∈I Di
fractional in case 1 and LF
is fractional or strictly less than R in case 2.
j1

Proof

Case 1: The facilities have identical capacity values.

To prove the proposition, consider the summation of constraints (2.15) over all facilities j ∈ J:

XXX X
aip Di zp ≤ k tj ∀(t, z) ∈ SSP P . (2.19)
j∈J p∈Pj i∈I j∈J

If we then substitute constraint (2.14) in (2.19) and divide by k, we get:

P
X Di
i∈I
tj ≥ ∀(t, z) ∈ SSP P . (2.20)
k
j∈J

P lP m
Di
The tj variables are binary, so j∈J tj must be integer. Therefore, with R =
i∈I
k ,

constraint (2.18) is valid. Furthermore, if the right hand side of constraint (2.20) is fractional,

then constraint (2.18) has CG rank 1. Otherwise, its rank is 0, since it is the non-negative linear

combination of the original inequalities.

Case 2: The facilities have different capacity values.

To generalize this statement to problems in which the facility capacities might be different, we

54
2.3. SIMPLE VALID INEQUALITIES

reorder the facilities as t[1] , t[2] , .., t[J] such that LF[1] ≥ LF[2] ≥ .. ≥ LF[J] . Then, inequality (2.19)

becomes

XXX X
aip Di zp ≤ LF[j] t[j] ∀(t, z) ∈ SSP P . (2.21)
j∈J p∈Pj i∈I j∈J

Again, if we substitute in (2.21) using constraint (2.14) and then divide each side by LF[1] , we get:

P
LF[2] LF[J] i∈I Di
t[1] + t[2] + .. + t[J] ≥ ∀(t, z) ∈ SSP P . (2.22)
LF[1] LF[1] LF[1]

If we let
&P '
i∈I Di
R2 = ,
LF[1]

then, by rounding up the coefficients of inequality (2.22), we get:

X
tj ≥ R2 ∀(t, z) ∈ SSP P . (2.23)
j∈J

The value of R2 is calculated based on the assumption that all facilities have a capacity equal to the

largest of the facility capacity values. However, in the procedure for calculating R for problems

that have different capacity values (case 2), it is considered that the second facility might not have

a capacity as large as the first facility, the third facility might not have a capacity as large as the

second facility and so on. Therefore, R ≥ R2 and constraint (2.18) dominates constraint (2.23).
µP ¶
i∈I Di
If LF
is fractional or R2 < R, then the CG rank of constraint (2.18) is 1.
[1]

Due to the similarity in structure between the (SPP-LRSP) and the CFLP, we might consider

adding the set of constraints:

zp ≤ tj ∀j ∈ J, p ∈ Pj , ∀(t, z) ∈ SSP P . (2.24)

55
2.3. SIMPLE VALID INEQUALITIES

In the CFLP, constraints similar to inequality (2.24) induce facets of the CFL polytope if the

demands of the customers in the constraints are less than the capacity of the facilities (Leung and

Magnanti (1989)). In addition, Daskin (1995) remarks that constraints in the CFLP similar to

(2.24) considerably improve the LP bound.

Constraints (2.24) ensure that a pairing of a facility can be selected only if that facility is se-

lected. Although constraints (2.24) are implied by constraints (2.15) when integrality is enforced,

their presence may improve the LP relaxation of the model as in the case of CFLP. However, since

there is potentially an exponential number of such constraints, we would like to find a smaller set

of constraints to add.

Proposition 2.2 Constraints

X
aip zp ≤ tj ∀j ∈ J, i ∈ I, ∀(t, z) ∈ SSP P , (2.25)
p∈Pj

have a CG rank of 1 and dominate constraints (2.24).

Proof In any feasible solution, each customer node i ∈ I is assigned to exactly one facility,
P
so p∈Pj aip zp = 1 for exactly one facility j and 0 for all others. For any j ∈ J such that
P
p∈Pj aip zp = 1, tj must be 1, since pairings can only be for open facilities (stated by constraints

2.15 and integrality restrictions).

Also, when the coefficients of the variables in inequality (2.25) and constraints (2.24) are

compared, it is easy to see that inequality (2.25) dominates constraint (2.24). For example, if we

take zp0 ≤ tj 0 from constraint set (2.24) for j = j 0 and p = p0 , we can compare it with a constraint
P
in set (2.25) for which j = j 0 and i = i0 such that ai0 p0 = 1, i.e., zp0 + p∈P 0 \{p0 } ai0 p zp ≤
j

tj 0 . Because of the terms involving variables other than zp0 on the left side of the constraint,
P
p∈Pj 0 ai0 p zp ≤ tj 0 dominates zp0 ≤ tj 0 .
P
Now we compute the rank of the inequality p∈P 0 ai0 p zp ≤ tj 0 for some i0 ∈ I and j 0 ∈ J.
j
P
Let D(k) = i∈I aik Di be the total demand of a pairing k ∈ Pj 0 . Using constraints (2.14) and

56
2.3. SIMPLE VALID INEQUALITIES

(2.15), we can derive the inequalities:

X X X
zp + zp ≤ 1 ∀(t, z) ∈ SSP P , (2.26)
{p∈Pj 0 |ai0 p =1} {j∈J\{j 0 }} {p∈Pj |ai0 p =1}
X X
D(p)zp + D(p)zp − LFj0 tj 0 ≤ 0 ∀(t, z) ∈ SSP P . (2.27)
{p∈Pj 0 |ai0 p =1} {p∈Pj 0 |ai0 p =0}

Let Dmin = min{p∈Pj 0 |ai0 p =1} {D(p)}. Multiplying inequality (2.26) by (LFj0 − Dmin ), adding the

resulting inequality to the inequality (2.27), dividing the obtained inequality by LFj0 and rounding,

we obtain:
$ % $ %
X LFj0 − Dmin + D(p) X X LFj0 − Dmin
zp + zp
LFj0 LFj0
{p∈Pj 0 |ai0 p =1} {j∈J/{j 0 }} {p∈Pj |ai0 p =1}
$ %
X D(p)
+ zp − tj 0
LFj0
{p∈Pj 0 |ai0 p =0}
$ %
LFj0 − Dmin
≤ ∀(t, z) ∈ SSP P . (2.28)
LFj0

Note that we assume that D(p) ≤ LFj0 ∀p ∈ Pj 0 . Because if the total demand of the pairing is

greater than the capacity of the facility, this pairing cannot be chosen in a feasible solution for this

facility and thus we can set zp = 0 and remove the variable from the constraint.

In inequality (2.28), the coefficients in the first summation are in the range [1, 2), the coeffi-

cients in the second summation are in [0, 1), and the coefficients in the third summation are in the

range [0, 1) if D(p) < LFj0 for all p ∈ Pj . Furthermore, the right hand side is in the range [0, 1).

Thus, when we round down the coefficients, we get

X
ai0 p zp ≤ tj 0 ∀(t, z) ∈ SSP P .
p∈Pj 0

Since we have obtained inequality (2.25) by using the original constraints and the CG inequality

procedure, the CG rank of the inequality is 1. Also, this inequality is not dominated by any linear

57
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

combination of the original constraints, thus its CG rank cannot be 0. If D(p) = LFj0 for some

p ∈ Pj 0 in the third summation, when we round down, the coefficient for this variable in (2.28) is

1. Thus, we can add the variables zp for all p ∈ Pj 0 and satisfying D(p) = LFj0 and ai0 p = 0 to

inequality (2.25) to get a stronger constraint.

Experiments evaluating the effectiveness of inequalities (2.18), (2.24) and (2.25) are presented in

Section 5.3.

2.4 Relationship Between Graph- and Set Partitioning-based For-

mulations

In this section, we explore the relationship between the formulations of the LRSP presented earlier

in the chapter. We prove that (SPP-LRSP) can be obtained by applying Dantzig-Wolfe decompo-

sition (DWD) to the graph-based formulation of the LRSP (G-LRSP). In Section 1.2.9, the DWD

methodology is explained for a general MILP model. Here, we provide the details of the procedure

for the LRSP.

2.4.1 Dantzig-Wolfe Decomposition of the Graph-based Formulation

To reformulate (G-LRSP) using DWD, we must decompose the formulation into two subsystems:

master problem and subproblem. The master problem is defined by constraints (2.3) and (2.4),

while the subproblem is defined by constraints (2.2), and (2.5) - (2.12). Recall that the vehicle

index set, H is the union of the vehicle index sets for each facility, i.e., H = ∪j∈J Hj . Thus,

all constraints of the subproblem are indexed by the set J and it is therefore immediate that the

subproblem decomposes by facility.

For each candidate facility j ∈ J, the set of integer solutions of the decomposed subproblem

58
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

can be represented by the set of vectors in the set

{ (x, t, v)j ∈ Z(|A|×|Hj |)+{0,1}+(|Hj |) , y j ∈ R(|A|×|Hj |) | (x, y, t, v)j satisfies

(2.2)j , (2.5)j , (2.6)j , (2.7)j , (2.8)j , (2.9)j , (2.10)j , (2.11)j , (2.12)j }, (2.29)

which is described only by those constraints associated with facility j and specifies values only

for those variables associated with vehicles assigned to facility j. Hence, the index j for the

constraints indicates the set of constraints associated with facility j. Set (2.29) is bounded for all

facilities. We let E be a set indexing the solutions in all of the above sets, with Ej the indices for

vectors associated only with facility j, so that E = ∪j∈J Ej . For a facility j and an index q ∈ Ej ,

the corresponding member (xq , yq , tq , vq ) of the above set is then a vector with the following

interpretation:


 1 if vehicle h travels on arc (i, k) in solution q, ∀(i, k) ∈ A, and h ∈ Hj ,
xqikh =

 0 otherwise,

yqikh = flow on arc (i, k) carried by vehicle h in solution q, ∀(i, k) ∈ A, and h ∈ Hj ,



 1 if facility j is selected in solution q,
q
t =

 0 otherwise,


 1 if vehicle h is used in solution q, ∀h ∈ Hj ,
q
vh =

 0 otherwise.

Note that the variable t indicating whether the facility is open does not appear in any of the linear

constraints of the subproblem and can hence be set to either 0 or 1 without affecting feasibility.

Following the DWD methodology, we can reformulate (G-LRSP) by replacing the subproblem

constraints with a requirement that solutions are convex combinations of the feasible points of the

subproblem. Because the subproblem decomposes as described above, solutions to the original

problem can be seen as vectors obtained by “recomposing” convex combinations of the members

59
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

of Ej for each j ∈ J. In other words, we express solutions (x, y, t, v) to formulation (G-LRSP) as

X
xikh = xqikh θq ∀(i, k) ∈ A, h ∈ H, (2.30)
q∈E
X
yikh = yqikh θq ∀(i, k) ∈ A, h ∈ H, (2.31)
q∈E
X
tj = tq θq ∀j ∈ J, (2.32)
q∈Ej
X
vh = vhq θq ∀h ∈ H, (2.33)
q∈E
X
θq = 1 ∀j ∈ J, (2.34)
q∈Ej
θq ≥ 0 ∀q ∈ E, (2.35)

where


 1 if solution vector (xq , yq , tq , vq ) for subproblem is used, ∀q ∈ E,
θq =

 0 otherwise.

Then, we rewrite the objective function (2.1), and the master problem constraints (2.3) and (2.4)

based on the relationship between solutions (x, y, t, v) and the set of solutions of the subproblem

defined by (2.30) - (2.35). The following linear master problem can be obtained.

X X XX X
Min (CqF tq + C V vqh + C O Tik xqikh )θq
q∈E h∈H i∈N k∈N h∈H
X XX q
(GMLP0 ) s.t. ( xkih )θq = 1 ∀i ∈ I, (2.36)
q∈E h∈H k∈N
X X X q X
− ( yjih )θq + LFj tq θq ≥ 0 ∀j ∈ J, (2.37)
q∈Ej h∈Hj i∈I q∈Ej
X
θq = 1 ∀j ∈ J, (2.38)
q∈Ej
θq ≥ 0 ∀q ∈ E. (2.39)

60
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

We can simplify the formulation (GMLP0 ) using some substitutions. First, we rewrite constraint

(2.37) by deriving a relationship between xq and yq based on (G-LRSP). We sum flow conservation

constraints (2.6) over all i ∈ I and observe that variables representing the flow between two

customer nodes (yikh ∀i, k ∈ I and h ∈ H) cancel each other out. In addition, by using the

information that in any feasible solution, the vehicle returns to the facility (variable yijh for some

facility j and vehicle h and for all i ∈ I) with zero load because of the characteristics of equations

(2.6), we obtain the equation

X X X
yjih = Di xikh ∀j ∈ J, h ∈ Hj , (2.40)
i∈I i∈I k∈N

which can be used to derive the equation

XX XX X
yqjih θq = Di xqikh θq ∀j ∈ J. (2.41)
q∈Ej i∈I q∈Ej i∈I k∈N

Substituting equality (2.41) and defining additional notation, we simplify (GMLP0 ) further:

X
Min C̃q θq
q∈E
X
(GMLP ) s.t. biq θq = 1 ∀i ∈ I, (2.42)
q∈E
XX X
− biq Di θq + LFj tq θq ≥ 0 ∀j ∈ J, (2.43)
q∈Ej i∈I q∈Ej
X
θq = 1 ∀j ∈ J, (2.44)
q∈Ej
θq ≥ 0 ∀q ∈ E, (2.45)

61
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

where

XX
biq = xqkih , ∀i ∈ I, q ∈ E,
h∈H k∈N
X XX X
C̃q = CqF tq + C V vqh + C O Tik xqikh , ∀q ∈ E,
h∈H i∈N k∈N h∈H

where CqF = CjF when q ∈ Ej for j ∈ J. Here, biq can be interpreted as the number of times

customer i is visited in solution q, and C̃q is the cost of solution q (including facility fixed cost and

vehicle fixed costs) for all q ∈ E.

Any feasible solution θ̄ to (GMLP ) can be transformed into vector (x̄, ȳ, t̄, v̄) using (2.30) -

(2.33). However, this solution may not be integer feasible for (G-LRSP). To be able to obtain a

true reformulation of (G-LRSP) using DWD, we also need to satisfy integrality restrictions for

(x, t, v). We can do that by adding integrality restriction to θ in (GMLP ). We call the following

problem the integer master problem and refer to it as (GM ).

X
Min C̃q θq
q∈E
X
(GM ) s.t. biq θq = 1 ∀i ∈ I,
q∈E
XX X
− biq Di θq + LFj tq θq ≥ 0 ∀j ∈ J,
q∈Ej i∈I q∈Ej
X
θq = 1 ∀j ∈ J,
q∈Ej
θq ∈ {0, 1} ∀q ∈ E.

To solve (GM ), we can apply a branch-and-price algorithm (see Section 1.2.10). The LP re-

laxation of the (GM ), which is (GMLP ), is solved using a column generation algorithm (described

in Section 1.2.8) to obtain a lower bound. Let (π, µ, α)> be a dual vector such that π is the dual

variable associated with constraint (2.42), µ is the dual variable associated with (2.43), and α is

the dual variable associated with (2.44). The reduced cost of a column with index q ∈ Ej for some

62
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

j ∈ J corresponding to variable θq is

à !
X X
Ĉq = C̃q − biq πi + biq Di − LFj tq µj − αj . (2.46)
i∈I i∈I

Using the definitions of C̃, we can rewrite (2.46):

X
Ĉq = (CjF − LFj µj )tq + C V vqh −
h∈Hj
X X X
( (C O Tki − πi + µj Di )xqkih + C O Tkj xqkjh ) − αj . (2.47)
h∈Hj k∈N i∈I

The column generation algorithm tries to determine the column of (GM ) with the most negative

reduced cost, as defined by (2.47), where the candidate columns are the set for facility j defined

in (2.29). This problem is called the subproblem for facility j ∈ J (referred to as (SPj )), and can

be formulated explicitly as follows:

(SPj )

X X X X
Min fˆj tj + C V vh − ĉki xkjh − αj
h∈Hj h∈Hj k∈M i∈M
X X
s.t. xikh − xkih = 0 ∀i ∈ M, h ∈ Hj , (2.48)
k∈M k∈M
V
yikh − L xikh ≤ 0 ∀i ∈ M, k ∈ M, h ∈ Hj , (2.49)
X X X
yikh − ykih + Di xikh = 0 ∀i ∈ I, h ∈ Hj , (2.50)
k∈M k∈M k∈M
X X
Tik xikh − LT vh ≤ 0 ∀h ∈ Hj , (2.51)
i∈M k∈M
xikh ∈ {0, 1} ∀i ∈ M, k ∈ M, h ∈ Hj , (2.52)

yikh ≥ 0 ∀i ∈ M, k ∈ M, h ∈ Hj , (2.53)

tj ∈ {0, 1}, and (2.54)

vh ∈ {0, 1} ∀h ∈ Hj , (2.55)

63
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

where M = I ∪ {j} and

fˆj = CjF − LFj µj ,



 C O Tki − πi + µj Di if k ∈ M , and i ∈ I,
ĉki =

 C O Tkj if k ∈ I.

The objective of each of the resulting single-facility subproblems (SPj ) is to generate sets of paths

with least-cost assigned to all vehicles located at the facility, though without the constraint that

customers appear exactly once. Because of this condition, we avoid calling these paths “routes”

for now.

Now we prove that solving (GM ) is equivalent to solving (SPP-LRSP). The similar forms of

(SPP-LRSP) and (GM ) should now be evident, but to rigorously show their equivalence, we need

to dissect the relationship between set Ej (used to define (GM )) and Pj (used to define (SPP-

LRSP)) for a given facility j ∈ J.

A member of set Ej , a feasible solution to (SPj ), consists of a collection of sets of paths

(starting at facility j and ending at facility j) assigned to all vehicles located at facility j. It may

be possible that some customers are visited multiple times by multiple vehicles. A member of

set Pj , on the other hand, is a single set of routes (pairing) that can be assigned to any vehicle at

facility j. A customer can only appear at most once in a pairing assigned to a vehicle.

Certain members of set Ej for j ∈ J can be constructed by associating at most |Hj | members

of set Pj and some number of empty pairings (zero vectors representing vehicles that are not used)

to the vehicles assigned to facility j. The members of Ej that include sets of paths assigned to a

single vehicle and visiting some customers multiple times cannot be constructed using members

of Pj .

By utilizing the integrality requirements from the original problem and carefully eliminating

the indices of symmetric solutions from Ej , we get a much smaller set that we will show is in

64
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

one-to-one correspondence with collections of members of Pj of cardinality at most |Hj |. We

proceed as follows:

1. Let q ∈ E be a member in which a vehicle visits at least one customer multiple times.

Thus, xqik1 h1 = 1 and xqik2 h1 = 1 for some customer i ∈ I, vehicle h1 ∈ H and nodes
P
k1 , k2 ∈ M such that k1 6= k2 . This means that k∈M xqikh1 > 1 and therefore biq > 1.

Because of constraint (2.42) and the binary restriction for θq , a solution to (GM ) for which

the corresponding solution in (G-LRSP) is feasible, θq must be 0. Therefore, we can remove

such members from E. In order not to generate these members, we can add the following

constraint to (SPj ),

X
xikh ≤ 1 ∀i ∈ I, h ∈ Hj . (2.56)
k∈M

When we add constraint (2.56) to (SPj ), the feasible solutions are restricted to sets of routes

each of which assigned to a vehicle and includes any customer at most once. Therefore,

based on the definition of a pairing in Section 2.2.2, we can say that the feasible solutions

to augmented (SPj ) are sets of pairings.

2. As the vehicles associated with a given facility j are identical, a set of pairings can be

assigned to vehicles in any arbitrary order. Hence, we obtain different members of Ej that

are all equivalent from the standpoint of both feasibility and cost. To eliminate superfluous

equivalent members of Ej , we divide the members of set Ej into equivalence classes, where

two members of Ej are considered equivalent if the set of customers assigned to the facility

and the partition of that set of customers defined by the pairings (set of routes assigned to

a vehicle) are identical. It is clear that any two members of Ej that are equivalent by this

definition will have exactly the same impact on both cost and feasibility.

In each equivalence class of Ej there can be up to |Hj |! members depending on the number

of vehicles serving to non zero pairings in the solutions. For example, let q1 and q2 be two

65
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

members of Ej and be in the same equivalence class. Then there exists a pair of vehicles h1

and h2 such that xqikh

1
1
= xqikh
2
2
, xqikh
1
2
q2
= xikh 1
∀i, k ∈ M and at least one of h1 and h2 is

serving a non zero pairing.

We form equivalence classes from which all but one member may safely be eliminated from

Ej .

3. Index q can be removed from Ej if there exist vehicles h1 , h2 ∈ Hj such that xqikh1 = xqikh2

∀i, k ∈ M , where xqikh1 > 0 for some i, k ∈ M (i.e., this is not an empty route). In this

case, vehicles h1 and h2 define exactly the same sets of routes, which means that biq > 1

for some i ∈ I. Because of constraint (2.42), such a solution must have θ̂q = 0.

4. From (2.32) and (2.43), we can conclude that if θ̂q = 1 and we have xqikh > 0 for some

i, k ∈ M and h ∈ H, then tq must be 1. Hence, we can eliminate any solution for which

tq = 0 that does not correspond to a zero solution (i.e., closed facility). All of this allows us

to rewrite (2.43) in the form

XX X
− biq Di θq + LFj θq ≥ 0 ∀j ∈ J. (2.57)
q∈Ej i∈I q∈Ej

5. A member q ∈ Ej can be removed if there exists a vehicle h ∈ Hj such that vqh = 1 and

xqikh = 0 for all i, k ∈ M (meaning that zero vector is assigned to vehicle h even though

we pay the fixed cost of that vehicle). In an optimal solution to (GM ), θq0 ≥ θq = 0 where
0
q, q 0 ∈ Ej , xq = xq , yq = yq , tq = tq , and vq = vq except that vqh = 0 since C̃q ≥ C˜q0 .
0 0 0 0

6. Finally, we can remove index q corresponding to the zero column (i.e., xq = 0, yq = 0,

vq = 0, and tq = 0) from Ej . Thus, we can update constraint (2.44):

X
θq ≤ 1 ∀j ∈ J. (2.58)
q∈Ej

If we restrict set Ej according to the rules described above and call the restricted set Ēj for
66
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

each j ∈ J, then we can finally conclude the following.

Proposition 2.3 There is a one-to-one correspondence between nonempty subsets of Pj with car-

dinality less than |Hj | and members of Ēj .

Proof The proof follows easily from the definitions of sets Pj and Ēj , and the restriction rules.

By replacing set Ej with Ēj for all j ∈ J in (GM ), as well as replacing (2.43) with (2.57)

and (2.44) with (2.58) we obtain a new (equivalent) formulation (GM 0 ). Then, we finally have the

equivalence of (SPP-LRSP) and (GM ) as follows.

Proposition 2.4 There is a one-to-one correspondence between solutions to (SPP-LRSP) and so-

lutions to (GM 0 ) such that corresponding solutions also have the same objective function value.

Thus, (SPP-LRSP), (GM 0 ), and (GM ) are all equivalent.

Proof First, we construct a one-to-one mapping to establish the correspondence between solution

sets of (SPP-LRSP) and (GM 0 ). We define mapping Ψ : (t, z) → θ. Let (t̄, z̄) be a feasible

solution to (SPP-LRSP) where t̄j = 1 if facility j is open, z̄p = 1 if pairing p is selected and

assigned to a vehicle for all p ∈ Pj and j ∈ J, then θ̄ = Ψ(t̄, z̄) be the corresponding solution in

(GM 0 ) where θ̄q = 1 if solution q is selected for all q ∈ Ēj , j ∈ J.

Since we can decompose variables in both formulations by facility, we can decompose map-

ping Ψ by facility and define mapping Ψj indexed by facility. Ψ = (Ψj )j∈J and θ̄Ēj = Ψj (t̄j , z̄Pj )

where notation xS represents a vector composed of the components of vector x indexed with set

S. Let J¯ = {j ∈ J|t̄j = 1}, P̄j = {p ∈ Pj |z¯p = 1} ∀j ∈ J¯ and P̄j = ∅ ∀j ∈ J\J¯ (because of

constraint (2.15)). For some facility j ∈ J, we have three types of solutions, which changes the

definition of θ̄Ēj = Ψj (t̄j , z̄Pj ).

i. Zero solution, t̄j = 0 and P̄j = ∅, thus z̄p = 0 for all p ∈ Pj . Then, Ψj (t̄, z̄Pj ) = θ̄Ēj = 0,

i.e., θ̄q = 0 for all q ∈ Ēj (since we removed zero solution from Ēj in reduction rule 6).

67
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

ii. Nonzero but empty solution, t̄j = 1 and P̄j = ∅. Then, θ̄i = 1 such that xi = 0, yi = 0,

vi = 0 and ti = t̄j = 1, and θ̄q = 0 for all q ∈ Ēj and q 6= i.

iii. t̄j = 1 and P̄j 6= ∅. For each facility we can choose at most |I| = |Hj | pairings, i.e.,
P
p∈Pj z̄p = |P̄j | ≤ |Hj |. Based on Proposition 2.3, there is one member k ∈ Ēj (in which

|P̄j | vehicles assigned one pairing from set P̄j , |Hj | − |P̄j | vehicles assigned zero solution,

and tkj = 1) corresponds to subset P̄j . Thus, the corresponding solution in (GM 0 ) is θ̄k = 1

and θ̄q = 0 for all q ∈ Ēj and q 6= k. We use notation k ­ P̄j to define this one-to-one

correspondence between member k and pairing set P̄j .

It follows that Ψj is one-to-one from construction, therefore Ψ is also one-to-one. It is clear that

we can easily construct the one-to-one reverse mapping, (t̄j , z̄Pj ) = Ψ−1
j (θ̄Ēj ) in a similar way.

From the construction of Ψj for each solution type, we can write the following relations be-

tween two formulations’ variables and parameters.

X
tj = θq ∀j ∈ J, (2.59)
q∈Ēj
X
biq = aip ∀i ∈ I, ∀q ∈ Ēj , (2.60)
p∈S:q­S
X
C̃q = CjF + Cp ∀q ∈ Ēj , ∀j ∈ J. (2.61)
p∈S:q­S

If (t̄, z̄) is feasible, then θ̄ = Ψ(t̄, z̄) is also feasible. This can be seen from the similarity

between the structures of (SPP-LRSP) and (GM 0 ) and from (2.59) and (2.60). Finally, the equiva-

lence of objective function values follows from (2.61).

Proposition 2.4 proves that applying DWD to (G-LRSP) yields (SPP-LRSP).

68
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

2.4.2 Comparing Alternative Formulations

(G-LRSP) is a three-index commodity flow formulation which can require up to O(|I|4 ) binary

variables and O(|I|4 ) constraints based on the assumption that the number of candidate facilities

is at most the number of customers. In addition, because of the facility and vehicle capacity

restrictions, we do not know a priori the number of vehicles required at each facility, a naive upper

bound is |I|.

On the other hand, (SPP-LRSP) includes O(|I|) number of constraints while the number of
S
variables depends on the size of the pairing sets (i.e., | j∈J Pj |), which can be exponentially

large. However, most of these variables will have value zero in an optimal solution. Therefore, in-

telligent enumeration techniques which may use some logical conditions, information from good

solutions found, and information from experience can be developed to avoid generating all possi-

ble columns.

Observe that the three-index formulation (G-LRSP) exhibits a high degree of symmetry under

the assumption that the vehicle fleet assigned to each facility is homogeneous. This is due to the

fact that the assignment of pairings to a specific vehicle is essentially arbitrary, i.e., the cost of

a given solution to (G-LRSP) is invariant under permutation of the indices assigned to specific

vehicles. In addition, from the definition of the variables in formulation (G-LRSP), symmetric

distance matrix results in solutions with same cost but including routes constructed in different

directions.

(SPP-LRSP) formulation eliminates the symmetry that arises in (G-LRSP) because of the iden-

tical vehicles. In (SPP-LRSP), we enumerate one set of feasible pairings that can be assigned to

any vehicle located at a facility instead of duplicating this set for each vehicle located at a facility.

Furthermore, while generating the pairings using a subproblem, depending on the solution proce-

dure, we can use undirected links between nodes in case of symmetric distance matrix. Therefore,

we can remove direction in the routes.

The lower bound obtained by solving the LP relaxation of (SPP-LRSP) is stronger than the

69
2.4. RELATIONSHIP BETWEEN GRAPH- AND SET PARTITIONING-BASED FORMULATIONS

bound obtained by solving the LP relaxation of (G-LRSP). This can be demonstrated both theo-
G be the LP relaxation of formulation (G-LRSP) and z SP P be the
retically and empirically. Let zLP LP

LP relaxation of formulation SPP-LRSP. In Proposition 2.4 we proved that (GM 0 ) obtained by ap-

plying Dantzig-Wolfe Decomposition to (G-LRSP) is equivalent to (SPP-LRSP). Therefore using

the DWD methodology we can theoretically compare the LP relaxation bounds of (G-LRSP) and
M be the LP relaxation of the formulation (G 0 ). Then, we have the following
(SPP-LRSP). Let zLP M
result.

G ≤ z M = z SP P .
Proposition 2.5 zLP LP LP

G is obtained by optimizing the objective function (2.1) on the polyhedron defined by

Proof zLP

constraints (2.2) - (2.8) ((G-LRSP) without the integrality constraints).

M is obtained by optimizing the objective function (2.1) on the poly-
On the other hand, zLP

hedron defined by the intersection of constraints (2.3) and (2.4) with the convex hull of of the

solutions that are constructed with composing the points in the set (2.29) for all facilities, which

is the convex hull defined by (2.2), and (2.5) - (2.12).

The proof follows from the fact that the subproblems to generate the set (2.29) for all facilities

do not have the integrality property. The polyhedron defined by the constraints (2.2) and (2.5) -

(2.8) is greater or equal to the convex hull of the same set of constraints.

The inequality in Proposition 2.5 can also be demonstrated empirically. In Section 5.2, we
G and z SP P .
present computational results that compare zLP LP

Finally, in terms of inequalities valid for the model, we can see additional differences between

working with formulation (G-LRSP) and formulation (SPP-LRSP). In (G-LRSP), the variables

are indexed with vehicle h, which enables us to define valid inequalities associated with either

a single vehicle (i.e., only the variables associated with vehicle h has nonzero coefficient) or a

set of vehicles. Let uh = (xikh , yikh , vh )(i,k)∈A represent the variable vector that is associated

with vehicle h ∈ H. We may derive cut (α, β) either in the form of αuh̄ ≥ β for some vehicle
P
h̄ ∈ H, or in the form of h∈K αh uh ≥ β for some set K such that K ⊆ H. In addition, we
70
2.5. COMPLEXITY

can multiply a cut generated for a single vehicle h for all vehicles in set H. A cut (α, β) which

is derived for vehicle h̄ and is in the form of αuh̄ ≥ β, can also be written for other vehicles,

e.g., αuh ≥ β for all h ∈ H. This can prevent jumps between the symmetrical solutions because

of the homogeneous vehicle fleet. Let ū1 and ū2 be the current LP relaxation solution vectors

corresponding to vehicle 1 and 2 which are in index set Hj for some facility j ∈ J, respectively.

When we only add αuh̄ ≥ β to the current LP relaxation and resolve, we may obtain an LP

relaxation solution such that ū01 = ū2 and ū02 = ū1 where ū01 and ū02 are the most recent solution

vectors corresponding to vehicle 1 and 2, respectively.

In formulation (SPP-LRSP), a pairing variable for a facility is associated with any vehicle

located at the facility, thus we can not differentiate the vehicles. We can only use cutting planes

associated with all vehicles located at a facility. Solution process for (SPP-LRSP) which employs

column generation may become more complicated with cutting planes with a certain structure.

Since the structure of the cutting planes searched may effect the efficiency of the solution algo-

rithm, the class of cutting planes considered may be more restrictive for (SPP-LRSP). However,

one advantage of (SPP-LRSP) is that some cutting planes valid for (G-LRSP) are already implied

by formulation (SPP-LRSP). This is evident from the Proposition 2.5 and the definition of the

columns of (SPP-LRSP). This will further be investigated in Chapter 4.

2.5 Complexity

In terms of formal computational complexity, the LRSP is in the complexity class NP-hard. To

show this, we consider an instance of the LRSP in which there is at most one vehicle at each

facility, there are no facility and vehicle fixed costs, and the vehicle time limit is unrestricted.

Such a special case is equivalent to an instance of the MDVRP, which was shown to be NP-hard

by Lenstra and Kan (1981), and Bodin et al. (1983). The MDVRP aims to find a minimum-cost

set of vehicle routes, each one starting from a depot and terminating at the same depot. Each

customer must be served exactly once and there are specified lower and upper bounds for the

71
2.5. COMPLEXITY

number of capacitated vehicles at each depot. In the created LRSP instance, location decision and

facility costs are eliminated by setting the fixed cost of facilities to zero, and scheduling part is

eliminated by setting the time limit to infinity. Solution of this instance results a set of routes with

minimum total operating cost. This solution can easily be translated to a solution for the MDVRP.

In the solution of the MDVRP, the allocation of the customers to the facilities, the vehicle routes

and the total cost are the same as in the solution of the LRSP. Since we have shown that the

MDVRP is a special case of the LRSP and the MDVRP is NP-hard, we can conclude that the

LRSP is also NP-hard.

72
Chapter 3

Branch-and-Price Algorithm

In this chapter, we describe our solution methodology for the LRSP. Although it would be possible

to develop a branch-and-cut algorithm for (G-LRSP), we suspect that such an approach would not

yield positive results, given the inherent symmetry in the formulation, the weak LP relaxation

bound and the known difficulty associated with solving the associated relaxations. Instead, we use

the set partitioning-based formulation, which is similar to those that have been successful for other

hard combinatorial problems. In our set partitioning-based formulation of the LRSP, each column

corresponds to a feasible pairing. For medium- and large-scale instances, the number of feasible

pairings can be extremely large, making it unlikely that we can efficiently solve instances that

explicitly include all feasible pairings. We have therefore developed a branch-and-price algorithm

in which we dynamically generate only a subset of the feasible pairings at each node of the tree,

as is typical in such cases.

In the following sections, we first give an overview of the branch-and-price algorithm and de-

scribe some notation. We then describe the components of our branch-and-price algorithm. This

involves describing the pricing problem and discussing solution approaches designed to solve the

pricing problem, as well as describing the methods of branching used to partition the feasible re-

gion of the original problem. Finally, we present some implementation details including initializa-

tion, the column generation algorithm consisting of different pricing algorithms in combination,

73
3.1. NOTATION AND OVERVIEW

upper bounding procedures and different versions of the branch-and-price algorithms providing

exact and heuristic solutions for the LRSP.

3.1 Notation and Overview

Recall that branch and price is an approach that combines branch and bound with a column gener-

ation algorithm to solve the LP relaxations arising during the enumeration. We have outlined the

main steps of a column generation algorithm (COLGEN) in Figure 1.4 and of a branch-and-price

algorithm (BRANCH-PRICE) in Figure 1.5.

As noted before, we developed the branch-and-price algorithm for the set partitioning-based

formulation of the LRSP. Because we have shown in Section 2.3 that adding both constraints

(2.18) and (2.25) to (SPP-LRSP) strengthens the formulation, we use the following augmented

(SPP-LRSP) formulation, denoted by (ASPP-LRSP) as the basis for our solution algorithm:

X XX
∗
zLRSP = Min CjF tj + Cp z p (3.1)
j∈J j∈J p∈Pj
XX
(ASPP-LRSP) s.t. aip zp = 1 ∀i ∈ I, (3.2)
j∈J p∈Pj
XX
aip Di zp − LFj tj ≤ 0 ∀j ∈ J,
p∈Pj i∈I
X
tj ≥ R,
j∈J
X
aip zp − tj ≤ 0 ∀j ∈ J, i ∈ I,
p∈Pj
tj ∈ {0, 1} ∀j ∈ J, and

zp ∈ {0, 1} ∀p ∈ Pj , j ∈ J.

In the remainder of this section, we provide an overview of the notation and the steps of the

branch-and-price algorithm, BRANCH-PRICE (Figure 1.5). Recall that to initialize BRANCH-

PRICE, we need to generate an initial set of columns to initialize the global column set denoted by

74
3.1. NOTATION AND OVERVIEW

CG . We describe the initialization approach to generating a subset of pairings in Section 3.4.1. The
∗
heuristic we apply for the initialization also provides an upper bound for zLRSP . The formulation

(ASPP-LRSP) is the original problem considered and corresponds to P0 in the algorithm. The set

of feasible solutions to (ASPP-LRSP) is denoted by S 0 . Refer back to Chapter 1 for the notation

used in this chapter.

The column generation algorithm for Pi begins by solving RMPi0 and then, using the resulting

dual solution information, formulates the pricing problem with the objective of either identifying

new columns to add to RMPi0 or proving optimality of the current solution. If the pricing algorithm

identifies at least one column with negative reduced cost, then we add the identified columns to

RMPi0 and continue to iteratively re-optimize the current restricted LP relaxation and solve the

pricing problem. If the pricing problem fails to find any column with negative reduced cost, the

current solution is optimal for the LP relaxation of Pi which we refer to as zLP

i . We describe

the pricing problem and several solution approaches used in our column generation algorithm

in Section 3.2. In Section 3.4, we present some implementation details of the branch-and-price

algorithm for the LRSP. We present two variations of the branch-and-price algorithm that are used

in our implementations.

Having briefly reviewed the branch-and-price algorithm for the LRSP, we discuss some chal-

lenges of the algorithm.

• The main challenge of applying column generation is solving the pricing problem, which
i
is generally an integer programming model itself. One call to COLGEN (to obtain zLP for

some Pi ) may require many calls to the oracle that solves the pricing problem. Therefore,

an efficient solution method for the pricing problem is crucial for the efficiency of the whole

algorithm. In order to improve the performance of COLGEN, one may opt to solve a more

tractable relaxation of the pricing problem when possible. Replacing the original pricing

problem with a relaxation yields a bound that is lower than it would otherwise. Alternatively,

one can solve the pricing problem heuristically. Indeed, Step 3 of COLGEN does not require

to find a column with the probably minimum cost in intermediate iterations. However, to
75
3.2. PRICING PROBLEM

terminate the COLGEN in Step 3 (b) with a valid lower bound for the optimal LP relaxation

value, one must find a column with the minimum cost, thus one must solve the pricing

problem to optimality at least once in COLGEN.

• Another challenge related to implementing COLGEN is that solving RMPik at any iteration

k may become time consuming as |CG | gets larger. The column management may become

a crucial part of the algorithm.

• In column generation algorithms, it is also very common to iterate between Steps 1 and

3 many times before termination. When there are alternative dual optimal solutions, the

algorithm may jump many times from one primal LP solution to another without much

improvement in the solution value.

• Besides the challenges inherent in the column generation algorithm, the main challenge in

developing a branch-and-price algorithm is developing effective branching rules that do not

destroy the structure of the pricing problem. At a non-root node of the tree, say node i,

we have the subproblem Pi defined over the feasible set S i which is obtained by applying

terms of the branching disjunctions defining node i. The pricing problem at the root node is

constructed based on the constraint set of P0 in order to generate columns feasible for S 0 .

Therefore, the pricing problem constructed at the root node may need to be modified for

S i . Therefore, for an efficient algorithm, branching rules must not complicate the pricing

problem and can easily be incorporated in the solution algorithm for the pricing problem.

In the following sections, we describe the components of the branch-and-price algorithm for the

LRSP in more detail.

3.2 Pricing Problem

In this section, we describe the model-specific components of the column generation algorithm

that is used in our branch-and-price algorithm. First, we describe the pricing problem in detail,
76
3.2. PRICING PROBLEM

and then present various solution algorithms. Later in Section 3.4.2, we provide the algorithm-

specific details of the overall column generation algorithm used in our branch-and-price algorithm.

3.2.1 Description

Formulation

The pricing problem aims to generate new feasible pairings, so the constraints are simply the linear

inequalities defining the feasibility conditions for pairings presented in Section 2.1. As explained

in Section 2.2.2, the set of pairings for each facility are independent, so we have a separate pricing

problem for each facility. In iteration k in node i, the objective of the pricing problem is to obtain

a pairing whose associated column has the smallest reduced cost with respect to the dual solution

of RMPik . To construct the objective function of the pricing problem for facility j, we consider

the formulation Pi and obtain an expression for the reduced cost of a pairing. In this section,

we describe the pricing problem and the solution algorithms used at the root node considering

formulation (ASPP-LRSP). Later, when we describe branching disjunctions, we will discuss the

necessary modifications to the pricing problem.

Let πl be the dual variable associated with constraint (2.14) for customer l, µj be the dual

variable associated with constraint (2.15) for facility j, and σjl be the dual variable for linking

constraint (2.25) for facility j and customer l. For variable zp for p ∈ Pj , the reduced cost Ĉp can

be written as

X X X
Ĉp = Cp − alp πl + alp Dl µj + alp σjl , (3.3)
l∈I l∈I l∈I

where Cp is the cost of pairing p. Thus, we can express the pricing problem for facility j as the

problem of minimizing Ĉp subject to the previously detailed constraints (in Section 2.1) defining

feasible pairings. A mathematical programming formulation of the problem for facility j is then

77
3.2. PRICING PROBLEM

as follows:

X X
Min C V + ĉik xik (3.4)
i∈M k∈M
X
(SPp ) s.t. xik ≤ 1 ∀i ∈ I, (3.5)
k∈M
X X
xik − xki = 0 ∀i ∈ M, (3.6)
k∈M k∈M
V
yik − L xik ≤ 0 ∀i ∈ M, k ∈ M, (3.7)
X X X
yik − yki + Di xik = 0 ∀i ∈ I, (3.8)
k∈M k∈M k∈M
X X
Tik xik ≤ LT , (3.9)
i∈M k∈M
xik ∈ {0, 1} ∀i ∈ M, k ∈ M, (3.10)

yik ≥ 0 ∀i ∈ M, k ∈ M, (3.11)

where M = I ∪ {j}, xik = 1 if arc (i, k) is used in the solution and 0 otherwise; yik is equal to

the amount of flow on arc (i, k); and ĉik is the reduced cost of arc (i, k) defined as follows.


 C O Tik − πk + Dk µj + σjk if i ∈ M , and k ∈ I,
ĉik = (3.12)

 C O Tij if i ∈ I.

Recall the subproblem (SPj ) for facility j that is obtained when DWD is applied to (G-LRSP)

(presented in Section 2.4). Notice that (SPp ) is equivalent to the problem that is obtained by

decomposing (SPj ) by vehicle.

To get more insight into the objective function (3.4), note that in equation (3.3), the cost Cp is

the sum of the fixed cost of a vehicle and the variable costs of the routes included in the pairing

which are equal to the sum of the costs of the arcs in the route. We account for the additional
P P P
terms − l∈I alp πl + l∈I alp Dl µj + l∈I alp σjl from (3.3) in (3.4) by modifying the original

arc costs of the arcs incoming to customer k by the values πk , Dk µj and σjk associated with

customer k. With this modification, the reduced cost of a pairing can be computed as the sum of

78
3.2. PRICING PROBLEM

the reduced costs of the arcs in the pairing.

Elementary Shortest Path Problem With Resource Constraints (ESPPRC)

As we have just seen, the pricing problem (SPp ) can be formulated as a generic mixed integer

programming model. By solving it this way, we do not take advantage of the special structure

of the problem. Upon closer inspection, the pricing problem can be interpreted as an elementary

shortest path problem with resource constraints (ESPPRC) on a modified network. In an ESPPRC,

the objective is to find a shortest path between the source node and the sink node such that each

other node is visited at most once and the resource constraints are not violated.

To properly represent our pricing problem as an ESPPRC, we need to construct a network,

determine the resources, and determine the resource consumptions and costs of the arcs, so that a

feasible path from the source node to the sink node corresponds to a feasible pairing that would be

generated by solving (SPp ). To define the pricing problem for facility j, we construct a network

with |I| + 2 nodes: |I| nodes for the customers, a node for facility j as the source node and a copy

of facility j as the sink node. Let Aj be the arc set of the network for facility j. Then Aj consists

of arcs from the source to customer nodes, arcs between customer nodes and the sink node, and

arcs between pairs of customer nodes. Note that there are no arcs from customer nodes back to

the source but there are arcs from the sink to customer nodes, this enables us to generate pairings

including multiple routes when we generate paths from the source node to the sink node. In this

network, in a path originating at the source node, an arc from the sink node to a customer node

represents the beginning of a new route, whereas an arc from a customer node to the sink node

represents the end of a route. Figure 3.1 shows an example network and a path on the network.

In the example, the path corresponds to pairing (facilityj - 3 - facilityj - 1 - 2 - facilityj ) which

includes two routes.

We set the length and the cost of arc (i, k) to Tik and ĉik (defined in (3.12)), respectively,

for all (i, k) ∈ Aj . We let each customer node have a demand equal to its demand in the LRSP

instance, the source have zero demand, and the sink have minus vehicle capacity demand. We use

79
3.2. PRICING PROBLEM

D1 D2
1 2

0 -LV
Source Sink

Facility j Facility j

3
D3

Figure 3.1: Example of a network and a path from source to sink (dashed line)

time and vehicle capacity as the two resources in the problem. As we construct a path, when we

include an arc, the length of the arc and the demand of the destination node are used to update the

elapsed time and the vehicle load (resources consumed), respectively. In addition, the total cost

of the path is equal to the cost of the pairing associated with the path because we use the same

cost parameter used in (SPp ). The process of checking the resource constraints and updating the

resource consumption at each node will be discussed in detail in Section 3.2.2. In this network and

with this design, an elementary resource constrained path from the source node to the sink node

corresponds to a feasible pairing.

3.2.2 Exact Methods

Because the pricing problem can be formulated as an ESPPRC, we can draw on the extensive

literature on this problem for potential solution algorithms. One particular special case occurs

when the underlying network is acyclic. In this case the “elementary” property of the paths can

be ignored, because in this case, any solution of the shortest path problem with resource con-

straints (SPPRC) is guaranteed to be feasible for the ESPPRC. Desrochers (1988) developed a la-

bel correcting algorithm for SPPRC based on dynamic programming that has a pseudo-polynomial

complexity that depends on the size of the resource windows and the number of resources. The

algorithm has been shown to be successful when there are tight resource constraints, but it be-

comes more time-consuming as the number of resources increases. As the algorithm builds partial

80
3.2. PRICING PROBLEM

paths, it assigns a label to each, indicating the resource consumption. To eliminate the problem

of an exponentially increasing number of labels, the algorithm applies dominance rules to delete

dominated labels.

If the network has negative arc costs, which is generally the case in pricing problems, the

elementary property of the paths must be considered for feasibility. Beasley and Christofides

(1989) describe how to find elementary paths in a graph by adding an extra binary resource for

each node showing whether the node is visited or not. However, this increases the number of

resources and possible combinations exponentially. In fact, the ESPPRC was proved to be NP-

Hard by Dror (1994). Feillet et al. (2004) adapt Desrochers’s label correcting algorithm to find

elementary paths and strengthen the algorithm by developing additional domination rules for the

labels. We adapted Feillet et al. (2004)’s label correcting algorithm for our pricing problem.

We developed two solution approaches. One approach is to solve a single ESPPRC on the

network described in Section 3.2.1. An alternative approach is to extend the first approach to a

two-phase pricing problem in which each phase is also formulated and solved separately as an

ESPPRC.

One-Phase Pricing Algorithm (1p-ESPPRC)

In this first approach, we use Feillet et al. (2004)’s algorithm to solve one ESPPRC on the con-

structed network. In the Feillet et al. (2004) algorithm, each path from the source to a node in

the network is assigned a label, which is a vector of the reduced cost of that (partial) path, the

resources consumed (vehicle load and elapsed time), the customer nodes included in the path, and

the customer nodes that cannot be visited due to the resource constraints. Customers are con-

sidered unreachable if they have already been visited in the path or if they cannot be served by

the vehicle without violating a resource constraint. The algorithm fans out from the source node,

repeatedly examining the nodes of the graph. Each time it examines a node, the algorithm ex-

tends each of the paths to each possible successor node, continuing until no further extensions are

feasible.

81
3.2. PRICING PROBLEM

To explain our algorithm that is based on Feillet et al. (2004)’s algorithm more specifically,

we let lri be a label that corresponds to path r from source to node i. We use vector lri =

(Rci , Rvi , Rti , Rni , V1i , .., V|I|

i ) to store the information about path r where

Rci = reduced cost of the path (so far),

Rvi = total demand of the path (so far),

Rti = time consumption of the path (so far),

Rni = total number of unreachable customer nodes (so far),

 

 −2 if customer k is temporarily unreachable for 


 


 


 


 the path 


 


 


 −1 if customer k is unreachable for the path 


 

Vki = because of the resource limits, ∀k ∈ I.

 


 


 0 if customer k is not in the path and it is 


 


 


 


 reachable for the path with the resources, 


 


 s > 0 if customer k is the sth node in the path, 


Note that the dimension of the vector V is |I|, so that there is one member for each customer node

in the network. In label lri , if customer k is visited in the path, Vki keeps the place of the customer

in the path and remarks that customer k is unreachable for this label because of the elementary

property. In Feillet et al.’s algorithm, if customer k is not visited in the path r, Vki is set to either

0 or -1. Even though we have two resource constraints (vehicle capacity and time limit) in the

problem, we only check the time limit constraint to mark a customer as “unreachable”, i.e., to set

Vki = −1, because if vehicle load exceeds vehicle capacity by adding customer k to the path, that

does not mean that we cannot add customer k to this path in the future. It may be possible that a

path extended from path r stops at the sink node and would have enough capacity to visit customer

k in further extensions. We explain the condition when Vki is set to -2 (which is used to define a

new domination rule) later in this section.

82
3.2. PRICING PROBLEM

We do not need an element in V for the source, because every path starts from the source

and can never return to it. Similarly, we do not need an element for the sink, because the sink

must always be reachable for every path. Extending a path to the sink is always resource feasible

because (1) demand of the sink is the negation of vehicle capacity (and hence total demand would

be less than the vehicle capacity), and (2) while constructing label r associated with a non sink

node i (lri = (Rci , Rvi , Rti , Rni , V )), the time limit constraint is checked to ensure that the path is

further extendable to the sink, i.e., Rti + Ti,sink ≤ LT .

Our algorithm based on the Feillet et al.’s algorithm for solving the ESPPRC is referred to as

1p-ESPPRC and includes three main parts:

1p-ESPPRC:

1. Initialization step

(a) To begin the algorithm, a label list is created for each node to keep the generated labels

(partial paths to the node).

(b) All label lists except the list for the source node are initially empty. The source has

one label, which is a zero vector, i.e., resource consumptions and the reduced cost are

zero, and all customers are reachable.

2. Extension step

(a) For each node, the label list is searched and all unprocessed labels are extended to all

possible successors of the node. Unprocessed means that the label was not extended

in previous iterations. The label lri for node i can be extended to node k if

i. node k is reachable: k = sink or Vki = 0, and

ii. the path is resource feasible: Rvi + Dk ≤ LV and Rti + Ti,k + Tk,sink ≤ LT .

(b) If it is feasible to extend lri to node k, a new label is created for node k by updating the

83
3.2. PRICING PROBLEM

resource consumptions. Let lpk be the label extended from lri . Then

Rck = Rci + ĉik

Rvk = max{Rvi + Dk , 0},

Rtk = Rti + Ti,k ,

 

 N if g = k and N = # of nodes in the path, 


 


 


 −1 if V i = 0 and Rk + Tk,g + Tg,sink > LT , 

k g t
Vg = ∀g ∈ I,

 

 Vgi if Vgi 6= 0,
 


 


 0 

otherwise,
X
Rnk = 1.
g∈I:Vgi 6=0

(c) Mark the labels that were extended as processed.

3. Elimination step

(a) After extending unprocessed labels of a node to its successors, and constructing new

labels for the successors, these new labels are combined with their previously created

label lists. Let node i be a node with new labels. To combine the label list of i with

the list of new labels created for i, each label is compared with the others in the list

and only the non-dominated labels are retained. Let l1i and l2i be two labels for node i.

Label l1i dominates label l2i if all of the following conditions are satisfied:

i. the current travel time and the vehicle load of l1i are less than or equal to those of

l2i ,

ii. the reduced cost of l1i is less than or equal to the reduced cost of l2i , and

iii. the number of total unreachable customer nodes of l1i are less than or equal to

those of l2i , and for all nodes g ∈ I, g is reachable only in label l1i or g is unreach-

able in both labels l1i and l2i .

All dominated labels are deleted from the label lists of the nodes.
84
3.2. PRICING PROBLEM

(b) If new labels were added to the list of any node after the domination algorithm was

applied, then the algorithm returns to the extension step. When all the labels are

extended and no new labels are added to the lists of nodes, the algorithm terminates.

Additional Domination Rule. We extended Feillet et al.’s algorithm by adding a new domina-

tion rule. Executing the described 1p-ESPPRC on the constructed network leads to paths from the

source to the sink that correspond to pairings including sets of ordered vehicle routes. Because

the ordering of the routes within a pairing is arbitrary, we would like to avoid generating multiple

paths that are different permutations of a set of routes and correspond to the same pairing. For

example, a call to 1p-ESPPRC may generate some or all of the following paths:

• (source - 1 - 2 - sink - 3 - 4 - sink - 5 - 6 - sink),

• (source - 1 - 2 - sink - 5 - 6 - sink - 3 - 4 - sink),

• (source - 3 - 4 - sink - 1 - 2 - sink - 5 - 6 - sink),

• (source - 3 - 4 - sink - 5 - 6 - sink - 1 - 2 - sink),

• (source - 5 - 6 - sink - 1 - 2 - sink - 3 - 4 - sink), and

• (source - 5 - 6 - sink - 3 - 4 - sink - 1 - 4 - sink).

However, those six paths are equivalent with respect to total cost and time, and they correspond

to the same pairing composed of the following three routes: (facility - 1 - 2 - facility), (facility - 3

- 4 - facility) and (facility - 5 - 7 - facility). To eliminate generation of all possible ordered sets,

we introduce a new domination rule in which we force a specific order for routes in the pairing,

which decreases the state space of the algorithm significantly. We add the following condition to

the algorithm.

Rule The index of the first customer node of each route should be less than the index of the first

customer node in the following route.

85
3.2. PRICING PROBLEM

The modified algorithm only generates paths from source to sink in which the first node of each

route is less than the first node of the following route. Thus, the algorithm with this rule generates

only path (source - 1 - 2 - sink - 3 - 4 - sink - 5 - 6 - sink).

To apply this rule, we add variable Rfi to keep the index of the first customer node in the

current route in the path to the definition of label lri , where lri = (Rci , Rvi , Rti , Rni , Rfi , V1i , .., V|I|
i ).

We make the following changes in 1p-ESPPRC:

• In the extension step of the algorithm, in part (b), while constructing label lpk from lri , we set

Rfk = k if i is the source or the sink (k will be the first node in the new route), and Rfk = Rfi

(still in the same route, thus the first node is the same) otherwise. We also need to modify

the procedure for updating vector V k .

 

 N if g = k and N = # of nodes in the path, 


 


 


 Vi if Vgi 6= 0 and Vgi 6= −2, 


 

 g 
Vgk = −1 if Vgi = 0 or Vgi = −2, and Rtk + Tk,g + Tg,sink > LT ,  ∀g ∈ I,

 

 


 −2 if k is sink, Vgi = 0 and g < Rfk , 


 


 

 
0 otherwise,
X
Rnk = 1.
g∈I:Vg 6=0 or Vg 6=−2
i i

Here, Vgk = −2 means that node g is temporarily unreachable because, based on the new

rule, it cannot be the first node in the new route. However, it may become part of the path

during further extensions. Therefore, when we calculate number of unreachable nodes Rnk ,

we do not count temporarily unreachable nodes. We treat them as reachable nodes. In

addition, while extending a path to other nodes after the first node in the route, we change

the condition temporarily unreachable to reachable by setting Vgk to 0.

• In the elimination step of the algorithm, the steps and the conditions that should be satisfied

to decide dominated labels are the same. While executing condition (iii) in part (a), if there

86
3.2. PRICING PROBLEM

are temporarily unreachable nodes, we count them as reachable. The only modification to

this step occurs when each label dominates the other. In the previous version, we were

allowed to pick either one of l1i or l2i arbitrarily and delete the other one. However, in this

version, if l1i dominates l2i and l2i dominates l1i based on conditions (i), (ii) and (iii), we keep

the one with smallest Rfi (first node in the current route).

At the end of a pricing algorithm, each label at the sink node corresponds to a feasible pairing

with an associated reduced cost. The costs of paths at the sink node are updated with the vehicle

fixed cost that must be included in the reduced cost of a variable, and any path with negative

reduced cost is a candidate pairing to add to RMP0k at iteration k. If the algorithm concludes with

no negative reduced cost labels on the sink node, then we can conclude that the RMP0k is optimal
0 .
and the solution value is equal to the zLP

Two-Phase Pricing Algorithm (2p-ESPPR)

An alternative approach to solving the exact pricing problem is to extend the first approach (1p-

ESPPRC) to a two-phase pricing problem: first, generate a set of feasible vehicle routes and

then combine them to construct a feasible pairing of minimum reduced cost. Each phase is also

formulated as a network problem and solved as an ESPPRC. Since a pairing is a set of routes, we

can express the variable part of the reduced cost of a pairing as the sum of the reduced costs of the

routes forming the pairing. For there to exist a negative reduced cost composed of more than one

route, there must exist routes with negative reduced cost. Otherwise, the pairing including only

the single route with the lowest reduced cost would be the lowest cost pairing overall.

Phase One: Generating Routes. To generate only feasible routes instead of pairings, we solve

a regular ESPPRC problem on the constructed network with |I| + 2 nodes. We delete the arcs

between the sink and the customers and terminate each path at the sink and we use algorithm

1p-ESPPRC without the additional domination rule. At the end of the algorithm, each label at the

sink node corresponds to a feasible vehicle route with an associated reduced cost.

87
3.2. PRICING PROBLEM

Let LS denote the set of feasible routes with negative reduced cost at the sink. If LS is non-

empty, we continue with phase two. Otherwise, we add the fixed cost from expression (3.3) to

the cost of the smallest reduced cost route to obtain the reduced cost of that single-route pairing.

If this reduced cost is negative, then we have a pairing that we can add to the restricted model.

Note the implication that, even in the case where there are no routes with negative reduced cost,

we may still get a negative reduced cost pairing (consisting of a single route) when the fixed cost

associated with the vehicle is negative. This can occur after applying certain branching rules (see

Section 3.3.1).

Phase Two: Generating Pairings. The objective of phase two is to generate feasible pairings

with negative reduced cost by combining the routes in set LS . Two (or more) vehicle routes can

be combined to form a feasible pairing if they do not have customer nodes in common and they

can both (all) be completed within the vehicle’s time limit. Finding a feasible pairing with the

minimum reduced cost again corresponds to solving an ESPPRC with a single resource constraint

(time).

We construct a network with 2+|LS | nodes: a source node, a sink node and one “intermediate”

node to represent each of the vehicle routes in set LS . The network includes an arc from the source

node to each intermediate node, arcs between pairs of intermediate nodes, and an arc from each

intermediate node to the sink node. Note that, because the sequence of routes in a pairing is

arbitrary, we only include an arc from an intermediate node to another intermediate node with

higher index. Associated with each arc inbound to an intermediate node are two values: the time

required to complete the corresponding route and the reduced cost of the route. The time and cost

for arcs inbound to the sink are zero. To find a feasible pairing with the minimum reduced cost,

we apply 1p-ESPPRC. The costs of pairings at the sink node are updated with the fixed cost that

must be included in the reduced cost of a variable, and any pairing with negative reduced cost is a

candidate pairing to add to the restricted model in the column generation.

Since we solve the exact ESPPRC algorithm using 1p-ESPPRC in each phase of the algorithm,

88
3.2. PRICING PROBLEM

Phase 1: 1 2
Route 1(r1): source − 1 − sink
Route 2 (r2): source − 2 − sink
Source Sink Route 3 (r3): source − 3 − sink
Route 4 (r4): source − 1 − 2 − sink
Route 5 (r5): source − 1 − 3 − sink
3 Route 6 (r6): source − 2 − 3 − sink

Phase 2: r1

r2

Source r3 Sink

r4

r5

r6

Figure 3.2: Demonstration of Phase 1 and Phase 2 Networks on a Small Example

the overall pricing algorithm is exact and results in a column with the most negative reduced cost,

if any. We refer to this exact pricing algorithm as 2p-ESPPRC. To clarify the algorithm, we

demonstrate how we construct the networks for each phase in Figure 3.2.

Generation of Multiple Pairings

In general, at the completion of 1p-ESPPRC or the second phase of 2p-ESPPRC, there is more

than one label at the sink. We may have a set of pairings with negative reduced cost that includes

a pairing with the most negative cost. Therefore, we may add multiple columns to the RMPik at

iteration k of the column generation algorithm for node i. If we want to add at most n columns

in any iteration, we can sort the labels on the sink node in ascending order based on the reduced

cost and choose the first n or less number of labels with negative reduced cost in order to add to

RMPik .

89
3.2. PRICING PROBLEM

In Section 5.4.1, we present computational results comparing the performance of 1p-ESPPRC,

1p-ESPPRC with the additional domination rule and 2p-ESPPRC, and computational results eval-

uating the effect of adding multiple columns instead of a single column to the performance of the

column generation algorithm.

3.2.3 Heuristic Methods

An exact algorithm is guaranteed to generate a column with the minimum reduced cost. The

downside of the approach is that it may not be very efficient for large problems. As the number of

customers, the vehicle capacity and the time limit increase, the state space of the algorithm, i.e.,

the number of feasible partial paths, also increases, so the number of labels in the label list of each

node in the algorithm increases. Thus, processing all labels and terminating the algorithm takes

substantially more time.

A common approach in the literature to overcoming this problem is to use heuristics to supple-

ment the exact procedure for the pricing problem to improve the efficiency of a branch-and-price

algorithm. To find feasible pairings more quickly, here we define two heuristic versions of a la-

beling algorithm that we use in our solution algorithm. We can use these heuristic modifications

for both 1p-ESPPRC and 2p-ESPPRC.

ESPPRC with Label Limit (ESPPRC-LL)

This heuristic version was first proposed by Dumitrescu (2002) and restricts the state space of the

exact algorithm by directly restricting the number of labels. In particular, during 1p-ESPPRC, or

either phase of 2p-ESPPRC, we limit the maximum number of non-processed labels allowed at

each node. During the algorithm, as we update the list of labels in the elimination step, at the end

of the domination algorithm, we order the non-dominated labels in increasing order of reduced

cost and we keep at most the first label limit number of non-processed labels. We refer to these

heuristic versions of pricing algorithms as 1p-ESPPRC-LL(l) and 2p-ESPPRC-LL(l), where l is

the value of the label limit. For small label limits, the algorithm terminates very quickly. And as l

90
3.2. PRICING PROBLEM

gets larger, the heuristic version approaches to the exact algorithm.

ESPPRC for a Subset of Customers (ESPPRC-CS)

This second heuristic restricts the state space by considering only a subset of the customers. In this

way, the total number of labels to be generated is restricted in advance. In each pricing iteration

for a facility, we determine the two customers that have the lowest reduced costs from the facility.

Let j ∈ J be the facility index, we choose customers i1 and i2 such that ĉji1 ≤ ĉji2 ≤ ĉji for all

i ∈ I\{i1 , i2 }. Let Iu = {i1 , i2 } be the subset. We initialize the subset with these two customers,

since we expect to have at least two routes in a pairing. Then, by exploring the arcs from the

subset of customers to the remaining customers, we repeatedly add one customer that is connected

by a cheapest arc to the subset until the size of the subset is equal to the determined subset size.

Particularly, we let Iu = Iu ∪{k} where k = argmink∈I\Iu {mini∈Iu {ĉik }} We stop when |Iu | = u

where u is the determined subset size.

After we determine Iu , we run the exact pricing algorithm with this smaller customer set.

We refer to these heuristic versions of pricing algorithms as 1p-ESPPRC-CS(u) and 2p-ESPPRC-

CS(u), where u is the subset size. In 2p-ESPPRC-CS(u), we use the selected customers in Phase 1,

while the state space of Phase 2 reduced as a result of the reduction in the state space of Phase 1.

If a heuristic version of the pricing algorithm identifies columns with negative reduced cost,

they are feasible and can be added to RMP0k . However, these columns are not necessarily the

columns with most negative reduced cost for a given dual vector. Therefore, if a heuristic ver-

sion of the pricing algorithm does not identify any columns with negative reduced cost, the exact

pricing algorithm is applied at least once in order to conclude whether or not there exist columns
0 . In Section 5.4.2, we present
with negative reduced cost or to evaluate a lower bound for zLP

some computational results evaluating the effect of applying heuristic versions of the pricing algo-

rithm in addition to the exact pricing algorithm and comparing the performance of both heuristic

approaches.

91
3.3. BRANCHING

3.3 Branching

An important step in developing a branch-and-price algorithm is devising good branching rules for

partitioning the original feasible solution set. In this section, we first list the disjunctions used to

partition S 0 in our branch-and-price algorithm. We describe how these disjunctions are imposed

to produce new subproblems and how the pricing problem is affected by the imposition of these

disjunctions. Because the disjunctions applied may affect the structure of the pricing problem

and may make the pricing algorithm more difficult, our main objective is to find disjunctions that

eliminate as much of the integer infeasible region as possible but that minimize the increase in

difficulty of solving the pricing problem. We refer to the disjunctions used as branching rules

or branching decisions. Second, we discuss the completeness of the branching rules in order to
∗
provide zLRSP at the termination of the branch-and-price algorithm.

3.3.1 Methods

We apply four branching rules in our branch-and-price algorithm:

1. Branching on the facility location variables t: At any node of the branch-and-price tree,

if the corresponding LP solution has at least one fractional facility location variable, then we

apply standard variable dichotomy branching for the facility location variables. We select

one of the fractional facility location variables, say tj , as our branching variable.

• In the left node, we set tj = 1, which forces facility location j to be open. The pricing

problem for the corresponding node does not change, since variable t is not present in

the constraints of the pricing problem (SPp ), and the pricing problem for each facility

is independent from the others.

• In the right node, we set tj = 0, which forces facility j to be closed. Since the facility

is closed, no pairings associated with the facility are allowed in the solution, so there

is no need to solve the pricing problem. In addition, we set zp = 0 ∀p ∈ Pj .

92
3.3. BRANCHING

2. Branching on the total number of vehicles at each facility: By definition, a pairing is as-

sociated with a single facility and corresponds to one vehicle. Thus, in any integer solution,

the number of vehicles (pairings) used at any facility must be integer. Let wj be the number

of vehicles used at facility j. Then,

X
wj = zp ∀j ∈ J. (3.13)
p∈Pj

If we add the equalities (3.13) to (ASPP-LRSP), we can branch explicitly on the wj vari-

ables. Assume that in a fractional solution wj = b, where b is not integer for some j ∈ J.

We create two new nodes:

wj ≤ bbc (the left node),

wj ≥ dbe (the right node).

P
Branching on wj instead of implicitly branching on p∈Pj zp is an example of explicit

constraint branching, an idea developed by Appleget and Wood (2000) to improve the per-

formance of branch-and-bound algorithms for mixed integer linear problems. These new

variables make it easier to identify the necessary updates to the pricing problem. Let βj be

the dual variable associated with constraint (3.13) for facility j. Then, the reduced cost of

the variable zp changes to the following:

X X X
Ĉp = Cp − aip πi + aip Di µj + aip σji − βj . (3.14)
i∈I i∈I i∈I

In equation (3.14), we can interpret βj as a fixed cost for facility j that is independent of

pairing p and the included customers. Therefore, we do not need to change our network

structure or the pricing problem structure to incorporate βj . Instead, we can calculate the

cost of a path in the modified network in the same way as before and then add the fixed cost

93
3.3. BRANCHING

−βj to the cost of the solution to calculate the correct reduced cost of the column.

3. Branching on assignment of a customer to a specific facility: At any node of the branch-

and-price tree, if the solution is fractional and Rules 1 and 2 are not applicable, then we

consider the assignment of a customer to a specific facility. In a fractional solution, a cus-

tomer may be partially served by two or more pairings assigned to different facilities. In

this case, we can create a branching rule to force a customer to be assigned to a specific

facility. To do so, for a customer k served by more than one facility and a facility j such
P
that p∈Pj akp zp is fractional, we create two nodes.

• In the right node (say node i), we force customer node k to be served by facility j by

adding the inequality

X
akp zp ≥ 1 (3.15)
p∈Pj

to Pi . Let γjk be the dual variable associated with inequality (3.15) in RMPi0 . To

incorporate this into the pricing problem for facility j, we modify the reduced cost of

any arc entering node k with −γjk (i.e., ĉlk = ĉlk − γjk ∀l ∈ M ) and solve the pricing

problem as before. In addition, we can remove customer k from the network for the

pricing problems of other facilities.

• In the left node (say node i + 1), we forbid the assignment of customer node k to

facility j by adding to Pi+1 the constraint

X
akp zp ≤ 0.
p∈Pj

To incorporate this restriction into the pricing problem for facility j, we remove node

k from the network of facility j. The pricing problems for the other facilities do not

change.

94
3.3. BRANCHING

4. Branching on flow between pairs of customer nodes: At any node of the branch-and-

price tree, if the solution is fractional and Rules 1, 2 and 3 are not applicable, then we

may encounter the case where a customer is covered by two or more (fractional) pairings

associated with the same facility. Then, we branch on flows on single arcs between pairs

of customer nodes using a modified version of the Ryan and Foster (1981) branching rule

first suggested by Desrochers and Soumis (1989). In the original Ryan and Foster (1981)

branching rule, a pair of set partitioning constraints is selected. In one branch, the rule is

to branch on the sum of the variables that satisfy both constraints and, in the other branch,

the rule is to branch on the sum of the variables that do not satisfy both constraints simul-

taneously. In the VRP context, set partitioning constraints are associated with customers,

and Desrochers and Soumis (1989) modify the Ryan and Foster branching rule by looking

only at the variables in which the selected pair of customers are in consecutive order or not.

This modification makes it possible to incorporate this branching rule to the pricing problem

without changing the structure or the solution method.

To apply the modified Ryan and Foster branching rule, we select a pair of customer nodes

(n1 , n2 ) such that

X X
0< zp < 1, (3.16)
j∈J {p∈Pj |(n1 →n2 )∈p}

where {p ∈ Pj |(n1 → n2 ) ∈ p} is the set of pairings (of facility j) in which n1 is followed

directly by n2 for some n1 , n2 ∈ I.

In one branch, we require that the flow on the directed arc between the selected pair of

customers is one in all possible solutions. In the other branch, we forbid the flow on the

95
3.3. BRANCHING

directed arc between the selected pair of customers. We then create two new nodes:

X X
zp ≥ 1 (the left node),
j∈J {p∈Pj |(n1 →n2 )∈p}
X X
zp ≤ 0 (the right node).
j∈J {p∈Pj |(n1 →n2 )∈p}

• In the left node (say node i), we are specifying that either node n1 is followed by node

n2 in a pairing or that neither node appears. To enforce this rule for the pairings to be

generated, we update the arcs in the network for the pricing problem. In the pricing

problem, all arcs leaving node n1 and all arcs entering node n2 are deleted except the

arc (n1 , n2 ). Therefore, if a pairing includes one of the nodes, then it includes the arc

(n1 , n2 ). To delete an arc in the network, we set the cost of the arc to infinity. In the

RMPi0 , to satisfy the branching rule for the previously created pairings, we set to zero

the variables corresponding to pairings that include only one of the nodes, n1 or n2 ,

or that include both nodes but in which n2 does not immediately follow n1 .

• In the right node (say node i + 1), we are specifying that n2 cannot follow n1 in a

feasible pairing. The feasible pairings may include n1 or n2 only or may include both

of them as long as n2 does not follow n1 directly. To apply this rule for the pairings

to be generated, we simply delete the arc (n1 , n2 ) from the network in the pricing

problem. In the RMPi+1

0 , we set to zero the previously created pairings in which n1 is

followed by n2 .

These four branching rules guarantee that the branch-and-price algorithm terminates with
∗
zLRSP . We further investigate this statement in the next section.

3.3.2 Completeness

In a standard branch-and-price algorithm, nodes satisfying some conditions can be pruned. For

each node i in the tree, Step 3 of the BRANCH-PRICE (Figure 1.5) checks these conditions which

96
3.3. BRANCHING

are:

i
1. Is zLP ≥ zU B ?

i
2. Is the LP relaxation solution associated with zLP (denoted by (t̄i , w̄i , z̄ i )) feasible for P0 ?

If node i satisfies one of these conditions, it can be pruned. In our branch-and-price algorithm,

there may exist some nodes that satisfy none of the conditions listed above, but to which none

of the four branching rules described in Section 3.3.1 is applicable. This issue brings to question

whether the branching rules listed in the previous section are complete. Does our branch-and-price

algorithm for the LRSP always guarantee an optimal solution?

In our branch-and-price algorithm, if node i cannot be pruned based on the conditions (1) and

(2) and none of the branching rules is applicable, then node i satisfies the following:

i. (t̄i , w̄i , z̄ i ) is feasible but not integer, and zLP

i < zU B ,

ii. t̄i ∈ Z|J| ,

iii. w̄i ∈ Z|J| ,

P
iv. p∈P̄j akp z̄pi ∈ Z ∀j ∈ J, k ∈ I, where P̄j is the set of pairings for j ∈ J in CG , and

P P i
v. j∈J {p∈P̄j |(n1 →n2 )∈p} z̄p ∈ Z ∀n1 , n2 ∈ I.

We will show that the nodes satisfying these conditions can still be fathomed based on fathom-

ing condition (2). We can construct an integer solution from (t̄i , w̄i , z̄ i ) with the same objective

function value and may update zU B . Therefore, we will conclude that our branching rules are

complete.

In the LP solution vector satisfying conditions (ii) to (v), each open facility is fully utilized

and the set of customers is divided into disjoint non-empty subsets, each of which is assigned to a

single open facility. Therefore, (t̄i , w̄i , z̄ i ) can be decomposed into smaller solution vectors, one

for each facility. Let (t̄ij , w̄ji , z̄ji ) be the vector associated with facility j ∈ J. Each of these vectors

corresponds to a solution to a single facility LRSP instance and thus can be investigated separately.
97
3.3. BRANCHING

Based on this observation, we investigate the solution vector associated with a single facility. The

statements are true for a solution vector corresponding to any facility. In a fractional (t̄ij , w̄ji , z̄ji ),

because t̄ij and w̄ji are integer, at least two pairing variables must be fractional. By modifying the

solution values based on the structure of the pairings associated with fractional variables, we can

change this solution to an integer solution associated with the same facility. To explain how, we

will show that the pairings associated with the fractional variables satisfying conditions (ii) - (v)

have a certain structure.

Let r1 and r2 be two routes that can be included in a feasible pairing, and let I1 and I2 be the

set of customers visited by routes r1 and r2 , respectively.

Definition 3.1 Node disjoint: Routes r1 and r2 are node disjoint if I1 ∩ I2 = ∅.

Definition 3.2 Identical (Same): Routes r1 and r2 are identical if they both include exactly the

same set of customers as well as the same set of arcs between these customers.

We can represent a pairing by specifying the combination of routes included in the pairing. The

routes included in a pairing are node disjoint by construction because in a pairing, a customer

may be visited at most once. However, in a non-integer solution vector for a facility, at least two

variables associated with pairings are fractional and thus one or more customer nodes may be

included in more than one pairing. And, in fact, at least two pairings corresponding to fractional

variables must include either the same route(s), different but non-node disjoint routes, or both.

We illustrate this result using an example. Solution 1 is an example of a fractional solution (for

a single facility) that includes pairings that share only the identical routes and all other routes are

node disjoint. Solution 2 is an example of a fractional solution that includes pairings that have

non-node disjoint routes.

98
3.3. BRANCHING

Example Solution 1: Same or node disjoint routes:

z1 = 0.5 : facility1 − 1 − 2 − facility1

z2 = 0.5 : facility1 − 1 − 2 − facility1 − 3 − 4 − 5 − facility1

z3 = 0.5 : facility1 − 6 − 7 − facility1 − 3 − 4 − 5 − facility1

z4 = 0.5 : facility1 − 6 − 7 − facility1

Solution 2: Different and non-disjoint paths:

z1 = 0.5 : facility1 − 1 − 2 − facility1 − 3 − 4 − 5 − facility1

z2 = 0.5 : facility1 − 4 − 6 − 7 − facility1

z3 = 0.5 : facility1 − 5 − facility1 − 7 − 1 − 2 − 3 − facility1

Both solution 1 and 2 are fractional since at least two variables associated with pairings includ-

ing the same customers are selected. However, in solution 1, selected pairings not only include the

same customers but also include the same routes such as route facility1 - 1 - 2 - facility1 . On the

other hand, in solution 2, pairings do not include the same routes but there are at least two pairings

(such as pairings 1 and 2) that include non-node disjoint routes.

A fractional solution for a facility may be similar to either solution 1 or solution 2, or it can be

a combination of these two. However, for a solution vector satisfying conditions (i) - (v) we can

state the following result.

Lemma 3.1 In an LP relaxation solution vector associated with node i (say (t̄i , w̄i , z̄ i )) in the

branch-and-price algorithm and satisfying conditions (i) - (v), there exists at least one facility

with an associated fractional solution vector. Furthermore, a fractional solution vector for such

facility corresponds to pairings including identical or node disjoint routes (as in solution 1). At

least two pairings have at least one identical route. Finally, all non identical routes in the pairings

99
3.3. BRANCHING

are node disjoint.

Proof We have noted before that (t̄i , w̄i , z̄ i ) can be decomposed by facility because of conditions

(ii) - (iv). There must exist at least one facility (say j 0 ) with an associated fractional solution vector

because of condition (i). Let (t̄ij 0 , w̄ji 0 , z̄ji 0 ) be the fraction solution vector corresponding to facility

j 0 . Because of conditions (ii) and (iii), there must be at least two fractional pairing variables (say

z̄ji 0 1 and z̄ji 0 2 ), and thus there must exist at least one customer (say k) visited by both pairings.

Let r1 and r2 be the routes visiting customer k in pairings represented by z̄ji 0 1 and z̄ji 0 2 , re-

spectively. Because r1 and r2 are both visiting customer k, they must be either the same route or

they are not node disjoint. If they are not identical and not node disjoint, then there exists at least

one pair of customers n1 and n2 such that arc (n1 , n2 ) is either in route r1 or in r2 , but not in

both. However, this contradicts with the satisfaction of condition (v). Therefore, r1 and r2 must

include the same set of arcs and thus be identical. Therefore, if two fractional pairings visit the

same customers, they have to include exactly the same route(s) covering the same customers. This

concludes that a fractional set of pairings must have at least one route included in at least two

pairings, and non identical routes must be node disjoint.

Lemma 3.2 If (t̄i , w̄i , z̄ i ) satisfies conditions (i) - (v), then there is an integer solution of the same

value.

Proof Let (t̄ij , w̄ji , z̄ji ) be the vector obtained from (t̄i , w̄i , z̄ i ) and associated with facility j ∈ J.

Based on Lemma 3.1, the pairings corresponding to the fractional pairing variables in (t̄ij , w̄ji , z̄ji )

include only identical routes and node disjoint routes ∀j ∈ J. By eliminating identical routes

from all but one of the pairings, we can obtain a set of pairings that includes each route exactly

once. This set of pairings corresponds to an integer solution. The objective function value of the

new solution is the same with (t̄i , w̄i , z̄ i ) because both solutions have exactly the same set of node

disjoint routes and the flow on each node is 1.

Therefore, based on Lemma 3.2, by setting the LP solution value equal to the value of an

incumbent solution, we can fathom a node satisfying conditions (i) - (v) based on condition (2).
100
3.4. IMPLEMENTATION DETAILS

This proves the following statement.

Proposition 3.1 The branching rules listed in Section 3.3.1 are complete and guarantee that the

branch-and-price algorithm provides an optimal solution.

We conclude this section with the following example of an integer solution that can be obtained

from solution 1 which satisfies conditions (i) - (v).

Example Solution 3: Integer solution from solution 1 based on Lemma 3.2:

z1 = 0 : facility1 − 1 − 2 − facility1

z2 = 1 : facility1 − 1 − 2 − facility1 − 3 − 4 − 5 − facility1

z3 = 0 : facility1 − 6 − 7 − facility1 − 3 − 4 − 5 − facility1

z4 = 1 : facility1 − 6 − 7 − facility1

3.4 Implementation Details

In this section, we discuss some implementation details for the algorithm BRANCH-PRICE de-

signed to solve the LRSP.

3.4.1 Initialization

To initialize BRANCH-PRICE, we need to initialize the column set CG including location vari-

ables and an initial set of pairing variables in order to construct the initial RMP00 . In addition, the

solution of RMP00 must be feasible in order to be able to call the column generation algorithm. In

this section, we describe the implementations used to initialize CG .

101
3.4. IMPLEMENTATION DETAILS

Artificial variables

We introduce artificial variables to (ASPP-LRSP) in order to guarantee an initially feasible RMP00 .

We create an imaginary facility (refer to as artificial facility) with |I| pairings, one for each cus-

tomer. Each pairing (column) visits one customer. Let ta be the decision variable representing the

artificial facility and be equal to 1 if it is open, and 0 otherwise. Let za1 , .., za|I| be the pairing

variables associated with the artificial facility. zak represents the pairing visiting customer k. We

modify (ASPP-LRSP) by adding these artificial variables:

P
1. Add CaF ta + i∈I Ca zai to the objective function (3.1), where CaF and Ca are costs of the

artificial variables.

2. Modify constraints (3.2) by adding the artificial pairing variables:

XX
aip zp + zai = 1 ∀i ∈ I. (3.17)
j∈J p∈Pj

3. Add the following constraints related with new variables:

zai − ta ≤ 0 ∀i ∈ I (3.18)

0 ≤ zai ≤ 1 ∀i ∈ I, (3.19)

0 ≤ ta ≤ 1. (3.20)

Note that we do not actually need to add a variable for the artificial facility or to add constraints

(3.18); we do that for purposes of interpretation and consistency with the model variables. We

assign large costs to the artificial variables to discourage them being nonzero in the solution,
P
unless it is infeasible. We let CaF = (R + 1)maxj∈J {CjF } and Ca = 2C O i∈I T1,i + C V .

The model with artificial variables results in a feasible RMP00 . In addition, these artificial

variables eliminate some of the cases in which the addition of the branching rules results in an

initially infeasible RMPi0 (without artificial variables) at any node i.

102
3.4. IMPLEMENTATION DETAILS

Initial Heuristic

From previous computational experience with column generation algorithms, we expect a reduc-

tion in the number of pricing iterations required to solve the LP relaxation when starting from a

good set of columns rather than an easily constructed or arbitrary solution. To be able to improve

the performance of the first column generation algorithm in our branch-and-price algorithm, we

developed a heuristic algorithm in order to generate a set of columns and a feasible solution for

the LRSP.

In order to construct feasible solutions for the integrated LRSP, we combine a facility location

heuristic, several well-known VRP heuristics (Clark and Wright, nearest neighbor, nearest inser-

tion and sweep heuristics), and a bin-packing heuristic (best fit decreasing). We construct pairings

for a given subset of facilities by determining the following decisions sequentially:

• Allocation of customers to facilities,

• design of feasible vehicle routes, and

• assignment of routes to vehicles.

The total cost of such a sequentially constructed solution for a given subset of facilities is calcu-

lated considering the routing costs, vehicle and facility fixed costs.

To decide a subset of open facilities that would yield such a solution of lowest cost, the heuris-

tic algorithm starts with all facilities open and iteratively closes the facility whose elimination

results in the greatest cost savings until the problem becomes infeasible (this is called the DROP

heuristic in the context of facility location problems).

We refer to the overall heuristic algorithm as I-Heur. I-Heur produces a set of columns as

well as a feasible solution for the LRSP. I-Heur is composed of two main functions denoted by

DROP(J) and TCOST(J). DROP(J) returns a subset of facility set J that results in the minimum

cost calculated using function TCOST. TCOST(Js ) returns cost of a feasible solution to an LRSP

instance in which all facilities in set Js are open. TCOST(Js ), first allocates the customers to open

103
3.4. IMPLEMENTATION DETAILS

facilities in set Js , then constructs vehicle routes for each facility to cover the customers assigned

to each facility, and assigns routes to vehicles.

In order to present an outline of I-Heur, we define some notation. Let CW(j, Ij ) denotes the

application of the Clark and Wright algorithm for facility j and customers in set Ij . We use

notation R = CW(j, Ij ) to denote the set of vehicle routes generated at the end of the algorithm.

Similarly, we use notation NN(j, Ij ), NI(j, Ij ), SW(j, Ij ) to represent the applications of nearest

neighbor, nearest insertion and sweep heuristics, respectively. Let (P, C) = BFD(R) denotes the

application of the best fit decreasing heuristic to the routes in set R, which returns the set of

pairings P and the associated cost (C) including routing and vehicle fixed cost. The outline of the

algorithm I-Heur is presented in Figure 3.3.

I-Heur returns a set of columns for facilities that are selected to be open. We add these columns

to RMP00 and initialize the zU B be the cost of the solution obtained from I-Heur. To balance the

number of columns added for each facility, for the facilities that are fixed to be closed in the I-

Heur solution, we determine a set of customers that can be served from the facility within the time

limit and use function CW and BFD to determine a set of pairings for those facilities. We update

column set CG with these columns.

3.4.2 Overview

We implemented the algorithm in MINTO 3.1 using CPLEX 9.1 as the LP solver. MINTO (Savels-

bergh and Nemhauser (1996)) is a solver for mixed integer linear programs that uses a branch-

and-bound algorithm with an LP solver. MINTO has built-in applications such as preprocessing,

constraint generation and primal heuristics, but it also allows the user to design his/her own appli-

cation code such as a branch-and-cut or a branch-and-price that is specific to a problem. Within

MINTO, we customized several routines and added new ones for our algorithm. We have writ-

ten routines such as to generate an initial set of columns, do column generation and to select a

branching variable. In the following, we explain some parts of the algorithm in detail.

104
3.4. IMPLEMENTATION DETAILS

I-Heur
Input: An instance of the LRSP defined by parameters and sets in Section 2.2.1.
Output: An upper bound for the optimal solution to the instance and a set of pairings.

DROP(J):

1. Jo ← J (set of open facilities).

2. For all k ∈ Jo :
i. Close facility k. J k ← Jo \{k}.
ii. Call TCOST(J k ).
3. k ∗ ← argmink∈Jo {TCOST(J k )}.
∗
4. If TCOST(J k )< ∞, then Jo ← Jo \{k ∗ } and return to step 2.
5. Return set of open facilities, Jo , total cost TCOST(Jo ) and set of pairings con-
structed in function TCOST(Jo ).

TCOST(Jt ):

1. Allocate customers, I, to facilities in set Jt . Assign each customer to the nearest

facility whose capacity is not exceeded.
Let Ij be the partition of I for all j ∈ Jt .
P
2. COST ← j∈Jt C F (fixed cost of open facilities).
3. For all j ∈ Jt :
(a)R1 ← CW(j, Ij ) and (P1 , C1 ) ← BFD(R1 ).
(b)R2 ← NN(j, Ij ) and (P2 , C2 ) ← BFD(R2 ).
(c)R3 ← NI(j, Ij ) and (P3 , C3 ) ← BFD(R3 ).
(d)R4 ← SW(j, Ij ) and (P4 , C4 ) ← BFD(R4 ).
(e)COST ← COST + min{C1 , C2 , C3 , C4 } and Pj ← Pargmin{C1 ,C2 ,C3 ,C4 } .
4. Return COST, and set of pairings Pj for all j ∈ Jt .

Figure 3.3: Outline of the Initial Heuristic I-Heur

Column Generation. In our column generation algorithm, we prefer the exact pricing algo-

rithm 2p-ESPPRC to 1p-ESPPRC (described in Section 3.2.2). Our computational experiments in

Section 5.4.1 show that 2p-ESPPRC is more efficient than 1p-ESPPRC. Therefore, in our branch-

and-price algorithm, we use 2p-ESPPRC as an exact solution approach for the pricing problem.

In addition, computational experiments in Section 5.4.1 suggest that adding multiple columns

105
3.4. IMPLEMENTATION DETAILS

instead of a single best column at each call to 2p-ESPPRC performs better. Among the values

we tested, addition of at most 40 columns (per facility) that have the smallest negative reduced

cost performed better than the others. Therefore, in our implementation, at each iteration of the

column generation algorithm, we generate at most 40 columns. Recall that pricing problem for

each facility is solved independently. Therefore, in our column generation algorithm, the pricing

algorithms are called sequentially for each open facility.

As discussed in Section 3.2.3, solving the pricing problem exactly by generating and process-

ing all the labels can be time-consuming. To reduce the time spent, we use the two previously

described heuristic versions of the pricing algorithm before employing the exact pricing algo-

rithm. The benefit of applying heuristic pricing algorithms before 2p-ESPPRC is demonstrated

computationally in Section 5.4.2. As in exact pricing algorithm, we add at most 40 columns at

each call to the heuristic pricing algorithms.

To generalize the process, we present the steps of the overall column generation algorithm

which we refer to as E-COLGEN(u, U , l, L), where u, U , l and L are input parameters associated

with heuristic pricing algorithms in order to define their usage in the column generation algorithm.

Recall that in the heuristic pricing algorithm 2p-ESPPRC-CS, we need to set the subset size.

We use notation P =2p-ESPPRC-CS(u) to represent a call to 2p-ESPPRC-CS with subset size

u, where P is the set of pairings generated. Similarly, let P =2p-ESPPRC-LL(l) denote a call to

2p-ESPPRC-LL with a label limit l. Based on the given notation, we generalize the steps of the

E-COLGEN in our branch-and-price algorithm in Figure 3.4.

In addition, as we mentioned before, in E-COLGEN each of the pricing algorithms 2p-

ESPPRC-CS, 2p-ESPPRC-LL and 2p-ESPPRC is called sequentially for each open facility. Steps

2 to 5 are executed independently for each open facility, thus the label limit for each facility at any

iteration may be different. For a facility, if 2p-ESPPRC does not return any columns, then we skip

generating columns for that facility until P = ∅ for all non skipped facilities. Then, 2p-ESPPRC

is called for all skipped facilities. If new columns are generated for any of them, we keep calling

2p-ESPPRC until P = ∅ for all facilities.

106
3.4. IMPLEMENTATION DETAILS

E-COLGEN(u, U , l, L)
Input: An instance Pi with column set CG and parameters u, U , l and L.
i
Output: zLP and column set CG updated with new columns.

1. Do P ←2p-ESPPRC-CS(u) until P = ∅ or number of iterations ≥ U .

2. If l ≤ L, do P ←2p-ESPPRC-LL(l) until P = ∅.
Otherwise, go to step 6.

3. Let l = 2l.

4. Execute P ←2p-ESPPRC.

5. If P = ∅, then l = L + 1 and STOP. Otherwise, return to Step 2.

6. Execute P ←2p-ESPPRC until P = ∅.

Figure 3.4: Column Generation Algorithm for 1S-EBP and 2S-EBP

Branch Selection. Given a fractional solution at a node of the branch-and-price tree, we apply

the branching rules in the order that they are described in Section 3.3.1. Our computational exper-

iments showed that the order of the application of the branching rules affects the performance of

the algorithm significantly. For each branching rule, we choose the most fractional disjunction in

each branching rule. In our computational experiments, choosing the most fractional disjunction

in rule 3 and 4 results in slightly better than choosing the first disjunction that is found.

Primal Heuristic. Having good primal feasible solutions can reduce the size of the explored

tree. We use I-Heur, MINTO’s built-in primal heuristic routines, and a standard branch-and-bound

algorithm to update the upper bound in the algorithm. As we mentioned earlier, we initialize the

branch-and-price algorithm with the upper bound obtained from I-Heur. MINTO’s primal heuristic

routines search for an upper bound automatically at each node of the tree. We may also apply a

standard branch-and-bound algorithm to problem P0 with the current set of columns. This will

provide an upper bound for P0 . We use this method only at the root node and subject to a one-hour

time limit. In the next section, we will present another approach to find a primal feasible solution.

We run a heuristic version of a branch-and-price algorithm (that includes only heuristic pricing

107
3.4. IMPLEMENTATION DETAILS

algorithms) which is a more powerful procedure to improve the upper bound. In Section 5.4.3, we

present some computational experiments comparing the performance of the heuristic approaches

used and the tightness of the upper bounds obtained in each.

3.4.3 Variants

We designed two versions of the branch-and-price algorithm for the LRSP: one-stage and two-

stage branch-and-price algorithm. Figure 3.6 summarizes these two variants of the algorithm. The

two-stage branch-and-price algorithm is an extended version of the one-stage algorithm, and it is

designed based on the idea of “restart”. The two-stage branch-and-price algorithm aims to solve

larger instances. In this section, we present details of these two versions of the solution algorithm

for the LRSP. In Section 5.4.4, we present computational experiments comparing the performance

of the one-stage and the two-stage branch-and-price algorithms.

One-Stage Branch-and-Price Algorithm (1S-EBP). For small or medium size instances, we

use a one-stage solution algorithm that is an exact branch-and-price algorithm, i.e., if it is run
∗
without a time limit, the algorithm terminates with zLRSP . We refer to this version of the solu-

tion algorithm as 1S-EBP. 1S-EBP is initialized by using the upper bound and columns that are

generated by I-Heur and employs E-COLGEN at each node of the branch-and-price tree.

Two-Stage Branch-and-Price Algorithm (2S-EBP). For larger instances, we designed a two-

stage solution algorithm which is obtained by augmenting the one-stage branch-and-price algo-

rithm with an improved version of initialization. We refer to this algorithm as 2S-EBP. To initialize

the two-stage branch-and-price algorithm, we designed a heuristic version of a branch-and-price

algorithm (which we refer to as HBP) which provides a good upper bound and a good set of

columns. We describe HBP next in the section. The first stage of the two-stage branch-and-price

algorithm employs HBP. Then, using the upper bound and the columns that are obtained from

108
3.4. IMPLEMENTATION DETAILS

H-COLGEN(u, U , l, L)
Input: An instance Pi with column set CG and parameters u, U , l and L.
Output: An upper bound for zLPi and column set CG updated with new columns.

1. Do P ←2p-ESPPRC-CS(u) until P = ∅ or number of iterations ≥ U.

2. Do P ←2p-ESPPRC-LL(l) until P = ∅.

3. Let l = 2l.

4. If l ≤ L, go to step 2. Otherwise, STOP.

Figure 3.5: Column Generation Algorithm for HBP

HBP, an exact branch-and-price algorithm employing E-COLGEN at each node of the tree is ex-

ecuted as a second stage of the 2S-EBP. We can define 2S-EBP as a combination of HBP and

1S-EBP.

Heuristic Branch-and-Price Algorithm (HBP). In HBP, we construct an initial branch-and-

price tree subject to a time limit, employing only the heuristic pricing algorithms 2p-ESPPRC-LL

and 2p-ESPPRC-CS. The algorithm is initialized with I-Heur. When the time limit is reached (or

the algorithm terminates), we have a set of generated columns that are available to be used and an

upper bound for the problem. Using the notation that is given in the previous section, we present

the main steps of the column generation algorithm (referred to as H-COLGEN) that is employed

in HBP in Figure 3.5.

In H-COLGEN, we use smaller U , and smaller l and L compared to E-COLGEN. Small

parameter values shorten the total execution time of H-COLGEN. HBP generates very good upper

bounds for the LRSP instances. In Section 5.4.3, we present computational results evaluating the

upper bounds obtained in HBP compared to other heuristic solution algorithms, and discuss the
0 and z ∗
strength of the bounds with respect to zLP LRSP .

109
3.4. IMPLEMENTATION DETAILS

A set of pairings One−Stage Branch−And−Price Algorithm

Initial +
Heuristic An upper bound

Heuristic A set of pairings

Branch & Price
+
An upper bound
H−COLGEN

Exact Optimal Solution

Branch & Price
Two−Stage Branch−And−Price Algorithm Or Best Gap
E−COLGEN

Figure 3.6: Variants of the Branch-and-Price Algorithm

110
Chapter 4

Valid Inequalities

In this chapter, we investigate classes of inequalities valid for the feasible region of the LRSP

whose generation can be incorporated into our branch-and-price algorithm. Extending a branch-

and-price algorithm to include methods for dynamically generating valid inequalities is a promis-

ing approach in the effort to improve the overall performance of the solution algorithm. The

addition of cutting planes to the LP relaxation can improve bounds, but this comes at the expense

of the additional time required to generate them and to solve the resulting larger LP relaxations,

as well as the possible performance reduction in the column generation procedure caused by the

incorporation of the cuts. Hence, efficiency of an extended algorithm depends on factors related

to the branch-and-price algorithm other than the strength of the new valid inequalities.

The goal of this chapter is to provide an overview of related classes of valid inequalities,

describe how to incorporate their generation into our branch-and-price algorithm and investigate

the performance of the overall algorithm. In order to find cuts that are valid for the LRSP, we

explore classes of valid inequalities derived from known relaxations of the LRSP, such as the

knapsack problem, the set partitioning problem, the LRP and the VRP.

The main contributions of this chapter are to investigate the most common cuts valid for poly-

topes related to the LRSP polytope; discuss their validity for the LRSP; describe their integration

into the (SPP-LRSP) model; describe the necessary updates to our branch-and-price algorithm;

111
4.1. NOTATION AND OVERVIEW

and assess the value of these cuts to the solution algorithm. In addition, we explore possible ways

of adapting the disjunctive cut procedure, commonly used in branch-and-cut algorithms. We are

not aware of any publication that discusses the use of general disjunctive cuts in a branch-and-price

algorithm. This chapter also provides some theoretical results about the strength of (SPP-LRSP)

by showing that some of the cuts that can be derived using the graph-based formulation are nec-

essarily satisfied by solutions to the LP relaxation of (SPP-LRSP).

4.1 Notation and Overview

Recall the steps of the algorithm referred to as BRANCH-CUT-PRICE presented in Figure 1.6.

We use the same notation presented in Chapter 1. Since we want to extend the branch-and-price

algorithm developed based on formulation (ASPP-LRSP) to a branch-cut-and-price algorithm,

P0 in the algorithm corresponds to (ASPP-LRSP). Therefore, we may use S 0 to denote the set of

feasible solutions to problem (ASPP-LRSP). For ease of notation, we refer to the overall procedure

that generates valid inequalities based on a given solution xik as CUTGEN(xik ). This procedure

returns a set of valid inequalities separating the solution xik from the polyhedron defined by the

constraints of RMPik . Note that, in a generic BRANCH-CUT-PRICE, the cut generation procedure

can be called at any time during the column generation algorithm. In the solution procedure for the

LRSP, we only allow cut generation at the end of a complete column generation phase. Therefore,

at Step 3(b) of BRANCH-CUT-PRICE, at iteration k of the column generation algorithm, if Cki =

∅ and if we want to search for valid inequalities, we proceed to Step 4 and call CUTGEN(xik ).

Recall that the pricing problem in the column generation algorithm is defined with respect

to a particular LP relaxation of the problem and a particular dual solution to the current RMP.

Therefore, whenever we add to Pi the set of valid inequalities generated, the pricing algorithm

needs to be modified based on the classes of valid inequalities added to the subproblem. We

do not express this modification explicitly in the steps of BRANCH-CUT-PRICE. However, in

our BRANCH-CUT-PRICE algorithm, we only generate the valid inequalities for which we can

112
4.1. NOTATION AND OVERVIEW

efficiently modify the solution algorithm for the pricing problem. Thus, it is important to discuss

the characteristics of classes of valid inequalities that we can easily use in BRANCH-CUT-PRICE

by modifying the pricing algorithm and the factors affecting the efficiency of the modified pricing

algorithm. As we describe the valid inequalities for the LRSP in the remainder of this chapter,

we describe particular modifications required for the pricing algorithm and point out which of the

general categories detailed below each class falls into.

Classification of Valid Inequalities. Consider an inequality (α, β) valid for subproblem Pi of

the form

X X
αj tj + αp zp ≥ β ∀(t, z) ∈ S i . (4.1)
j∈J p∈P

|J|+|P |
Let P̄ be the set of columns generated so far and (t̄, z̄) ∈ R+ be an optimal fractional solution

to the LP relaxation of Pi . Since we use only a restricted version of the LP relaxation of Pi in order

to obtain (t̄, z̄), z̄p = 0 for all p ∈ P \P̄ . In the cut generation procedure, while we generate an

inequality (α, β) that is violated by (t̄, z̄), in order to guarantee that the inequality is valid for S i ,

we need to consider and be able to calculate the coefficients of the columns in set P \P̄ in the

inequality. We may assume that αp = 0 for all p ∈ P \P̄ while we generate (α, β), however, we

need to be able to modify the coefficients α, β as we generate new columns.

To consider the non zero α coefficients for the columns to be generated, we need to modify

the pricing problem and the pricing algorithm. In order not to reduce the overall efficiency of

the branch-cut-and-price algorithm, we are interested in the question of for which type of valid

inequalities we can modify the pricing algorithm without increasing its complexity significantly.

Spoorendonk and Desaulniers (2008a) list two properties for a valid inequality that make it fa-

vorable in a column generation algorithm: (i) we can determine αp for column p ∈ P \P̄ in a

reasonable time, and (ii) we can easily modify the pricing algorithm in order to consider the dual

variable associated with the inequality in the cost function of the pricing problem.

113
 4.1. NOTATION AND OVERVIEW

We classify the inequalities of the form (4.1) based on the structure of α to identify the cases

in which we can easily incorporate a class of valid inequalities into our branch-cut-and-price

algorithm. For a given solution, we let γ be the dual variable corresponding to the new inequality.

We define three groups:

P 0
Group 1 For all p ∈ P \P̄ , αp = (c,k)∈A lckp αck , where A is the set of arcs in the problem, lckp =1
0 is fixed and known for all (c, k) ∈ A.
if pairing p visits arc (c, k) and αck
0 γ, we can easily handle the additional
By modifying the cost of arc (c, k) by adding αck

dual variable in the reduced cost of zp . None of the pricing algorithms we use changes

(conditions we check, domination rules, etc.) because we know the exact reduced cost of

any partial path generated during the algorithm and how it will compare to other paths in

future extensions. The inequalities that satisfy this condition can be used in our branch-cut-

and-price algorithm without any significant changes in the pricing algorithm.

Group 2 For all p ∈ P \P̄ , αp = K, where K is a known constant.

The inequalities satisfying this condition can easily be used in our branch-cut-and-price

algorithm, since the reduced cost of any column is modified by a fixed amount Kγ. This

modification can be done after the pricing algorithm is terminated.

Group 3 An inequality that does not satisfy the conditions of Groups 1 and 2 is a Group 3 inequality.

As in Groups 1 and 2, the reduced cost of variable zp needs to be modified by adding

αp γ. However, the value of αp for all p ∈ P \P̄ may depend on the satisfaction of some

conditions and may be different for any pairing p. In the pricing algorithm, consider two

partial paths, p1 and p2 . Let αp1 and αp2 be the coefficients, and ĉp1 and ĉp2 be the reduced

costs calculated for each, respectively. We can compare ĉp1 and ĉp2 at the current iteration,

but we do not know how these values will change in comparison to each other for extensions

of these paths in future iterations, since we do not know what the coefficients of future

extensions will be. Therefore, we cannot use the same domination rules that we are currently

using to decide whether to keep or delete partial paths in any iteration of the algorithm. We
114
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

need to modify the domination rules in order to handle additional conditions that need to be

considered.

Note that even if it is valid to set αp = 0 for all p ∈ P \P̄ , the pricing problem still needs

to be modified. Let q such that αq > 0 be one of the columns in CG . Unless, we forbid

generation of a pairing that is in the same symmetry group (pairings including the same set

of routes) with pairing q, the column generation would generate a new column and replace

zq with it. In that case, the effect of the inequality on the LP relaxation bound is eliminated.

It is usually very difficult (except in very simple cases) to forbid generation of a particular

pairing.

Furthermore, for each class of inequalities (depending on the α) new domination rules need

to be designed. We do not have an efficient way of handling any class of inequalities that can

be counted in this group in the branch-cut-and-price algorithm. However, there are some

studies in the literature (will be discussed later) using some classes of inequalities in this

group with column generation algorithm.

Next, we investigate classes of valid inequalities for the LRSP. We discuss whether we can use

them in BRANCH-PRICE-CUT and describe procedures to be embedded in CUTGEN procedures

for the classes that can be included in BRANCH-PRICE-CUT.

4.2 Valid Inequalities For Set Partitioning-Based Formulations

In this section, we discuss some valid inequalities derived based on the set partitioning-based

formulation of the LRSP (SPP-LRSP). In Chapter 2, we tightened this formulation with two sets

of valid inequalities, (2.18) and (2.25), that are shown to be effective computationally in Section

5.3. Here, we study the LRSP polytope, defined as

P
|Pj |
PSPP = conv{t ∈ {0, 1}|J| , z ∈ {0, 1} j∈J |t, z satisfies (2.14), (2.15)}. (4.2)

115
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

This study is two-fold: first, we investigate the valid inequalities for relaxations of polytope PSPP

and second, we investigate valid inequalities for reformulations of PSPP .

4.2.1 Relaxations

In this section, we investigate two relaxations: the knapsack problem and the set partitioning prob-

lem. We list the most common classes of valid inequalities for these problems (from the literature)

based on our notation for (SPP-LRSP). Although we do not generate any of these classes of valid

inequalities in our branch-cut-and-price algorithm (because they are group 3 inequalities), we pro-

vide this review of the literature and discuss possible implementation difficulties for completeness.

Knapsack Polytope. In PSPP , if we relax constraints (2.15) by dropping the variable t, then

each constraint is a 0/1 knapsack constraint. Since we assume that the column sets for each

facility, i.e., the items available for each knapsack, are disjoint from each other, these constraints

decompose into |J| single knapsack polytopes rather than one multiple knapsack polytope. We let
j
PKNP be the knapsack polytope that corresponds to facility j ∈ J, so that

j
X
PKNP = conv{z ∈ {0, 1}|Pj | | D(p)zp ≤ LFj } ∀j ∈ J, (4.3)
p∈Pj

P j
where D(p) = i∈I aip Di , the total demand of customers in pairing p. For each j ∈ J, PKNP
j
can be seen as a relaxation of PSPP . Thus, we can investigate inequalities valid for PKNP ,

as these are also valid for (SPP-LRSP). Balas (1975), Padberg (1980), Crowder et al. (1983)

and Weismantel (1997) are some of the authors who have studied the polyhedral structure of

knapsack problems. Here, we evaluate the applicability of the following valid inequalities that are

investigated in those studies to our algorithm.

j
One set of well-known nontrivial inequalities that are valid for PKNP are the cover inequali-
j P
ties. Based on our notation and definition of PKNP , a set Rj ⊆ Pj is called a cover if p∈Rj D(p)

116
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

j
> LFj . If the set Rj is minimal, then the following inequality is valid for PKNP :

X j
zp ≤ |Rj | − 1 ∀z ∈ PKNP . (4.4)
p∈Rj

j
Another set of nontrivial inequalities defined for PKNP is the (1-k)-configuration inequalities.
P
Let Rj ⊆ Pj and p∈Rj D(p) ≤ LFj . If R0 ∪ {q} is a minimal cover for all R0 ⊆ Rj and
j
q ∈ Pj \Rj , then, the following inequality is valid for PKNP :

X j
(r − k + 1)zq + zp ≤ r ∀z ∈ PKNP , (4.5)
p∈Rj r

where r is some integer satisfying k ≤ r ≤ |Rj |, and Rjr is a subset of Rj such that |Rjr | = r.

When k = |Rj |, the inequality (4.5) is a cover inequality; therefore, cover inequalities are also

(1-k)-configuration inequalities.

The inequalities (4.4) and (4.5) are group 3 inequalities. At this time, it is not clear how to

update pricing algorithms efficiently in order to handle these, so we did not utilize these in our

branch-cut-and-price algorithm. If an efficient way of handling inequalities in the form of (4.4)

and (4.5) could be found, we conjecture they would improve the performance of our algorithm as
j
they have been very successful in the solution of PKNP .

Set Partitioning Polytope. Constraints (2.14) and the integrality constraints in PSPP define a

set partitioning polytope as follows:

P XX
|Pj |
PSP = conv{z ∈ {0, 1} j∈J | aip zp = 1 ∀i ∈ I}. (4.6)
j∈J p∈Pj

Because PSP is the intersection of set packing and set covering polytopes, valid inequalities for

the set partitioning problem can be derived by investigating the polyhedral structure of these two

polytopes (see Balas and Padberg (1976) for a review of polyhedral structures of all these poly-

topes). A common class of valid inequalities used in the literature is the clique inequalities. Let
117
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

P = ∪j∈J Pj and Ip = {i ∈ I|aip = 1} for p ∈ P . For R ⊆ P such that Ip1 ∩ Ip2 6= ∅ for all p1 ,

p2 ∈ R, the following inequality is called a clique inequality:

X
zp ≤ 1 ∀z ∈ PSP . (4.7)
p∈R

Inequality (4.7) enforces that only one pairing from such a set R may be selected. In order to

generate violated inequalities of this form, the notion of a clique is used. A clique is a complete

subgraph (a graph in which every pair of nodes are connected by an edge) of a given graph. In

order to derive clique inequalities for PSP based on this definition, a graph (generally called the

intersection graph or the conflict graph) including one node for each variable of PSP can be

constructed. In this graph, there is an edge between a pair of nodes p1 and p2 if Ip1 ∩ Ip1 6= ∅

for all p1 , p2 ∈ P . A valid inequality in the form of (4.7) can be written for each set of variables

corresponding to a clique in this graph.

Another common class of valid inequalities used for PSP is the odd-hole inequalities that are

constructed from chordless cycles with an odd number of nodes in an intersection graph associated

with an instance of the set partitioning problem. If C ⊆ P is the set of variables representing nodes

in an odd cycle, then the following is valid for PSP :

X j |C| k
zp ≤ ∀z ∈ PSP . (4.8)
2
p∈C

There are other types of inequalities such as anti-hole, webs, and wheels (see Eso (1999) for

details) that are derived from the intersection graphs in a similar way and have structure similar

to (4.7) and (4.8). Here, our aim is to provide a short review of valid inequalities for the set

partitioning problem. Since these inequalities are group 3 inequalities, we did not implement

procedures for generating any of these cuts. We do not know of a way of modifying the pricing

algorithm efficiently to incorporate these valid inequalities.

118
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

Spoorendonk and Desaulniers (2008b) developed a methodology in order to use clique in-

equalities in the column generation algorithm for the VRP. Jepsen et al. (2008) also developed

a branch-cut-and-price algorithm for a VRP in which they provided a procedure to modify the

pricing problem in order to use subset-row inequalities introduced based on the idea of clique and

odd-hole inequalities. We review this work in order to further clarify the reasons that updating

the pricing problem for inequalities in the form of (4.4) to (4.8) is not trivial and the structure of

the cut is very important for a branch-cut-and-price algorithm. They constructed these inequali-

ties based on information from the rows of the set partitioning polytope instead of considering a

certain set of variables, as in clique and odd-hole inequalities. Based on our notation, Jepsen et al.

(2008)’s subset-row inequalities are as follows

X X j1 X k j |R| k
aip zp ≤ ∀z ∈ PSP , (4.9)
k k
j∈J p∈Pj i∈R

where R ⊆ I (rows in set partitioning constraints) and 0 < k ≤ |R|. A column corresponding to

a pairing visiting at least k of the customers in set R has a nonzero coefficient in inequality (4.9);

otherwise, the coefficient is zero. They used this condition to modify the reduced cost calculation

for a column in the pricing problem, which is also an ESPPRC.

jP k
aip
In this case, the reduced cost of a column should be augmented by the term −γ i∈R
k

where γ is the dual variable that corresponds to constraint (4.9). To do that, they maintained

a value ml = (|R ∩ Vl | mod k) for each label l in the ESPPRC algorithm where Vl is the set

of customers in label l. They then modified the domination rules based on the value of ml for

each label l. They showed that these inequalities were effective and improved the performance of

their branch-cut-and-price algorithm. We continue our investigation by looking at the polytopes

obtained by reformulating the (SPP-LRSP).

119
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

4.2.2 Reformulations

In order to overcome the difficulty of incorporation of the valid inequalities for the relaxations

of PSPP reviewed in the previous section into the pricing problem, we propose here a change of
P
variables. We define vji = p∈Pj aip zp for all j ∈ J and i ∈ I and replace the variables z. The
r
polytope PSPP is then

X
r
PSPP = conv {(t, v) ∈ {0, 1}|J|+|J|×|I| | vji = 1 ∀i ∈ I, (4.10)
j∈J
X
Di vji − LFj tj ≤ 0 ∀j ∈ J}. (4.11)
i∈I

We can interpret vji as the fraction of the demand of customer i ∈ I assigned to facility j ∈
r
J. Again, we investigate some relaxations of the polytope PSPP to derive valid inequalities.

The valid inequalities obtained can then be rewritten in terms of the original variables and used

in our algorithm. We implemented two classes of the valid inequalities that will be presented

next in this section: lifted cover inequalities and generalized assignment polytope inequalities.

We implemented these in our branch-cut-and-price algorithm and presented some computational

results in Section 5.5.1.

Multiple Knapsack Polytope. If we replace variable tj with its upper bound 1 in inequalities

(4.11) and relax the equality restriction in (4.10), the resulting inequalities along with the binary

restriction for variable v define a multiple knapsack polytope:

X
PMKN = conv {v ∈ {0, 1}|J|×|I| | vji ≤ 1 ∀i ∈ I, (4.12)
j∈J
X
Di vji ≤ LFj ∀j ∈ J}. (4.13)
i∈I

Valid inequalities for the single knapsack polytope are also valid for PMKN . The cover inequali-

ties and (1-k) configuration inequalities described in Section 4.2.1 can be applied here.
P P
Let R ⊆ I be a minimal cover for facility j ( i∈R Di > LFj and i∈R\{k} Di ≤ LFj for all
120
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

k ∈ R). Then, we have

X
vji ≤ |R| − 1, (4.14)
i∈R

r
valid for PMKN and PSPP . The inequalities (4.14) can be strengthened using the lifting proce-

dure described by Nemhauser and Wolsey (1989) and Gu et al. (1998), which we implemented in

our branch-cut-and-price algorithm. MINTO has a built-in function to detect violated cover in-

equalities for 0-1 integer programs and lift them. This function implements the heuristic algorithm

proposed by Gu et al. (1998) to search for the lifted cover inequalities of the form

X X X X
vji + γji vji + αji vji ≤ |R1 | − 1 + γji ∀v ∈ PMKN , (4.15)
i∈R1 i∈R2 i∈I\R i∈R2

where R = (R1 ∪ R2 ) ⊆ I is a minimal cover for j and R1 6= ∅. We employed the MINTO

function in our algorithm to detect lifted cover inequalities violated by the portion of a given

solution consisting of the variables associated with a single facility j ∈ J.

An inequality of the form (4.15) can rewritten in terms of the variables of PSPP as follows:

X X X X X X X
aip zp + γji aip zp + αji aip zp ≤ |R1 | − 1 + γji , (4.16)
i∈R1 p∈Pj i∈R2 p∈Pj i∈I\R p∈Pj i∈R2

for all (R1 ∪ R2 ) ⊆ I, j ∈ J and ∀(t, z) ∈ PSPP . An inequality of the form (4.16) is a group

1 inequality and can easily be added to our branch-cut-and-price algorithm without destroying
j
the pricing problem structure. Let ωR be the dual variable corresponding to inequality (4.16)

associated with customer set R and facility j. To modify our algorithm to accommodate these

inequalities, we modify the cost of incoming arcs for customer i in sets R1 , R2 and I\R at facility
j j j
j by ωR , γji ωR and αji ωR , respectively. These additional costs for single customers result in a

modified total reduced cost for a column with a nonzero coefficient in inequality (4.16).
P
|J|+ j∈J |P̄j |
In our cut generation procedure, for a given fractional solution (t̄, z̄) ∈ R+ to

121
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

RMPik , where P̄j is the current set of pairings for facility j ∈ J, we do the following.

|I|
• For each facility j ∈ J, if t̄j = 1, we find v̄ j ∈ R+ using the transformation v̄jl =
P
p∈P̄j alp z̄p for all l ∈ I.

• For a given v̄ j , we try to generate a violated lifted cover inequality in the form of (4.15)

using MINTO’s function.

• If the function returns a cut, then we rewrite it in the form of (4.16) and add it to Fki .

As mentioned earlier, computational results for the implementation of this set of valid inequalities

are presented in Section 5.5.1. The (1-k) configuration inequalities are also valid for the single

knapsack polytope; thus, it would also be possible to derive (1-k) configuration inequalities for
r
PSPP in terms of variable v. In a procedure similar to that for cover inequalities, (1-k) configu-

ration inequalities in terms of variable v could be generated in a branch-cut-and-price algorithm

without requiring significant changes in the pricing algorithms. Based on the computational exper-

iments for the lifted cover inequalities, which are a special case of (1-k) configuration inequalities,

however, we decided not to implement (1-k) configuration inequalities in our study.

Generalized Assignment Polytope. The generalized assignment problem (GAP) is a general-

ization of the multiple knapsack problem. The generalization is that the GAP may be formulated

with different weights for each item (coefficients in (4.13)) for each knapsack. The generalized

assignment polytope can be defined as follows:

X
PGAP = conv{v ∈ {0, 1}|J|×|I| |v satisfy (4.10) and Di vji ≤ LFj ∀j ∈ J}.
i∈I

Valid inequalities for the knapsack polytopes are also valid for PGAP . Besides the inequalities for

the knapsack polytope, there are only a few papers investigating the facets and valid inequalities of

GAP. Some of these are Gottlieb and Rao (1990) and de Farias and Nemhauser (2001). Both pa-

pers relaxed the set partitioning constraints (4.10) and replaced them with set packing constraints.

122
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

In this relaxed GAP, de Farias and Nemhauser (2001) added the condition that for all i ∈ I, there

is at most one j ∈ J such that vji > 0. The following theorem from de Farias and Nemhauser

(2001) presents a valid inequality for GAP.

P
Theorem 4.1 [de Farias and Nemhauser (2001)] Let R ⊆ I and k ∈ J. Suppose that i∈R Di >
P
LFk and i∈R−{r} di < LFk for some r ∈ R. Then,

 
X X X X
Di vki + max{0, LFk − Da } vji ≤ LFk ∀v ∈ PGAP (4.17)
i∈R i∈R a∈R−{i} j∈J−{k}

induces a facet of the relaxed GAP polytope.

The interpretation of (4.17) is simple: if the total demand of the customers in set R exceeds

the capacity of facility k, then there must be a customer in R that is served by a facility other

than k. If the second part of the left hand side of (4.17) is zero, it is equivalent to the facility

capacity constraint for k, which is one of the trivial inequalities. However, in a non integer solution
|J|+|J|×|I| P
(t̄, v̄) ∈ R+ , if i∈I Di v̄ki = LFk and v̄ki > 0 for all i ∈ R, then there exists some h ∈ J
P
such that v̄hi > 0 for some i ∈ R. In this case, if LFk − a∈R−{r} Da > 0 for some r ∈ R

satisfying v̄hr > 0, then the inequality (4.17) is violated. We can rewrite inequality (4.17) for

PSPP :

 
X X X X X X
Di aip zp + max{0, LFk − Da } aip zp ≤ LFk , (4.18)
i∈R p∈Pk i∈R a∈R−{i} j∈J−{k} p∈Pj

P
where i∈R Di > LFk . Since the inequalities of the form (4.18) are in group 1, they can be added

to the algorithm without destroying the structure of the pricing problem.

We implemented this class of inequalities, which we refer to as GAP inequalities, and modified

our branch-cut-and-price algorithm to include the generation of valid inequalities of the form of

(4.18). We can incorporate GAP inequalities into our branch-cut-and-price algorithm by slightly
k be the dual variable corresponding to
modifying the cost function in the pricing problem. Let γR

123
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

the inequality (4.18) for customer set R and facility k. Then we modify the cost of the incoming
k (for facility k) or (max{0, LF −
P k
arcs of customers in R by adding either Di γR k a∈R−{i} Da })γR

(for all facilities other than k).

de Farias and Nemhauser (2001) present a separation heuristic for this class of inequalities.

Figure 4.1 describes the heuristic (referred to as GAP-Heur) in terms of our notation. We modified

GAP-Heur slightly in our implementation by requiring in step 1(i) that the facility variable is equal
P
|J|+ j∈J |P̄j |
to 1. To generate GAP inequalities violated by a given fractional solution (t̄, z̄) ∈ R+ ,

where P̄j is the current set of pairings for facility j ∈ J, we do the following.
P
• We determine (t̄, v̄) ∈ R|J|+|J|×|I| using the transformation v̄ji = p∈P̄j aip z̄p for all i ∈ I

and j ∈ J.

• We call GAP-Heur to separate cuts in the form of inequality (4.17).

• We rewrite the inequality in the form of (4.18).

In our experience with the LRSP instances, GAP-Heur fails to find violated cuts most of the time.

The reason for this failure is condition (4.20) in Figure 4.1. Generally, condition (4.20) is not

satisfied even though the facility capacity constraint is binding and condition (4.19) is satisfied. In

the LRSP, we are dealing with the assignment of pairings to the facilities. Therefore, fractional

pairing variables lead to multiple customers (rather than a single customer) that are served by

more than one facility. In a fractional solution, for some facility, even though the capacity is fully

utilized, it is not enough to remove the assignment of a single customer to this facility in order

to force the total demand of the remaining customers to be less than the facility capacity since

more than one customer is fractionally served by that facility. To increase the effectiveness of the

procedure, we modified the method of determining the set R in the step 1 (ii) of the GAP-Heur, so

that we may shrink this set in order to obtain a violated cut.

In our algorithm, we first execute steps 1 (ii) and 1 (iii) as in the original GAP-Heur procedure

by setting R = {i ∈ I|vki > 0}. If we cannot find a violation, we reduce the size of set R by

removing one customer at a time in the order of non increasing demand until a violation is found
124
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

GAP-Heur [de Farias and Nemhauser (2001)]

|J|×|I|
Input: An instance of the GAP and a fractional solution to its LP relaxation, v̄ ∈ R+ .
Output: A set of valid inequalities of the form (4.17) violated by v̄.

1. For each k ∈ J:
P
i. If i∈I Di v̄ki = LFk , then proceed. Otherwise go to step 1 and pick the next
facility.
P
ii. Let R = {i ∈ I|v̄ki > 0} and D(R) = i∈R Di .
If D(R) > LFk , then for each c ∈ R, check the following conditions:
– Is c also served by another facility?
X
1 > v̄kc > 0, and v̄jc > 0 (4.19)
j∈J\{k}

– When we remove c from the assignment set of facility k, does the full as-
signment of the rest of the customers to k become feasible?

D(R\{c}) < LFk (4.20)

iii. If there is at least one c ∈ R satisfying both (4.19) and (4.20), then (4.17) is
violated for facility k and customer set R. Return a GAP cut for facility k.

Figure 4.1: Heuristic Separation Algorithm for GAP inequalities [de Farias and Nemhauser
(2001)]

or D(R) ≤ LFk . The steps of the updated algorithm are listed in Figure 4.2. As mentioned earlier,

computational results for the implementation of this class of valid inequalities are presented in

Section 5.5.1.

Capacitated Facility Location Polytope. If we relax the integrality of variable v in the con-
r
straints of PSPP , we get the capacitated facility location polytope:

PCFL = conv {t ∈ {0, 1}|J| , v ∈ R|J|×|I| |(t, v) satisfies (4.10), (4.11),

and vji ∈ [0, 1] ∀j ∈ J, i ∈ I}.

125
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

GAP-Heur-2
Input: An instance of the GAP and a fractional solution to its LP relaxation, (t̄, v̄) ∈
|J|+|J|×|I|
R+ .
Output: A set of valid inequalities of the form (4.18) violated by (t̄, v̄).

1. For each facility k ∈ J:

P
i. If t̄k = 1 and i∈I Di v̄ki = LFk , then proceed. Otherwise return to step 1.
ii. Let R = {i ∈ I|v̄ki > 0}. If D(R) > LFk , then for each c ∈ R, check the
conditions (4.19) and (4.20).
iii. If there is at least one c ∈ R satisfying (4.19) and (4.20), then return a cut for
facility k and set R and return to step 1.
iv. Sort the customers in R, i1 < i2 < .. < i|R| such that Dip ≥ Diq for all
p, q = 1, .., |R| and ip < iq .
v. For e = 1 to |R|:
a. Let R = R − {ie }. If D(R) ≤ LFk , then return to step 1.
b. If there is at least one c ∈ R satisfying (4.19) and (4.20), then return a cut
for facility k and set R and return to step 1.

Figure 4.2: Outline of the Modified Heuristic Separation Algorithm for GAP inequalities for the
LRSP

Therefore, if the valid inequalities for PCFL are rewritten in terms of variables (t, z), they are
r
valid for PSPP . Indeed, PSPP , the capacitated facility location polytope with binary restriction

on variable v, is called the single source capacitated facility location polytope. We are not aware

of any studies that investigate specifically the polyhedral structure of CFLPSS. Thus, we review

the studies for PCFL , of which there are a few. Aardal et al. (1995), Leung and Magnanti (1989)

and Padberg et al. (1985) investigated valid inequalities and gave conditions for these inequalities

to be facet-inducing for PCFL . We did not implement generation procedures for any of these

classes of valid inequalities, but here we give a brief list of common ones.

One class of valid inequalities is the knapsack cover inequality, defined in terms of the fa-
P
cility variables t. Aardal et al. (1995) call a subset M of the facilities a cover if j∈M LFj >
P F
j∈J Lj − D(I). Based on this definition, they stated the following theorem:

126
4.2. VALID INEQUALITIES FOR SET PARTITIONING-BASED FORMULATIONS

Theorem 4.2 [Aardal et al. (1995)] If M is a minimal cover with respect to J and I, LFmin =
P
minj∈M {LFj }, and j∈J\M LFj + LFmin > D(I), then the knapsack cover inequality

X
tj ≥ 1 ∀(t, v) ∈ PCFL (4.21)
j∈M

is facet-inducing for PCFL ∩ {t ∈ {0, 1}|J| : tj = 1 ∀j ∈ J\M }.

This theorem requires that at least one facility from set M must be open because the total ca-

pacity of the facilities in the set J\M is not enough to cover the total demand of customers. Since
r
(4.21) is valid for PCFL , then it is also valid for PSPP and PSPP . Since the inequality defined in

Theorem (4.2) only has nonzero coefficients for the t variables, we can add this inequality without

affecting the structure of the pricing problem.

We did not implement the separation algorithms for this inequality in our study because of

the effect of inequality (2.18) in (ASPP-LRSP). In the LRSP instances used for our computational

results, the facility capacity values are all equal. Based on this, when all possible inequalities of
P
the form (4.21) summed, we get j∈J tj ≥ m where m ∈ R+ . With some algebra, we can

easily show that the minimum number of required facilities R (in (2.18)) is greater than m, if

1 < R < |J| (which is the case in our instances). Thus, the inequality (2.18) requiring that

the total number of facilities must be greater than or equal to R is stronger than the inequality

obtained by summing all possible inequalities in the form of (4.21). In addition, we do not add

the knapsack cover inequalities to (ASPP-LRSP), since the addition of inequality (2.18) generally

results in integer values for the facility variables in the LP relaxation solution to (ASPP-LRSP)

and its subproblems in the tree.

Another class of valid inequality for PCFL is the flow cover inequalities. Aardal et al. (1995)
P
and Padberg et al. (1985) defined M as a flow cover if M ⊂ J and j∈M LFj > D(I). Aardal

et al. (1995) and Padberg et al. (1985) presented the following theorem (expressed in terms of the

notation in this section) that describes the flow cover inequalities:

127
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

Theorem 4.3 [Aardal et al. (1995)] Let M ⊂ J be a flow cover with respect to J and I and
P F F
P F F
j∈M Lj = D(I) + λ. If maxj∈M {Lj } > λ and j∈J Lj > D(I) + Lr ∀r ∈ M , then the

flow cover inequality

XX X
Di vji + (LFj − λ)+ (1 − tj ) ≤ D(I) ∀(t, v) ∈ PCFL (4.22)
j∈M i∈I j∈M

is facet-inducing for PCFL .

If all customers are assigned to the facilities in M , the excess capacity λ leads at least one facility

in set M to be fractional in any solution to the LP relaxation associated with PCFL . The aim of

the inequality (4.22) is to eliminate the effect of excess capacity. The second part in the left-hand

side of the inequality forces all facilities in set M to take value close to 1. In inequality (4.22), if
P
we replace the variable vji with p∈Pj aip zp , we have the following corresponding inequality for

PSPP :

X XX X
Di aip zp + (LFj − λ)+ (1 − tj ) ≤ D(I) ∀(t, z) ∈ PSPP . (4.23)
j∈M p∈Pj i∈I j∈M

In a procedure similar to that which we described before, we can incorporate the inequality (4.23)

into a branch-cut-and-price algorithm since it is a group 1 inequality. However, we did not imple-

ment procedures for generating the class of this inequalities in our study because in most of our
|J|
LP relaxation solutions t̄ ∈ Z+ .

4.3 Valid Inequalities Based on Graph-Based Formulations

Another approach to deriving problem-specific classes of valid inequalities for the LRSP is to

study the LRSP polytope PGBP defined as the convex hull of the set of feasible solutions to

(G-LRSP). When a valid inequality is derived in terms of the variables in (G-LRSP), then this

inequality can be transformed to one valid for (SPP-LRSP) by transformation of variables using

128
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

the equalities (2.30)- (2.33) defined in the DWD of (G-LRSP) in Section 2.4.

The literature does not include any studies devoted to polyhedral structure of the LRSP. How-

ever, there are studies that investigate valid inequalities for the VRP and the LRP, each of which

are relaxations of the LRSP. In this section, we investigate some sets of valid inequalities derived

for the VRP or the LRP and originally defined in terms of graph-based variables. We discuss how

we can incorporate the generation of these inequalities into our algorithm for solving (SPP-LRSP)

and whether these inequalities have a positive impact. In fact, we either learned from the literature

or proved herein that most of these inequalities are already satisfied by all solutions to the LP

relaxation of (SPP-LRSP).

The most commonly used formulations in studies of the polyhedral structure of the VRP and

the LRP are two-index vehicle flow formulations. These formulations use binary variables as-

sociated with the arcs between pairs of customer or facility locations as in Section 2.2.1. Valid

inequalities for the two-index formulations can be transformed to one valid for (SPP-LRSP) by

using the transformation

XX
xik = likp zp ∀(i, k) ∈ A, (4.24)
j∈J p∈Pj

where xik is the aforementioned binary variable and likp = 1 if arc (i, k) is in pairing p. In

some graph-based formulations, like (G-LRSP), variable x has also a third index associated with

the vehicles in the problem. By summing the three-index variables associated with an arc for all

vehicles in the problem, we can obtain the corresponding two-index variable associated with the

same arc.

Certain classes of valid inequalities used in branch-and-cut algorithms for the VRP or the

LRP are satisfied by all solutions to the LP relaxations of the set partitioning-based formulations

of these problems that are analogous to our formulation. In the literature, there are a limited

number of studies investigating the theoretical relationship between set partitioning-based formu-

lations and graph-based formulations of the VRP, but there are none for the LRP. Letchford and

129
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

González (2006) proved theoretically that a solution to the LP relaxation of a set partitioning-

based formulation for the VRP (with elementary resource constrained paths) implies some of the

valid inequalities (such as knapsack large multistar, subtour elimination, generalized large mul-

tistar, hypotour-like inequalities) that are derived for a graph-based formulation in the two index

space. This result is also applicable to the LRSP, which means that LP relaxation solutions to

(SPP-LRSP) satisfy these type of cuts. Ropke and Cordeau (2009) did a similar study for the

VRPTW with pickup and delivery, investigating the relationship between a branch-and-cut and

branch-cut-and-price algorithm for that problem. They showed that some of the valid inequali-

ties for the problem (e.g., fork inequalities, reachability inequalities, precedence inequalities) are

implied by the set partitioning-based formulation of the problem obtained with a pricing problem

that generates elementary resource constrained paths.

We started our study by investigating two of the most common classes of inequalities for the

VRP: fractional capacity inequalities and rounded capacity inequalities.

Fractional Capacity Inequalities. Letchford and González (2006) noted that fractional capac-

ity inequalities are dominated by generalized large multistar (GLM) inequalities and GLM in-

equalities are a subclass of the knapsack large multistar inequalities, which are proved to be im-

plied by the set partitioning-based formulation of the VRP. Therefore, the results of Letchford and

González (2006) prove that fractional capacity inequalities are redundant for the set partitioning

formulation. Here, we re-derive this result. Let PVRP be the VRP polytope based on the two-

index vehicle flow formulation. Then the fractional capacity inequality for R ⊆ I is as follows:

P
X X
i∈R Di
xik ≥ ∀x ∈ PVRP . (4.25)
LV
i∈R k∈V \R

In other words, the total capacity of the vehicles leaving R must be greater than the total demand

130
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

of customers in R. Using equation (4.24), we can rewrite inequality (4.25) for the (SPP-LRSP) as

P
XXX X XX
i∈R Di
likp zp = αp z p ≥ ∀(t, z) ∈ PSPP , (4.26)
LV
j∈J p∈Pj i∈R k∈V \R j∈J p∈Pj

where αp is the coefficient that corresponds to zp for all p ∈ ∪j∈J Pj . Here, αp is the number of

arcs from set R to set V \R visited in pairing p. As before, we let Ip = {i ∈ I|aip = 1}. For

ease of notation, define nK as a bound on the minimum number of vehicle routes that are needed
P
Di
to cover customer demand in set K, i.e., nK = d i∈K
LV
e. Since a pairing (say p) represents a

feasible vehicle schedule, by construction, the total number of vehicles leaving customers in set

K ⊆ Ip is at least nK . The following lemma derives from this fact.

Lemma 4.1 For pairing p, let K = Ip ∩ R, where R is the set of customers with respect to which

an inequality of the form (4.25) is stated. If K 6= ∅, then the contribution (coefficient) of variable

zp to the left hand side of the inequality is at least nK . In other words, αp ≥ nK .

|J|+|∪j∈J P̄j |
Let (t̄, z̄) ∈ R+ be a solution to the LP relaxation of (SPP-LRSP) where P̄j is the current

set of pairings for facility j ∈ J.

Proposition 4.1 If (t̄, z̄) satisfies (2.14) and (2.15), then it also satisfies (4.26).

Proof Let us first multiply constraint (2.14) for each customer i by its demand Di for all i ∈ R

and then sum over all customers in set R:

XXX X
Di aip z̄p = Di ,
j∈J p∈P̄j i∈R i∈R
XX X
D(K)z̄p = D(R),
j∈J K⊆R p∈P̄jK

where P̄jK ⊆ P̄j such that Ip ∩ R = K for all p ∈ P̄jK , K ⊆ R and j ∈ J. Note that

P̄jK1 ∩ P̄jK2 = ∅ for all K1 , K2 ⊆ R such that K1 6= K2 and K1 , K2 are nonempty. Since

the pairings are feasible with respect to vehicle capacity by construction, nK LV ≥ D(K) for all
131
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

K ⊆ R. Therefore,

XX X XX X
nK LV z̄p ≥ D(K)z̄p = D(R),
j∈J K⊆R p∈P̄jK j∈J K⊆R p∈P̄jK
XX X D(R)
nK z̄p ≥ .
LV
j∈J K⊆R p∈P̄ K
j

From lemma 4.1:

XX X XX X D(R)
αp z̄p ≥ nK z̄p ≥ .
LV
j∈J K⊆R p∈P̄ K j∈J K⊆R p∈P̄ K
j j

Therefore, inequality (4.26) is satisfied for any solution to the LP relaxation of the (SPP-LRSP).

Rounded Capacity Inequalities. Another common class of inequalities is rounded capacity

inequalities. These inequalities dominate the fractional capacity inequalities and they have the

following form:

X X
xik ≥ nR ∀x ∈ PVRP , (4.27)
i∈R k∈V \R

P
Di
where nR = d i∈R
LV
e for some customer set R. Here, we investigate whether the inequalities

obtained by mapping these rounded capacity inequalities to the variable space of PSPP can im-

prove the LP relaxation associated with PSPP . Based on equation (4.24), if we rewrite (4.27) in

terms of z variables, we obtain:

XXX X XX
likp zp = αp zp ≥ nR ∀(t, z) ∈ PSPP . (4.28)
j∈J p∈Pj i∈R k∈V \R j∈J p∈Pj

132
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

|J|+|∪j∈J P̄j |
Let (t̄, z̄) ∈ R+ be a solution to the LP relaxation of (SPP-LRSP), where P̄j is the

current set of pairings for facility j ∈ J.

Lemma 4.2 If nR = 1 or nR = |R|, then (t̄, z̄) satisfies inequality (4.28).

Proof If nR = 1, inequality (4.28) is equal to a subtour elimination inequality, which is implied

by the constraints of the set partitioning formulation (Letchford and González (2006)).

If nR = |R|, then no two customers in R can be visited in the same route. Thus, if Lemma

4.1 is applied, the coefficient of pairing p that visits customers in set K ⊆ R in (4.28) is at least

nK = |K|. If we sum the constraints (2.14) for all customers in set R, the validity of inequality

(4.28) follows:

XXX
aip z̄p = |R|,
j∈J p∈P̄j i∈R
XX X XX X
αp z̄p ≥ |K|z̄p = |R|.
j∈J K⊆R p∈P̄ K j∈J K⊆R p∈P̄ K
j j

If the minimum number of vehicle routes that are needed to serve customers in R is greater than

1 and less than |R|, then the rounded capacity inequalities may be violated by some fractional LP

solutions of SPP-LRSP. An example of such a solution for a small instance follows:

Example Instance with four customers and two facilities:

• Let I = {1, 2, 3, 4}, J = {1, 2}.

• Demand is 20 for all customers. Vehicle capacity is 50, and facility capacity is 70.

• Let (t̄, z̄) be a solution to the LP relaxation of (SPP-LRSP).

133
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

• Nonzero variable values and the pairings represented by pairing variables are:

t̄1 = t̄2 = 1

z̄1 = 1 : facility1 − 4 − facility1

z̄2 = 0.5 : facility2 − 3 − 1 − facility2

z̄3 = 0.5 : facility2 − 1 − 2 − facility2

z̄4 = 0.5 : facility2 − 3 − 2 − facility2

• Rounded capacity inequality for R = {1, 2, 3} is violated,

X X XXX X
xik = likp z̄p = 1.5 < 2.
i∈R k∈V \R j∈J p∈Pj i∈R k∈V \R

Although it is possible to have an impact with these inequalities, we have decided to leave the

implementation of these inequalities in our branch-cut-and-price algorithm as future work since

we have reasons to believe that the probability of finding a violated rounded capacity inequality

is low for the LRSP instances because of the multiple trips to the facility in a pairing. For a pair-
P P P P
ing p including multiple routes, i∈R k∈V \R likp ≥ i∈R k∈V \R likq , where q is a pairing

composed of single route. Consider the following solution based on the example presented before.

The rounded capacity inequality for R = {1, 2, 3} is not violated for this solution.

t̄1 = t̄2 = 1

z̄1 = 1 : facility1 − 4 − facility1

z̄2 = 0.5 : facility2 − 3 − 1 − facility2 − 2 − facility2

z̄3 = 0.5 : facility2 − 1 − 2 − facility2 − 3 − facility2

134
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

Next, we investigate some cutting planes used for the LRP, discuss their validity for the LRSP,

and assess their impact with respect to the formulation (SPP-LRSP).

Chain Barring Inequalities. Laporte et al. (1986) introduced chain barring constraints to elim-

inate paths from one depot to another in a solution of the LRP. Let PLRP be the LRP polytope

based on a two-index formulation of the problem. Let r = (j1 , i1 , i2 , i3 , .., ik , j2 ) be a path from

node j1 to j2 where j1 and j2 are two depots and i1 , i2 , .., ik are customer nodes. To eliminate

path r, the following chain barring inequalities are defined:

xj1 i1 + 3xi1 i2 + xi2 j2 ≤ 4, if k = 2, ∀x ∈ PLRP , (4.29)

k−1
X
xj1 i1 + 2 xit it+1 + xik j2 ≤ 2k − 1, if k ≥ 3, ∀x ∈ PLRP . (4.30)
t=1

We will prove that the LP relaxation of (SPP-LRSP) satisfies the following chain barring inequal-

ities that are obtained by rewriting (4.29) and (4.30) in terms of the variables defining PSPP :

X XX X
lj1 i1 p zp + 3 li1 i2 p zp + li2 j2 p zp ≤ 4, if k = 2, (4.31)
p∈Pj1 j∈J p∈Pj p∈Pj2

X k−1
XXX X
lj1 i1 p zp + 2 lit it+1 p zp + lik j2 p zp ≤ 2k − 1, if k ≥ 3, (4.32)
p∈Pj1 j∈J p∈Pj t=1 p∈Pj2

∀(t, z) ∈ PSPP . For ease of notation, let αp be the coefficient corresponding to variable zp for

all p ∈ ∪j∈J Pj . Let (t̄, z̄) be a solution to the LP relaxation of (SPP-LRSP) and in this solution

r = (j1 , i1 , i2 , i3 , .., ik , j2 ) be a path from depot j1 to j2 including k customers.

Lemma 4.3 If pairing p ∈ ∪j∈J Pj visits exactly s customers in path r, then it can visit at most s

arcs in path r. Depending on the number of customers in the path, the contribution (coefficient)

of z̄p to the left hand side of inequalities (4.31) and (4.32) is as follows:

i. If k = 2, the contribution of p in (4.31) is 3(s − 1) + 1 = 3s − 2.

ii. If k ≥ 3, the contribution of p in (4.32) is 2(s − 1) + 1 = 2s − 1.

135
4.3. VALID INEQUALITIES BASED ON GRAPH-BASED FORMULATIONS

Proof If p visits one customer in the path, then it can use at most one arc in r. If p only visits one

of the customers in {i2 , .., ik−1 }, it cannot include any arc included in r. If p visits either i1 or ik ,

then it can include either the arc (j1 , i1 ) or (ik , j2 ). The path cannot include both arcs since from

the construction of p, p can only visit one of the facilities. Therefore, the contribution of variable

z̄p to both (4.31) and (4.32) can be at most one because of either (j1 , i1 ) or (ik , j2 ).

If p visits s (s > 1) customers included in r, then it can use the largest number of arcs in r

only if it visits these customers consecutively by using s − 1 arcs. And if p visits the first (i1 ) or

last customer (ik ) in the path, then it can also include either the arc (j1 , i1 ) or (ik , j2 ). Therefore, p

can use at most s arcs in r. The contribution of p to (4.31) and (4.32) follows from the coefficients

in the inequalities and the fact that at most s − 1 of s arcs are between two customers.

Proposition 4.2 If a vector (t̄, z̄) satisfies (2.14) and (2.15) then it also satisfies the chain barring

inequalities (4.31) and (4.32).

Proof Let r = (j1 , i1 , i2 , i3 , .., ik , j2 ) be a path from depot j1 to j2 including customers i1 , i2 , .., ik

and assume that z̄ violates either (4.31) or (4.32). Let P̄js be the set of pairings visiting exactly s

of the customers in r for facility j, for all j ∈ J and s = 1, .., k.

We sum the constraints (2.14) for customers in set R = {i1 , i2 , .., ik }:

XXX
aip z̄p = k,
j∈J p∈P̄j i∈R
k X
XX
sz̄p = k. (4.33)
j∈J s=1 p∈P̄js

For k = 2, using lemma 4.3 part (i):

XX 2
XX X XX XX
αp zp ≤ (3s − 2) z̄p = z̄p + 4 z̄p
j∈J p∈P̄j j∈J s=1 p∈P̄js j∈J p∈P̄j1 j∈J p∈P̄j2
XX XX
≤ 2( z̄p + 2 z̄p ) = 4. (4.34)
j∈J p∈P̄j1 j∈J p∈P̄j2

136
4.4. DISJUNCTIVE VALID INEQUALITIES

The equality in (4.34) follows from the equality (4.33) for k = 2. Therefore, (4.31) cannot be

violated by any solution of the LP relaxation of (SPP-LRSP).

For k ≥ 3, using lemma 4.3 part (ii), we have:

XX k
X XX
αp zp ≤ (2s − 1) z̄p
j∈J p∈P̄j s=1 j∈J p∈P̄js

k
X (2k − 1) X X
≤ (s ) z̄p = 2k − 1. (4.35)
k
s=1 j∈J p∈P̄j
s

The second inequality in (4.35) follows from the fact that (2s − 1) ≤ s (2k−1)
k for all s = 1..k.

To obtain the equality, multiply (4.33) by (2k − 1)/k. The right hand side of (4.33) after the

multiplication with (2k − 1)/k is equal to 2k − 1. Therefore, (4.32) cannot be violated by any

solution of the LP relaxation of (SPP-LRSP).

4.4 Disjunctive Valid Inequalities

In this section, we discuss the generation of disjunctive valid inequalities in the context of a branch-

cut-and-price algorithm. As discussed in Section 1.2.5, many general-purpose valid inequalities

(such as Gomory cuts and mixed integer rounding cuts) for mixed integer programming models

can be derived based on disjunctions and can be considered to be disjunctive valid inequalities.

Although generation of these general-purpose valid inequalities is very common in branch-and-cut

algorithms, we haven’t yet seen a study in the literature in which disjunctive valid inequalities are

utilized in a branch-cut-and-price algorithm, with the exception of a recent presentation describing

the application of disjunctive valid inequalities in a branch-cut-and-price algorithm developed for

the capacitated VRP (Ropke (2008)). Here, we first describe how we can incorporate disjunc-

tive valid inequalities into a branch-cut-and-price algorithm in general and then describe some

procedures for generating disjunctive valid inequalities in the context of our branch-cut-and-price

algorithm for the LRSP.

137
4.4. DISJUNCTIVE VALID INEQUALITIES

The theory behind and the main steps of the procedure to generate disjunctive valid inequalities

are explained in Section 1.2.5. We briefly summarize the procedure here. Let (P) be the MILP

for which we want to generate disjunctive valid inequalities. We refer to (P) as the base model in

the cut generation procedure. The vector x̄ is such that it is not contained in S (possibly x̄ is a

solution to the LP relaxation of (P)). In order to generate an inequality valid for S but violated by

x̄, we determine a disjunction valid for S but violated by x̄. Here, we consider a disjunction with

only two terms: B 1 x̄ ≥ b1 ∨ B 2 x̄ ≥ b2 . For a given base problem (P), a vector x̄ and a binary

disjunction, we construct the cut generating LP (CGLP) that is defined by constraints (1.3) - (1.6)

and objective (1.2) and solve it. If the optimal solution value for the CGLP is negative, the result

is an inequality (α, β) that is valid for S and violated by x̄.

In a branch-cut-and-price algorithm, at any node of the tree (say node i), in order to generate

disjunctive inequalities valid for S i , we need to determine what our base model should be. At

node i, we have only a restricted version of the subproblem Pi defined by a subset of columns

CG . If we try to generate a valid inequality based on the restricted version of Pi , we will have

problems. First, it is not clear how we can find a disjunction valid for S i in the cut generation

procedure. Second, even if we are able to define a disjunction, the inequality generated based on

the restricted version of Pi is not necessarily valid for S i . We need to modify the inequality as we

generate new columns, therefore the inequality would not have a certain structure and would be a

group 3 inequality which is difficult to include in a branch-cut-and-price algorithm.

Instead of using a restricted version of Pi as a base model in the disjunctive cut generation

procedure, we may use an alternative model that has a well-defined relationship with Pi . The

basic idea that we develop is to use the alternative model as a base model to generate disjunctive

valid inequalities and rewrite the generated inequality using the defined relationship between the

alternative model and Pi . We have already presented several such models in previous sections of

the thesis. Recall that we have derived the relationship between (G-LRSP) and (SPP-LRSP) by

applying the DWD methodology to (G-LRSP) in Section 2.4. In Section 4.2.2, we present some

reformulations of (SPP-LRSP) that have a small number of variables for which solution of the LP

138
4.4. DISJUNCTIVE VALID INEQUALITIES

relaxation does not require column generation. Later in this section, we define a reformulation of

a relaxation of (G-LRSP) for which there is a well-defined relationship with (SPP-LRSP).

Let PiR be a reformulation of Pi . To simplify the presentation of our disjunctive cut generation

procedure, we let Y i ≡ Ψ(xi ) denote the mapping of xi (in the variable space of Pi ) to Y i

(in the variable space of PiR ), which is a set of equivalent solutions in the variable space of PiR .

We use set Y i to indicate that the mapping might be one-to-many, e.g., because of symmetry

in the formulation of PiR . For y ∈ Y i , we may generate a disjunctive valid inequality (αy , β y ),

maximally violated by y in the variable space of PiR and generated using the CGLP based on PiR . If

|Y i | = 1, we may rewrite (αy , β y ), the valid inequality obtained by using y ∈ Y i , as an inequality

(αx , β x ) valid for Pi . However, if |Y i | > 0, we try to come up with a valid inequality (αy , β y )

that is attained by combining the disjunctive valid inequalities obtained by using each member of

set Y i individually. For example, in a reformulation in which the variables are indexed by vehicle,

such as formulation (G-LRSP), the fact that the vehicles are homogeneous is one reason for a one-

to-many mapping. Requiring identical coefficients for vehicles located at the same facility solves

this problem. To clarify this idea, we describe some implementations below which |Y i | > 0.

In Figure 4.3, using this notation, we outline the main steps of the disjunctive cut generation

procedure (referred to as DISJCUTGEN) for node i in the branch-cut-and-price algorithm. DIS-

JCUTGEN generates a set of valid inequalities for a given instance of the problem defined based

on the current set of columns and a fractional solution to its LP relaxation.

Next, we describe four implementations of the disjunctive cut generation procedure applied in

our branch-cut-and-price algorithm for the LRSP. Each implementation chooses a different base

model and different disjunctions. The details of each implementation are as follows.

Implementation 1. In this implementation, we use the reformulation of (SPP-LRSP) defined in

Section 4.2.2. As we noted before, we use (ASPP-LRSP) in the branch-cut-and-price algorithm

and we can rewrite and add valid inequalities (2.18) and (2.25) to the reformulation. The LP

139
4.4. DISJUNCTIVE VALID INEQUALITIES

DISJCUTGEN
Input: A subproblem Pi with column set CG and xiLP .
Output: An inequality (αx , β x ) valid for Pi and violated by xiLP if any.

1. Let PiR be the reformulation of Pi . Set Y i ← Ψ(xiLP ), where Y i is the set of solutions
in the variable space of PiR obtained by mapping Ψ.

2. Generate a valid inequality (αy , β y ) by combining disjunctive valid inequalities ob-

tained for y (using the CGLP based on PiR , y and some disjunction violated by y) for
all y ∈ Y i .

3. If (αy , β y ) can be rewritten in the variable space of Pi , obtain (αx , β x ). Otherwise,

STOP.

4. If (αx , β x ) is violated by xiLP , return the inequality. STOP.

Figure 4.3: Outline of the Disjunctive Cut Generation Procedure for Branch-Cut-and-Price

relaxation of the reformulation used in this implementation (referred to as (M1 )) is as follows:

X
(M1 ) vji = 1 ∀i ∈ I,
j∈J
X
Di vji − LFj tj ≤ 0 ∀j ∈ J,
i∈I
X
tj ≥ R,
j∈J
vji − tj ≤ 0 ∀j ∈ J, ∀i ∈ I,

0 ≤ tj ≤ 1 ∀j ∈ J,

0 ≤ vji ≤ 1 ∀j ∈ J, ∀i ∈ I.

P
|J|+ j∈J |P̄j |
Let (t̄, z̄) ∈ R+ be a fractional solution to the LP relaxation of Pi at node i, where P̄j is
|J|+|J|×|I|
the set of pairings generated so far for facility j ∈ J. We can transform vector (t̄, v̄) ∈ R+
P
in the variable space of (M1 ) using v̄ji = p∈P̄j aip z̄p for all i ∈ I, j ∈ J.

If (t̄, v̄) is fractional, we consider the following disjunctions in order to determine a disjunction

violated by (t̄, v̄).

140
4.4. DISJUNCTIVE VALID INEQUALITIES

1. Flow from a facility to a customer must be either one or zero:

vji ≤ 0 ∨ vji ≥ 1 ∀j ∈ J, i ∈ I. (4.36)

2. Total number of customers assigned to a facility must be an integer:

X X
vji ≤ gj ∨ vji ≥ gj + 1 ∀j ∈ J, (4.37)
i∈I i∈I

for some integer gj and for all j ∈ J.

After discovering a violated disjunction of the form (4.36) or (4.37), we write the CGLP based on

(M1 ) and investigate a disjunctive valid inequality that is maximally violated by (t̄, v̄) in the form

X XX
αjt tj + v
αji r
vji ≥ β ∀(t, v) ∈ PSPP . (4.38)
j∈J j∈J i∈I

P P P
The objective function of the CGLP is to minimize j∈J αjt t̄j + j∈J i∈I
v v̄ − β. If the
αji ji

optimal solution value is negative, the CGLP results in a valid inequality (α, β) ∈ R|J|+|J|×|I|+1 .

We can rewrite this inequality in the variable space of PSPP as

X XXX
αjt tj + v
αji aip zp ≥ β ∀(t, z) ∈ PSPP . (4.39)
j∈J j∈J p∈Pj i∈I

A valid inequality in the form of (4.39) is a group 1 inequality and thus can easily be added to

the branch-cut-and-price algorithm by modifying the costs of the incoming arcs for customer i
v ω for all i ∈ I in the pricing algorithm for facility j ∈ J. ω is the dual variable
by adding αji

corresponding to (4.39).

Implementation 2. In the second implementation of DISJCUTGEN, we chose a graph-based

formulation that is a reformulation of a relaxation of (G-LRSP). In (SPP-LRSP), each pairing

corresponds to a possible vehicle schedule associated with any vehicle. Thus, the vehicles are

141
4.4. DISJUNCTIVE VALID INEQUALITIES

not differentiated with an index. In light of this information, to be able to relate a generated valid

inequality to (SPP-LRSP), we reformulate a relaxation of (G-LRSP) by mapping its variables with

P
vehicle index to another variable space without the vehicle index. Let ukij = h∈Hj xkih for all
P
(k, i) ∈ A, j ∈ J, and wj = h∈Hj vh for all j ∈ J. Variable ukij will take value 1 if any vehicle
r
located at facility j travels arc (k, i). wj is the number of vehicles used at facility j. Let PGBP

denote the polytope defined by the feasible solutions to this reformulation. We use the following

LP relaxation formulation referred to as (M2 ) in the disjunctive cut generation procedure.

XX
(M2 ) ukij = 1 ∀i ∈ I, (4.40)
j∈J k∈V
X X
− Di ukij + CjF tj ≥ 0 ∀j ∈ J, (4.41)
i∈I k∈V
X X
uikj − ukij = 0 ∀i ∈ I, j ∈ J, (4.42)
k∈V k∈V
XX
T Lwj − tki ukij ≥ 0 ∀j ∈ J, (4.43)
k∈V i∈V
0 ≤ tj ≤ 1 ∀j ∈ J, (4.44)

0 ≤ uikj ≤ 1 ∀i, k ∈ V, j ∈ J, (4.45)

wj ≥ 0 ∀j ∈ J. (4.46)

As with (G-LRSP), constraints (4.40) require that each customer must be served. Constraints

(4.41) are facility capacity constraints. Constraints (4.42) equate the number of vehicles in and

out of a customer node. In contrast, (G-LRSP) has constraints (2.2) which are similar to con-

straints (4.42) but disaggregated by vehicle. We consider an aggregated version of (2.2) in the

reformulation. Similarly, constraints (4.43) are the reformulation of the aggregated version of the

time limit constraints (2.7) in (G-LRSP). While constructing (M2 ), we dropped some constraints

of (G-LRSP) that include the flow variables (represented by vector y). Therefore, (M2 ) is a refor-

mulation of a relaxation of (G-LRSP).

Variable y in (G-LRSP) indicates the amount of vehicle load between pairs of locations in the

142
4.4. DISJUNCTIVE VALID INEQUALITIES

problem. The reason for leaving out the constraints including flow variables is that we want to

avoid generating valid inequalities including nonzero coefficients for flow variables. Based on the

DWD procedure, we know how to rewrite the flow variables in terms of variables of (SPP-LRSP)

using equation (2.31). However, we do not know how to handle valid inequalities of this type in the

pricing algorithm. By using (M2 ), we aim to generate disjunctive valid inequalities not including

original flow variables or any reformulation of them. (M2 ) enables a search for disjunctive valid

inequalities of the form

X X XXX
αjt tj + αjw wj + u
αkij r
ukij ≥ β ∀(t, w, u) ∈ PGBP . (4.47)
j∈J j∈J j∈J k∈V i∈V

P
|J|+ j∈J |P̄j |
Let (t̄, z̄) ∈ R+ be a fractional solution for the LP relaxation of Pi . Define (t̄, w̄, ū) ∈
|J|+|J|+|J|×|A| P P
R+ using w̄j = p∈P̄j z̄p and ūkij = p∈P̄j lkip z̄p for all (k, i) ∈ A, j ∈ J.

If (t̄, w̄, ū) is fractional, then we consider the following disjunctions in order to determine a

disjunction violated by (t̄, w̄, ū).

1. A facility is either open or closed.

tj ≤ 0 ∨ tj ≥ 1 ∀j ∈ J. (4.48)

2. Total number of vehicles used at a facility is an integer.

wj ≤ gj ∨ wj ≥ gj + 1 ∀j ∈ J, (4.49)

for some integer gj .

3. Number of links into a set of customers is integer. Let S be a customer set such that 2 ≤

143
4.4. DISJUNCTIVE VALID INEQUALITIES

|S| ≤ 3:

X X X X X
( ukij + uikj ) ≤ g
j∈J k∈V \S i∈S k∈V \S i∈S
∨ (4.50)
X X X X X
( ukij + uikj ) ≥ g + 1,
j∈J k∈V \S i∈S k∈V \S i∈S

for some integer g.

By solving the CGLP constructed using (M2 ) and one of the disjunctions violated by (t̄, w̄, ū),

we are searching for a valid inequality that is maximally violated by (t̄, w̄, ū) in the form (4.47).

An optimal solution to the CGLP with negative objective value provides such an inequality (α, β) ∈

R|J|+|J|×|A|+1 . We can rewrite the inequality in terms of variables of (SPP-LRSP).

X X XX X
αjt tj + + αjw wj + u
αikj likp zp ≥ β ∀(t, z) ∈ PSPP . (4.51)
j∈J j∈J j∈J p∈Pj (i,k)∈A

Note that we include w in (ASPP-LRSP) in order to allow implementation of one of our branching

rules. A valid inequality in the form of (4.51) is a group 1 inequality. In order to use this class

of inequalities in the branch-cut-and-price algorithm, in the pricing algorithm for facility j, we

u ω for all (i, k) ∈ A and the total cost of a pairing with a
modify the cost of arc (i, k) with αikj

fixed cost of αjw ω for all j ∈ J, where ω is the dual variable corresponding to (4.51).

144
4.4. DISJUNCTIVE VALID INEQUALITIES

Implementation 3. In this implementation, the CGLP is formulated based on the LP relaxation

of (G-LRSP). We refer to this LP relaxation as (M3 ). Recall that (M3 ) is as follows:

XX
(M3 ) xikh = 1 ∀i ∈ I,
h∈H k∈N
X X
xikh − xkih = 0 ∀i ∈ N, h ∈ H,
k∈N k∈N
X X
yjkh − LFj tj ≤ 0 ∀j ∈ J,
h∈Hj k∈I

yikh − LV xikh ≤ 0 ∀(i, k) ∈ A, h ∈ H,

X X X
yikh − ykih + Di xikh = 0 ∀i ∈ I, h ∈ H,
k∈N k∈N k∈N
X
Tik xikh − LT vh ≤ 0 ∀h ∈ H,
(i,k)∈A
0 ≤ tj ≤ 1 ∀j ∈ J,

0 ≤ xikh ≤ 1 ∀i, k ∈ V, h ∈ H,

0 ≤ vh ≤ 1 ∀h ∈ H,

yikh ≥ 0 ∀i, k ∈ V, h ∈ H.

P
|J|+ j∈J |P̄j |
Let (t̄, z̄) ∈ R+ be a fractional solution to the LP relaxation of Pi . Let P̂j = {p ∈
|J|+|J|×|H|+|H|×|A|+|H|×|A|
P̄j |z̄p > 0} for all j ∈ J. We can generate a vector (t̄, v̄, x̄, ȳ) ∈ R+

associated with solution (t̄, z̄) using a mapping of (t, z) to (t, v, x, y). Since we have identical

vehicles, we can assign pairings associated with nonzero variables from 1 to |P̂j | for all j ∈

J to vehicles in (M3 ) arbitrarily, starting from the first vehicle. Let (h ' p) represent such

correspondence where p is the hth smallest index in P̂j and pairing p is assigned to vehicle h

for all h = 1, .., |P̂j | and j ∈ J. Note that a vector (t̄, z̄) may lead to multiple corresponding

(t̄, v̄, x̄, ȳ) vectors under the identical vehicles assumption. By defining the correspondence (')

between h and p, we restrict the mapping of vectors (t̄, z̄) and (t̄, v̄, x̄, ȳ) to be one-to-one. To

145
4.4. DISJUNCTIVE VALID INEQUALITIES

construct (t̄, v̄, x̄, ȳ), we use the following transformations based on ('):


 z̄p If (h ' p), ∀p ∈ P̂j , h ∈ Hj , j ∈ J,
v̄h =

 0 otherwise.


 likp z̄p If (h ' p), ∀p ∈ P̂j , h ∈ Hj , j ∈ J,
x̄ikh =

 0 otherwise.


 fikp z̄p If (h ' p), ∀p ∈ P̂j , h ∈ Hj , j ∈ J,
ȳikh =

 0 otherwise,

where fikp is equal to the vehicle load on arc (i, k) for pairing p, for all (i, k) ∈ A, p ∈ Pj and

j ∈ J.

Using (M3 ) and one of the disjunctions (4.48) - (4.50) that is violated by (t̄, v̄, x̄, ȳ), we for-

mulate the CGLP. Disjunction (4.50) needs to be rewritten in terms of x variables:

X X X X X X X
( xkih + xkih ) ≤ g
j∈J k∈V \S i∈S h∈Hj k∈V \S i∈S h∈Hj
∨
X X X X X X X
( xkih + xkih ) ≥ g + 1,
j∈J k∈V \S i∈S h∈Hj k∈V \S i∈S h∈Hj

for some set of customers S. The CGLP based on (M3 ) enables a search for a disjunctive inequal-

ity valid ∀(t, v, x, y) ∈ PGBP in the form of

X X X X y
X X
αjt tj + αhv vh + αkih ykih + x
αkih xkih ≥ β. (4.52)
j∈J h∈H (k,i)∈A h∈H (k,i)∈A h∈H

Since we cannot handle valid inequalities with nonzero coefficients for flow variables y, we add

the following constraint to the CGLP to fix the coefficients associated with y to zero:

y
αkih = 0 ∀(i, k) ∈ A, h ∈ H. (4.53)

146
4.4. DISJUNCTIVE VALID INEQUALITIES

After a valid inequality in the form of (4.52) and satisfying (4.53) is generated for a given

fractional solution, (t̄, v̄, x̄, ȳ), we can rewrite more valid inequalities by circulating coefficients
x and αv over all h ∈ H for all j ∈ J based on the assumption of identical vehicles. In
αikh h j

other words, we assume that we obtain multiple (t̄, v̄, x̄, ȳ) vectors from one (t̄, z̄) vector (one-to-

many correspondence between solution vectors). Then we can combine the coefficients αx and

αv indexed by h ∈ Hj for all j ∈ J and rewrite the inequality (4.52) valid ∀(t, v, x, y) ∈ PGBP

in the form of

X X X X X X
αjt tj + αjvc vh + xc
αkij xkih ≥ β c (4.54)
j∈J j∈J h∈Hj j∈J (k,i)∈A h∈Hj
X
such that αjvc = αhv ,
h∈Hj
X
xc x
αikj = αikh ,
h∈Hj
β c = |I|β,

where |Hj | = |I| for all j ∈ J. Finally, we can rewrite this inequality in the (ASPP-LRSP)

variable space:

X X X X X
αjt tj + αjvc wj + xc
αikj likp zp ≥ β c ∀(t, z) ∈ PSPP . (4.55)
j∈J j∈J j∈J (i,k)∈A p∈Pj

Inequalities in the form of (4.55) are in group 1 and can be easily used in the branch-cut-and-price

algorithm by modifying the arc costs in the pricing algorithm.

In summary, we solve the CGLP to find a valid inequality in the form of (4.52) that is violated

by (t̄, v̄, x̄, ȳ). Then we modify the coefficients of (4.52) to get a valid inequality in the form of

(4.54). Note that since the CGLP is solved for (4.52) and not for (4.54), the fractional solution

(t̄, z̄) may not violate (4.55). In order to overcome this problem, we designed implementation 4.

147
4.4. DISJUNCTIVE VALID INEQUALITIES

Implementation 4. This implementation is designed to improve implementation 3. As in imple-

mentation 3, we use (M3 ) to derive the disjunctive valid inequalities. The steps of this implemen-

tation are the same as those of implementation 3. However, we directly search for an inequality

that is maximally violated by (t̄, v̄, x̄, ȳ) and in the form of (4.54). Recall that implementation 3

searches for an inequality in the form of (4.52). In order to search for an inequality in the form

of (4.54), we add constraints (4.53) to the CGLP. In addition, to be able to obtain the combined

(aggregated for all vehicles located at the same facility) coefficients for x and v, we set the α

coefficients for variables xikh and vh to be equal for all h ∈ Hj for j ∈ J.

αhv = αjvc , ∀h ∈ Hj , ∀j ∈ J,
x xc
αikh = αikj , ∀h ∈ Hj , ∀j ∈ J.

In Section 5.5.2, we provide computational results to evaluate the performance of each of these

implementations and discuss some issues regarding.their effectiveness.

148
Chapter 5

Computational Experiments

In this chapter, we present computational experiments to empirically validate the procedures dis-

cussed in the previous three chapters. To begin, we first describe the details of the instances used

in our experiments. We then present computational results related to Chapter 2 that compare the

LP relaxation bounds for the two main formulations for the LRSP and assess the effect of the sim-

ple valid inequalities that are presented there. Next, we describe a set of experiments to evaluate a

number of variants of the branch-and-price algorithms for the LRSP described in Chapter 3. Later,

we discuss the effect of adding the procedures for dynamic cut generation described in Chapter

4. Finally, we present computational results that assess the cost benefits of solving the integrated

LRSP as opposed to sequentially optimizing the location, routing and scheduling decisions.

In some of the experiments, we use performance profiles (introduced by Dolan and More

(2002)) to aid visualization and evaluation of the performance of the algorithmic variants used in

the experiments. Performance profile graphs consist of the cumulative distribution function for

the ratio of a comparison criteria for an algorithm to the best corresponding value obtained in any

of the algorithms in the pool. In our experiments, we generally use time and percentage gap as

comparison criteria.

149
5.1. DESCRIPTION OF THE INSTANCES

5.1 Description of the Instances

The only LRSP instances available in the literature are those in Lin et al. (2002) of which we were

only able to obtain four small instances. We describe those instances and associated results in

Section 5.4.6. Because they are very limited in size, we generated several sets of new instances

for testing purposes. We describe these next.

Location and Demand Data. We generated three sets of instances with 20, 25 and 40 customers

based on the MDVRP instances of Cordeau et al. (1997). All instances include five candidate

facility locations. Each of our instances contains a subset of the customer locations and demand

data from one of the Cordeau et al. (1997) instances. In particular, we selected instances p01 (50

customers), p03 (75 customers), and p07 (100 customers) from which to generate our three sets of

instances.

Table 5.1 presents the sets of customers used to construct the LRSP instances in each group.

The first column specifies the marker letter used to label the instances in our set (the details of

how we label our instances is explained later in this section), the identity of the corresponding

Cordeau et al. instance (denoted as Ref) and the indices of customers from the reference instance

included in our instance. In addition to the customer indices, Table 5.2 lists the coordinates of the

five facility locations in each instance group differentiated with each reference instance.

Travel Times. Rounded Euclidean distances between each pair of nodes were used in the al-

gorithm as a proxy for travel times. Based on the assumption that the distances are in miles and

vehicles travel 20 miles per hour, we can convert distances into travel times in order to justify the

values of cost and time limit parameters with respect to real life data.

Capacity and Cost Parameters. Based on the demand information, we assumed that the capac-

ity of all facilities was 250, which means that at least two facilities are required to be open in any

feasible solution for each instance.

150
5.1. DESCRIPTION OF THE INSTANCES

Table 5.1: Location & Demand Data for LRSP instances generated from Cordeau et al. (1997)

Marker Indices of the Cust. Coord. from Ref.

Letter Ref 20-cust. 25-cust. 40-cust.
f p01 1 - 20 1 - 25 1-40
l p01 31 - 50 26 - 50 11-50
f p03 1 - 20 1 - 25 1-40
m p03 28 - 47 26 - 50 16-35, 41-60
l p03 56 - 75 51 - 75 36-75
f p07 1 - 20 1 - 25 1-40
s p07 21 - 40 26 - 50 31-70
t p07 31 - 50 51 - 75 1-20, 51-70

Table 5.2: Facility Locations for LRSP instances

Instance Group Ref Facility Coordinates (X,Y)

20- and 25-customer p01 (20,20), (30,40), (15,40), (45,40) (55,50)
p03 (40,40), (50,22),(25,45), (20,20), (55,55)
p07 (15,35), (55,35),(35,20), (35,50), (25,45)
40-customer p01 (20,20), (30,40), (15,40), (50, 30), (55,55)
p03 (40,40), (50,22), (25,45), (20,20), (55,55)
p07 (15,35), (55,35),(35,20), (35,50), (25,45)

For each instance, we considered two time limit values and two vehicle capacity values. Time

limits of 7 and 8 hours (which can be converted to distance limits of 140 and 160 miles) were

used. In all instances, 140 is more than twice the distance between each customer and its farthest

candidate facility to ensure that any customer can be assigned to any facility, and the feasible

solution set is not restricted because of an infeasible customer-facility pair. We used “t1” or “t2”

to mark the instances based on the distance limit; “t1” was used for instances with distance limit

140 and “t2” was used for instances with distance limit 160 (the role of these markers in the

labeling strategy used for our instances is explained later).

We chose two vehicle capacity values. To determine these values, we used the fact that at least

two facilities are required to be open (due to the capacity) and used the following equation similar

151
5.1. DESCRIPTION OF THE INSTANCES

to that used by Laporte et al. (1986):

P
i∈I Di
B = (1 − α)maxi∈I {Di } + α , (5.1)
2

with α equal to 0.1 and 0.2. Small α values were considered since a schedule employing multiple

trips for a single vehicle is only realistic if the vehicle capacity is small compared to the demands

of the customers. We used the 25-customer instances to determine vehicle capacities and used the

same values in the 20- and 40-customer instances. For each α value, we took the average of the

calculated values of B for the instances generated from the same Cordeau et al. (1997) instance

(for example, all instances generated using p01), divided it by 10, rounded it up, and multiplied

by 10. The purpose of averaging all instances that were generated from the same reference is to

generalize the value of the parameter that is calculated, because the value of the equation (5.1)

depends on the demand data and the demand data in instances that are generated using the same

reference is from the same distribution.

We obtained vehicle capacity values 50 (for α = 0.1) and 70 (for α = 0.2) for the instances

generated from p01 and p07, and 60 (for α = 0.1) and 80 (for α = 0.2) for the instances generated

from p03. Vehicles with the smaller capacity values can cover at least one customer plus the

customer that has the largest demand in the instance. We used “v1” or “v2” in the labels of the

instances to differentiate the instances based on vehicle capacity.

Based on the cost values reported by Burlett (2002) and Kenworth Truck Company (2003),

fixed costs for the facilities were generated using a uniform distribution on the interval [1500,2300],

the fixed cost and the operating cost of a mid-size vehicle were estimated to be $225 per day and

$1 per mile traveled, respectively. The same cost parameters were used in all instances.

Table 5.3 summarizes the capacity and cost parameters. In the table, as in Chapter 2, LV and

LT represent the vehicle capacity and distance limit, respectively. CjF denotes the fixed cost of

facility j for all j ∈ J, while C V and C O denote the vehicle fixed and operating costs, respectively.

152
5.2. COMPARING THE LP RELAXATIONS OF ALTERNATIVE FORMULATIONS

Table 5.3: Summary of Capacity and Cost Parameters

Generated Using LV1 LV2 LT1 LT2 Fac. cap.
p01,p07 50 70 140 160 250
p03 60 80 140 160 250

C1F C2F C3F C4F C5F CV CO

$1615 /day $1845 /day $1781 /day $1975 /day $1753 /day $225 /day $ 1 /mile

Labels for the LRSP Instances. We used a labeling strategy to demonstrate the characteris-

tics of each instance that was used in our computational experiments. We labeled our instances

considering the reference Cordeau et al. (1997) instances, the set of customers included from the

corresponding reference (the marker letters listed in Table 5.1 used to indicate this information),

the number of customers, vehicle capacity (one of the markers v1 or v2) and distance limit values

(one of t1 or t2). For example, we used label “p01-f25-v1t1” for the LRSP instance (in the 25-

customer instance group) generated using customers 1 to 25 (based on letter “f”) in p01. The “v1”

and “t1” markers indicate that the smaller vehicle capacity and distance limit values (50 and 140,

respectively, based on Table 5.3) were used in the instance. Another example label is “p03-m40-

v1t2”. This label denotes that the instance is one of the 40-customer instances and is generated

from reference p03 using customers 16 to 35, 41 to 60. Vehicle capacity and distance limits are

60 and 160, respectively. Note that the only difference between instance “p03-m40-v1t1” and

instance “p03-m40-v1t2” is the time limit value used.

5.2 Comparing the LP Relaxations of Alternative Formulations

The goal of the computational experiments in this section is to compare the LP relaxation bounds

of (G-LRSP) and (SPP-LRSP). In Section 2.4, it is shown that the application of Dantzig-Wolfe

Decomposition to the graph-based model (G-LRSP) results in the set partitioning-based model

(SPP-LRSP), and using this result, in Section 2.4.2, we state that the LP relaxation bound of

(SPP-LRSP) is at least as strong as the LP relaxation bound of (G-LRSP). In this section, we

153
5.2. COMPARING THE LP RELAXATIONS OF ALTERNATIVE FORMULATIONS

validate this result and empirically demonstrate the strength of the LP relaxation of (SPP-LRSP)

as compared to (G-LRSP).

We used the instances with 25 customers and 40 customers in this experiment. The LP relax-

ation of (G-LRSP) was solved using CPLEX 9.1. In (G-LRSP), the upper bound for the number of

vehicles at each facility is the number of customers. However, this increases the number of vari-

ables and the LP solution times significantly. Therefore, we set a smaller upper bound value for

the number of vehicles at each facility. For the 25-customer instances, we used the upper bound

value 10 and for the 40-customer instances, we used the upper bound value 14. These maximum

number of vehicle values are at least 40% larger than a simple estimate that can be calculated con-

sidering the facts that at least two facilities have to be open and vehicle capacity is large enough

to cover at least two customers. These facts are not enough to decide a valid upper bound for the

maximum number of vehicles at each facility. However, in the heuristic solutions to the LRSP

instances obtained using I-Heur (see Section 3.4.1), the number of vehicles used at each open

facility is much smaller than the upper bounds that we set. Therefore, we are confident that the

number of vehicles used in the optimal solution does not exceed these bounds. The LP relaxation

of (SPP-LRSP) (constraints (2.13) through (2.17)) was obtained by using the column generation

algorithm E-COLGEN(0,0,50,350) described in Section 3.4.2. All instances in both experiment

groups were run with a time limit of 8 CPU hours.

The results for the 25- and 40-customer instances are provided in Tables 5.4 and 5.5. The

tables report the LP relaxation values and the solution times for (G-LRSP) and (SPP-LRSP) for

each instance. Based on Tables 5.4 and 5.5, the LP relaxations obtained from (SPP-LRSP) are 5%

to 10% stronger than the LP relaxations obtained from (G-LRSP). In addition, the computation

times for (SPP-LRSP) are much better than the times for (G-LRSP) in all instances except one.

For both the 25- and the 40-customer instances, solution times of (SPP-LRSP) are an order of

magnitude better than solution times of (G-LRSP).

154
5.2. COMPARING THE LP RELAXATIONS OF ALTERNATIVE FORMULATIONS

Table 5.4: LP Relaxation Bounds of (G-LRSP) and (SPP-LRSP): 25-Customer Instances

(SPP-LRSP) G-LRSP (a-b)/b*100

Data File LP (a) CPU (s)(c) LP (b) CPU (s)(d) % Gap c/d
p01-f25-v1t1 3959.89 8.28 3729.764 280.54 6.17 0.03
p01-f25-v1t2 3891.71 23.15 3664.293 272.44 6.21 0.08
p01-f25-v2t1 3836.89 20.88 3636.819 291.65 5.5 0.07
p01-f25-v2t2 3762.65 54.99 3575.846 299.38 5.22 0.18
p01-l25-v1t1 3749.46 3.52 3402.078 301.2 10.21 0.01
p01-l25-v1t2 3651.02 10.53 3329.01 316.47 9.67 0.03
p01-l25-v2t1 3593.77 11.08 3279.46 488.59 9.58 0.02
p01-l25-v2t2 3523.2 24.16 3214.12 426.4 9.62 0.06
p03-f25-v1t1 4633.06 4.75 4291.424 290.04 7.96 0.02
p03-f25-v1t2 4556.52 19.88 4217.155 347.01 8.05 0.06
p03-f25-v2t1 4465.42 8.95 4204.282 390.64 6.21 0.02
p03-f25-v2t2 4377.68 23.06 4136.160 513.38 5.84 0.04
p03-m25-v1t1 4481.74 11.43 4153.278 481.04 7.91 0.02
p03-m25-v1t2 4398.71 29.03 4085.718 285.73 7.66 0.1
p03-m25-v2t1 4334.34 27.31 4055.503 521.65 6.88 0.05
p03-m25-v2t2 4246.11 79.62 3994.734 466.8 6.29 0.17
p03-l25-v1t1 3951.49 4.36 3619.805 387.24 9.16 0.01
p03-l25-v1t2 3839.15 9.64 3545.993 343.01 8.27 0.03
p03-l25-v2t1 3826.99 10.32 3535.861 370.51 8.23 0.03
p03-l25-v2t2 3731.76 24.38 3466.029 348.45 7.67 0.07
p07-f25-v1t1 3418.42 7.86 3158.331 402.82 8.23 0.02
p07-f25-v1t2 3360.13 15.53 3092.431 285.36 8.66 0.05
p07-f25-v2t1 3317 22.38 3076.413 358.08 7.82 0.06
p07-f25-v2t2 3214.84 59.78 3014.934 260.57 6.63 0.23
p07-s25-v1t1 4136.93 1.98 3772.090 340.85 9.67 0.01
p07-s25-v1t2 4016.72 4.03 3688.345 471.77 8.9 0.01
p07-s25-v2t1 3966.49 4.58 3608.132 556.86 9.93 0.01
p07-s25-v2t2 3817.5 11.6 3534.823 561.57 8 0.02
p07-t25-v1t1 3754.74 5.97 3425.342 659.82 9.62 0.01
p07-t25-v1t2 3682.45 19.61 3359.481 621.15 9.61 0.03
p07-t25-v2t1 3631.42 12.31 3300.849 625.36 10.01 0.02
p07-t25-v2t2 3552.1 34.09 3240.682 374.81 9.61 0.09
Average 19.03 404.41 8.09 0.05

155
5.2. COMPARING THE LP RELAXATIONS OF ALTERNATIVE FORMULATIONS

Table 5.5: LP Relaxation Bounds of (G-LRSP) and (SPP-LRSP): 40-Customer Instances

SPP-LRSP (G-LRSP) (a-b)/b*100

Data File LP (a) CPU (s) (c) LP (b) CPU (s) (d) % Gap c/d
p01-f40-v1t1 6019.5 139.06 5661.25 10701.91 6.33 0.01
p01-f40-v1t2 5940.94 613.25 5565.86 11683.6 6.74 0.05
p01-f40-v2t1 5851.26 475.53 5495.23 8037.53 6.48 0.06
p01-f40-v2t2 5750.61 1189.53 5406.40 13603.01 6.37 0.09
p01-l40-v1t1 5968.39 82.26 5575.07 11841.91 7.06 0.01
p01-l40-v1t2 5869.81 270.39 5479.82 11609.7 7.12 0.02
p01-l40-v2t1 5745.9 246.4 5423.90 13155.77 5.94 0.02
p01-l40-v2t2 5625.51 1297.07 5337.46 9286.94 5.4 0.14
p03-f40-v1t1 7255.42 273.61 6846.76 6190.5 5.97 0.04
p03-f40-v1t2 7137.05 1989.77 6748.45 6203.96 5.76 0.32
p03-f40-v2t1 7019.46 1121.65 6683.24 10388.86 5.03 0.11
p03-f40-v2t2 6945.08 5709.81 6595.75 14661.17 5.3 0.39
p03-m40-v1t1 6767.07 76.77 6380.18 10594.08 6.06 0.01
p03-m40-v1t2 6692.17 347.77 6281.08 10420.49 6.54 0.03
p03-m40-v2t1 6614.35 198.52 6222.73 6173.39 6.29 0.03
p03-m40-v2t2 6506.1 1177.49 6133.50 8741.29 6.07 0.13
p03-l40-v1t1 6208.81 154.05 5844.60 7394.12 6.23 0.02
p03-l40-v1t2 6127.31 350.96 5746.85 10487.14 6.62 0.03
p03-l40-v2t1 6034 324.94 5722.65 12068.38 5.44 0.03
p03-l40-v2t2 5942.4 2497.59 5633.57 10434.82 5.48 0.24
p07-f40-v1t1 5710.64 59.22 5308.25 5869.88 7.58 0.01
p07-f40-v1t2 5555.82 254.36 5203.41 5540.26 6.77 0.05
p07-f40-v2t1 5432.94 319.35 5123.49 8257.54 6.04 0.04
p07-f40-v2t2 5323.22 1303.25 5027.03 7140.7 5.89 0.18
p07-s40-v1t1 6113.67 32.63 5643.66 6115.72 8.33 0.01
p07-s40-v1t2 5949.65 75.07 5527.65 5337.76 7.63 0.01
p07-s40-v2t1 5801.4 88.14 5392.51 9352.83 7.58 0.01
p07-s40-v2t2 5661.45 380.95 5292.29 6786.14 6.98 0.06
p07-t40-v1t1 5482.41 205.45 5089.73 6323.65 7.72 0.03
p07-t40-v1t2 5369.25 639.62 4993.55 8390.05 7.52 0.08
p07-t40-v2t1 5223.99 2028.13 4916.37 7456.19 6.26 0.27
p07-t40-v2t2 5118.89 10607.65 4828.02 7506.42 6.02 1.41
Average 1079.07 8992.37 6.45 0.12

156
5.3. ASSESSING THE STRENGTH OF SIMPLE VALID INEQUALITIES

5.3 Assessing the Strength of Simple Valid Inequalities

In this section, we assess the strength of the valid inequalities described in Section 2.3. Consider

the base (SPP-LRSP) formulation containing only constraints (2.14) - (2.17). We construct (SPP-

LRSP1 ) by adding constraints (2.24), (SPP-LRSP2 ) by adding constraints (2.25), (SPP-LRSP3 )

by adding constraint (2.18), and (ASPP-LRSP) by adding both constraints (2.18) and (2.25) to

(SPP-LRSP).

In this experiment, we solved the LP relaxations of formulations (SPP-LRSP), (SPP-LRSP1 ),

(SPP-LRSP2 ), (SPP-LRSP3 ) and (ASPP-LRSP) for 20-, 25- and 40-customer instances. We cal-

culated the percentage LP relaxation gap associated with each formulations. We compared the

LP relaxation gap of (SPP-LRSP) with the LP relaxation gaps of (SPP-LRSP1 ), (SPP-LRSP2 )

and (SPP-LRSP3 ) to evaluate the strength of constraints (2.24), (2.25) and (2.18). Comparison

between the results of (SPP-LRSP1 ) and (SPP-LRSP2 ) validates why we prefer constraints (2.25)

to (2.24), and comparison between the results of (SPP-LRSP3 ) and (ASPP-LRSP) evaluates the

impact of adding constraint (2.18) alone and in combination with constraints (2.25). For the 20-

customer instances, we enumerated all of the feasible pairings and solved the instances using a

standard branch-and-bound algorithm without column generation. The reason for avoiding col-

umn generation and employing enumeration is that we could not incorporate constraints (2.24)

into the column generation algorithm. For the 25- and 40-customer instances, we solved the LP

relaxations of all formulations except (SPP-LRSP1 ) using a column generation algorithm. Due to

the large number of feasible pairings, we could not use enumeration for these instances, thus, we

could not evaluate the effect of constraints (2.24) for the 25- and 40-customer instances.

Figure 5.1 presents performance profiles for the LP relaxation gap for each of the formula-

tions (SPP-LRSP), (SPP-LRSP1 ), (SPP-LRSP2 ), (SPP-LRSP3 ) and (ASPP-LRSP). Performance

profiles indicate that formulations (SPP-LRSP3 ) and (ASPP-LRSP) perform significantly better

than the others with respect to the LP relaxation gap. For all instances, (ASPP-LRSP) is the best

157
5.3. ASSESSING THE STRENGTH OF SIMPLE VALID INEQUALITIES

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
spp-lrsp
spp-lrsp2 spp-lrsp
spp-lrsp3 spp-lrsp2
aspp-lrsp spp-lrsp3
spp-lrsp1 aspp-lrsp
0 0
0 10 20 30 40 50 0 5 10 15 20 25 30

(a) 20-customer Instances (b) 25-customer Instances

0.8

0.6

0.4

0.2

spp-lrsp
spp-lrsp2
spp-lrsp3
aspp-lrsp
0
0 20 40 60 80 100 120

(c) 40-customer Instances

Figure 5.1: Performance profiles for LP relaxation gap of (SPP-LRSP), (SPP-LRSP1 ), (SPP-
LRSP2 ), (SPP-LRSP3 ) and (ASPP-LRSP)

formulation. The graphs show that in all instances, we can improve the LP relaxation gap with for-

mulation (SPP-LRSP2 ) compared to (SPP-LRSP), and Figure 5.1(a) shows that constraints (2.24)

do not improve the performance of (SPP-LRSP).

In addition to the performance profiles, in Tables A.1, A.2 and A.3 in Appendix A, for each

instance, we report the percentage LP relaxation gap associated with each of the formulations.

For the 20-customer instances (in Tables A.1), we also report the total number of feasible pairings

and the integer optimal objective value. For one instance, the set of feasible pairings could not be

158
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

generated, so we did not use it in the experiments. In Tables A.2 and A.3, we omited the column

reporting optimal or best known objective value for each instance because we report these values

later in this chapter.

Table A.1 indicates that adding inequalities (2.24) in (SPP-LRSP1 ) made the problem difficult

to solve and yielded at most only a small improvement in the LP relaxation bound for each solved

instance. Based on Tables A.1, A.2 and A.3, adding inequalities (2.25) in (SPP-LRSP2 ) improved

the LP relaxation bound by a moderate amount. In all instances, adding inequality (2.18) in (SPP-

LRSP3 ) led to a dramatic reduction in the gap as compared to adding (2.25) alone. Together, the

two inequalities (2.18) and (2.25) in (ASPP-LRSP) further reduced the gap for all but 5 instances.

The average gaps for (SPP-LRSP3 ) are 3.54, 4.47 and 2.96 in the 20-, 25- and 40-customer in-

stances, respectively. These corresponding averages for (ASPP-LRSP) are 2.28, 2.68 and 2.04.

These results support the theoretical results in Section 2.3 and suggest that adding both in-

equalities (2.18) and (2.25) to the (SPP-LRSP) yields a stronger formulation which we refer to as

(ASPP-LRSP) and used to obtain the results in the next section.

5.4 Computational Experiments With Branch and Price

The goal of the computational experiments in this section is to evaluate the effect of varying the

individual components of the branch-and-price algorithm described in Chapter 3 and to explain

the rationale for extensions of the basic branch-and-price algorithm used to improve the overall

performance. In particular, we evaluate the different pricing algorithms used in the column gener-

ation process; the effect of modifying the column generation by using heuristic pricing algorithms;

the strength of the primal heuristics to update the upper bound; and the effect of the method used

to generate the initial set of columns. Finally, the overall performance of the branch-and-price

algorithm is demonstrated for the 25- and the 40-customer instances, and some instances from

literature.

159
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

5.4.1 Performance of Exact Pricing Algorithms

In this section, we compare the performance of the exact pricing algorithms. The algorithms

tested are the one-phase pricing algorithm (referred to as 1p-ESPPRC), the one-phase pricing

algorithm with new domination rule (presented in Section 3.2.2), and the two-phase pricing algo-

rithm (referred to as 2p-ESPPRC). We use notation 1p-ESPPR-NwD to denote pricing algorithm

with new domination rule. In Section 3.2.2, we present the details of these pricing algorithms

and we demonstrate empirically here that the new domination rule improves the performance of

1p-ESPPRC, and that 2p-ESPPRC performs better than 1p-ESPPR-NwD.

To compare the methods, we solved the LP relaxation of (ASPP-LRSP) at the root node using

the column generation algorithm (COLGEN in Figure 1.4) with each one of the pricing algo-

rithms: 1p-ESPPRC, 1p-ESPPRC-NwD and 2p-ESPPRC. In the experiments, at each iteration

of COLGEN, for each facility, at most one column (the one with the most negative reduced cost)

was added to the current restricted LP relaxation (RMP) of the problem. We ran the experiments

with a time limit of 3 CPU hours and using the 25- and the 40-customer instances.

Figure 5.2 presents the performance profiles for the root node solution time using each of the

pricing algorithms. The results clearly show that 2p-ESPPRC significantly outperforms both 1p-

ESPPRC and 1p-ESPPRC-NwD. The effect of employing 2p-ESPPRC is even more significant

for the 40-customer instances. Within the time limit of 3 CPU hours, COLGEN with 1p-ESPPRC

and 1p-ESPPRC-NwD could solve the root node in fewer than 25% of the 40-customer instances,

while 2p-ESPPRC could solve about 82% of the instances.

Tables A.4 and A.5 in Appendix A present details of the results from the experiments. For

each 25- and 40-customer instance, and for each implementation of COLGEN, tables report the

total CPU time (in seconds) to solve the LP relaxation of the problem (the root node solution time)

and the average time for each COLGEN iteration (total COLGEN time / # of column generation

iterations).

160
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

1pessprc 1pessprc-40
1pessprc-nwd 1pessprc-nwd-40
2pessprc 2pessprc-40
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25

(a) 25-customer Instances (b) 40-customer Instances

0.8

0.6

0.4

0.2

1pessprc-all
1pessprc-nwd-all
2pessprc-all
0
0 5 10 15 20 25 30

(c) 25- & 40-customer Instances

Figure 5.2: Performance profile for the root node solution time using 1p-ESPPRC, 1p-ESPPRC-
NwD and 2p-ESPPRC

The results in Tables A.4 and A.5 show that for the instance set with 25 customers, the to-

tal COLGEN time was improved 36% on average with 1p-ESPPRC-NwD and 89% with 2p-

ESPPRC, as compared to 1p-ESPPRC. The corresponding average values for the 40-customer

instances solved within the time limit are 33% and 93%. In the 25-customer instances, the av-

erage time per iteration improved 43% with 1p-ESPPRC-NwD and 94% with 2p-ESPPRC. For

40-customer instances, the corresponding improvements were 36% and 70%, respectively.

In addition, we tested the effect of adding multiple columns at each iteration of COLGEN. At

161
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

each iteration of COLGEN and for each facility, we evaluated the addition of at most 1, 20, 40 and

100 of the columns with the most negative reduced costs. We use 2p-ESPPRC, 2p-ESPPRC20 ,

2p-ESPPRC40 and 2p-ESPPRC100 to denote the exact pricing algorithms in which we add at

most 1, 20, 40 and 100 of the columns per iteration, respectively. For these experiments, we used

COLGEN with 2p-ESPPRC, 2p-ESPPRC20 , 2p-ESPPRC40 and 2p-ESPPRC100 .

Figure 5.3 compares the root node solution times obtained with the variants of COLGEN em-

ploying 2p-ESPPRC, 2p-ESPPRC20 , 2p-ESPPRC40 and 2p-ESPPRC100 . The performance

profile indicates a performance improvement with the addition of multiple columns. The perfor-

mance of each of 2p-ESPPRC20 , 2p-ESPPRC40 and 2p-ESPPRC100 are similar, however, though

2p-ESPPRC40 is better by a slim margin.

Tables A.6 and A.7 in Appendix A present an instance-by-instance comparison of COLGEN

with 2p-ESPPRC and 2p-ESPPRC40 . For each of these methods and for each instance, Tables A.6

and A.7 present the total CPU time (in seconds) to solve the root node, the number of COLGEN

iterations, and the average time for each COLGEN iteration. Adding multiple columns at each

iteration instead of one column reduced the number of COLGEN iterations in every instance. On

average, for the 25-customer instances, the total time for 2p-ESPPRC40 was 47% better than the

time for 2p-ESPPRC. For the 40-customer instances, the time improvement was consistent, it was

50%, on average.

In summary, for the 25- and 40-customer instances, COLGEN with 2p-ESPPRC performed

significantly better than the variants with 1p-ESPPRC and 1p-ESPPRC-NwD. In addition, the

benefit of adding multiple columns instead of a single best column at each iteration of COLGEN

was empirically demonstrated. In the rest of the experiments, we used 2p-ESPPRC40 as exact

pricing algorithm.

5.4.2 Effect of Using Heuristic Pricing Algorithms

In this section, we empirically demonstrate the performance changes for our branch-and-price

algorithm with the addition of the heuristic pricing algorithms to the overall column generation

162
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

0.8

0.6

0.4

0.2

2pessprc
2pessprc20
2pessprc40
2pessprc100
0
0 2 4 6 8 10 12

Figure 5.3: Performance profile for the root node solution time using 2p-ESPPRC and 2p-
ESPPRC with multiple columns (20, 40 &100 columns per iteration)

scheme. In Section 3.2.3, we describe two heuristic versions of the pricing algorithm: ESPPRC

with label limit (referred to as ESPPRC-LL) and ESPPRC for a subset of customers (referred to

as ESPPRC-CS). In the preceding section, we observed that adding at most 40 columns at each

iteration of the exact pricing algorithm resulted in the best performance. Therefore, we chose to

add at most 40 columns at each iteration of the heuristic pricing algorithms. For ease of notation,

we drop subindex 40 from the notation of the algorithms. The aim of this section is to evaluate the

effect on the overall branch-and-price algorithm of employing 2p-ESPPRC-LL and 2p-ESPPRC-

CS with 2p-ESPPRC in the column generation procedure.

Experiments Using the One-Stage Branch-and-Price Algorithm. In the computational ex-

periments, we applied one-stage branch-and-price algorithms (1S-EBP) using the E-COLGEN

procedure outlined in Figure 3.4 for a given set of input parameters. These parameters deter-

mine whether E-COLGEN calls 2p-ESPPRC-LL and 2p-ESPPRC-CS before 2p-ESPPRC and

also includes the input parameters required for 2p-ESPPRC-LL and 2p-ESPPRC-CS themselves.

Recall that E-COLGEN takes four parameters:

• u = the subset size in 2p-ESPPRC-CS,

• U = the maximum number of iterations for 2p-ESPPRC-CS,

163
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

• l = the initial label limit for 2p-ESPPRC-LL, and

• L = the label limit upper bound for 2p-ESPPRC-LL.

We refer to the algorithm with a given set of parameters as E-COLGEN(u, U , l, L). We compared

four variants of 1S-EBP:

1. In the first experiment, we used 1S-EBP with E-COLGEN(0, 0, 0, 0), i.e., employing only

exact pricing 2p-ESPPRC.

2. In the second experiment, we used 1S-EBP with E-COLGEN(0, 0, l, L), for some positive

l and L, i.e., at each node, we only employed 2p-ESPPRC-LL before 2p-ESPPRC.

3. In the third experiment, we used 1S-EBP with E-COLGEN(u,U ,0,0), for some positive u

and U , i.e., we only employed 2p-ESPPRC-CS before 2p-ESPPRC.

4. Finally, we used both 2p-ESPPRC-CS and 2p-ESPPRC-LL before applying 2p-ESPPRC

at each node of 1S-EBP.

In each experiment, we set a time limit of 8 CPU hours for each instance (with 25 and 40

customers). Based on preliminary computational experiments, we decided to set l= 50 and L =

350 in all experiments in this section whenever 2p-ESPPRC-LL was employed. The choice of

l = 50 and L = 350 results in iterations of E-COLGEN in which 2p-ESPPRC-LL with label limits

of 50, 100 and 200 is run. For small label limit values, 2p-ESPPRC-LL terminates quickly, but the

generation of a column with the smallest reduced cost is less probable than 2p-ESPPRC-LL with

larger label limit that can search a larger set of columns. In general, the number of columns with

negative reduced cost is larger in early iterations of the column generation algorithm. Therefore,

the number of labels at each node in the pricing algorithm tends to increase rapidly in the early

iterations. Increasing the label limit gradually, we aim to pass the earlier iterations of E-COLGEN

quickly, and explore a larger search space in later iterations to be able to get a column with smallest

cost.

164
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

After some preliminary experiments, we decided to set U = 12 in all experiments, and u = 12

for the 25-customer instances and u = 15 for the 40-customer instances whenever 2p-ESPPRC-

CS was employed. For large values of u, the heuristic becomes slow. However, if u is not large

enough, the quality of the generated columns may not be sufficient (i.e., corresponding to pairings

including less number of customers than they can actually include). Note that, the value of u

differs based on the number of customers in the instance. We chose these values of u by looking

at the solutions we found for the 25- and the 40-customer instances. We observed that a vehicle

serves at most 11 customers in optimal solutions to the 25-customer instances, while a vehicle

serves at most 14 customers in optimal or best solutions to the 40-customer instances. In addition,

our experiments showed that U values greater than 20 reduced the overall performance of the

algorithm.

Figure 5.4 presents the performance profiles for the total time spent in each variant of 1S-EBP:

1S-EBP that uses only the exact pricing algorithm (referred to as 2pespprc in the figure), 1S-EBP

that also calls 2p-ESPPRC-LL (referred to as 2pespprcLL), 1S-EBP that calls 2p-ESPPRC-CS

before 2p-ESPPRC (referred to as 2pespprcCS), and 1S-EBP that calls both 2p-ESPPRC-LL and

2p-ESPPRC-CS (referred to as 2pespprcLLCS).

Based on Figures 5.4(a) and 5.4(b), for the 25-customer instances, the variant that employs

2p-ESPPRC-LL outperforms all others. The variant that uses only 2p-ESPPRC seems to be better

than the ones that call only 2p-ESPPRC-CS or both 2p-ESPPRC-CS and 2p-ESPPRC-LL. We

conclude that the use of 2p-ESPPRC-CS does not improve the total solution time of the algorithm

for the 25-customer instances. Tables A.8 and A.9 in Appendix A present some details of the

experiments for the 25-customer instances. In Tables A.8 and A.9, for each instance and for each

variant of 1S-EBP, we report the total time, # of evaluated nodes, the average COLGEN time

(total COLGEN time / # of evaluated nodes). In addition, in both tables, we calculate the fraction

of the total time in each variant to the total time of the one with only 2p-ESPPRC. For the 25-

customer instances, Table A.8 shows that on average, applying 2p-ESPPRC-LL improves the total

165
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

2pespprc 2pespprc
2pespprcLL 2pespprcLL
2pespprcCS 2pespprcCS
2pespprcLLCS 2pespprcLLCS
0 0
1 1.5 2 2.5 3 3.5 4 1 2 4

(a) 25-customer Instances (b) 25-customer Instances, log2 scale

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

2pespprc-40 2pespprc-40
2pespprcLL-40 2pespprcLL-40
2pespprcCS-40 2pespprcCS-40
2pespprcLLCS-40 2pespprcLLCS-40
0 0
1 1.5 2 2.5 3 1 2

(c) 40-customer Instances (d) 40-customer Instances, log2 scale

Figure 5.4: Performance profile for the total time of 1S-EBP employing only 2p-ESPPRC or
performing one or two of 2p-ESPPRC-LL and 2p-ESPPRC-CS before 2p-ESPPRC

execution time 5%, while employing 2p-ESPPRC-CS or both 2p-ESPPRC-CS and 2p-ESPPRC-

LL increases solution time 18%, on average.

For the 40-customer instances, Figures 5.4(c) and 5.4(d) show that the variants of 1S-EBP

that employ 2p-ESPPRC-LL outperform those without 2p-ESPPRC-LL. The variant that calls

both 2p-ESPPRC-LL and 2p-ESPPRC-CS is better than the variant that calls only 2p-ESPPRC-

LL in 55% of the instances. However, the variant that calls only 2p-ESPPRC-LL solves more

instances within the time limit. Tables A.10 and A.11 in Appendix A present the total time; # of

166
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

evaluated nodes; the average COLGEN time; the integrality gap at the end of the algorithm; and

the fraction of the total time in each version of 1S-EBP to the total time of the 1S-EBP without

any heuristic pricing for each instance, and for each variant of 1S-EBP. On average, the addition

of the 2p-ESPPRC-LL to the algorithm reduces the total time 13% and the addition of both

2p-ESPPRC-LL and 2p-ESPPRC-CS reduces the total time 16%. In instance p07-t40-v2t2, the

1S-EBP without any heuristic pricing routines could not even solve the root node within the time

limit.

Overall, the computational results show that the addition of 2p-ESPPRC-LL to E-COLGEN

in a branch-and-price algorithm improves the performance of the algorithm for both the 25- and

40-customer LRSP instances. Even though the addition of 2p-ESPPRC-CS does not cause any

positive improvement for the 25-customer instances, we observed that the variant of 1S-EBP

that includes both 2p-ESPPRC-LL and 2p-ESPPRC-CS performed well for the 40-customer

instances. Therefore, we decided to investigate the performance of the 2p-ESPPRC-CS further in

the larger LRSP instances, we implemented two sets of additional experiments using the instances

with 40 customers: an analysis of the variants of the heuristic branch-and-price algorithm and the

two-stage branch-and-price algorithm.

Experiments Using the Heuristic Branch-and-Price Algorithm. Based on the results of our

previous experiments with the variants of 1S-EBP, we decided to investigate the effect of 2p-

ESPPRC-CS in the heuristic branch-and-price algorithm framework described in Section 3.4.3.

The overall algorithm is referred to as HBP and it is either used to provide an upper bound or to

initialize the two-stage branch-and-price algorithm. In HBP, at each node of the branch-and-price

tree, the heuristic column generation algorithm ( H-COLGEN) presented in Figure 3.5 is used. We

constructed two variants of HBP using two implementations of H-COLGEN:

1. In the first variant, apply only 2p-ESPPRC-LL as a pricing algorithm in H-COLGEN, i.e.,

use H-COLGEN(0, 0, l, L).

167
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

0.8

0.6

0.4

0.2

HBP_LL
HBP_CS12LL
HBP_CS20LL
0
1 2 3 4 5 6 7 8 9

Figure 5.5: Performance profile for total time of HBP employing only 2p-ESPPRC-LL, and both
2p-ESPPRC-LL and 2p-ESPPRC-CS with U = 12 or 20

2. In the second variant, employ both 2p-ESPPRC-CS and 2p-ESPPRC-LL, i.e., use H-

COLGEN(u, U , l, L).

In the experiments, we set l = 5 and L=15, and u = 15. We varied the second design by using

two different values (12 and 20) for U . Recall that we set U =12 in 1S-EBP whenever we used

2p-ESPPRC-CS. By setting a higher value for U here, we hoped to get a better upper bound at

the end of HBP at the expense of longer running times. HBP was run with a time limit of 2 CPU

hours in each case.

Figure 5.5 compares the time spent in each variant of HBP. In the figure, HBP LL refers to the

first variant, HBP CS12LL and HBP CS20LL refer to the variants of HBP with 2p-ESPPRC-CS.

The figure shows that HBP with only 2p-ESPPRC-LL is better than the others in 70% of the

instances. However, the variants that employ 2p-ESPPRC-CS solve more of the instances within

the time limit. Since the objective of applying HBP before an exact branch-and-price algorithm is

to start with a good upper bound as well as a good set of columns, the quality of the upper bounds

obtained in each variant of HBP is also important.

We evaluate the quality of the upper bounds in each HBP in Figure 5.6. We drew the perfor-

mance profiles based on the percentage gap between the upper bound at the end of the algorithm

168
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

HBP_LL HBP_LL
HBP_CS12LL HBP_CS12LL
HBP_CS20LL HBP_CS20LL
0 0
0 5000 10000 15000 20000 25000 1 4 16 64 256 1024 4096 16384

(a) 40-customer instances, HBP (%Gap)×100 (b) 40-customer Instances, HBP %Gap, log2 scale

Figure 5.6: Performance profile for the gap between zU B & best known IP

(zU B ) and the optimal or best known solution (IP). To generate the performance profiles, we re-

placed the 0% gaps with 0.0001% in order to avoid division by zero while calculating the ratio

of performance measures. Figure 5.6(a) compares the 100×% gap, and Figure 5.6(b) compares

the gap by using a log2 scale. Both figures show that the variant that uses 2p-ESPPRC-CS with

U =12 performs slightly better than the others in terms of the quality of upper bound. Figures 5.5

and 5.6 indicate that we may prefer to use either the variant of HBP with only 2p-ESPPRC-LL or

the variant with 2p-ESPPRC-CS in which U =12.

Table A.12 in Appendix A presents detailed results with three HBP variants: HBP with only

2p-ESPPRC-LL (columns 3 to 5); HBP with 2p-ESPPRC-CS and U set to 12 in addition to 2p-

ESPPRC-LL (columns 6 to 8); and HBP that is the same as the second variant except that U is set

to 20 (columns 9 to 11). In Table A.12, for each instance and for each HBP variant, we report the

optimal or best known solution (Best IP), the upper bound that is obtained with each variant of

HBP (zU B ), % gap between the zU B and the best IP, and the total time. Based on the table, HBP

with U = 12 results in the best average % gap and the execution time, however, there are some

instances for which one variant of HBP results in a better gap or a lower execution time than the

others.

169
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

Experiments Using the Two-Stage Branch-and-Price Algorithm. Next, we aim to investigate

the effect of 2p-ESPPRC-CS in the two-stage branch-and-price algorithm (2S-EBP) framework

described in Section 3.4.3. Our experiments here evaluate the variants of 2S-EBP similar to those

of HBP evaluated in the previous set of experiments. The second stage of 2S-EBP is an 1S-EBP

initialized using the upper bound and the columns from HBP. In particular, we used three different

variants of HBP that were compared in the previous set of experiments as the first stages of a

2S-EBP algorithm. Thus, we obtained corresponding three variants of the 2S-EBP algorithm.

We employed E-COLGEN(0, 0, 50, 350) in the second stage of each 2S-EBP algorithm. In our

experiments, we used a time limit of 6 CPU hours in the second step. We did not employ 2p-

ESPPRC-CS in E-COLGEN because of the results of our previous experiments. The aim of this

set of experiments is to evaluate the effect of using 2p-ESPPRC-CS in addition to 2p-ESPPRC-

LL in the first stage of 2S-EBP.

Figure 5.7 includes two sub-figures that show the performance of different 2S-EBP variants in

terms of execution time. Figure 5.7(a) compares the total time for the second stage, while Figure

5.7(b) compares the total time for both stages. Figure 5.7 indicates that the variant that employs

2p-ESPPRC-CS in addition to 2p-ESPPRC-LL in the first stage performs better than the one that

uses only 2p-ESPPRC-LL. The variant that uses 2p-ESPPRC-CS with U = 20 is slightly better

than the one that employs 2p-ESPPRC-CS with U =12. For each instance, Table A.13 in Appendix

A presents results for each variant of 2S-EBP.

Figure 5.8 evaluates the performance of the variants of the two-stage algorithms in terms of

the final optimality gap. To generate the graph, we replaced the 0% optimality gaps with value

0.0001% and use a log2 scale as before. Based on the figure, the two-stage variants that use 2p-

ESPPRC-CS perform better than the one that uses only 2p-ESPPRC-LL in the first stage. When

we compare the optimality gap instance by instance (see Table A.13 in Appendix A), we observed

that for most of the instances, the observed gaps are very similar for the variants of 2S-EBP,

though the average gaps are slightly better in the two-stage algorithm with 2p-ESPPRC-CS. The

best average is obtained with the design that sets U to 20.

170
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

HBP_LL HBP_LL
HBP_CS12LL HBP_CS12LL
HBP_CS20LL HBP_CS20LL
0 0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1 1.2 1.4 1.6 1.8 2 2.2

(a) The 1S-EBP time in 2S-EBP (b) Total 2S-EBP time

Figure 5.7: Performance profile for execution time of 2S-EBP

0.8

0.6

0.4

0.2

HBP_LL
HBP_CS12LL
HBP_CS20LL
0
1 4 16 64 256 1024

Figure 5.8: Performance profile for the optimality gap in 2S-EBP, log2

Overall, the test results showed that the addition of 2p-ESPPRC-LL to E-COLGEN improves

the performance of the algorithm. The addition of 2p-ESPPRC-CS to the E-COLGEN, how-

ever, does not have a positive effect for the 25-customer instances. For the 40-customer instances,

employing both 2p-ESPPRC-LL and 2p-ESPPRC-CS in E-COLGEN has good results. Fur-

thermore, employing 2p-ESPPRC-CS in the first stage of the two-stage algorithm improves the

performance.

171
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

5.4.3 Performance of Heuristics

As described in Section 3.4, we implemented three approaches to finding primal feasible solu-

tions: a simple construction-based heuristic (referred to as I-Heur and described in Figure 3.3);

a standard branch-and-bound algorithm; and a heuristic branch-and-price algorithm, HBP (de-

scribed in Section 3.4.3). In this section, we empirically demonstrate the quality of the upper

bound obtained using each heuristic approach and compare the time spent in each approach. The

branch-and-bound approach (referred to as BB) may be used at the root node of either HBP or

1S-EBP. In the BB approach, the IP model that includes columns which are present at the end of

processing the root node of a branch-and-price tree is solved.

In the first experiment, we compared I-Heur, BB at the root node of HBP, HBP that calls BB

at the root node and HBP that does not employ BB. We set a time limit of 1 CPU hour for BB and

each of the two variants of HBP was run with a time limit of 2 CPU hours. In the HBP variant for

the 25-customer instances, we used H-COLGEN(0, 0, 5, 15), and for the 40-customer instances,

we used H-COLGEN(15, 20, 5, 15).

Figure 5.9 compares the execution times for the two variants of HBP. The figure indicates that

for the 40-customer instances, the HBP without BB performs better than the HBP with BB. For the

25-customer instances, HBP without BB is slightly better for about 80% of the instances. Tables

A.14 and A.15 in Appendix A compare the time spent and upper bounds obtained by each of the

four heuristic approaches. The time used by I-Heur is significantly smaller than the other three

approaches in all instances. The BB time which includes the root node solution time of HBP in

addition to the solution time of an integer program, is smaller than the times spent in two variants

of HBP. The average time in BB is about 85% better than the first HBP algorithm, while the second

HBP time is 9% better than the first HBP variant.

In addition to the time, the strength of the upper bounds obtained by four heuristic approaches

must be compared. We compared both the % gap between the LP relaxation and the zU B and

the % gap between the optimal (or best) solution and the zU B obtained in each of the heuristics.

172
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

HBP_BB HBP_BB
HBP_noBB HBP_noBB
0 0
1 2 3 4 5 6 7 8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8

(a) 25-customer instances (b) 40-customer instances

Figure 5.9: Performance profile for the execution time in HBP with and without BB

Figure 5.10 presents the performance profiles comparing these gaps; Figures 5.10(a) and 5.10(b)

compare the % gap between the LP relaxation and the zU B in the 25- and 40-customer instances,

respectively, and Figures 5.10(c) and 5.10(d) compare the 100×%gap between the zU B and the

best IP. The two variants of HBP are significantly better than the other heuristics. Even though

the performance of the HBP variants are very close to each other, the HBP without a BB seems to

perform slightly better. In all instances, BB performs better than I-Heur.

In addition to performance profiles, Tables A.16 and A.17 in Appendix A report the % gap

between the optimal (or best) solution and the zU B and % gap between the LP relaxation and the

zU B , for each variant and instance. As seen from Tables A.16 and A.17, I-Heur and BB provide

upper bounds that are between 1.5% and 11.7% away from the LP relaxation bounds. In some

instances, the BB bounds are slightly better than the I-Heur bounds. This difference is larger

in the 25-customer instances; on average, BB provides bounds that are 5.5% away from the LP

relaxation bound while the average gap is 7% in I-Heur. The two HBP variants provide stronger

bounds than I-Heur and BB. The average gap between the upper bounds obtained in each of the

HBP variant and the LP relaxation bounds is 2.9% in all instances, and the average values are

similar in both 25- and 40-customer instances.

173
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

HBP_BB HBP_BB
HBP_noBB HBP_noBB
BB BB
IHeur IHeur
0 0
0 2 4 6 8 10 0 2 4 6 8 10 12 14 16

(a) Gap btw LP: 25-customer instances (b) Gap btw LP: 40-customer instances

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

HBP_BB HBP_BB
HBP_noBB HBP_noBB
BB BB
IHeur IHeur
0 0
0 10000 20000 30000 40000 50000 60000 70000 0 10000 20000 30000 40000 50000 60000 70000

(c) 100×(%Gap btw IP): 25-customer instances (d) 100×(%Gap btw IP): 40-customer instances

Figure 5.10: Performance profile for the gap between zU B , and the LP bound and the optimal
solution (IP)

Tables A.16 and A.17 also evaluate the strength of the primal heuristic bounds with respect

to the best solution. I-Heur provides bounds that are about 4.6% from the optimal solution, on

average. The effect of BB is more significant for the 25-customer instances compared to the

40-customer instances: the average gap for the 25-customer instances is 2.5%, while it is 3.6%

for the 40-customer instances. Both HBP variants provide similar results and the upper bounds

provided are very strong in almost all instances. The first HBP variant provides either optimal or

best solution in 41 instances out of 64, while the second HBP variant provides either optimal or

174
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

best solution in 43 instances. For both variants, the average gap between the optimal or the best

solution is 0.5%.

These results suggest that all of the bounding procedures perform well and provide good upper

bounds with respect to the LP relaxation or the best solution. The execution times for I-Heur are

very small compared to the time limits set for the branch-and-price algorithm. Therefore, I-Heur is

employed always. The execution times for the HBP variants are not very small, therefore, we can

employ one of the HBP variants if the exact branch-and-price algorithm does not terminate very

quickly, in particularly for larger instances. We evaluate the effect of HBP to the exact branch-

and-price algorithm further in the next section. Since the two variants of HBP perform similar and

the bounds obtained by BB are stronger than I-Heur we can employ BB at the root node of HBP.

In this section, we evaluated BB at the root node of HBP. We may also employ BB at the root node

of the exact branch-and-price algorithm, which we investigate next.

Evaluation of BB in One-Stage Branch-and-Price Algorithm. We designed another set of

experiments to evaluate the effect of BB at the root node of the 1S-EBP algorithm: we compared

1S-EBP in which BB is employed at the root node with the 1S-EBP without BB. In both 1S-EBP

variant, we used E-COLGEN(0,0,50,350) and set the time limit to 8 CPU hours.

Figure 5.11 presents the performance profiles based on the time spent in the variants of 1S-EBP

for the 25- and 40-customer instances. The figure shows that 1S-EBP without BB outperforms 1S-

EBP with BB with respect to the execution time. Figure 5.12 compares the variants of 1S-EBP

with respect to the final optimality gap in the 40-customer instances. In more than 90% of the

instances, the variant with BB provides better gaps than the one without BB.

Tables A.18 and A.19 present the results that compare the variants of 1S-EBP. Tables report

the gap between the zU B obtained using BB and the known optimal (or best) solution (column

BBIP-Opt Gap), the number of evaluated branch-and-price nodes (column Nds) and the total time

(column Time) for each variant of 1S-EBP, as well as the fraction of execution time in the variant

with BB to the execution time of the variant without BB (column Time BB/WOBB). In addition,

175
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

EBP_BB EBP_BB
EBP_noBB EBP_noBB
0 0
1 1.2 1.4 1.6 1.8 2 1 1.1 1.2 1.3 1.4 1.5

(a) 25-customer instances (b) 40-customer instances

Figure 5.11: Performance profile for the execution time of 1S-EBP with and without BB at the
root node

Table A.19 reports the optimality gap at the end of each 1S-EBP algorithm (column Gap). Gap

column is omitted from Table A.18 since in all 25-customer instances that are terminated normally,

the optimality gaps are zero.

With the exception of two 25-customer instances, the total algorithm time increases in all 25-

and 40-customer instances, if BB is employed at the root node. On average, the algorithm time

increases about 13%. Even though the upper bound obtained using the BB procedure is about

1.8% away from the optimal (or best) solution on average, using BB at the root node of 1S-EBP

does not improve the algorithm execution time.

Overall, in cases where the algorithm terminates because of the time limit, employing the

primal heuristic BB at the root node of 1S-EBP reduces the optimality gaps; and in cases where

the algorithm terminates with zero gap, the solution times are deteriorated with the BB procedure.

5.4.4 Comparing One- and Two-Stage Branch And Price

In this section, we compare the performance of the one-stage branch-and-price algorithm (1S-

EBP) with the performance of the two-stage branch-and-price algorithm (2S-EBP). Recall that

HBP is used to initialize 2S-EBP (see Section 3.4.3) and refers to the first stage of 2S-EBP where

176
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

0.8

0.6

0.4

0.2

EBP_BB
EBP_noBB
0
10 20 30 40 50 60 70 80 90 100

Figure 5.12: Performance profile for the optimality gap in 1S-EBP with and without BB for 40-
customer instances

the second stage is an exact branch-and-price. Therefore, in this section, we evaluate the effect of

employing HBP to the overall branch-and-price algorithm.

We implemented two experiments. In the first experiment, the one-stage branch-and-price

algorithm employing E-COLGEN(0,0, 50, 350) was used. We set the time limit to 8 CPU hours.

In the second experiment, we used the two-stage branch-and-price algorithm. In the first stage,

we employed H-COLGEN(0,0,5,15) and set a time limit of 2 CPU hours. In the second stage, we

initialized 1S-EBP with the columns and the upper bound from HBP and used 1S-EBP employing

E-COLGEN(0, 0, 50, 350) with a time limit of 6 CPU hours. Thus, the total time limit for 2S-EBP

was also 8 hours.

Figure 5.13 compares 1S-EBP and 2S-EBP based on the total solution time. In the 25-

customer instances, 1S-EBP performs better than 2S-EBP, while in the 40-customer instances,

2S-EBP performs better than 1S-EBP. Figure 5.14 presents the performance profile comparing

1S-EBP and 2S-EBP with respect to optimality gap at the termination of the algorithm for the

40-customer instances. 2S-EBP is more successful in closing the gaps than 1S-EBP. Tables A.20

and A.21 present results of the experiments for the 25- and 40-customer instances, respectively,

including the number of evaluated nodes, the total solution time in both 1S-EBP and 2S-EBP, and

the execution time of 1S-EBP as a fraction of that of 2S-EBP. In the 25-customer instances, both

177
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

1S-EBP 1S-EBP
2S-EBP 2S-EBP
0 0
1 2 3 4 5 6 1 1.1 1.2 1.3 1.4 1.5 1.6

(a) 25-customer instances (b) 40-customer instances

Figure 5.13: Performance profile for the total solution time of 1S-EBP and 2S-EBP

0.8

0.6

0.4

0.2

1S-EBP
2S-EBP
0
1 4 16 64 256 1024 4096

Figure 5.14: Performance profile for the optimality gap of 1S-EBP and 2S-EBP, log2 scale

algorithms terminated with optimal solutions in all cases. However, in some of the 40-customer

instances, the algorithms terminated with a non-zero optimality gap. Thus, Table A.21 also reports

the % gap at the end of the algorithm. Table A.20 indicates that on average, 1S-EBP is about 13%

faster than 2S-EBP. However, based on Table A.21 1S-EBP is 7% slower than the 2S-EBP. If the

optimality gaps for the 40-customer instances at the end of each algorithm 1S-EBP and 2S-EBP

are compared instance by instance, 2S-EBP is more successful in closing the gaps. The strong

upper bound obtained in HBP has a significant role in this. Therefore, for large instances, we

prefer using 2S-EBP.

178
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

5.4.5 Performance on Generated Instances

In this section, we present solution and performance details of the exact branch-and-price algo-

rithms for the 25- and 40-customer instances. Based on the results of our previous experiments,

we propose below some default settings for our branch-and-price algorithm and present computa-

tional results for all 64 instances in the 25- and 40-customer instance sets.

Branch And Price for the 25-Customer Instances. For the 25-customer instances, we used

1S-EBP employing E-COLGEN(0,0, 50, 350) with a time limit of 6 CPU hours. Table 5.6 shows

the results of our experiments. For each instance, we report the upper bound obtained from I-

Heur (I UB), the LP relaxation value at the root node (LP), the optimal objective value (IP), the

integrality gap at the end of the running time (Gap %), the number of evaluated nodes (Nds),

and the total CPU time in minutes (CPU). The last three columns present some details about the

optimal solution: the indices of open facilities (Loc), the number of vehicles used at each open

facility (# Veh), and the total number of routes in each solution (# Rt).

Solution Time: Overall, the algorithm handled the 25-customer instances well; 29 of the 32

instances were solved to optimality in less than 25 minutes with the majority (24) solved in 10

minutes or less. For only one instance did the algorithm terminate due to the time limit before

proving optimality. The strengths of the LP relaxation and I-Heur bounds have been discussed

before in Sections 5.3 and 5.4.3, respectively.

Effect of Vehicle Capacity and Time Limit Parameters: As can be understood from the labels of

the instances, the computational tests can also measure the sensitivity of the algorithm to vehicle

capacity and time limit. As the labels of the instances indicate that there are 16 pairs of instances in

which the pairs differ only in vehicle capacity (instances with label x-v1t1 and x-v2t1) and 16 pairs

in which the pairs differ only in time limits (the instances with label x-v1t1 and x-v1t2). There

are also eight instances in which both vehicle capacity and time limit increases from a reference

instance (in an instance with label x-v2t2, both the vehicle capacity and the distance limit are

higher than the instance x-v1t1). As expected, because the pricing problem gets more difficult, the

179
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

solution time tends to increase with the vehicle capacity and the vehicle time limit. In addition, the

optimal cost does not increase when the vehicle capacity and/or the vehicle time limit increase(s).

In 10 out of 16 pairs the computation time increases when the vehicle capacity increases, in 14 out

of 16 pairs, the computation time increases when the distance limit increases. In all eight cases,

the computation time increases when both time and vehicle limits increase. However, we note that

there are also instances that favor the increase in time limit or capacity.

An increase in the vehicle capacity and time limit can reduce the cost through a change in

the routing cost together with a decrease in the number of vehicles and/or a change in the set of

selected facilities. On average, we calculated about 3.7% and 2.2% reduction in cost with vehicle

capacity increase and time limit increase, respectively. The cost reduces about 5.8% on average if

both vehicle and time limit increase.

Branch and Price for the 40-Customer Instances. For the 40-customer instance, we used 2S-

EBP. HBP employing H-COLGEN(15, 20, 5, 15) was run with a time limit of 2 CPU hours.

At the root node of HBP, BB was used. In the second stage of branch-and-price, we used E-

COLGEN(0,0,50,350) and set a time limit of 6 CPU hours. Table 5.7 shows the results. For each

instance, we report the upper bound obtained from I-Heur (I UB), the upper bound obtained at the

end of HBP (HBP UB), the LP relaxation value at the root node (LP), the optimal objective value

(IP), the integrality gap at the end of the running time (Gap %), the number of evaluated nodes

(Nds), the CPU times spent in HBP and 1S-EBP, and the total CPU time in minutes.

Solution Time and Quality: As expected, the solution times in the 40-customer instances were

larger than the 25-customer instances. Even though the algorithm terminated with a nonzero gap

in some cases, the algorithm was able to prove optimality most of the time (21 of the 32 instance).

For other instances, good solutions with small integrality gaps were obtained within the time limit.

The gap was at most 0.35% for four of the instances and between 1.26% and 4.18% for the other

seven instances. The strengths of the LP relaxation, I-Heur and HBP bounds have been discussed

before in Sections 5.3 and 5.4.3. For 23 instances out of 32, the HBP upper bound value equaled

180
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

Table 5.6: Results for instances with 25 customers and 5 facilities

Data File I UB LP IP1 Gap % Nds CPU (m) Loc # Veh # Rt
p01-f25-v1t1 4795 4504.4 4699 0 25 1.3 0,4 2,2 10
p01-f25-v1t2 4795 4425.8 4511 0 297 208.1 0,4 1,2 10
p01-f25-v2t1 4722 4366.1 4425 0 99 9.9 0,4 1,2 7
p01-f25-v2t2 4722 4295.4 4425 0 119 15.9 0,4 1,2 7
p01-l25-v1t1 5193 4849.8 4899 0 24 0.6 0,1 2,2 8
p01-l25-v1t2 4968 4725.4 4815 0 139 2.9 0,4 2,2 8
p01-l25-v2t1 4809 4669.0 4759 0 157 5.7 0,4 2,2 6
p01-l25-v2t2 4762 4537.1 4602 0 39 6.9 0,1 1,2 7
p03-f25-v1t1 5281 4810.0 5062 0 236 3.0 0,2 3,2 10
p03-f25-v1t2 5139 4701.1 4837 0 77 2.5 0,2 2,2 10
p03-f25-v2t1 5048 4606.1 4758 0 393 4.6 0,4 2,2 8
p03-f25-v2t2 4831 4502.5 4742 0 52 5.3 0,4 2,2 8
p03-m25-v1t1 4995 4697.4 4785 0 509 5.2 0,4 2,2 10
p03-m25-v1t2 4869 4578.2 4785 0.1 3790 360.0 0,4 2,2 10
p03-m25-v2t1 4992 4485.8 4689 0 59 2.9 0,4 2,2 7
p03-m25-v2t2 4767 4396.1 4479 0 9 8.7 0,4 1,2 7
p03-l25-v1t1 5186 4795.3 4842 0 17 0.6 0,2 2,2 9
p03-l25-v1t2 4871 4670.2 4803 0 37 1.9 0,4 2,2 7
p03-l25-v2t1 4805 4620.9 4741 0 41 2.5 0,4 2,2 6
p03-l25-v2t2 4805 4507.6 4655 0 37 7.5 2,4 1,2 7
p07-f25-v1t1 4870 4642.6 4761 0 55 1.4 0,2 1,3 7
p07-f25-v1t2 4823 4544.5 4685 0 67 131.2 2,4 2,1 9
p07-f25-v2t1 4721 4485.6 4703 0 229 8.8 0,2 2,2 6
p07-f25-v2t2 4721 4388.3 4482 0 197 15.1 0,2 1,2 6
p07-s25-v1t1 5412 5000.1 5098 0 599 1.5 0,4 2,3 9
p07-s25-v1t2 5238 4852.2 4889 0 123 2.9 0,4 2,2 9
p07-s25-v2t1 5159 4739.4 4787 0 21 0.6 0,4 2,2 7
p07-s25-v2t2 4934 4609.8 4781 0 100 3.6 0,4 2,2 8
p07-t25-v1t1 5134 4836.5 4898 0 35 0.9 0,1 2,2 8
p07-t25-v1t2 5162 4703.1 4886 0 1345 25.0 0,2 2,2 8
p07-t25-v2t1 5025 4646.2 4767 0 777 18.8 0,2 2,2 6
p07-t25-v2t2 4769 4513.1 4592 0 91 12.6 0,1 1,2 6
1 Optimal or best IP found by the algorithm within the time limit.

181
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

that of the optimal (or best integer) solution.

Effect of Vehicle Capacity and Time Limit Parameters: As with the 25-customer instances we

consider the change in the performance of the algorithm with vehicle capacity and time limit val-

ues. In seven out of 16 pairs, the computation time increases when the vehicle capacity increases,

in 12 out of 16 pairs, the computation time increases when the time limit increase. In six cases out

of eight, the computation time increases when both time and vehicle limits increase. Because in

a third of the 40-customer instances, the algorithm terminated due to the time limit, the solution

times are not very sensitive to the changes in the capacity parameters. On average, we calculated

about 3.8% and 1.6% reduction in cost with vehicle capacity increase and time limit increase,

respectively. The cost is reduced about 5.3%, on average, if both vehicle and time limit increase.

These results are in line with the corresponding averages for the 25-customer instances.

Discussion of the Solution Details. The last three columns of Table 5.6 and Table 5.9 present

some details about the optimal solutions: the indices of open facilities (Loc), the number of vehi-

cles used at each open facility (# Veh), and the total number of routes in each solution (# Rt) for

the 25- and 40-customer instances, respectively.

In all 25-customer instances, the number of facilities selected to be used was two and the total

number of vehicles used was between three and five, while the number of vehicles per facility was

between one and three. In Table 5.8, we present some summary statistics about the solutions of all

32 instances, such as the number of vehicles used, the average number of routes per vehicle, the

average number of customers per route and per vehicle, and the maximum number of customers

per vehicle. Based on the table, we can see that each vehicle’s schedule covers about two routes

(2.1 routes per vehicle on average) and some vehicles cover more than three routes (average # of

routes is between 1.5 to 3.3) including 11 customers at most. These statistics give us an intuition

about the difficulty of the instances.

In all 40-customer instances, three to four facilities are selected to be open, and the total

number of vehicles used changes between four to six. Similar to Table 5.8, Table 5.10 presents

182
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

Table 5.7: Results for instances with 40 customers and 5 facilities

I HBP Gap CPU(m)
Data file UB UB LP IP1 % Nds HBP EBP Total
p01-f40-v1t1 7170 7131 6976.0 7131 0 807 53.2 150.0 203.2
p01-f40-v1t2 7170 7170 6826.2 6999 0.32 1065 120.0 360.0 480.0
p01-f40-v2t1 7064 6866 6700.2 6866 0 203 4.9 113.0 117.9
p01-f40-v2t2 6984 6821 6581.0 6821 1.26 35 35.4 360.7 396.1
p01-l40-v1t1 7449 7183 6899.7 7183 3.22 573 120.1 360.4 480.5
p01-l40-v1t2 7236 6934 6757.4 6934 0 43 5.6 48.0 53.6
p01-l40-v2t1 7166 6823 6616.7 6823 0 129 2.4 103.2 105.6
p01-l40-v2t2 7166 6633 6513.6 6612 0 25 8.4 187.6 196.0
p03-f40-v1t1 9253 8780 8697.4 8780 0 97 9.6 97.8 107.4
p03-f40-v1t2 9028 8753 8578.6 8753 0 41 28.0 154.6 182.6
p03-f40-v2t1 8943 8663 8439.0 8440 0 3 7.8 8.8 16.6
p03-f40-v2t2 8943 8438 8390.4 8438 0 9 11.5 138.2 149.7
p03-m40-v1t1 7827 7740 7050.9 7253 0 2524 120.0 360.1 480.1
p03-m40-v1t2 7544 7236 6907.7 7236 2.95 24 84.0 360.4 444.4
p03-m40-v2t1 7479 7132 6793.9 7132 2.79 27 120.0 360.1 480.1
p03-m40-v2t2 7430 6947 6655.5 6947 4.18 4 37.2 360.3 397.5
p03-l40-v1t1 7386 7167 6920.4 7167 0 183 1.5 55.6 57.1
p03-l40-v1t2 7212 6950 6771.6 6950 0.05 2121 120.0 360.0 480.0
p03-l40-v2t1 7121 6849 6662.9 6849 0 159 3.9 113.8 117.7
p03-l40-v2t2 7018 6846 6544.7 6639 0 53 39.0 290.9 329.9
p07-f40-v1t1 7424 7167 7053.1 7167 0 95 3.3 22.1 25.3
p07-f40-v1t2 7340 7001 6906.1 7001 0 433 11.0 75.5 86.5
p07-f40-v2t1 7174 6881 6745.5 6881 0 317 8.5 149.0 157.5
p07-f40-v2t2 7000 6845 6629.4 6845 2.34 43 20.3 360.1 380.4
p07-s40-v1t1 7860 7343 7256.3 7343 0 79 0.8 4.4 5.2
p07-s40-v1t2 7656 7124 7091.1 7117 0 29 1.8 6.5 8.3
p07-s40-v2t1 7466 7005 6924.5 7000 0 277 3.5 128.6 132.2
p07-s40-v2t2 7241 6944 6762.8 6944 0 251 5.7 211.2 216.9
p07-t40-v1t1 7283 7013 6940.5 7012 0 206 10.2 147.8 158.1
p07-t40-v1t2 7058 6972 6806.6 6972 0.35 66 76.2 360.5 436.7
p07-t40-v2t1 6945 6857 6649.8 6665 0.04 19 68.2 391.1 459.2
p07-t40-v2t2 6945 6658 6535.4 6658 1.64 6 26.1 363.3 389.5
1 Optimal or best IP found by the algorithm within the time limit.

183
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

Table 5.8: More Solution Details for 25-customer instances

Case Changes btw [min - max] Avg.
# of vehicles per facility 1-3
Total # of vehicles 3-5
Avg. # of routes per vehicle 1.5 - 3.3 2.1
Avg. # of customers per route 2.5 - 4.2 3.3
Avg. # of customers per vehicle 5 - 8.3 6.9
Max # of customers per vehicle 7 - 11 8.5

some summary statistics about the solutions of the 40-customer instances. In the best solutions

reported, each facility employed up to three vehicles, each of which was serving at least two routes

and may serve up to four routes. On average, a vehicle covered 8.3 customers and the maximum

number of customers per vehicle was between seven and 14.

The solution information provided in Tables 5.6 and 5.9 highlights the value of solving the

integrated problem instead of a series of problems. Although the same number of facilities is

selected within an instance group (meaning a fixed set of customer locations and a set of candidate

facility locations), the particular facilities selected may differ depending on the vehicle capacity

and time limit values. Which facilities are selected impacts which customers are served by each

facility and hence the sequence of the routes and the structure of the pairings. From the solution

information, we can see that, within an instance group, increases in the vehicle capacity and the

time limit reduce the total cost through changes in facility, decreases in the number of vehicles

required, and changes in the underlying route sequences.

In all 25- and 40-customer instances, only the minimum number of required facilities (based

on the facility capacity) were selected to be open. This is expected because on average, the total

facility fixed costs are 77% of the total cost. The second major cost component is the total vehicle

fixed costs. On average, 17% of the total cost is the total vehicle fixed cost. Thus, the operating

costs is about 6% of the total cost. In the 25-customer instances, the average vehicle fixed cost

is about 17% of the average facility cost, while in the 40-customer instances, this percentage is

about 25. In all instances, on average, the total operating cost is 8% of the total facility cost.

184
5.4. COMPUTATIONAL EXPERIMENTS WITH BRANCH AND PRICE

Table 5.9: Solution details for instances with 40 customers and 5 facilities
Data file Loc # Veh # Rt Data file Loc # Veh # Rt
p01-f40-v1t1 0,3,4 2,2,1 14 p03-l40-v1t1 0,2,4 2,2,2 13
p01-f40-v1t2 0,2,4 1,2,2 14 p03-l40-v1t2 0,2,4 2,1,2 13
p01-f40-v2t1 0,2,4 1,2,2 11 p03-l40-v2t1 0,2,4 1,2,2 10
p01-f40-v2t2 0,3,4 2,1,1 10 p03-l40-v2t2 0,2,4 1,1,2 10
p01-l40-v1t1 0,2,4 1,2,3 15 p07-f40-v1t1 0,2,3 1,2,2 13
p01-l40-v1t2 0,2,4 1,2,2 15 p07-f40-v1t2 0,2,4 1,2,2 13
p01-l40-v2t1 0,2,4 1,2,2 10 p07-f40-v2t1 0,2,4 1,2,2 10
p01-l40-v2t2 0,2,4 1,1,2 10 p07-f40-v2t2 1,2,4 1,1,2 10
p03-f40-v1t1 0,1,2,4 2,1,1,1 16 p07-s40-v1t1 0,1,4 2,2,2 14
p03-f40-v1t2 0,1,2,4 1,1,2,1 16 p07-s40-v1t2 0,1,4 1,2,2 13
p03-f40-v2t1 0,1,2,4 1,1,1,1 11 p07-s40-v2t1 0,1,4 1,2,2 10
p03-f40-v2t2 0,1,2,4 1,1,1,1 11 p07-s40-v2t2 0,2,4 1,2,2 10
p03-m40-v1t1 0,2,4 2,2,2 15 p07-t40-v1t1 0,1,4 1,2,2 13
p03-m40-v1t2 0,2,4 2,2,2 15 p07-t40-v1t2 0,2,4 1,2,2 13
p03-m40-v2t1 0,2,4 2,2,2 11 p07-t40-v2t1 0,1,4 1,1,2 10
p03-m40-v2t2 0,1,4 2,1,2 12 p07-t40-v2t2 0,2,4 1,1,2 9

Table 5.10: More Solution Details for 40-customer Instances

Case Changes btw [min - max] Avg.
# of vehicles per facility 1-3
Total # of vehicles 4-6
Avg. # of routes per vehicle 1.8 - 3.2 2.5
Avg. # of customers per route 2.5 - 4.4 3.4
Avg. # of customers per vehicle 6.7 - 10 8.3
Max # of customers per vehicle 7 - 14 10.2

185
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

Table 5.11: Comparison of Lin et al. (2002) results and branch-and-price results

Instance Lin B & B Lin Heuristic 1S-EBP

# Cust # Fac Objective CPU (s) Objective CPU(s) Objective CPU(s)
10 4 309,817 1155 309,817 0.44 309,817 0.66
10 4 309,808 982 309,808 0.49 309,808 0.10
12 4 312,036* > 10,000 312,036 0.82 312,036 0.05
27 4 - - 625,752.5 6 625,750.2 57.83

5.4.6 Performance on Instances from the Literature

The only LRSP instances available in the literature are those of Lin et al. (2002) which presents

six instances: a base instance that includes 27 customers and four facility locations in Hong Kong

plus five instances generated from the original by considering three subsets of 10 customers and

two subsets of 12 customers while keeping the same four facility locations. We used only the four

instances available to us for testing.

Lin et al. (2002) solved these instances with a branch-and-bound algorithm and with their

heuristic algorithm. Within a time limit of 10,000 seconds, their branch-and-bound algorithm was

able to prove optimality only for the 10-customer instances and was not able to find a feasible solu-

tion for the original 27-customer problem. The heuristic quickly found solutions for all instances.

Table 5.11 compares their reported results to results obtained with 1S-EBP. As shown in Table

5.11, our algorithm quickly provided optimal solutions for all four instances. (The * indicates the

best solution found in 10,000 CPU seconds on a Pentium III machine.)

5.5 Computational Experiments for Cut Generation

The goal of the computational experiments in this section is to test the performance of the branch-

and-price algorithm when augmented with some of the cutting plane generation procedures de-

scribed in Chapter 4. We empirically show the effect of each implemented procedure. We im-

plemented lifted cover inequalities, generalized assignment polytope (GAP) cuts and disjunctive

186
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

valid inequalities.

5.5.1 Lifted Cover and GAP Inequalities

In this section, we evaluate the effect of lifted cover inequalities (inequalities of the form 4.16))

and GAP cuts (inequalities of the form (4.18)) which are valid for (SPP-LRSP) as described in

Section 4.2.2.

For the experiments in this section, we solved the 25-customer instances with the one-stage

branch-and-price algorithm (1S-EBP) using E-COLGEN(12, 12, 50, 350). In our implementa-

tions, to generate valid inequalities in 1S-EBP we ran cut generation procedure (referred to as

CUTGEN) at each node after the column generation algorithm. For any node, we called CUT-

GEN if the depth of the node was less than 9. At node i of the branch-and-price algorithm, the

steps of the bounding procedure were as follows:

• Call E-COLGEN.

• If xiLP is fractional and the depth ≤ 8, call CUTGEN procedure. Otherwise STOP.

• If a valid inequality is generated and added to Pi , call E-COLGEN procedure. Otherwise

STOP.

We compared four variants of 1S-EBP:

1. 1S-EBP without any call to CUTGEN.

2. 1S-EBP employing the CUTGEN procedure that searches for violated lifted cover inequal-

ities.

3. 1S-EBP employing the CUTGEN procedure that searches for violated GAP inequalities.

4. 1S-EBP employing the CUTGEN procedure that searches for both violated lifted cover and

GAP inequalities.

187
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

1S-EBP-nocutgen 1S-EBP-nocutgen
1S-EBP-gap 1S-EBP-gap
1S-EBP-knap 1S-EBP-knap
1S-EBP-knap-gap 1S-EBP-knap-gap
0 0
1 1.05 1.1 1.15 1.2 1.25 1.3 1 1.1 1.2 1.3 1.4 1.5

(a) Total execution time (b) # of evaluated nodes

Figure 5.15: Performance profile comparing 1S-EBP without any cuts, with CUTGENK ,
CUTGENG and CUTGENK+G

We use notations CUTGENK , CUTGENG and CUTGENK+G to denote the cut generation

procedures applied in the second, third and forth variants of 1S-EBP, respectively. We used a time

limit of 6 CPU hours for all experiments.

Figure 5.15 compares the four variants of 1S-EBP based on the total solution time and the

number of evaluated nodes in the branch-and-price tree in order to evaluate the effect of the CUT-

GEN procedures to the performance of the branch-and-price algorithm. Figure 5.15(a) indicates

that 1S-EBP without any cuts and 1S-EBP employing CUTGENG perform better than the other

variants of 1S-EBP with respect to the execution time. CUTGENG slightly improves the perfor-

mance of 1S-EBP with respect to the execution time for some of the instances. However, we can

conclude that the branch-and-price algorithm without any cut generation is better than the ones

employing any of the cut generation procedures. Figure 5.15(b) indicates that the variants with

lifted cover inequalities result in less number of node evaluations in the algorithm than the variant

without any cut or the variant with CUTGENG .

Tables A.22 and A.23 in Appendix A present some details about the computational experi-

ments comparing the four variants of the one-stage branch-and-price algorithm. Tables report the

number of evaluated nodes (Nds) and the number of cuts generated (# cuts) in each variant of

188
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

1S-EBP. In addition, tables report the ratio of the time of a variant using a CUTGEN procedure

to the time of the variant without CUTGEN (Time W/WO) to compare the performance of each

variant with a CUTGEN against the variant without any cuts. Tables show that the CUTGEN

procedure only effects the performance of the algorithm slightly. On average, the computational

time of the variants with CUTGEN is almost the same as the variant without CUTGEN. On each

instance, the time in the algorithms with CUTGEN is either slightly lower or slightly higher than

the time for the base algorithm. In most of the instances, if the cut generation was successful, the

number of evaluated nodes either stayed the same or decreased. Compared to 1S-EBP without

any cut generation, in three out of 31 instances, the number of evaluated nodes increased in the

variant with CUTGENG . These numbers were seven and six for the variants with CUTGENK

and CUTGENK+G , respectively.

As a summary, even though when we investigate the output in details, we can see that the

LP relaxation bound improves slightly with the generated cuts at the relevant nodes and the total

time spent in CUTGEN procedures is very small, the CUTGEN procedure does not contribute

significantly to the performance of the overall algorithm.

We also tested the CUTGENK and CUTGENG in some of the 40-customer instances. We

did not see promising results for larger instances either. In addition, we had some computational

difficulties in most of the instances. Because of these new cuts, the size of the array that keeps the

nonzero coefficients in constraints and cuts increases quickly and exceeds the predetermined limit

of about 80000. We conclude that the improvement in the LP relaxation bound obtained with cuts

is not worth the overall performance deterioration in the algorithm.

5.5.2 Disjunctive Valid Inequalities

In this section, we present some computational results for the four implementations of the disjunc-

tive cut generation algorithm described in Section 4.4. We refer to the cut generation procedures

for each of the four implementations as DISJCUTGENI1 , DISJCUTGENI2 , DISJCUTGENI3 ,

and DISJCUTGENI4 .

189
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

For the experiments in this section, we solved the 25-customer instances with the one-stage

branch-and-price algorithm (1S-EBP) using E-COLGEN(12, 12, 50, 350). We used a time limit

of 6 CPU hours. At each node, when the depth was 8 or less, we called one of the DISJCUTGEN

procedure after the column generation algorithm. The steps of the bounding procedure followed

at each node of the branch-and-price tree were the same as those listed in the previous section

except that we called a DISJCUTGEN procedure. All disjunctive valid inequalities generated

were added as local for the corresponding node.

We compared five variants of 1S-EBP:

1. 1S-EBP without any call to DISJCUTGEN.

2. 1S-EBP employing DISJCUTGENI1 in order to identify violated disjunctive valid inequal-

ities based on implementation 1.

3. 1S-EBP employing DISJCUTGENI2 in order to identify violated disjunctive valid inequal-

ities based on implementation 2.

4. 1S-EBP employing DISJCUTGENI3 in order to identify violated disjunctive valid inequal-

ities based on implementation 3.

5. 1S-EBP employing DISJCUTGENI4 in order to identify violated disjunctive valid inequal-

ities based on implementation 4.

In variants with DISJCUTGENI2 , DISJCUTGENI3 and DISJCUTGENI4 , we added at most

one disjunctive cut at each node. In the branch-and-price algorithm employing DISJCUTGENI1 ,

at each node, we may generate at most four disjunctive valid inequalities.

Figure 5.16 compares the performance of five variants of 1S-EBP based on the execution

time and the number of evaluated nodes. Figure 5.16(a) indicates that we obtain the best per-

formance (with respect to the execution time) with the one-stage branch-and-price without any

cut generation calls. However, the performance of the variants employing DISJCUTGENI1 and

DISJCUTGENI2 are relatively close to the variant without cut generation calls. DISJCUTGENI4
190
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
1S-EBP-nocutgen 1S-EBP-nocutgen
1S-EBP-dsj1 1S-EBP-dsj1
1S-EBP-dsj2 1S-EBP-dsj2
1S-EBP-dsj3 1S-EBP-dsj3
1S-EBP-dsj4 1S-EBP-dsj4
0 0
0 10 20 30 40 50 60 70 80 2 4 6 8 10 12

(a) Total execution time (b) # of evaluated nodes

Figure 5.16: Performance profile comparing 1S-EBP without any cuts, with DISJCUTGENI1 ,
DISJCUTGENI2 , DISJCUTGENI3 , DISJCUTGENI4

procedure reduces the performance of 1S-EBP significantly compared to other cut generation pro-

cedures in terms of computation time. Figure 5.16(b) indicates that the variants with

DISJCUTGENI1 and DISJCUTGENI3 result in an increase in the number of evaluated nodes.

To evaluate the performance reduction with the DISJCUTGEN procedures further, we com-

pared the time spent in the cut generation procedures in each of the variant. Figure 5.17 compares

four implementations of disjunctive cut generation procedure. The time spent in the DISJCUT-

GEN procedure is mostly the time to solve the CGLPs. Since the CGLP in implementation 1 is

very small compared to other implementations, the time spent in DISJCUTGENI1 is negligible.

The CGLP in implementation 2 is larger than implementation 1, but much smaller than the CGLPs

in implementations 3 and 4 with respect to the number of constraints and variables. The total time

spent in the DISJCUTGEN procedure increased significantly in implementation 3 and implemen-

tation 4. The number of variables and the main set of constraints in the CGLPs for implementation

3 and implementation 4 are the same. However, in implementation 4, we also added some restric-

tions for the coefficients of variables in the generated cut, which made the CGLP more difficult to

solve and resulted in a more significant increase in time.

191
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

0.8

0.6

0.4

0.2

1S-EBP-dsj1
1S-EBP-dsj2
1S-EBP-dsj3
1S-EBP-dsj4
0
0 2000 4000 6000 8000 10000

Figure 5.17: Performance profile for the cut generation time with DISJCUTGENI1 ,
DISJCUTGENI2 , DISJCUTGENI3 , DISJCUTGENI4

The presented results showed that the DISJCUTGEN procedure degrades the algorithm per-

formance. Especially, the cut generation procedures for implementation 3 and implementation 4

are significantly inefficient. Implementation 1 and implementation 2 seem the most promising

ones with respect to time. The performance of the variants using implementation 1 and implemen-

tation 2 are closest to the algorithm without any cut generation. However, even though the total

execution time for the algorithm increases, are the disjunctive valid inequalities generated in our

algorithm helping somehow by improving the LP relaxation bound at the related nodes? To be

able answer this question, we looked at the optimal LP solutions before and after adding the cuts

at each node in each implementation. Let x̄ be the fractional solution to the LP relaxation of Pi .

First, we evaluate implementation 3 and implementation 4.

In implementation 3, almost none of the inequalities generated is violated by x̄. The reason

is that the cut coefficients are obtained by combining the coefficients found by solving the CGLP.

As explained in Section 4.4, for implementation 3, x̄ can be mapped to a set of variables (say Y )

associated with the arcs in the problem. Recall that we define a one-to-one mapping in order to

represent the set Y with a single member (say ȳ). A disjunctive valid inequality violated by ȳ is

found first by solving the CGLP. We combine some coefficients considering the other members of

Y and then rewrite the obtained inequality in the variable space of Pi . Therefore, the inequality

192
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

with combined coefficients is not necessarily violated by ȳ and so the corresponding reformulation

of the inequality may not be violated by x̄. Hence, implementation 3 performs very poorly with

respect to LP bound improvement. Therefore, we can easily say that DISJCUTGENI3 is not a

promising cut generation procedure.

Implementation 4 was designed to overcome the problems implementation 3 had with gener-

ating cuts that are not violated by ȳ. Instead of combining the coefficients after the cut is gener-

ated, implementation 4 adds new conditions to the CGLP to eliminate the combination procedure.

Therefore, the generated cut must be violated by ȳ. Implementation 4 is successful for small in-

stances. However, our computational tests with the 25-customer instances showed that the addition

of new conditions to the CGLP increases the solution time significantly. In addition to this draw-

back, by investigating the solution to (RMPik ) before and after adding the inequality, we found that

there exists some nodes in which even though the generated inequality improves the LP relaxation

value at first, the additional column generation steps move the LP solution to a point which has an

objective value that is the same or only slightly better than the previous objective value. As seen

from the form of the generated valid inequality (4.55), there are coefficients corresponding to the

directed arcs in the problem. Because the distance matrix is symmetric, the column generation al-

gorithm generates columns that include different orientations of routes than those that are already

a part of the columns in the previous LP solution. The new LP solution satisfies the inequality, but

has the same objective value as before because of these new equivalent columns. Therefore, the

LP solution value does not change while the set of nonzero pairing variables changes. We provide

an example to illustrate this problem.

Example The branch-cut-and-price algorithm with cuts from DISJCUTGENI4 was run for in-

stance p01-f25-v1t1. In one of the nodes, the LP relaxation value before the cut was equal to

the LP relaxation value after the cut was added and additional column generation was done. We

compared two LP solutions and only listed the non zero variables which were different in these

two solutions.

193
5.5. COMPUTATIONAL EXPERIMENTS FOR CUT GENERATION

Before cut is added:

z̄1 = 0.188 : f0 − 4 − 11 − f0 − 22 − 6 − 23 − f0 − 13 − 24 − f0 − 17 − f0

z̄2 = 0.16 : f0 − 3 − 16 − 18 − 12 − f0 − 22 − 6 − 5 − f0 − 13 − 24 − f0

z̄3 = 0.028 : f0 − 22 − 6 − 23 − f0 − 3 − 18 − 12 − f0 − 13 − 5 − f0

After cut is added and column generation is done:

z̄2 = 0.131 : f0 − 3 − 16 − 18 − 12 − f0 − 22 − 6 − 5 − f0 − 13 − 24 − f0

z̄4 = 0.188 : f0 − 11 − 4 − f0 − 23 − 6 − 22 − f0 − 24 − 13 − f0 − 17 − f0

z̄5 = 0.029 : f0 − 12 − 18 − 16 − 3 − f0 − 5 − 6 − 22 − f0 − 24 − 13 − f0

z̄6 = 0.028 : f0 − 23 − 6 − 22 − f0 − 12 − 18 − 3 − f0 − 13 − 5 − f0

After the cut was added, the column generation generated z̄4 and z̄6 and replaced z̄1 and z̄3 with

them. As the sequences of customers in the corresponding pairings are investigated, equivalence

of z̄1 and z̄4 , and z̄3 and z̄6 are obvious. In addition, z̄5 which is equal to z̄2 took nonzero value in

the solution.

Even though there are some nodes in which the disjunctive cuts from DISJCUTGENI4 strictly

improve the LP solution values, we can conclude that the branch-and-price algorithm without

DISJCUTGENI4 is better of than the branch-cut-and-price algorithm with DISJCUTGENI4 . The

performance of the overall algorithm is reduced with DISJCUTGENI4 because of the cases in

which an improvement in the LP solution cannot be obtained with cuts (similar to the example)

and the large CGLP solution times in DISJCUTGENI4 .

The form of the cuts generated in implementation 2 is the same as the forms of the cuts in

implementations 3 and 4. The implementation provides coefficients corresponding to arcs in the

problem. Therefore, we face the same problem in implementation 4, in some of the nodes, the

LP relaxation bound does not improve with the addition of the cut. The symmetric nature of the
194
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

problem allows the additional column generation to provide columns that are equal to the existing

columns.

In implementation 1, we obtained improvement in the LP relaxation bounds in almost all

cuts. The new LP solution does not include new equivalent columns (as in the example) since the

information passed from the cuts is not symmetric in nature. However, we observed that in some

of the nodes the LP relaxation bound improvement is very small (less than 1). Therefore, for these

nodes, there is almost no benefit of adding the new cuts, but the cost of searching for the cuts and

additional column generation time is significant. This explains why the computational results for

implementation 1 do not show an overall performance improvement even though the time spent in

DISJCUTGENI1 procedure is small. Another observation from this implementation is that, the

number of times that the function searching for disjunctive cuts fails may be large. More clearly,

even though we have a fractional xiLP , we may not be able to find a disjunctive cut. The reason

is that the cuts are generated based on a reformulation of the integer model considered at node i.

The fractional solution of this reformulation obtained from xiLP may not be an extreme point of

the reformulation polytope. It may be an interior point or on a face of the polytope, which results

in a failure in the CGLP that aims to find an inequality violated by that solution.

As a summary, implementation 1 seems to be the best approach compared to other imple-

mentations of the disjunctive cut generation procedure with respect to the total time spent in cut

generation, total execution time and the LP bound improvement. However, the algorithm with

DISJCUTGENI1 terminates abnormally because of the constraint matrix size restriction in more

than 20% of the instances while this percentage is at most five for the other variants of the algo-

rithm.

5.6 Value of Integrating Scheduling Decision

In this thesis, we focused on the integrated location, routing and scheduling problem (LRSP).

The difference between the LRSP and the integrated location and routing problem (LRP) is the

195
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

incorporation of the scheduling decision into the optimization model. The LRP models assume

that each vehicle covers at most one route. However, the LRSP models remove this classical

assumption and consider the possibility of assigning multiple routes to a single vehicle. At the

beginning of the thesis, in Section 1, we claimed that optimizing scheduling decision with the

facility location and vehicle routing decisions results in more efficient system designs, and more

accurate and cost-effective solutions. We also explained the motivating factors for this statement.

In this section, our aim is to empirically show the effect of including scheduling decision into the

model by comparing an LRSP model against an LRP model. One can always find the optimal

assignment of routes in the solution to vehicles based on the time limit after solving an LRP. We

experimentally show the effect of solving an LSRP as opposed to solving an LRP and a scheduling

problem sequentially. We focus on the effect of the vehicle capacity.

Models to be Compared. We compare an LRSP with capacity constraints with an LRP with

capacity constraints. In the LRP, to be able to satisfy the time limit constraint for each route, we use

an LRP with distance constraints (LRPDC). After solving an LRPDC, a scheduling problem for

the generated routes can be solved to minimize the number of vehicles. By solving a bin packing

problem (BPP), the minimum number of vehicles required to cover all routes, and assignment of

routes to vehicles can be found. The capacity of each bin is equal to the time limit, and the length

of each route is the space that is allocated by the route in a bin that corresponds to a vehicle. For

each facility j ∈ J, we solve the BPP using the following integer programming model.

X
minimize yi (5.2)
i∈K
X
subject to tr xri ≤ LT yi ∀i ∈ K, (5.3)
r∈R
X
xri = 1 ∀r ∈ R, (5.4)
i∈K
xri ∈ {0, 1} ∀r ∈ R, i ∈ K, (5.5)

yi ∈ {0, 1} ∀i ∈ K, (5.6)

196
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

where R is the set of routes for facility j, K is the set of vehicles and tr is the trip time for route r.

yi is equal to 1 if the vehicle i is used, and xri is 1 if route r is assigned to vehicle i for all r ∈ R

and i ∈ K.

To solve the LRPDC, we use a set partitioning-based formulation which has the same structure

as the formulation for the LRSP with the only difference being the definition of their column sets.

In the LRPDC, a column that is generated using the column generation algorithm represents a

single feasible vehicle route satisfying the vehicle capacity and the time limit constraints. In the

LRPDC, the pricing problem is an ESPPRC, where the vehicle capacity and the time limit are

the resources. To solve the pricing problem in the LRPDC, we use only the first phase of the

2p-ESPPRC (described in Section 3.2.2). By using the first phase of the 2p-ESPPRC-LL and

2p-ESPPRC-CS, we can obtain heuristic pricing algorithms for the LRPDC. In Chapter 6, we

describe a branch-and-price algorithm to solve an LRP in detail.

Implementation Details and Discussion of the Results. In our experiments, we used the 25-

customer instances with smaller time limit value (instances labeled with “t1”). Recall that two

different vehicle capacities calculated with α = 0.1 and 0.2 were used in these instances. These

α values were specifically chosen to be small for the LRSP in order to result in smaller vehicle

capacity values requiring multiple trips to a facility. We expect to see that as the vehicle capacity

gets larger, the benefit of solving the LRSP instead of the LRPDC is reduced. Therefore, we

populated the set of instances with two more capacity values referred to as LV3 and LV4 that were

calculated by setting α = 0.5 and α = 0.8. We indicate these capacity values using “v3” and “v4”

in the labels of the instances. Table 5.12 lists all vehicle capacity values for each instance group.

We obtained 32 instances for this experiment. For each model (LRSP and LRPDC) and instance,

we used the one-stage branch-and-price algorithm with E-COLGEN(0,0,50,350) with a time limit

of 8 CPU hours.

Figure 5.18 compares the total solution time for each problem. Figure indicates that the

branch-and-price algorithm for the LRDC performs better than the branch-and-price algorithm

197
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Table 5.12: Vehicle capacity values

Ref LV1 (α = 0.1) LV2 (α = 0.2) LV3 (α = 0.5) LV4 (α = 0.8)

p01 50 70 120 170
p03 60 80 140 190
p07 50 70 110 150

0.8

0.6

0.4

0.2

LRPDC
LRSP
0
10 20 30 40 50 60 70 80 90

Figure 5.18: Performance profile comparing the execution time of 1S-EBP for LRSP and LRPDC

for the LRSP with respect to computational time. This is expected because the pricing problem

used to solve the LRSP is more difficult than the pricing problem used in the solution of the LR-

PDC. For the set of instances that were used in the experiment, we calculated that the column

generation time in the LRSP was about 2.9 times larger on average than the column generation

time in the LRPDC.

In Tables 5.13 and 5.14, we report the solution details for each instance and for each model,

including indices of facilities that are open in the solution (Fac. Idx), number of vehicles (# vec.),

total facility fixed cost (Fac. Cost), total operating cost (Route Cost), and total cost (TCost). The

fourth column in Table 5.13 (# vec. w. BPP) reports the number of vehicles obtained by solving

the BPP for the LRPDC solution to reduce the number of vehicles and the seventh column (TCost

w. BPP) for the LRPDC reports the total cost that was calculated with the reduced number of

vehicles after the BPP.

First, we compared the optimal total costs of the problems reported in columns labeled with

198
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Table 5.13: Solution details for LRPDC + BPP

Data File Fac. # # vec. Fac. Route TCost

Idx vec. w. BPP Cost Cost w. BPPt
p01-f25-v1t1 0,4 9 4 3368 447 4715
p01-f25-v2t1 0,4 7 4 3368 377 4645
p01-f25-v3t1 0,4 4 4 3368 307 4575
p01-f25-v4t1 0,4 4 4 3368 299 4567
p01-l25-v1t1 0,4 8 5 3368 547 5040
p01-l25-v2t1 0,4 6 5 3368 472 4965
p01-l25-v3t1 0,4 3 3 3368 366 4409
p01-l25-v4t1 0,4 3 3 3368 354 4397
p03-f25-v1t1 0,2 10 5 3396 541 5062
p03-f25-v2t1 0,4 8 5 3368 474 4967
p03-f25-v3t1 0,4 4 4 3368 363 4631
p03-f25-v4t1 0,4 4 4 3368 361 4629
p03-m25-v1t1 0,4 9 5 3368 524 5017
p03-m25-v2t1 0,4 7 4 3368 421 4689
p03-m25-v3t1 0,4 6 4 3368 336 4604
p03-m25-v4t1 0,4 4 4 3368 368 4636
p03-l25-v1t1 0,4 7 6 3368 535 5253
p03-l25-v2t1 0,4 5 5 3368 491 4984
p03-l25-v3t1 0,4 4 4 3368 396 4664
p03-l25-v4t1 0,4 3 3 3368 395 4438
p07-f25-v1t1 0,2 7 4 3396 465 4761
p07-f25-v2t1 0,2 5 4 3396 412 4708
p07-f25-v3t1 0,4 4 3 3368 362 4405
p07-f25-v4t1 0,4 3 3 3368 354 4397
p07-s25-v1t1 0,4 8 7 3368 647 5590
p07-s25-v2t1 0,4 6 5 3368 513 5006
p07-s25-v3t1 0,4 4 4 3368 437 4705
p07-s25-v4t1 0,4 3 3 3368 397 4440
p07-t25-v1t1 0,2 8 5 3396 573 5094
p07-t25-v2t1 0,2 6 5 3396 457 4978
p07-t25-v3t1 0,4 4 4 3368 407 4675
p07-t25-v4t1 0,4 3 3 3368 357 4400
average 5.5 4.22 3372.38 429.84 4751.44

199
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Table 5.14: Solution details for LRSP

Data File Fac. # Fac. Route

Idx vec. Cost Cost TCost
p01-f25-v1t1 0,4 4 3368 431 4699
p01-f25-v2t1 0,4 3 3368 382 4425
p01-f25-v3t1 0,4 3 3368 307 4350
p01-f25-v4t1 0,4 3 3368 300 4343
p01-l25-v1t1 0,1 4 3460 539 4899
p01-l25-v2t1 0,4 4 3368 491 4759
p01-l25-v3t1 0,4 3 3368 366 4409
p01-l25-v4t1 0,4 3 3368 354 4397
p03-f25-v1t1 0,2 5 3396 541 5062
p03-f25-v2t1 0,4 4 3368 490 4758
p03-f25-v3t1 0,4 4 3368 363 4631
p03-f25-v4t1 0,4 3 3368 366 4409
p03-m25-v1t1 0,4 4 3368 517 4785
p03-m25-v2t1 0,4 4 3368 421 4689
p03-m25-v3t1 0,4 3 3368 361 4404
p03-m25-v4t1 0,4 3 3368 339 4382
p03-l25-v1t1 0,2 4 3396 546 4842
p03-l25-v2t1 0,4 4 3368 473 4741
p03-l25-v3t1 0,4 3 3368 408 4451
p03-l25-v4t1 0,4 3 3368 395 4438
p07-f25-v1t1 0,2 4 3396 465 4761
p07-f25-v2t1 0,2 4 3396 407 4703
p07-f25-v3t1 0,4 3 3368 362 4405
p07-f25-v4t1 0,4 3 3368 354 4397
p07-s25-v1t1 0,4 5 3368 605 5098
p07-s25-v2t1 0,4 4 3368 519 4787
p07-s25-v3t1 0,4 4 3368 427 4695
p07-s25-v4t1 0,4 3 3368 397 4440
p07-t25-v1t1 0,1 4 3460 538 4898
p07-t25-v2t1 0,2 4 3396 471 4767
p07-t25-v3t1 0,4 3 3368 407 4450
p07-t25-v4t1 0,4 3 3368 357 4400
average 3.59 3378.13 428.09 4614.81

200
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Figure 5.19: % Improvement in the total cost with LRSP compared to LRPDC + BPP

TCost in Tables 5.13 and 5.14. The total cost in the LRPDC with the BPP is generally larger than

the cost obtained by solving the LRSP, with eight instances, having equal costs. For each instance,

we calculated the % cost improvement obtained by solving an LRSP. Figure 5.19 shows the %

improvement calculated for each instance. On average, the total cost obtained by solving an LRSP

is 2.9% better than the cost obtained by solving the LRPDC and the BPP sequentially. Figure

5.19 also indicates the instances in which the set of open facilities in the LRPDC solution differs

from the set in the LRSP solution using bars in such cases. As it can be seen, in three instances,

the set of open facilities is different. This shows that integration of the scheduling decision may

effect the facility location decision. In Tables 5.13 and 5.14, we can observe that the routing costs

in the LRPDC and the LRSP may also differ. In eight instances, the routing cost in the LRPDC

is smaller than the routing cost in the LRSP solution. Since the weight of the routing cost is

small with respect to facility and vehicle fixed costs, the major cost difference between these two

problems is because of the difference in the number of vehicles.

Figure 5.20 compares the number of vehicles in the LRPDC solution before and after solving

the resulting BPP with the number of vehicles in the LRSP solution. In all but seven instances,

201
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Table 5.15: Average number of vehicles for each vehicle capacity value
Capacity average # of vehicles cost saving /year
LRPDC LRP+BPP LRSP Diff
LV1 8.25 5.13 4.25 0.88 51187
LV2 6.25 4.63 3.88 0.75 43875
L3V 4.13 3.75 3.25 0.5 29250
LV4 3.38 3.38 3 0.38 21937.5
Average 5.5 4.22 3.59 0.63 36562.5

the number of vehicles in the LRPDC solution before solving the resulting BPP is larger than the

number of vehicles in the LRSP solution. In all of the instances with vehicle capacity “v1” or

“v2”, the number of vehicles in the LRPDC is strictly larger than the number of vehicles in the

LRSP. As the vehicle capacity gets larger, the number of vehicles in the LRPDC gets closer to the

number of vehicles in the LRSP.

Table 5.15 compares and summarizes the average number of vehicles as well as the cost saving

for each vehicle capacity and for each model. On average, the number of vehicles in the LRPDC

is 5.5, while this average is 3.59 in the LRSP. Solving the BPP after the LRPDC reduces the

number of vehicles that are needed. The BPP is able to reduce the number of vehicles when

vehicle capacities are smaller (“v1” and “v2”). As the vehicle capacity gets larger, solving the

BPP does not improve the number of vehicles used. For instance, solving the BPP is not able to

reduce the number of vehicles in any of the instances that are labeled with “v4”, only in two of the

instances that are labeled with “v3”, the number of vehicles is improved. On average, the number

of vehicles after the BPP is 4.22, which is still larger than the average number of vehicles in the

LRSP. This difference results in about $142 vehicle cost difference per day with a vehicle fixed

cost of $225. Thus, on average, solving an LRSP results in a reduction of about $36600 in the

total vehicle fixed cost per year (based on 260 days). When we only consider instances with small

vehicle capacities such as instances that are labeled with “v1” and “v2”, on average, the difference

between the number of vehicles in the LRPS and the number of vehicles in the LRPDC and the

BPP is 0.82, which results in more than $47000 savings per year.

202
5.6. VALUE OF INTEGRATING SCHEDULING DECISION

Figure 5.20: # of vehicles with LRPDC, LRPD + BPP and LRSP

These experiments computationally demonstrate the difference between solving an LRSP and

an LRPDC. The results indicate the benefit of integrating scheduling decision with the location

and routing decisions. This action may even change the facility locations that are selected. The

number of vehicles in the LRSP and the LRPDC differ significantly, especially with small vehicle

capacities. Even if we solve a BPP after solving an LRPDC, the number of vehicles may be larger

than that in the LRSP solution. Thus, we can conclude that solving an LRSP is more preferable

than using an LRPDC model when the vehicle fixed cost are large. The instances that are used in

the experiments include only 25 customers. The number of vehicles used, as well as the total cost,

may be improved even more substantially in larger instances. The difference between the routing

costs are not significant. Therefore, in the absence of the vehicle fixed costs or in the absence of

a restriction on the number of vehicles, the difference between solving an LRPDC and an LRSP

may not be significant.

203
5.7. COMPARING INTEGRATED AND SEQUENTIAL OPTIMIZATION

5.7 Comparing Integrated and Sequential Optimization

In this section, we aim to empirically compare the integrated optimization (LRSP) and the opti-

mization of facility location, vehicle routing and vehicle scheduling decisions sequentially, which

we refer to as sequential optimization.

Models to be Compared. In sequential optimization, we first solve a capacitated facility loca-

tion problem in which each customer can only be assigned to a single facility (CFLPSS). We solve

the following optimization problem:

X XX
Min CjF tj + C O (Tji + Tij )vji
j∈J j∈J i∈I
X
s.t. vji = 1 ∀i ∈ I,
j∈J
X
Di vji − LFj tj ≤ 0 ∀j ∈ J,
i∈I
tj ∈ {0, 1}, vji ∈ {0, 1} ∀j ∈ J, i ∈ I,

where vji = 1 if customer i is assigned to facility j for all i ∈ I, j ∈ J. Then, we solve the

capacitated LRPDC problem in which only the facilities that are open in the solution of CFLPSS

are set to be open. Finally, as in the previous section, one BPP is solved for each open facility

to optimize the scheduling of vehicles in the solution to LRPDC. Comparing the solution to the

LRSP with the solution to the sequential optimization would evaluate the value of optimizing

facility location, vehicle routing and vehicle scheduling decisions interdependently.

Implementation Details and Discussion of the Results. In our experiments, we used the 32

25-customer instances that are labeled with “v1t1”, “v1t2”, “v2t1” and “v2t2”. For each approach

(LRSP and sequential optimization) and instance, we used the one-stage branch-and-price algo-

rithm with E-COLGEN(0,0,50,350) with a time limit of 6 CPU hours. In sequential optimization,

the branch-and-price was used to solve the LRPDC; CFLPSS and BPP were solved using CPLEX

204
5.7. COMPARING INTEGRATED AND SEQUENTIAL OPTIMIZATION

0.8

0.6

0.4

0.2

SEQ
LRSP
0
0 100 200 300 400 500 600

Figure 5.21: Performance profile comparing the execution time of LRSP and Sequential optimiza-
tion

9.1. Figure 5.21 compares the total solution time for each approach. Figure indicates that total

time in the sequential approach is much better than the total time required to solve the LRSP.

In Tables 5.13 and 5.14, we report the solution details for each instance and for each solution

approach, including indices of facilities that are open in the solution (Fac. Idx), number of vehicles

(# vec.), total facility fixed cost (Fac. Cost), total operating cost (Route Cost), and total cost

(TCost). The total cost obtained using the sequential approach is generally larger than the cost

obtained by solving the LRSP, with five out of 32 instances, having equal costs. For each instance,

we calculated the % cost improvement obtained by solving an LRSP and visualize these values in

Figure 5.22. On average, the total cost obtained by solving an LRSP is 3.8% better than the cost

obtained from the sequential optimization. Average total cost saving per year is about $47000.

Figure 5.22 also indicates the instances in which the sets of open facilities are different for two

solution approaches. In 18 out of 32 instances, the sets of open facilities are different. This shows

that the integrated optimization decision may effect the facility location decision. In Figure 5.23,

we compare the vehicle fixed costs and routing costs for the LRSP and the sequential optimization.

Figure 5.23(a) presents the % improvement in total vehicle fixed cost obtained by solving an LRSP

as opposed to applying the sequential optimization for each instance. On average, the integrated

optimization results in %21 improvement in the total vehicle fixed cost which is equal to $44000

205
5.7. COMPARING INTEGRATED AND SEQUENTIAL OPTIMIZATION

Figure 5.22: % Improvement in the total cost with LRSP compared to sequential optimization

(a) % Improvement in the vehicle fixed cost (b) % Improvement in the routing cost

Figure 5.23: LRSP versus sequential optimization

saving per year. Figure 5.23(b) presents the % improvement in total routing cost. In 18 out of

32 instances, sequential optimization approach results in a lower routing cost than the LRSP. On

average, the routing cost is %1.4 higher in the LRSP.

These experiments computationally demonstrate the value of solving an LRSP instead of op-

timizing each decision sequentially. In conclusion, solution to an LRSP may result in large cost

saving.

206
5.7. COMPARING INTEGRATED AND SEQUENTIAL OPTIMIZATION

Table 5.16: Solution details for Sequential Optimization

Data File Fac. # Fac. Route

Idx vec. Cost Cost TCost
p01-f25-v1t1 0,4 4 3368 447 4715
p01-f25-v1t2 0,4 4 3368 447 4715
p01-f25-v2t1 0,4 4 3368 377 4645
p01-f25-v2t2 0,4 4 3368 377 4645
p01-l25-v1t1 0,4 5 3368 547 5040
p01-l25-v1t2 0,4 4 3368 547 4815
p01-l25-v2t1 0,4 5 3368 472 4965
p01-l25-v2t2 0,4 5 3368 472 4965
p03-f25-v1t1 0,2 5 3396 541 5062
p03-f25-v1t2 0,2 4 3396 541 4837
p03-f25-v2t1 0,2 4 3396 463 4759
p03-f25-v2t2 0,2 4 3396 463 4759
p03-m25-v1t1 0,4 5 3368 524 5017
p03-m25-v1t2 0,4 4 3368 524 4792
p03-m25-v2t1 0,4 4 3368 421 4689
p03-m25-v2t2 0,4 4 3368 421 4689
p03-l25-v1t1 0,4 6 3368 535 5253
p03-l25-v1t2 0,4 4 3368 535 4803
p03-l25-v2t1 0,4 5 3368 491 4984
p03-l25-v2t2 0,4 4 3368 482 4750
p07-f25-v1t1 0,1 5 3460 457 5042
p07-f25-v1t2 0,1 4 3460 457 4817
p07-f25-v2t1 0,1 4 3460 398 4758
p07-f25-v2t2 0,1 4 3460 398 4758
p07-s25-v1t1 0,1 7 3460 657 5692
p07-s25-v1t2 0,1 6 3460 657 5467
p07-s25-v2t1 0,1 5 3460 475 5060
p07-s25-v2t2 0,1 4 3460 475 4835
p07-t25-v1t1 2,4 5 3534 524 5183
p07-t25-v1t2 2,4 4 3534 524 4958
p07-t25-v2t1 2,4 5 3534 428 5087
p07-t25-v2t2 2,4 4 3534 428 4862
Average 4.53 3415.25 484.53 4919.31

207
5.7. COMPARING INTEGRATED AND SEQUENTIAL OPTIMIZATION

Table 5.17: Solution details for LRSP

Data File Fac. # Fac. Route

Idx vec. Cost Cost TCost
p01-f25-v1t1 0,4 4 3368 431 4699
p01-f25-v1t2 0,4 3 3368 468 4511
p01-f25-v2t1 0,4 3 3368 382 4425
p01-f25-v2t2 0,4 3 3368 382 4425
p01-l25-v1t1 0,1 4 3460 539 4899
p01-l25-v1t2 0,4 4 3368 547 4815
p01-l25-v2t1 0,4 4 3368 491 4759
p01-l25-v2t2 0,1 3 3460 467 4602
p03-f25-v1t1 0,2 5 3396 541 5062
p03-f25-v1t2 0,2 4 3396 541 4837
p03-f25-v2t1 0,4 4 3368 490 4758
p03-f25-v2t2 0,4 4 3368 474 4742
p03-m25-v1t1 0,4 4 3368 517 4785
p03-m25-v1t2 0,4 4 3368 517 4785
p03-m25-v2t1 0,4 4 3368 421 4689
p03-m25-v2t2 0,4 3 3368 436 4479
p03-l25-v1t1 0,2 4 3396 546 4842
p03-l25-v1t2 0,4 4 3368 535 4803
p03-l25-v2t1 0,4 4 3368 473 4741
p03-l25-v2t2 2,4 3 3534 446 4655
p07-f25-v1t1 0,2 4 3396 465 4761
p07-f25-v1t2 2,4 3 3534 476 4685
p07-f25-v2t1 0,2 4 3396 407 4703
p07-f25-v2t2 0,2 3 3396 411 4482
p07-s25-v1t1 0,4 5 3368 605 5098
p07-s25-v1t2 0,4 4 3368 621 4889
p07-s25-v2t1 0,4 4 3368 519 4787
p07-s25-v2t2 0,4 4 3368 513 4781
p07-t25-v1t1 0,1 4 3460 538 4898
p07-t25-v1t2 0,2 4 3396 590 4886
p07-t25-v2t1 0,2 4 3396 471 4767
p07-t25-v2t2 0,1 3 3460 457 4592
Average 3.78 3396.88 491.16 4738.81

208
Chapter 6

Location and Routing Problems

In this chapter, we investigate the location and routing problem (LRP). Given a set of candidate

facility locations and a set of customer locations, the objective of the LRP is to determine the

number and location of facilities and construct a set of vehicle routes from facilities to customers

in such a way as to minimize total system cost. The system cost may include both the fixed and

operating costs of both facilities and vehicles. In this study, we consider an LRP with capacity

constraints on both the facilities and the vehicles.

As mentioned in Chapter 1, the LRP is a special case of the LRSP. To see this, consider an

LRSP instance in which there is one vehicle available at each facility and the time limit for the

vehicle is infinite. In addition, we set the fixed cost of a vehicle to zero, but the arc costs from

facilities to customers are modified by adding a vehicle fixed cost, if there is one. The solution of

the described LRSP instance gives a solution to the corresponding LRP instance.

Few papers have been devoted to combined location and routing models, and even fewer pa-

pers have been devoted to the study of exact algorithms for the LRP. The literature related to

the LRP was reviewed in Section 1.3.3. In this chapter, we modify the branch-and-price algo-

rithm that is developed for the LRSP to solve the LRP under capacity restrictions, and report on

our computational experience solving both randomly generated instances and instances from the

literature.

209
6.1. PROBLEM DEFINITION AND FORMULATIONS

The remainder of the chapter is organized as follows. Section 6.1 presents a formal description

of the problem and provides two formulations. Section 6.2 describes the heuristic and the exact

algorithms for the set partitioning formulation. Section 6.3 provides some computational results

evaluating the performance of the algorithms.

6.1 Problem Definition and Formulations

The objective of the LRP is to select a subset of facilities and construct an associated set of vehicle

routes serving customers at minimum total cost, where the cost includes the fixed and operating

costs of both facilities and vehicles. The constraints of the problem are as follows:

1. Each customer must be serviced by exactly one vehicle.

2. Each vehicle must be assigned to exactly one facility at which it must start and end its route.

3. The total demand of the customers assigned to a route must not exceed the vehicle capacity.

4. The total demand of the customers assigned to a facility must not exceed the capacity of the

facility.

In the literature, most of the exact methods developed for the described LRP or its special cases

are based on the two-index vehicle flow formulation of the problem. To the best of our knowledge,

an exact solution algorithm based on a set partitioning formulation has not been applied to the

case of the LRP with capacity constraints. Next, we present a set partitioning formulation of the

problem.

6.1.1 Set Partitioning Formulation

Here, we utilize a modified version of the set partitioning formulation for the LRSP (SPP-LRSP)

that is presented in Chapter 2. We define the set R to be a set indexing all vehicle routes feasible

with respect to vehicle capacity and originating from and returning to the same facility. We let

210
6.1. PROBLEM DEFINITION AND FORMULATIONS

Rj ⊆ R index routes associated with vehicles assigned to facility j for all j ∈ J. We use the

same notation that is presented in Section 2.1 to represent the facility fixed costs (C F ), the facility

capacities (C F ) and the demand (D).

The definitions of some parameters and variables that are used in (SPP-LRSP) change slightly

depending on the column index set R. We redefine the following parameters and variables for the

LRP formulation.

Parameters:

Cr = operating and fixed cost of route r ∈ R,



 1 if customer i ∈ I is assigned to route r ∈ R, and
air =

 0 otherwise.

Decision Variables:


 1 if route r ∈ R is selected, and
zr =

 0 otherwise,


 1 if facility j ∈ J is selected to be open, and
tj =

 0 otherwise,

The formulation referred to as (SPP-LRP) is as follows.

X X
Min CjF tj + Cr zr (6.1)
j∈J r∈R
X
(SPP-LRP) s. t. air zr = 1 ∀i ∈ I, (6.2)
r∈R
XX
Di air zr − CjF tj ≤ 0 ∀j ∈ J, (6.3)
r∈Rj i∈I
zr ∈ {0, 1} ∀r ∈ R, (6.4)

tj ∈ {0, 1} ∀j ∈ J. (6.5)

211
6.1. PROBLEM DEFINITION AND FORMULATIONS

The objective function (6.1) seeks to minimize the total cost, which includes the fixed cost of the

selected facilities and the fixed and operating costs of the vehicles. Constraints (6.2) guarantee

that each customer location is served by exactly one route. Constraints (6.3) ensure that the total

demand of the selected routes for a facility does not exceed the facility capacity. Constraints (6.4)

and (6.5) are standard binary restrictions on the variables.

6.1.2 Vehicle Flow Formulation

Here, we present a three-index vehicle flow formulation for the LRP similar to the model de-

veloped for the LRSP in Section 2.2.1 (G-LRSP). We use the same notation that is presented in

Section 2.1 and in addition, we define a variable uih that takes value 1 if vehicle h visits customer

i and 0, otherwise, ∀i ∈ I, h ∈ H. A vehicle flow formulation of the LRP that we refer to as

(G-LRP) is as follows:

X X X
Min CjF tj + C O Tik xikh (6.6)
j∈J h∈H (i,k)∈A
XX
(G-LRP) s.t. xikh = 1 ∀i ∈ I, (6.7)
h∈H k∈N
X X
xikh − xkih = 0 ∀i ∈ N, h ∈ H, (6.8)
k∈N k∈N
X
xkih − uih = 0 ∀i ∈ I, h ∈ H, (6.9)
k∈N
X X
uih − xklh ≤ 0 ∀S ⊆ I, i ∈ S, h ∈ H, (6.10)
k∈S l∈V \S
X
Di uih − C V ≤ 0 ∀h ∈ H, (6.11)
i∈I
X X
Di uih − CjF tj ≤ 0 ∀j ∈ J, (6.12)
h∈Hj i∈I
xikh ∈ {0, 1} ∀i, k ∈ V, h ∈ H, (6.13)

uih ∈ {0, 1} ∀i ∈ I, h ∈ H, (6.14)

tj ∈ {0, 1} ∀j ∈ J. (6.15)

212
6.2. SOLUTION ALGORITHM

In (G-LRP), the objective function (6.6) seeks to minimize the total cost, which includes the

fixed cost of the selected facilities and the operating cost of the vehicles. Vehicle fixed costs

can easily be incorporated into the model by increasing the cost of traveling from the facility to

each customer location by a fixed amount. Constraints (6.7) specify that exactly one vehicle must

service customer i. Constraints (6.8) require that each vehicle should enter and leave a location

the same number of times. Constraints (6.9) determine the assignment of customers to vehicles.

Constraints (6.10) eliminate subtours, i.e., routes that do not include a facility. Constraints (6.11)

are vehicle capacity constraints. Constraints (6.12) ensure that the capacity of a facility is not

exceeded by demand flows to customer locations. For notational convenience, we assume that

variables associated with travel between different facilities or travel between a customer and a

facility using a truck not associated with that facility are fixed to zero. Constraints (6.13), (6.14),

and (6.15) are the set of binary restrictions on the variables.

We can demonstrate the relationship between (G-LRP) and (SPP-LRP) by applying Dantzig-

Wolfe decomposition (DWD) to (G-LRP). This approach will yield a model that is equivalent to

(SPP-LRP). In the application of the DWD method, constraints (6.7) and (6.12) form the master

problem, while the remaining constraints define the subproblem. This subproblem can then be

decomposed by facility and further by vehicle. The equivalence of (SPP-LRP) and the DWD of

(G-LRP) follows after a further reduction in the subproblem column set based on symmetry and

constraints in the master problem. A similar study has already been done for the LRSP in Section

2.4 and deriving similar results for the LRP is straightforward. Therefore, we skip the details here.

6.2 Solution Algorithm

Comparing (SPP-LRP) and (SPP-LRSP), we see that the only difference between them is the

definition of the column set. The columns of (SPP-LRSP) correspond to the set of feasible pairings

for all facilities. Since the scheduling decision is not considered in an LRP, (SPP-LRP) is defined

over the set of feasible routes for all facilities (denoted by R). Based on this similarity, we can

213
6.2. SOLUTION ALGORITHM

easily modify our solution algorithms for the LRSP to solve the LRP. We alter our branch-and-

price implementation in order to utilize formulation (SPP-LRP) and generate routes instead of

pairings in the column generation part of the algorithm.

6.2.1 Strengthening the Set Partitioning-Based Formulation

As in the LRSP case (described in Section 2.3), we first strengthen (SPP-LRP) by adding the

following additional valid inequalities:

X
air zr − tj ≤ 0 ∀i ∈ I, j ∈ J, (6.16)
r∈Rj
X
tj ≥ N F , (6.17)
j∈J
X
zr = vj ∀j ∈ J, (6.18)
r∈Rj
X »P ¼
i∈I Di
vj ≥ , (6.19)
CV
j∈J
v j ≥ tj ∀j ∈ J, (6.20)

vj ∈ Z+ ∀j ∈ J, (6.21)

where vj represents the number of vehicles used at facility j, ∀j ∈ J and N F is a lower bound on

the number of facilities that must be open in any feasible solution, calculated as follows:

à l
!
X X
F
N = argmin{l=1,...,|J|} CjFt ≥ Di s.t. CjF1 ≥ CjF2 ≥ ... ≥ CjFn .
t=1 i∈I

Constraints (6.16) force a facility to be open if any customer is assigned to it. Constraint (6.17)

sets a lower bound on the total number of facilities required in any integer feasible solution. In the

LRSP case, constraints (6.16) and constraint (6.17) are shown computationally to improve the LP

relaxation of the model. As in the LRSP, constraints (6.18) are only added to facilitate branching

on the integrality of the number of vehicles at each facility in the branch-and-price algorithm.

214
6.2. SOLUTION ALGORITHM

Constraint (6.19) sets a lower bound for the total number of vehicles in the solution. Finally,

constraints (6.20) force the number of vehicles used at a facility to be at least 1 if the facility is

open. We refer to formulation (6.1)–(6.5) and (6.16)–(6.21) as (ASPP-LRP) for the rest of the

chapter.

6.2.2 Branch-and-Price Algorithm

Since we are modifying the branch-and-price algorithm used for the LRSP, which is described in

detail in Chapter 3, we only explain here the modified parts of the algorithm. The main struc-

ture of the algorithm BRANCH-PRICE, the COLGEN procedure, the branching rules, and the

versions of the branch-and-price algorithms designed are the same as for the LRSP case. The

formulation (ASPP-LRP) contains a variable for each possible vehicle route originating from each

facility. Therefore, we alter the solution algorithms for the pricing problem for the LRSP in order

to generate feasible routes instead of pairings.

Pricing Problem. The pricing problem (for each facility) for the LRP is also an ESPPRC. Con-

sider a network with |I| + 2 nodes, one node for each customer, one for the facility j as a source

node and a copy of the facility j as a sink node. We assign each customer node a demand equal to

its demand in the original problem and let the cost of each arc (i, l) in the network be equal to the

contribution of arc (i, l) to the reduced cost of a column that represents a route including arc (i, l).

The reduced cost function for the RMP in the LRP is similar to the reduced cost function in the

LRSP case. A shortest path from source to sink visiting a customer at most once and satisfying

the constraint that the total demand of customers included in the path does not exceed the vehicle

capacity then corresponds to a vehicle route. The total cost of the path plus the fixed costs (if any)

is the reduced cost of the associated column.

To solve the ESPPRC, we use Feillet et al. (2004)’s label setting algorithm with a single

resource (the vehicle capacity) which is described in Section 3.2.2 and referred to as 1p-ESPPRC.

Note that the algorithm described in Section 3.2.2 is simplified by removing the time limit from the

215
6.2. SOLUTION ALGORITHM

resource set in order to use it for the LRP, however, we still use notation 1p-ESPPRC to denote the

exact pricing algorithm used in this chapter. To enhance the efficiency of the overall branch-and-

price algorithm, we use the two heuristic extensions of the exact pricing algorithm: ESPPRC-LL

and ESPPRC-CS presented in Section 3.2.3. We refer to these algorithms as 1p-ESPPRC-LL(l)

and 1p-ESPPRC-CS(u).

As in the LRSP case, in the branch-and-price algorithm, we apply pricing algorithms 1p-

ESPPRC, 1p-ESPPRC-LL(l) and 1p-ESPPRC-CS(u) in combination in the column generation

algorithms to improve the overall performance. Two designs of the column generation algorithm,

E-COLGEN(u, U , l, L) (Figure 3.4) and H-COLGEN(u, U , l, L) (Figure 3.5), can be modified

for the LRP. In the modified versions, we call 1p-ESPPRC, 1p-ESPPRC-LL(l) and 1p-ESPPRC-

CS(u) instead of 2p-ESPPRC, 2p-ESPPRC-LL(l) and 2p-ESPPRC-CS(u) in combination. In this

chapter, we still use representations E-COLGEN(u, U , l, L) and H-COLGEN(u, U , l, L) to denote

the exact and heuristic column generation algorithms for the LRP.

In addition to the exact and heuristic approaches to solve the pricing problem, we designed

an algorithm that solves a relaxation of the pricing problem. We refer to this algorithm as 2-Cyc-

SPPRC-PE(C) where C is a subset of customers.

2-Cyc-SPPRC-PE(C). Recall that the shortest path problem with resource constraints (SPPRC)

is a relaxation of the ESPPRC in which the path may visit some customers more than once. The

SPPRC is solvable in pseudo-polynomial time, but use of this further relaxation of the column gen-

eration subproblem results in reduced bounds. Eliminating solutions containing cycles of length

two strengthens this relaxation of the pricing problem and improves the bound (for details of the

algorithm, see Irnich and Desaulniers (2005)). We refer to this pricing algorithm as 2-Cyc-SPPRC.

In addition, we can also generate paths that are elementary with respect to a given subset of cus-

tomers. We call the resulting algorithm 2-Cyc-SPPRC-PE(C), where C is the set of customers

that can be visited at most once in the resulting solutions.

216
6.2. SOLUTION ALGORITHM

R-COLGEN(m)
Input: An LRP instance, a subset of columns CG and an integer m (< |I|).
Output: A lower bound for the LP relaxation of the problem at the node and a
column set modified by adding new columns.

1. Do P → 2-Cyc-SPPRC.

2. If P = ∅, STOP. Otherwise, consider the column with minimum cost in P and

choose at most m customers that are visited more than once. Let C1 be the set
of these customers. If C1 = ∅ (the path is already elementary), then add the
column to set CG and return to step 1. Otherwise, proceed.

3. Do P → 2-Cyc-SPPRC-PE(C1 ).

4. If P = ∅, STOP. Otherwise, consider the column with minimum cost in P and

choose at most m customers that are visited more than once. Let C2 be the set
of these customers. If C2 = ∅, then add the column to set CG and return to
step 3. Otherwise, set C2 → C1 ∪ C2 .

5. Do P → 2-Cyc-SPPRC-PE(C2 ) until P = ∅.

Figure 6.1: The main steps of R-COLGEN(m)

Column Generation With the Relaxations of the Pricing Problem. We design a column gen-

eration algorithm that uses 2-Cyc-SPPRC and 2-Cyc-SPPRC-PE(C) in combination. The algo-

rithm is referred to as R-COLGEN(m) and described in Figure 6.1.

Variations of the Branch-and-Price Algorithm. For computational testing, we implemented

four variants of the branch-and-price algorithm based on the above pricing schemes. The first

three designs are also applied for the LRSP case.

• Heuristic Branch-and-Price (HBP): The purpose of this algorithm is to provide a good upper

bound. At each node of the tree, we use H-COLGEN(u, U , l, L) with small values of u, l

and L.

• One-Stage Exact Branch-and-Price (1S-EBP): The purpose of this algorithm is to prove the

optimality of the solution or provide an integrality gap. At each node of the tree, we use

E-COLGEN(u, U , l, L).
217
6.3. COMPUTATIONAL RESULTS

• Two-Stage Exact Branch-and-Price (2S-EBP): In this variant, HBP is run first to generate

initial columns and an upper bound. Then, in the second stage an exact branch-and-price

algorithm which is initiated with the columns and the upper bound obtained from HBP is

run using E-COLGEN(u, U , l, L).

• Non-elementary Exact Branch-and-Price (NEBP): This is similar to one-stage branch-and-

price algorithm except that R-COLGEN(m) is used to solve the pricing problem.

6.3 Computational Results

In this section, we discuss the performance of our branch-and-price algorithms for the LRP on

two sets of instances. The first set of instances are random instances that we generated to test the

performance of our 1S-EBP and 2S-EBP algorithms. We evaluate the effect of facility capacity

constraints and other parameters using this set of instances. The second set contains the LRP

instances used in Barreto et al. (2007) that are available from Barreto (2003). We used these

instances to test the performance of our HBP and NEBP algorithms. For all of our experiments,

we used a Linux-based workstation with a 1.8 GHz processor and 2GB RAM.

6.3.1 Generated Instances

On random instances, Laporte et al. (1986) provided computational results for an exact method

(branch-and-cut algorithm) for the capacitated LRP. Belenguer et al. (2006) also developed a

branch-and-cut algorithm for the capacitated LRP, but neither the details of the instances nor the

computational results are publicly available. Therefore, we evaluated our algorithm by gener-

ating random instances as in Laporte et al. (1986), where they generated instances with 10, 15

and 20 customers and 4 to 8 facilities. In addition to these, we generated instances with 30 and

40 customers and 5 facilities to test the performance of the algorithm on larger instances. The

coordinates of the customers and the facilities and the demand of each customer were generated

using a uniform distribution on [0,100]. We then calculated the Euclidean distance between each

218
6.3. COMPUTATIONAL RESULTS

Table 6.1: Details for the instances

# of Customers # of Facilities NF MF # of instances α

10 4 1 3 3 {0, 0.25, 0.5, 0.75, 1}
15 6 1 4 3 {0, 0.25, 0.5, 0.75, 1}
20 8 1 5 3 {0, 0.25, 0.5, 0.75, 1}

pair of customers and between customers and facilities and rounded the calculated distance to two

decimal places. Demand for each customer was rounded to the nearest integer. Vehicle capacity

C V was calculated as
X
C V = (1 − α)maxi∈I {Di } + α Di , (6.22)
i∈I

where α was a parameter and the values were chosen in set {0, 0.25, 0.5, 0.75, 1}.

Small and Medium Random Instances

Laporte et al. (1986) solved location and routing problems with capacitated vehicles, but they

did not have a facility capacity. Instead, they had a lower and upper bound for the total number

of facilities that could be open in a solution. In this experiment, in order to provide a better

comparison of our algorithm with that of Laporte et al. (1986), we removed constraint (6.3) from

(SPP-LRP). We set N F , the minimum number of open facilities in (6.17), to 1, and we added
P
constraint j∈J tj ≤ M F , where M F is the maximum number of facilities that can be open in

any solution. Facility and vehicle fixed costs were set to be zero. As in Laporte et al. (1986),

three groups of instances with different numbers of customers and facilities were available. For

each group, five different vehicle capacities were calculated by changing α. Some details about

the instances are listed in Table 6.1.

Tables 6.2, 6.3 and 6.4 present the results achieved with our branch-and-price algorithms. In-

stances listed in these tables are labeled with the number of customers and facilities and with

letters {a,b,c,d,e} based on the α value used. For example, the instance r10x4-a-1 has 10 cus-

tomers, 4 facilities and a vehicle capacity calculated with α = 0. The integer from 1 to 3 used

219
6.3. COMPUTATIONAL RESULTS

as the last character in the label represents the id of the instances within the same group. The ta-

bles present the name of each instance, the LP solution value at the root node, the optimal solution

value, the number of evaluated nodes, and the CPU time in seconds. In these instances, we first ran

the 1S-EBP algorithm for 5 minutes and if the algorithm did not terminate within 5 minutes, we

switched to the 2S-EBP algorithm. Table 6.2 presents the results for instances with 10 customers,

Table 6.3 presents the results for instances with 15 customers, and Table 6.4 presents the results

for instances with 20 customers. We marked the instances with a “+” sign if the 2S-EBP algorithm

was used. For instances with at least 20 customers, we needed to use the 2S-EBP algorithm. An

exact branch-and-price algorithm (one-stage or two-stage) was very successful in finding an op-

timal solution quickly. In general, our computation times were much smaller than those reported

by Laporte et al. (1986), but it is difficult to make a fair comparison, given advances in computing

technology. In most of the instances, the LP solution found by our algorithm at the root node was

an optimal integer solution.

The instances become more difficult when the vehicle capacity is increased (α increases) be-

cause the number of labels generated in the pricing problem depends directly on the vehicle capac-

ity. Laporte et al. (1986) observed the reverse effect with regard to their branch-and-cut algorithm.

The number of cuts generated increases and the problem gets more difficult when the vehicle ca-

pacity is small. This is most likely due to the fact that the problem structure becomes more like

that of the traveling salesman problem (TSP) as the capacity is increased and the TSP is much

easier to solve by branch and cut than as a capacitated routing problem.

To strictly differentiate the instances from those with a single depot, we experimented with

changing the value of parameter N F , the minimum number of open facilities, to 2 and ran the

r15x6 and r20x8 c and d instances. There were no significant changes in the computational times

or the number of evaluated nodes. We present these results in Table 6.5.

We then added facility capacities to the same set of problems and ran our algorithm again.

Table 6.6 presents the computational results, as well as the facility capacity values used for the

facilities. The capacity value was chosen in order to require at least two open facilities. The

220
6.3. COMPUTATIONAL RESULTS

Table 6.2: Performance of Exact Branch-and-Price for 10-customer instances

Instance LP OPT. # of Nodes CPU(s)

r10x4-a-1 472.11 472.11 1 0.00
r10x4-a-2 421.44 421.44 1 0.00
r10x4-a-3 548.28 548.28 1 0.02
r10x4-b-1 313.01 313.18 3 0.04
r10x4-b-2 297.57 305.27 19 0.08
r10x4-b-3 352.66 354.92 3 0.03
r10x4-c-1 257.25 257.25 1 0.06
r10x4-c-2 259.76 259.76 1 0.04
r10x4-c-3 296.82 296.82 1 0.05
r10x4-d-1 243.42 257.25 21 0.52
r10x4-d-2 250.04 250.04 1 0.04
r10x4-d-3 296.82 296.82 1 0.04
r10x4-e-1 226.46 226.46 1 0.32
r10x4-e-2 225.82 225.82 1 0.17
r10x4-e-3 272.85 272.85 1 0.31

computational results do not show any significant difference from those of the uncapacitated in-

stances. For larger instances, we expect that the LP solution times will tend to increase if the

master problem has facility capacity constraints. Adding a facility capacity to a two-index vehi-

cle flow formulation requires an additional set of constraints (Belenguer et al., 2006), the size of

which can be large. In a branch-and-cut algorithm, it may require additional time to generate this

set of constraints.

Laporte et al. (1986) report that adding facility fixed costs to the problem makes the problem

easier. We added facility fixed costs ($200 for each facility) to r15x6 and r20x8 c and d instances.

The performance of the branch-and-price algorithm was not affected in instances with 15 cus-

tomers. However, for some of the 20-customer instances, the computational times exceeded 1

CPU hour, which is more than the times in the experiments without the fixed cost. Results of this

experiment is reported in Table 6.7.

221
6.3. COMPUTATIONAL RESULTS

Table 6.3: Performance of Exact Branch-and-Price for 15-customer instances

Instance LP OPT. # of Nodes CPU(s)

r15x6-a-1 435.2 435.2 1 0.01
r15x6-a-2 663.32 663.32 1 0.01
r15x6-a-3 411.45 411.45 1 0.01
r15x6-b-1 313.46 313.46 1 0.31
r15x6-b-2 414.65 414.65 1 0.21
r15x6-b-3 285.01 285.01 1 0.12
r15x6-c-1 313.46 313.46 1 3.1
r15x6-c-2 392.75 392.75 1 1.98
r15x6-c-3 279.82 279.82 1 5.37
r15x6-d-1 313.36 313.46 3 4.92
r15x6-d-2 378.76 378.76 1 9.41
r15x6-d-3 279.82 279.82 1 14.13
r15x6-e-1 305.86 312.18 5 9.82
r15x6-e-2+ 374.86 374.86 1 16.62+
r15x6-e-3 274.22 274.22 1 300.01
+ The 2S-EBP algorithm was used, total time is reported.

Large Random Instances

In this section, we present the results of our exact branch-and-price algorithms applied to solve

larger capacitated random instances. We generated 6 instances with 30 customers and 6 instances

with 40 customers. Each instance had 5 facilities with capacity constraints. The facilities had a

fixed cost of 100. The characteristics of each instance are listed in Table 6.8. The first column

includes the name of each instance which was set based on the number of customers, facilities

and the vehicle capacity. For instances in group “a” the vehicle capacity value (listed in the

fourth column) was chosen to be seven times the average demand and for group “b”, the vehicle

capacity value was 5.5 times the average demand. Facility capacity (listed in the second column)

was chosen based on the total demand such that at least two facilities (the minimum number of

facilities is listed in the third column) should be open in a feasible solution.

We used the 2S-EBP algorithm in which each stage was run with a time limit of 3 CPU hours.

Table 6.9 presents the results of both stages. The algorithm was very successful in finding optimal

or near-optimal solutions for these larger instances. Some details, such as the number of open

222
6.3. COMPUTATIONAL RESULTS

Table 6.4: Performance of Exact Branch-and-Price for 20-customer instances

Instance LP OPT. # of Nodes CPU(s)

r20x8-a-1 639.77 653.11 57 0.41
r20x8-a-2 542.23 551.58 25 0.98
r20x8-a-3 760.42 760.42 1 0.05
r20x8-b-1 415.3 417.13 7 8.35
r20x8-b-2 383.19 383.19 1 6.32
r20x8-b-3 439.18 447.72 121 37.32
r20x8-c-1+ 398.34 398.34 1 26.25+
r20x8-c-2+ 363.86 363.86 1 110.87+
r20x8-c-3+ 402.85 402.85 1 28.74+
r20x8-d-1+ 392.26 392.26 1 10.04+
r20x8-d-2 + 359.49 359.49 1 200.29+
r20x8-d-3 + 402.85 402.85 1 124.24+
r20x8-e-1+ 392.28 392.28 1 12.39+
r20x8-e-2+ 355.39 355.39 1 82.77+
r20x8-e-3+ 402.46 402.46 1 103.27+
+ The 2S-EBP algorithm is used, total time is reported.

facilities, the number of vehicles used at each open facility, the average number of customers in

each route (vehicle), and the number of customers in the longest route, are presented in Table 6.8.

6.3.2 Instances From the Literature

To the best of our knowledge, there are no benchmark instances available specifically for the LRP.

Barreto et al. (2007) used the instances in the literature available for other types of problems to

construct a set of LRP instances. They report lower bounds found by applying a branch-and-cut

algorithm to the two-index formulation of the problem (Barreto, 2004) and upper bounds found by

applying a sequential heuristic based on clustering techniques (Barreto et al., 2007). They listed

19 instances, three of which have more than 150 customers, too large for our approach to work

efficiently. We removed these three instances plus one more with 117 customers and fractional

demand, since we assume integer demands. The labels of the instances denote the source of the

instance and the number of customers and facilities in the instances (for more details about the

223
6.3. COMPUTATIONAL RESULTS

Table 6.5: Instances with N F ≥ 2

Instance LP OPT. # of Nodes Total Time
r15x6-c-1 313.46 313.46 1 1.51
r15x6-c-2 392.75 392.75 1 1.28
r15x6-c-3 279.82 279.82 1 4.02
r15x6-d-1 313.36 313.46 3 2.79
r15x6-d-2 378.76 378.76 1 9.45
r15x6-d-3 279.82 279.82 1 19.16
r20x8-c-1 398.34 398.34 1 27.86
r20x8-c-2 363.86 363.86 1 71.9
r20x8-c-3 402.85 402.85 1 19.52
r20x8-d-1 392.26 392.26 1 25.56
r20x8-d-2 359.49 359.49 1 254.83
r20x8-d-3 402.85 402.85 1 124.6

references, see Barreto (2003)).

We first ran HBP with a time limit of 3 CPU hours, focusing on producing quality upper

bounds. Table 6.10 presents the instances we tested and compares the results with the upper

bounds reported in Barreto et al. (2007). Since neither our HBP nor Barreto et al. (2007) can

provide a valid lower bound for the problem, we used the best lower bounds as found in Barreto

(2004) (second column in Table 6.10) to measure the quality of our upper bound. The “Gap” in

Table 6.10 is the percent gap between the upper bound and the lower bound listed in the second

column. HBP is capable of finding better upper bounds (usually optimal) for the instances of

small and medium size. In these cases, the computation time is also very short. However, for

larger instances (instances with more than 75 customers), the upper bounds reported by Barreto

et al. (2007) are generally better. In addition, their heuristic algorithm is very efficient—they

report that in most of the instances, it provides the result in less than one second.

Next, we used the 2S-EBP algorithm to test the ability to produce lower bounds and prove

optimality. With this algorithm, we could not solve all instances within the total time limit of 5

CPU hours (3 CPU hours for HBP and 2 CPU hours for 1S-EBP). We report the results of five

instances that we could solve. The lower bounds found by our algorithm, along with the best

integer solution found, optimality gap, and computation time are reported in Table 6.11. Note that

224
6.3. COMPUTATIONAL RESULTS

Table 6.6: Instances with Capacitated Facilities

Instance Fac. Cap LP OPT. # of Nodes Total Time

cr15x6-c-1 600 299.07 299.07 1 1.15
cr15x6-c-2 600 392.75 392.75 1 2.53
cr15x6-c-3 600 279.82 279.82 1 1.86
cr15x6-d-1 700 299.07 299.07 1 0.57
cr15x6-d-2 700 378.76 378.76 1 6.42
cr15x6-d-3 700 279.82 279.82 1 11.01
cr20x8-c-1 750 398.34 398.34 1 40.46
cr20x8-c-2 750 363.86 363.86 1 88.12
cr20x8-c-3 750 402.85 402.85 1 15.25
cr20x8-d-1 900 392.29 392.29 1 29.1
cr20x8-d-2 750 359.49 359.49 1 257.94
cr20x8-d-3 900 402.85 402.85 1 31.3

because Gaskell67-21x5 and Perl83-12x2 are very easy to solve, we used the 1S-EBP algorithm

instead of the 2S-EBP algorithm in these cases.

Finally, for instances that we could not provide lower bounds by using the 1S-EBP or the

2S-EBP algorithm, we used the NEBP algorithm with a time limit of 5 CPU hours or evaluated

node limit of 50 nodes. The results are reported in Table 6.12. For the instances C69-100x10,

Min92-134x8 and Dsk95-88x8, we could not even solve the root node within the time limit.

In general, the lower bounds found by using branch-and-cut (Barreto, 2004) are better than

our lower bounds. However, in some cases, their computation times are much larger than ours.

The HBP algorithm can provide good upper bounds, but for instances of medium and large size,

we need to improve our methods of computing lower bound, perhaps by adding dynamic cut

generation to our algorithm in order to close the optimality gaps.

225
6.3. COMPUTATIONAL RESULTS

Table 6.7: Instances with facility fixed cost

Instance Best IP %Gap # of Nodes Total Time

r15x6-c-1 612.82 0 1 3.33
r15x6-c-2 632.96 0 1 4.53
r15x6-c-3 527.91 0 1 6.9
r15x6-d-1 592.26 0 1 5.47
r15x6-d-2 612.18 0 21 43.67
r15x6-d-3 504.44 0 1 23.88
r20x8-c-1 655.9 0 29 1764.31+
r20x8-c-2 624.8 0 51 2536.09+
r20x8-c-3 666.59 1.42 108 5406.42+
r20x8-d-1 640.14 1.65 84 4700+
r20x8-d-2 605 0.87 25 4253.48+
r20x8-d-3 641.35 0 75 1720.78+
+ The 2S-EBP algorithm is used, total time is reported.

Table 6.8: Characteristics of the Instances and the Optimal Solutions

Instance OPT/BestIP Solution Info

Name Fac. N F LV # of # of Avg. # of # of Cust.
Cap. Fac. Vec. Cust/route longest route
cr30x5a-1 1000 2 350 2 3,2 6 8
cr30x5a-2 1000 2 350 2 3,2 6 7
cr30x5a-3 1000 2 350 2 3,3 5 7
cr30x5b-1 1000 2 275 2 3,2 6 8
cr30x5b-2 1000 2 275 2 4,3 4.29 7
cr30x5b-3 1000 2 275 2 3,4 4.29 6
cr40x5a-1 1750 2 340 2 3,3 6.67 8
cr40x5a-2 1750 2 390 2 3,4 5.71 8
cr40x5a-3 1750 2 370 2 3,3 6.67 7
cr40x5b-1 1750 2 275 2 3,5 5 7
cr40x5b-2 1750 2 275 2 3,5 5 7
cr40x5b-3 1750 2 325 2 5,3 5 7

226
6.3. COMPUTATIONAL RESULTS

Table 6.9: Performance of the 2S-EBP for Instances with 30 and 40 Customers, Capacitated Fa-
cilities and Facility Fixed Cost

Instance HBP Second Stage, Exact Branch and Price Total

IP CPU(s) LP IP Gap #N. CPU(s) CPU (s)
cr30x5a-1 819.53 43.5 810.29 819.5 0 33 993.3 1036.8
cr30x5a-2 823.49 7202.3 790.49 823.49 2.55 500 10806.5 18008.8
cr30x5a-3 702.29 44.2 687.72 702.19 0 51 917.9 962.1
cr30x5b-1 880.02 164.3 865.47 880.01 0 251 6420.6 6585
cr30x5b-2 825.32 8.3 815.95 825.3 0 7 33.2 41.5
cr30x5b-3 884.62 13.4 881.33 884.55 0 19 41.7 55.1
cr40x5a-1 928.11 631.8 911.39 928.11 1.49 11 10882.8 11514.6
cr40x5a-2 888.37 378.4 871.66 888.37 0.93 13 11052.9 11431.3
cr40x5a-3 947.24 173 939.54 947.24 0.18 28 10862 11035
cr40x5b-1 1052.07 257.3 1043.62 1052 0 627 8084.6 8342
cr40x5b-2 981.52 60.1 976.88 981.27 0 47 862.5 922.7
cr40x5b-3 964.32 62.6 959.05 964.23 0 45 963 1025.6

Table 6.10: Performance of Heuristic Branch-and-Price

Instance LB1 Barreto et al. (2007) Heuristic Branch-and-Price

UB2 %Gap UB Gap CPU(s)
Gaskell67-21x5 424.9 ∗ 435.9 2.59 424.9 0 1.2
Gaskell67-22x5 585.1∗ 591.5 1.09 585.1 0 41.5
Gaskell67-29x5 512.1∗ 512.1 0 512.1 0 67.7
Gaskell67-32x5 556.5 571.7 2.73 562.3 1.04 10801.8
Gaskell67-32x5-2 504.3∗ 511.4 1.41 505.8 0.3 85.6
Gaskell67-36x5 460.4 ∗ 470.7 2.24 460.4 0 1077.6
C69-50x5 549.4 582.7 6.06 565.6 2.95 239.4
C69-75x10 744.7 886.3 19.01 852.1 14.42 10802.3
C69-100x10 788.6 889.4 12.78 929.5 17.86 10836.6
Perl83-12x2 204∗ 204 0 204.0 0 0.2
Perl83-55x15 1074.8 1136.2 5.71 1121.8 4.37 10800.0
Perl83-85x7 1568.1 1656.9 5.66 1668.2 6.38 10813.8
Min92-27x5 3062 ∗ 3062 0 3062.0 0 4.2
Min92-134x8 - 6238 - 6421.6 - 10850.9
Dsk95-88x8 356.4 384.9 8 390.6 9.58 10808.9
1 Reported by Barreto (2004), found using branch-and-cut
2 Reported by Barreto et al. (2007), found using a heuristic
∗ Branch-and-cut used in Barreto (2004) proves the optimality

227
6.3. COMPUTATIONAL RESULTS

Table 6.11: Performance of Two-Stage Exact Branch-and-Price

Instance LB OPT/BestIP %Gap CPU(s) Total CPU(s)

Gaskell67-21x5 1 424.9 424.9 0 3.0 3.0
Gaskell67-22x5 585.1 585.1 0 2999.9 3041.4
Gaskell67-32x5 544.1 562.3 3.24 8453.1 19254.9
Perl83-12x21 204.0 204.0 0 0.2 0.2
Min92-27x5 3062.0 3062.0 0 833.6 837.8
1 1S-EBP algorithm is used

Table 6.12: Performance of Non-Elementary Exact Branch-and-Price

Instance LB OPT/BestIP %Gap CPU(s)

Gaskell67-29x5 441.2 512.1 13.85 14411.5
Gaskell67-32x5-2 494.4 505.8 2.26 5654.8
Gaskell67-36x5 455.5 460.4 1.05 100.1
C69-50x5 526.2 565.6 6.96 2634.1
C69-75x10 693.5 852.1 18.62 8785.8
Perl83-55x15 852.3 1121.8 24.02 319.5
Perl83-85x7 1272.4 1668.2 23.73 1379.6

228
Chapter 7

Conclusions and Future Directions

In this research, we focused on solving the integrated facility location, vehicle routing and vehicle

scheduling problem (LRSP). In jointly optimizing the decisions of facility location, vehicle rout-

ing, and route assignment, the LRSP provides solutions that more accurately capture the practical

aspects and costs of designing logistics systems. We considered the interdependency between the

facility location decision which is a strategic level decision, and route construction and route as-

signment decisions which are operational level decisions. This results in a better estimate of the

total cost that includes the total vehicle fixed and operating costs, which become crucial in the

long term.

We developed mathematical formulations for the LRSP and showed empirically that the set

partitioning-based formulation resulted in stronger lower bounds. This is one of the motivations

for us to develop the solution algorithm based on the set partitioning-based formulation. We could

achieve further reduction in the LP relaxation gap by introducing valid inequalities at the root

node. We developed a branch-and-price algorithm with several variants: one-stage exact branch-

and-price, two-stage exact branch-and-price and heuristic branch-and-price algorithms.

We computationally evaluated the performance of the branch-and-price algorithm and tried to

decide some strategies and parameter values in order to improve the overall performance of the

algorithm. As in general, the efficiency of the pricing algorithm significantly affects the overall

229
performance. We extended our pricing algorithm to a two-phase pricing algorithm and achieved

significant improvement (about 90% time reduction at the root node) in the efficiency. In addition,

we applied the strategies of adding multiple columns in any iteration of the column generation

algorithm and including the heuristic versions of the pricing algorithm and improved the perfor-

mance of the branch-and-price algorithm further. Furthermore, we presented computational results

for instances including 25 and 40 customers. The algorithm found optimal solutions for the ma-

jority of the instances and obtained good solutions with small integrality gaps for the rest. Finally,

we empirically demonstrated that two-stage exact branch-and-price algorithm was preferred to

one-stage exact branch-and-price algorithm while optimizing larger size instances, and the heuris-

tic branch-and-price algorithm was able to provide very strong upper bounds within reasonable

times.

We studied the extension of the branch-and-price algorithm to a branch-cut-and-price algo-

rithm using dynamic cut generation. We investigated some classes of valid inequalities that can be

used for the LRSP. Even though the cut generation procedures were fast and the valid inequalities

we generated improved the LP relaxation bounds at the associated nodes, in our computational

experiments we could not observe a performance improvement when we ran branch-cut-and-price

algorithms. Furthermore, we studied the disjunctive cut procedure and proposed a methodology

that eliminates the major challenges in the procedure that appear in a branch-cut-and-price al-

gorithm. We constructed four implementations of this methodology by choosing different base

models to derive valid inequalities and evaluated those computationally. We observed that the

efficiency of the disjunctive cut procedure implemented depended on the constructed cut gener-

ation linear program formulated based on the base model. We could not observe a performance

improvement when we generated disjunctive valid inequalities in the branch-cut-and-price algo-

rithm.

There were two practical implications of our study. First, we computationally demonstrated

that there is value in solving the integrated problem. Namely, for a fixed set of customers and

candidate facility locations, the integrated problem identifies opportunities to lower cost through

230
changes in facility selection, reduction in vehicle quantity, and/or changes in customer routing. We

could obtain lower costs by solving the LRSP than those obtained by using a sequential optimiza-

tion of the decisions. Second, since the LRSP is a generalization of other well-known problems,

such as the LRP, the MDVRP, the VRP, and the VRPMT, we could modify the algorithm for the

LRSP to solve these special cases. To solve the capacitated LRP, we developed a branch-and-price

algorithm, which to our knowledge is the first of its kind for this class of problem. We provided

computational results for randomly generated instances and instances from the literature. Our

experiments indicated that the algorithm is very effective at producing quality solutions and can

handle larger instances than previously suggested approaches, which have been primarily based on

two-index formulations. The approach, however, does not seem as effective at producing quality

lower bounds.

Since we expected that the addition of cutting plane routines would prove helpful in solving

larger instances, one area of future research is to explore new inequalities for the LRSP and to

explore new ideas to improve the methodology that generates disjunctive cuts in a branch-cut-and-

price algorithm. We believe that such a study would be valuable to the literature. Another research

avenue is to extend the computational study evaluating the value of integrated optimization in order

to generalize the results. As we studied the LRP, we plan to explore possible contributions of our

algorithm to the literature for other special cases such as the VRPMT and the MDVRP. In addition

to this future work, we plan to investigate possible approaches to improving the performance

of our branch-and-price algorithm for the LRSP. One approach is to add another set of branching

disjunctions to the set of disjunctions that we applied. We observed that branching on flow between

pairs of customer nodes (modified Ryan and Foster branching) was causing inefficiencies in the

algorithm. We may search alternative ways to the current method or develop additional branching

methods before applying this. Another approach is to develop additional heuristic or exact versions

of the pricing algorithm in order to obtain performance improvement. Finally, we have some

alternative formulations in mind which would result in better performance of the branch-and-price

or branch-cut-and-price algorithms.

231
Bibliography

K Aardal. Capacitated Facility Location: Separation Algorithms and Computational Experience.

Mathematical Programming, 81(2):149–175, 1998.

K Aardal, Y. Pochet, and L. A. Wolsey. Capacitated Facility Location: Valid Inequalities and

Facets. Mathematics of Operations Research, 20:562–582, 1995.

Y. Agarwal, K. Mathur, and H. M. Salkin. A Set-Partitioning-Based Exact Algorithm for the

Vehicle Routing Problem. Networks, 19:731–749, 1989.

R. K. Ahuja, J. B. Orlin, S. Pallottino, M. P. Scaparra, and M. G. Scutellá. A Multi-exchange

Heuristic for the Single-source Capacitated Facility Location Problem. Management Science,

50(6):749–760, 2004.

M. Albareda-Sambola, J. A. Diaz, and E. Fernandez. A Compact Model and Tight Bounds for a

Combined Location-routing Problem. Computers and Operations Research, 32:407–428, 2005.

R. Anbil, J. J. Forrest, and W. R. Pulleyblank. Column Generation and the Airline Crew Pairing

Problem. Proceedings of the International Congress of Mathematicians Berlin, Extra Volume,

III:667–686, 1998.

J. A. Appleget and R. K. Wood. Explicit-Constraint-Branching for Solving Mixed-Integer Pro-

grams. In M. Laguna and J.L. González-Velarde, editors, Computing Tools for Modeling, Op-

timization and Simulation, pages 245–261. Kluwer Academic Publishers, Boston, 2000.

232
BIBLIOGRAPHY

J. R. Araque. Lots of Combs of Different Size for Vehicle Routing, 1990. CORE Discussion

Paper.

J. R. Araque, L. A. Hall, and T. L. Magnanti. Capacitated Trees, Capacitated Routing and Asso-

ciated Polyhedra, 1990. CORE Discussion Paper.

P. Augerat, J. M. Belenguer, E. Benavent, A. Corberán, and D. Naddef. Separating Capacity

Inequalities in the CVRP Using Tabu Search. European Journal of Operational Research, 106:

546–557, 1999.

B. M. Baker. A Partial Dual Algorithm for the Capacitated Warehouse Location Problem. Euro-

pean Journal of Operational Research, 23:48–56, 1986.

E. Balas. Disjunctive Programming: Facets of the Convex Hull of Feasible Points. Technical

report No. 348, GSIA, Carnigie Mellon University, 1974.

E. Balas. Facets of the Knapsack Polytope. Mathematical Programming, 8:146–164, 1975.

E. Balas. Disjunctive Programming. Annals of Discrete Mathematics, 5:3–51, 1979.

E. Balas, S. Ceria, and G. Cornuéjols. A Lift-and-Project Cutting Plane Algorithm for Mixed 0-1

Programs. Mathematical Programming, 58:295–324, 1993.

E. Balas, S. Ceria, and G. Cornuéjols. Mixed 0-1 Programming by Lift-and-Project in a Branch-

and-Cut Framework. Management Science, 42(9):1229–1246, 1996.

E. Balas and M. Padberg. Set Partitioning: A Survey. SIAM Review, 18:710–760, 1976.

E. Balas and A. Saxena. Optimizing Over the Split Closure. Mathematical Programming, 113(2):

219–240, 2008.

M. Balinski and R. Quandt. On an Integer Program for a Delivery Problem. Operations Research,

12:300–304, 1964.

233
BIBLIOGRAPHY

C. Barnhart, E. L. Johnson, G.L. Nemhauser, M. W. P. Savelsbergh, and P. H. Vance. Branch and

Price: Column Generation for Solving Huge Integer Programs. Operations Research, 46(3):

316–329, 1998.

S. Barreto. LRP instances, 2003. URL http://sweet.ua.pt/iscfl43.

S. Barreto. Analysis and Modeling of Location-Routing Problems (in Portuguese). Phd thesis,

University of Aveiro, Aveiro, Portugal, 2004.

S. Barreto, C. Ferreira, J. Paixao, and B. S. Santos. Using Clustering Analysis in a Capacitated

Location-routing Problem. European Journal of Operational Research, 127(3):968–977, 2007.

J. E. Beasley. An Algorithm for Solving Large Capacitated Warehouse Location Problems. Euro-

pean Journal of Operational Research, 33:314–325, 1988.

J. E. Beasley and N. Christofides. An Algorithm for the Resource Constrained Shortest Path

Problem. Networks, 19:379–394, 1989.

J. M. Belenguer, E. Benavent, C. Prins, C. Prodhon, and R. Wolfler-Calvo. A Branch and Cut

Method for the Capacitated Location-Routing Problem. Service Systems and Service Manage-

ment, 2006 International Conference on, 2:1541–1546, 2006.

R. T. Berger. Location-Routing Models for Distribution System Design. PhD thesis, Department

of Industrial Engineering and Management Sciences, Northwestern University, 1997.

R. T. Berger, C. R. Coullard, and M. S. Daskin. Location-Routing Problems with Distance Con-

straints. Transportation Science, 41:29–43, 2007.

A. A. Bertossi, P. Carraresi, and G. Gallo. On Some Matching Problems Arising in Vehicle

Scheduling Models. Networks, 17:271–281, 1987.

L. Bianco, A. Mingozzi, and S. Ricciardelli. A Set Partitioning Approach to the Multiple Depot

Vehicle Scheduling Problem. Optimization Methods and Software, 3:163–194, 1994.

234
BIBLIOGRAPHY

O. Bilde and J. Krarup. Sharp Lower Bounds and Efficient Algorithms for the Simple Plant

Location Problem. Discrete Mathematics, 1:79–97, 1977.

A. Bixby. Polyhedral Analysis and Effective Algorithms for the Capacitated Vehicle Routing Prob-

lem. PhD thesis, Department of Industrial Engineering and Management Sciences, Northwest-

ern University, 1999.

A. Bixby, C. Coullard, and D. Simchi-Levi. The Capacitated Prize-collecting Traveling Salesman

Problem. Working paper, 1997.

U. Blasum and W. Hochstättler. Application of the Branch and Cut Method to the Vehicle Routing

Problem. Technical Report ZPR2000-386, Universität zu Köln, 2000.

L. Bodin and B. Golden. Classification in Vehicle Routing and Scheduling. Networks, 11(2):

97–108, 1981.

L. Bodin, B. Golden, A. Assad, and M. Ball. Routing and Scheduling of Vehicles and Crews: The

State of the Art. Computers and Operations Research, 10:63–211, 1983.

J. Brandao and A. Mercer. A Tabu Search Algorithm for the Multi-Trip Vehicle Routing and

Scheduling Problem. European Journal of Operational Research, 100:180–191, 1997.

J. Brandao and A. Mercer. The Multi-Trip Vehicle Routing Problem. Journal of the Operational

Research Society, 49:799–805, 1998.

L. Burlett. Regional Operating Expenses: A Case Study, 2002. Available at http://www.sb-

d.com/issues/winter2002/features/save.asp.

A. Caprara and A. N. Letchford. On the Separation of Split Cuts and Related Inequalities. Math-

ematical Programming, 94:279–294, 2003.

G. Carpaneto, M. Dell’Amico, M. Fischetti, and P. Toth. A Branch and Bound Algorithm for the

Multiple Depot Vehicle Scheduling Problem. Networks, 19:531–548, 1989.

235
BIBLIOGRAPHY

P. Carraresi and G. Gallo. Network Models for Vehicle and Crew Scheduling. European Journal

of Operational Research, 16:139–151, 1984.

I.M. Chao, B. L. Golden, and E. A. Wasil. A New Heuristic for the Multi-Depot Vehicle Routing

Problem that Improves Upon Best-Known Solutions. American Journal of Mathematical and

Management Sciences, 13:371–406, 1993.

D. C. Cho, E. L Johnson, M. W. Padberg, and M. R. Rao. On the Uncapacitated Plant Location

Problem. I. Valid Inequalities and Facets. Mathematics of Operations Research, 8:579–589,

1983 a.

D. C. Cho, M. W. Padberg, and M. R. Rao. On the Uncapacitated Plant Location Problem. II.

Facets and Lifting Theorems. Mathematics of Operations Research, 8:590–612, 1983 b.

N. Christofides. Vehicle Routing. In E. L. Lawler, J. F. Lenstra, A.H.G. Rinnooy Kan, and D. B.

Shmoys, editors, The Traveling Salesman Problem, pages 431–448. Wiley, Chichester, 1985.

V. Chvátal. Edmonds Polytopes and a Hierarchy of Combinatorial Problems. Discrete Mathemat-

ics, (4):305–337, 1973.

G. Clarke and J. Wright. Scheduling of Vehicles from a Central Depot to a Number of Delivery

Points. Operations Research, 12(4):568–581, 1964.

Kenworth Truck Company. White Paper On Life Cycle Cost, 2003.

W. Cook, , R. Kannan, and A. Schrijver. Chva´tal Closures For Mixed Integer Programming

Problems. Mathematical Programming, 47(2):155–174, 1990.

J. F. Cordeau, M. Gendreau, and G. Laporte. A Tabu Search Heuristic for Periodic and Multi-

Depot Vehicle Routing Problems. Networks, 30:105–119, 1997.

G. Cornuejols. Valid Inequalities For Mixed Integer Linear Programs. Mathematical Program-

ming B, 112:2008, 2006.

236
BIBLIOGRAPHY

G. Cornuejols, M. L. Fisher, and G. L. Nemhauser. Location of Bank Accounts to Optimize Float:

An Analytic Study of Exact and Approximate Algorithms. Management Science, 23:789–810,

1977.

G. Cornuejols and F. Harche. Polyhedral Study of the Capacitated Vehicle Routing Problem.

Mathematical Programming, 60:21–52, 1993.

G. Cornuejols and Y. Li. Elementary Closures For Integer Programs. Operations Research Letters,

28:1–8, 2001.

G. Cornuejols, G. L. Nemhauser, and L.A. Wolsey. The Uncapacitated Facility Location Problems.

In P. B. Mirchandani R. L. Francis, editor, Discrete Location Theory, pages 119–171. Wiley-

Interscience, New York, 1990.

G. Cornuejols and J. M. Thizy. Some Facets of the Simple Plant Location Polytope. Mathematical

Programming, 23:50–74, 1982.

H. Crowder, E. L. Johnson, and M. Padberg. Solving Large-Scale Zero-One Programming Prob-

lems. Operations Research, 31(5):803–834, 1983.

F. Cullen, J. Jarvis, and D. Ratliff. Set Partitioning Based Heuristics for Interactive Routing.

Networks, 11:125–144, 1981.

G. B. Dantzig and P. Wolfe. Decomposition Principle For Linear Programs. Operations Research,

8:101–111, 1960.

M. S. Daskin. Network and Discrete Location: Models, Algorithms and Applications. John Wiley

and Sons, Inc., New York, 1995.

I. R. de Farias and G. L. Nemhauser. A Family of Inequalities for the Generalized Assignment

Polytope. Operations Research Letters, 29:49–55, 2001.

M. Dell’Amico, M. Fischetti, and P. Toth. Heuristic Algorithms for the Multiple Depot Vehicle

Scheduling Problem. Management Science, 39(1):115–125, 1993.

237
BIBLIOGRAPHY

H. Delmaire, J. A. Diaz, E. Fernandez, and M.Ortega. Comparing New Heuristics for the Pure

Integer Capacitated Location Problem. Research Report DR97/10, Universidad Politecnica de

Catalunya, Spain, 1997.

M. Desrochers. An Algorithm for the Shortest Path Problem with Resource Constraints. Technical

Report G-88-27, GERAD, 1988.

M. Desrochers, J. Desrosiers, and M. M. Solomon. A New Optimization Algorithm for the Vehicle

Routing Problem with Time Windows. Operations Research, 40(2):342–354, 1992.

M. Desrochers and G. Laporte. Improvements and Extensions to Miller-Tucker-Zemlin Subtour

Elimination Constraints. Operations Research Letters, 10:27–36, 1991.

M. Desrochers and F. Soumis. A Column Generation Approach to the Urban Transit Crew

Scheduling Problem. Transportation Science, 23(1):1–13, 1989.

J. Desrosiers, Y. Dumas, M. Desrochers, F. Soumis, F. Sanso, and P. Trudeau. A Breakthrough in

Airline Crew Scheduling. Technical Report G-91-11, GERAD, 1991.

E. D. Dolan and J. J. More. Benchmarking Optimization Software With Performance Profiles.

Mathematical Programming, 91(2):201–213, 2002.

M. Dror. Note on the Complexity of the Shortest Path Models for Column Generation in vrptw.

Operations Research, 42(5):977–978, 1994.

I. Dumitrescu. Constrained Path and Cycle Problems. PhD thesis, Department of Mathematics

and Statistics, University of Melbourne, 2002.

S. Eilon. Management Perspectives in Physical Distribution. Omega, 5(4):437–462, 1977.

D. Erlenkotter. A Dual Based Procedure for Uncapacitated Facility Location. Operations Re-

search, 26:992–1009, 1978.

238
BIBLIOGRAPHY

M. Eso. Parallel Branch-and-cut for Set Partitioning. Phd thesis, Cornell University, Ithaca, NY,

1999.

D. Feillet, P. Dejax, M. Gendreau, and C. Gueguen. An Exact Algorithm for the Elementary Short-

est Path Problem with Resource Constraints: Application to Some Vehicle Routing Problems.

Networks, 44(3):216–229, 2004.

M. Fischetti, A. Lodi, S. Martello, and P. Toth. A Polyhedral Approach to Simplified Crew

Scheduling and Vehicle Scheduling Problems. Management Science, 47(6):833–850, 2001.

M. A. Forbes, J. N. Holt, and A. M. Watts. An Exact Algorithm for Multiple Depot Bus Schedul-

ing. European Journal of Operational Research, 72:115–124, 1994.

R. Fukasawa, H. Longo, J. Lysgaard, M. Poggi de Aragão, M. Reis, E. Uchoa, and R. F. Werneck.

Robust Branch-and-Cut-and-Price for the Capacitated Vehicle Routing Problem. Mathematical

Programming, 106(3):491–511, 2006.

R. Galvao. The Use of Lagrangean Relaxation in the Solution of Uncapacitated Facility Location

Problems. Location Science, 1:57–79, 1993.

W. M. Garvin, H. W. Crandall, J. B. John, and R. A. Spellman. Applications of Linear Program-

ming in the Oil Industry. Management Science, 3:407–430, 1957.

B. Gavish and S. C. Graves. The Traveling Salesman Problem and Related Problems. Working

Paper 7905, 1979.

M. Gendreau, A. Hertz, and G. Laporte. New Insertion and Post Optimization Procedures for the

Traveling Salesman Problem. Operations Research, 40:1086–1094, 1992.

B.E. Gillett and J.G. Johnson. Multi-Terminal Vehicle-Dispatch Algorithm. Omega, 4:711–718,

1976.

239
BIBLIOGRAPHY

R. E. Gomory. An Algorithm For Integer Solutions to Linear Programming. In R. L. Graves and

P. Wolfe, editors, Recent Advances in Mathematical Programming, pages 269–302, New York,

1963. McGraw-Hill.

E. S. Gottlieb and M. R. Rao. The Generalized Assignment Problem: Valid Inequalities and

Facets. Mathematical Programming, 46:31–52, 1990.

L. Gouveia. A Result On Projection for the Vehicle Routing Problem. European Journal of

Operational Research, 85(3):610–624, 1995.

Z. Gu, G. L. Nemhauser, and M. W. P. Savelsbergh. Lifted cover inequalities for 0–1 integer

programs: computation. INFORMS Journal on Computing, 10:427–438, 1998.

M. Guignard. Fractional Vertices, Cuts and Facets of the Simple Plant Location Problem. Mathe-

matical Programming Study, 12:150–162, 1980.

E. Hadjiconstantinou, N. Christofides, and A. Mingozzi. A New Exact Algorithm for the Vehicle

Routing Problem Based on q-Paths and k-Shortest Paths Relaxations. Annals of Operations

Research, 61:21–43, 1995.

S. L. Hakimi. Optimal Locations of Switching Centers and the Absolute Centers and Medians of

a Graph. Operations Research, 12:450–459, 1964.

S. L. Hakimi. Optimal Distribution of Switching Centers in a Communication Network and Some

Related Theoretic Graph Problems. Operations Research, 13:462–475, 1965.

P. H. Hansen, B. Hegedahl, S. Hjortkjaer, and B. Obel. A Heuristic Solution to the Warehouse

Location-Routing Problem. European Journal of Operational Research, 76:111–127, 1994.

K. Holmberg, M. Rönnqvist, and D. Yuan. An Exact Algorithm for the Capacitated Facility

Location Problem with Single Sourcing. European Journal of Operational Research, 113:544–

559, 1999.

240
BIBLIOGRAPHY

S. Irnich and G. Desaulniers. Column Generation, chapter Shortest Path Problems with Resource

Constraints, pages 33–65. GERAD 25th Anniversary Series. Springer, 2005.

M. Jepsen, B. Petersen, S. Spoorendonk, and D. Pisinger. Subset-Row Inequalities Applied to the

Vehicle-Routing Problem with Time Windows. Operations Research, 56:497–511, 2008.

O. Kariv and S.L. Hakimi. An Algorithmic Approach to Network Location Problems. Part ii: The

p-median. SIAM Journal of Applied Mathematics, 37:539–560, 1979.

J. Klincewicz and H. Luss. A Lagrangean Relaxation for Capacitated Facility Location with

Single-source Constraints. Journal of the Operational Research Society, 37:495–500, 1986.

A. Klose and A. Drexl. Facility Location Models for Distribution System Design. European

Journal of Operational Research, 162:4–29, 2005.

M. Körkel. On the Exact Solution of Large Scale Simple Plant Location Problems. European

Journal of Operational Research, 39:157–173, 1989.

M. Labbé and F. V. Louveaux. Location problems. In S. Martello M. Dell’Amico, F. Maffioli,

editor, Annotated Bibliographies in Combinatorial Optimization. J. Wiley & Sons, New York,

1997.

G. Laporte. Location Routing Problems. In B.L. Golden and A.A. Assad, editors, Vehicle Routing:

Methods and Studies, pages 293–318. North-Holland, Amsterdam, 1988.

G. Laporte. The Vehicle Routing Problem: An Overview of Exact and Approximate Algorithms.

European Journal of Operational Research, 59:345–358, 1992.

G. Laporte and Y. Nobert. An Exact Algorithm for Minimizing Routing and Operating Cost in

Depot Location. European Journal of Operational Research, 6:224–226, 1981a.

G. Laporte and Y. Nobert. Comb Inequalities for the Vehicle Routing Problem. Methods of

Operations Research, 51:271–, 1981b.

241
BIBLIOGRAPHY

G. Laporte and Y. Nobert. Exact Algorithms for the Vehicle Routing Problem. Annals of Discrete

Mathematics, 31:147–184, 1987.

G. Laporte, Y. Nobert, and D. Arpin. Optimal Solutions to Capacitated Multidepot Vehicle Rout-

ing Problems. Congressus Numerantium, 44:283–292, 1984.

G. Laporte, Y. Nobert, and D. Arpin. An Exact Algorithm for Solving a Capacitated Location-

Routing Problem. Annals of Operations Research, 6:293–310, 1986.

G. Laporte, Y. Nobert, and M. Desrouchers. Optimal Routing with Capacity and Distance Restric-

tions. Operations Research, 33:1050–1073, 1985.

G. Laporte, Y. Nobert, and Y. Pelletier. Hamiltonian Location Problems. European Journal of

Operational Research, 12:82–89, 1983.

G. Laporte, Y. Nobert, and S. Taillefer. Solving a Family of Multi-Depot Vehicle Routing and

Location-Routing Problems. Transportation Science, 22:161–172, 1988.

J. Lenstra and A. Rinnooy Kan. Complexity of Vehicle Routing and Scheduling Problems. Net-

works, 11:221–228, 1981.

A. N. Letchford. On Disjunctive Cuts For Combinatorial Optimization. Journal of Combinatorial

Optimization, 5:299–315, 2001.

A. N. Letchford and J. José S. González. Projection Results for Vehicle Routing. Mathematical

Programming, 105(2-3):251–274, 2006.

J. M. Y. Leung and T. L. Magnanti. Valid Inequalities and Facets of the Capacitated Plant Location

Problems. Mathematical Programming, 44:271–291, 1989.

C. K. Lin, C. K. Chew, and A. Chen. A Location-Routing-Loading Problem for Bill Delivery

Services. Computers and Industrial Engineering, 43:5–25, 2002.

242
BIBLIOGRAPHY

C. K. Lin and R. C. W. Kwok. Multi-Objective Metaheuristics for a Location-Routing Problem

with Multiple Use of Vehicles on Real Data and Simulated Data. European Journal of Opera-

tional Research, 175(3):1833–1849, 2006.

J. Linderoth. Lecture Notes for Integer Programming, Department of Industrial and Systems

Engineering, Lehigh University. 2005.

M. E. Lübbecke and J. Desrosiers. Selected Topics in Column Generation. Operations Research,

53(6):1007–1023, 2005.

F. Maranzana. On the Location of Supply Points to Minimize Transport Cost. Operations Research

Quarterly, 15:261–270, 1964.

T. McCormick, M. R. Rao, and G. Rinaldi. When is Min Cut with Negative Edges Easy to Solve?

Easy and Difficult Objective Functions for Max Cut. Research report, IASI-CNR, 2002.

C. E. Miller, A. W. Tucker, and R. A. Zemlin. Integer Programming Formulations and Traveling

Salesman Problems. Journal of ACM, 7(1):326–329, 1960.

H. Min. Consolidation Terminal Location-allocation and Consolidated Routing Problems. Journal

of Business Logistics, 17(2):235–263, 1996.

H. Min, V. Jayaraman, and R. Srivastava. Combined Location-Routing Problems: A Synthesis and

Future Research Directions. European Journal of Operational Research, 108(1):1–15, 1998.

G. Nagy and S. Salhi. Location-routing Issues, Models and Methods. European Journal of Oper-

ational Research, 177:649–672, 2007.

J. M. Nambiar, L. F. Gelders, and L. N. Van Wassenhove. A Large Scale Location Allocation

Problem in the Natural Rubber Industry. European Journal of Operational Research, 6:183–

189, 1981.

G. L. Nemhauser and L. A. Wolsey. Integer and Combinatorial Optimization. Wiley-Interscience

Series in Discrete Mathematics and Optimization. John Wiley & Sons, New York, 1989.
243
BIBLIOGRAPHY

G. L. Nemhauser and L. A. Wolsey. A Recursive Procedure to Generate All Cuts For 0-1 Mixed

Integer Programs. Mathematical Programming, 46(3):379–390, 1990.

A. Olivera and O. Viera. Adaptive Memory Programming for the Vehicle Routing Problem with

Multiple Trips. Computers and Operations Research, 34:28–47, 2007.

M. W. Padberg. (1,k)- Configurations and Facets for Packing Problems. Mathematical Program-

ming, 18:94–99, 1980.

M. W. Padberg, T. J. van Roy, and L. A. Wolsey. Valid Inequalities for Fixed Charge Problems.

Operations Research, 33:842–861, 1985.

J. Perl and M. S. Daskin. A Warehouse Location-Routing Model. Transportation Research, 19

(B):381396, 1985.

M. Perregaard and E. Balas. Generating Cuts from Multiple-term Disjunctions. In Integer

Programming and Combinatorial Optimization, pages 348–360. Springer Berlin / Heidelberg,

2001.

R. J. Petch and S. Salhi. A Multi-Phase Constructive Heuristic for the Vehicle Routing Problem

with Multiple Trips. Discrete Applied Mathematics, 133:69–92, 2004.

H. Pirkul. Efficient Algorithms for the Capacitated Concentrator Location Problem. Computers

and Operations Research, 14:197–208, 1987.

C. Prins, C. Prodhon, A. Ruiz, P. Soriano, and R. Wolfler Calvo. Solving the Capacitated Location-

Routing Problem by a Cooperative Lagrangean Relaxation-Granular Tabu Search Heuristic.

Transportation Science, 41(4):470–483, 2007.

T. K. R. Ralphs, L. Kopman, W. R. Pulleyblank, and L.E. Trotter. On the Capacitated Vehicle

Routing Problem. Mathematical Programming, 94:343–, 2003.

J. Renaud, G. Laporte, and F. F. Boctor. A Tabu Search Heuristic for the Multi-Depot Vehicle

Routing Problem. Computers and Operations Research, 23(3):229–235, 1996.

244
BIBLIOGRAPHY

MacroSy Research and Technology. Logistics Costs and U.S. Gross Domestic Product, 2005.

Prepared for Federal Highway Administration Department of Transportation. Available at

http://ops.fhwa.dot.gov/freight/freight analysis/econ methods/lcdp rep/index.htm.

C. S. ReVelle and H. A. Eiselt. Location Analysis: A Synthesis and Survey. European Journal of

Operational Research, 165:1–19, 2005.

C. S. ReVelle and R. W. Swain. Central Facilities Location. Geographical Analysis, 2:30–42,

1970.

C. Ribeiro and F. Soumis. A Column Generation Approach to the Multiple Depot Vehicle Schedul-

ing Problem. Operations Research, 42:41–52, 1994.

Y. Rochat and E. Taillard. Probabilistic Diversification and Intensification in Local Search for

Vehicle Routing. Heuristics, 1:147–167, 1995.

S. Ropke. Disjunctive Cuts in Branch-and-Cut-and-Price Algorithms. Application to the Capaci-

tated Vehicle Routing Problem. In International Workshop on Column Generation, June 2008.

S. Ropke and J. F. Cordeau. Branch and Cut and Price for Pickup and Delivery Problems With

Time Windows. Transportation Science, 43(3):267–286, 2009.

K. Rosing and C. S. ReVelle. Heuristic Concentration: Two Stage Solution Construction. Euro-

pean Journal of Operational Research, 97:75–86, 1997.

T. J. Van Roy. A Cross Decomposition Algorithm for Capacitated Facility Location. Operations

Research, 34:145–163, 1986.

D. Ryan and B. Foster. An Integer Programming Approach to Scheduling. In A. Wren, edi-

tor, Computer Scheduling of Public Transport Urban Passenger Vehicle and Crew Scheduling,

pages 269–280. North-Holland, 1981.

S. Salhi and G. K. Rand. The Effect of Ignoring Routes When Locating Depots. European Journal

of Operational Research, 39(2):150–156, 1989.

245
BIBLIOGRAPHY

M.W.P. Savelsbergh and G. Nemhauser. Functional Description of MINTO, A Mixed INTeger

Optimizer. Georgia Institute of Technology, 1996.

H. P. Simao and J. M. Thizy. A Dual Simplex Algorithm for the Canonical Representation of the

Uncapaciated Facility Location Problem. Operations Research Letters, 8:279–286, 1989.

S. Spoorendonk and G. Desaulniers. Clique Inequalities Applied to the Vehicle Routing Problem

With Time Windows. In International Workshop on Column Generation, June 2008a.

S. Spoorendonk and G. Desaulniers. Clique inequalities applied to the vehicle routing problem

with time windows. Les Cahiers des GERAD G-2008-72, GERAD, Canada, 2008b.

R. Sridharan. A Lagrangean Heuristic for the Capacitated Plant Location Problem with Single

Source Constraints. European Journal of Operational Research, 66:305–312, 1993.

R. Srivastava. Alternate Solution Procedures for the Location-routing Problem. Omega, 21(4):

497–506, 1993.

E. D. Taillard, G. Laporte, and M. Gendreau. Vehicle Routing with Multiple Use of Vehicles. The

Journal of Operational Research Society, 47(8):1065–1070, 1996.

M. Teitz and P. Bart. Heuristic Methods for Estimating the Generalized Vertex Median of a

Weighted Graph. Operations Research, 16:955–961, 1968.

R. R. Thomas. Lecture Notes for Linear and Integer Polyhedra, Department of Mathematics,

University of Washington. 2005.

F. A. Tillman. The Multiple Terminal Delivery Problem with Probabilistic Demands. Transporta-

tion Science, 3:192–204, 1969.

F. A. Tillman and T. M. Cain. An Upper Bound Algorithm for the Single and Multiple Terminal

Delivery Problem. Management Science, 18:664–682, 1972.

246
BIBLIOGRAPHY

P. Toth and D. Vigo. The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics

and Applications, 2002.

P. Toth and D. Vigo. The Granular Tabu Search and Its Application to the Vehicle-Routing Prob-

lem. INFORMS J. on Computing, 15(4):333–346, 2003.

H. P. Vance, C. Barnhart, E. Johnson, and G. L. Nemhauser. Airline Crew Scheduling: A New

Formulation and Decomposition Algorithm. Operations Research, 45(2):188–199, 1997.

F. Vanderbeck. Decomposition and Column Generation for Integer Programs. Phd thesis, Univer-

site Catholique de Louvain, 1994.

M. H. Webb. Cost Functions in the Location of Depots for Multiple-Delivery Journeys. Opera-

tional Research Quarterly, 19:311–320, 1968.

R. Weismantel. On the 0\1 Knapsack Polytope. Mathematical Programming, 77:49–68, 1997.

247
Appendix A

Tables of Computational Results

In tables:

• “-” represents the missing data caused by termination with time limit,

• “x” represents the missing data caused by termination prematurely since memory,

• “*” represents the missing data caused by the termination signal of condor since time,

• “**” refers to the case when the root node cannot be solved within time limit.

248
 Table A.1: LP relaxation bounds with simple valid inequalities: 20-cust. Inst.

# of SPP SPP1 SPP2 SPP3 ASPP

Data Pairings IP Obj % Gap % Gap % Gap % Gap % Gap
p01-f20-v1t1 67558 4420 24.54 24.49 19.79 2.48 1.78
p01-f20-v1t2 154903 4419 26.06 - 22.08 3.69 3.52
p01-f20-v2t1 147368 4378 26.17 - 23.24 3.80 3.44
p01-f20-v2t2 349885 4369 27.52 - 25.05 4.86 4.83
p01-l20-v1t1 26266 4780 33.51 33.38 27.98 5.37 2.16
p01-l20-v1t2 60211 4599 32.77 32.72 28.08 3.49 0.58
p01-l20-v2t1 51538 4605 33.64 33.51 29.91 4.30 1.57
p01-l20-v2t2 121027 4481 33.24 33.23 30.81 2.97 1.54
p03-f20-v1t1 52619 4468 15.64 15.63 14.08 0.72 0.43
p03-f20-v1t2 121933 4461 17.25 - 16.12 2.32 2.13
p03-f20-v2t1 107172 4421 18.04 - 17.42 2.88 2.75
p03-f20-v2t2 254863 4421 19.78 - 19.12 4.58 4.44
p03-m20-v1t1 37751 4543 18.6 18.42 15.83 3.20 3.02
p03-m20-v1t2 89465 4442 18.13 - 16.02 2.35 2.35
p03-m20-v2t1 78580 4383 19.15 - 17.59 3.19 3.19
p03-m20-v2t2 193509 4376 20.13 - 19.2 4.12 4.12
p03-l20-v1t1 32801 4617 31.27 30.84 26 4.74 1.72
p03-l20-v1t2 79215 4467 30.19 30.13 27.08 2.77 1.30
p03-l20-v2t1 51052 4404 29.84 29.84 28.35 1.45 1.04
p03-l20-v2t2 133853 4399 31.76 - 30.45 3.55 3.11
p07-f20-v1t1 68267 4426 38.22 37.96 34.78 2.81 1.13
p07-f20-v1t2 166551 4421 39.33 - 37.18 3.60 2.57
p07-f20-v2t1 133873 4380 39.45 - 38.55 3.25 2.58
p07-s20-v1t1 27623 4840 31.69 31.47 25.75 4.73 1.01
p07-s20-v1t2 61753 4801 32.79 32.72 27.97 5.61 2.91
p07-s20-v2t1 54964 4751 32.8 32.48 29.09 5.34 2.32
p07-s20-v2t2 127679 4543 31.15 - 28.1 2.43 0.68
p07-t20-v1t1 11267 4758 29.49 29.47 26.03 3.62 1.05
p07-t20-v1t2 28315 4670 30.44 30.06 28.58 3.77 2.08
p07-t20-v2t1 19283 4688 32.39 31.83 30.98 5.57 4.14
p07-t20-v2t2 53689 4460 30.47 30.28 29.2 2.11 1.33
Avg. 28.24 29.91 25.50 3.54 2.28

249
Table A.2: LP relaxation bounds with simple valid inequalities: 25-cust. Inst.

SPP-LRSP SPP-LRSP2 SPP-LRSP3 ASPP-LRSP

Data % Gap % Gap % Gap % Gap
p01-f25-v1t1 18.66 13.59 5.08 4.32
p01-f25-v1t2 15.91 11.89 2.4 1.93
p01-f25-v2t1 15.33 12.41 1.65 1.35
p01-f25-v2t2 17.6 14.98 3.09 3.02
p01-l25-v1t1 30.66 21.63 4.25 1.01
p01-l25-v1t2 31.88 24.04 4.65 1.9
p01-l25-v2t1 32.42 26.29 4.7 1.93
p01-l25-v2t2 30.62 26.58 2.89 1.43
p03-f25-v1t1 9.26 6.41 7.9 5.24
p03-f25-v1t2 6.16 4.09 4.82 2.89
p03-f25-v2t1 6.55 4.53 5.14 3.3
p03-f25-v2t2 8.32 6.6 6.83 5.32
p03-m25-v1t1 6.77 4.12 3.91 1.87
p03-m25-v1t2 8.78 6.98 5.75 4.52
p03-m25-v2t1 8.18 7.25 5.08 4.53
p03-m25-v2t2 5.48 4.54 2.44 1.89
p03-l25-v1t1 22.54 16.39 3.98 0.97
p03-l25-v1t2 25.11 19.31 5.55 2.84
p03-l25-v2t1 23.88 20.31 4.35 2.6
p03-l25-v2t2 24.74 21.75 4.48 3.27
p07-f25-v1t1 39.27 29.8 5.22 2.55
p07-f25-v1t2 39.43 32.18 4.79 3.09
p07-f25-v2t1 41.78 37.41 6.15 4.85
p07-f25-v2t2 39.42 34.58 3.34 2.13
p07-s25-v1t1 23.23 17.58 5.03 1.96
p07-s25-v1t2 21.72 17.25 3.28 0.76
p07-s25-v2t1 20.69 18.59 2.21 1
p07-s25-v2t2 25.24 22.55 5.42 3.71
p07-t25-v1t1 30.45 21.16 4.84 1.27
p07-t25-v1t2 32.68 25.56 6.22 3.89
p07-t25-v2t1 31.27 26.33 4.8 2.6
p07-t25-v2t2 29.28 25.39 2.74 1.75
Average 22.6 18.19 4.47 2.68

250
Table A.3: LP relaxation bounds with simple valid inequalities: 40-cust. Inst.

SPP-LRSP SPP-LRSP2 SPP-LRSP3 ASPP-LRSP

Data % Gap % Gap % Gap % Gap
p01-f40-v1t1 18.46 14.6 4.45 2.22
p01-f40-v1t2 17.54 14.86 3.46 2.3
p01-f40-v2t1 17.34 15.84 3.06 2.47
p01-f40-v2t2 18.56 17.67 3.74 3.6
p01-l40-v1t1 19.95 16.73 5.01 3.76
p01-l40-v1t2 18.13 16.01 3.01 2.61
p01-l40-v2t1 18.75 17.74 3.23 3.12
p01-l40-v2t2 17.54 16.64 1.51 1.51
p03-f40-v1t1 21.01 19 1.16 0.95
p03-f40-v1t2 22.64 20.92 2.06 2.03
p03-f40-v2t1 20.24 19.28 0.01 0.01
p03-f40-v2t2 21.5 21.08 0.55 0.55
p03-m40-v1t1 7.18 4.4 5.44 2.87
p03-m40-v1t2 5.72 4.01 4.02 2.42
p03-m40-v2t1 5.88 4.81 4.04 3.08
p03-m40-v2t2 6.78 6.18 4.85 4.38
p03-l40-v1t1 15.43 11.92 5.62 3.56
p03-l40-v1t2 13.43 11.16 3.77 2.64
p03-l40-v2t1 13.51 12.03 3.68 2.79
p03-l40-v2t2 11.72 11.03 1.55 1.44
p07-f40-v1t1 25.5 22.13 3.13 1.62
p07-f40-v1t2 26.01 22.77 2.93 1.37
p07-f40-v2t1 26.65 24.54 2.99 2.01
p07-f40-v2t2 26.03 24.78 2.06 1.2
p07-s40-v1t1 20.11 17.17 2.78 1.19
p07-s40-v1t2 19.62 17.03 1.98 0.37
p07-s40-v2t1 20.66 18.94 2.05 1.09
p07-s40-v2t2 22.65 21.53 3.4 2.68
p07-t40-v1t1 27.9 24.64 2.53 1.03
p07-t40-v1t2 29.85 27.27 3.49 2.43
p07-t40-v2t1 27.58 25.99 0.88 0.23
p07-t40-v2t2 30.07 28.83 2.35 1.88
Average 19.19 17.24 2.96 2.04

251
 Table A.4: Performance of 1p-ESPPRC, 1p-ESPPRC-NwD, 2p-ESSPRC: 25-customer instances
1p-ESPPRC 1p-ESPPRC-NwD 2p-ESPPRC
Tot.time Time per Tot.time (b/a) Time per Tot.time (c/a) Time per
Data File (s) (a) iter (s) (b) iter (s) (c) iter
p01-f25-v1t1 282.44 4.15 174.46 0.62 2.57 21.42 0.08 0.32
p01-f25-v1t2 2555.16 33.18 1238.2 0.48 16.08 91.44 0.04 0.82
p01-f25-v2t1 2001.41 21.75 849.4 0.42 11.48 116.77 0.06 1.27
p01-f25-v2t2 13467.95 134.68 6515.66 0.48 69.32 515.59 0.04 5.49
p01-l25-v1t1 134.5 2.36 76.97 0.57 1.48 14.37 0.11 0.29
p01-l25-v1t2 752.68 12.98 425.1 0.56 6.54 44.32 0.06 0.76
p01-l25-v2t1 451.86 9.41 268.62 0.59 5.6 68.69 0.15 0.85
p01-l25-v2t2 2127.41 31.75 1601.02 0.75 24.63 195.69 0.09 2.8
p03-f25-v1t1 154.57 2.53 77.29 0.5 1.55 15.19 0.1 0.24
p03-f25-v1t2 726.3 10.38 460.92 0.63 6.23 60.06 0.08 0.7
p03-f25-v2t1 616.07 8.21 290.1 0.47 4.4 45.19 0.07 0.65
p03-f25-v2t2 3273.45 32.73 2054.31 0.63 23.08 142.52 0.04 1.76
p03-m25-v1t1 362.8 6.6 178.94 0.49 3.14 24.59 0.07 0.38
p03-m25-v1t2 3238.9 46.27 1427.19 0.44 19.55 107.76 0.03 1.31

252
p03-m25-v2t1 1714.34 24.49 1117.63 0.65 16.68 66.95 0.04 1.01
p03-m25-v2t2 12680.8 162.57 6380.45 0.5 76.87 321.98 0.03 4.13
p03-l25-v1t1 84.76 1.3 58.64 0.69 1.01 15.32 0.18 0.25
p03-l25-v1t2 499.57 7.35 285.85 0.57 4.2 37.17 0.07 0.51
p03-l25-v2t1 187.57 3.03 145.72 0.78 2.21 41.57 0.22 0.7
p03-l25-v2t2 1417.49 19.96 962.73 0.68 13.56 114.12 0.08 1.52
p07-f25-v1t1 150.75 2.79 155.74 1.03 2.88 26.22 0.17 0.47
p07-f25-v1t2 1154.08 16.73 760.58 0.66 11.02 86.32 0.07 1.2
p07-f25-v2t1 421.47 6.8 327.5 0.78 5.2 98.88 0.23 1.73
p07-f25-v2t2 2231.63 31.88 1664.52 0.75 24.48 293.53 0.13 4.59
p07-s25-v1t1 29.38 0.61 22.13 0.75 0.37 6.28 0.21 0.14
p07-s25-v1t2 119.4 2.3 77.21 0.65 1.54 13.71 0.11 0.3
p07-s25-v2t1 51.11 1.14 44.11 0.86 0.98 19.45 0.38 0.41
p07-s25-v2t2 305.07 4.49 220.66 0.72 3.56 51.44 0.17 0.87
p07-t25-v1t1 162.02 2.7 126.93 0.78 1.98 22.49 0.14 0.39
p07-t25-v1t2 1505.23 21.2 976.62 0.65 14.15 50.21 0.03 0.85
p07-t25-v2t1 641.48 12.1 448.56 0.7 8.16 75.14 0.12 1.45
p07-t25-v2t2 4036.72 61.16 2492.33 0.62 39.56 223.61 0.06 3.29
Average 1798.07 23.11 997.07 0.64 13.25 94.62 0.11 1.3
 Table A.5: Performance of 1p-ESPPRC, 1p-ESPPRC-NwD, 2p-ESSPRC: 40-customer instances
1p-ESPPRC 1p-ESPPRC-NwD 2p-ESPPRC
Tot.time Time per Tot.time (b/a) Time per Tot.time (c/a) Time per
Data File (s) (a) iter (s) (b) iter (s) (c) iter
p01-f40-v1t1 13272.35 189.61 9252.84 0.7 118.63 421.91 0.03 5.15
p01-f40-v1t2 - - - - - 2071.86 - 23.02
p01-f40-v2t1 - - - - - 2403 - 24.77
p01-f40-v2t2 - - - - - - - -
p01-l40-v1t1 7822.41 87.89 4749.8 0.61 53.37 305.09 0.04 3.63
p01-l40-v1t2 - - - - - 1633.48 - 15.12
p01-l40-v2t1 - - - - - 2627.27 - 24.33
p01-l40-v2t2 - - - - - - - -
p03-f40-v1t1 - - - - - 967.01 - 12.56
p03-f40-v1t2 - - - - - 7304.94 - 73.05
p03-f40-v2t1 - - - - - 3665.82 - 29.56
p03-f40-v2t2 - - - - - - - -
p03-m40-v1t1 10858.44 116.76 6065.59 0.56 60.06 372.55 0.03 4.19
p03-m40-v1t2 - - - - - 2102.19 - 17.97

253
p03-m40-v2t1 - - - - - 2080.48 - 20.6
p03-m40-v2t2 - - - - - 8504.13 - 75.93
p03-l40-v1t1 - - - - - 728.35 - 6.39
p03-l40-v1t2 - - - - - 2451.58 - 20.43
p03-l40-v2t1 - - - - - 1898.16 - 20.63
p03-l40-v2t2 - - - - - - - -
p07-f40-v1t1 6334.35 86.77 4278.6 0.68 56.3 360.19 0.06 5.22
p07-f40-v1t2 - - - - - 1673.53 - 22.31
p07-f40-v2t1 - - - - - 2131.64 - 27.68
p07-f40-v2t2 - - - - - 11560.98 - 129.9
p07-s40-v1t1 1464.53 23.62 978.07 0.67 16.58 152.88 0.1 2.25
p07-s40-v1t2 9537.53 136.25 6579.21 0.69 96.75 580.9 0.06 7.64
p07-s40-v2t1 3811.24 61.47 2901.68 0.76 46.8 574.03 0.15 8.57
p07-s40-v2t2 - - - - - 3320.35 - 46.77
p07-t40-v1t1 - - - - - 1026.73 - 13.33
p07-t40-v1t2 - - - - - 10006.02 - 125.08
p07-t40-v2t1 - - - - - 5200.24 - 68.42
p07-t40-v2t2 - - - - - - - -
Average 7585.84 100.34 4972.26 0.67 64.07 2819.46 0.07 30.91
 Table A.6: Performance of 2p-ESSPRC and 2p-ESSPRC40 : 25-customer instances
2p-ESPPRC 2p-ESPPRC40
Tot.time # of Time per Tot.time (d/c) # of Time per
Data File (s) (c) iter iter (s) (d) iter iter
p01-f25-v1t1 21.42 68 0.32 12.87 0.6 38 0.34
p01-f25-v1t2 91.44 111 0.82 33.98 0.37 34 1
p01-f25-v2t1 116.77 92 1.27 44.24 0.38 48 0.92
p01-f25-v2t2 515.59 94 5.49 143.92 0.28 52 2.77
p01-l25-v1t1 14.37 49 0.29 7.36 0.51 22 0.33
p01-l25-v1t2 44.32 58 0.76 31.03 0.7 35 0.89
p01-l25-v2t1 68.69 81 0.85 29.54 0.43 27 1.09
p01-l25-v2t2 195.69 70 2.8 95.13 0.49 29 3.28
p03-f25-v1t1 15.19 63 0.24 10.3 0.68 41 0.25
p03-f25-v1t2 60.06 86 0.7 27.4 0.46 35 0.78
p03-f25-v2t1 45.19 69 0.65 27.85 0.62 33 0.84
p03-f25-v2t2 142.52 81 1.76 70.43 0.49 35 2.01
p03-m25-v1t1 24.59 64 0.38 12.88 0.52 25 0.52
p03-m25-v1t2 107.76 82 1.31 45.75 0.42 36 1.27

254
p03-m25-v2t1 66.95 66 1.01 44.76 0.67 29 1.54
p03-m25-v2t2 321.98 78 4.13 173.35 0.54 38 4.56
p03-l25-v1t1 15.32 62 0.25 7.67 0.5 30 0.26
p03-l25-v1t2 37.17 73 0.51 14.91 0.4 27 0.55
p03-l25-v2t1 41.57 59 0.7 31.49 0.76 31 1.02
p03-l25-v2t2 114.12 75 1.52 76.26 0.67 33 2.31
p07-f25-v1t1 26.22 56 0.47 13.41 0.51 26 0.52
p07-f25-v1t2 86.32 72 1.2 49.83 0.58 34 1.47
p07-f25-v2t1 98.88 57 1.73 45.55 0.46 21 2.17
p07-f25-v2t2 293.53 64 4.59 140.89 0.48 27 5.22
p07-s25-v1t1 6.28 45 0.14 3.26 0.52 18 0.18
p07-s25-v1t2 13.71 45 0.3 7.82 0.57 24 0.33
p07-s25-v2t1 19.45 47 0.41 7.29 0.37 18 0.41
p07-s25-v2t2 51.44 59 0.87 27.7 0.54 31 0.89
p07-t25-v1t1 22.49 57 0.39 12.01 0.53 26 0.46
p07-t25-v1t2 50.21 59 0.85 34.9 0.7 38 0.92
p07-t25-v2t1 75.14 52 1.45 38.68 0.51 22 1.76
p07-t25-v2t2 223.61 68 3.29 124.55 0.56 29 4.29
Average 94.62 67.56 1.3 45.22 0.53 31 1.41
 Table A.7: Performance of 2p-ESSPRC and 2p-ESSPRC40 : 40-customer instances

2p-ESPPRC 2p-ESSPRC40
Tot.time # of Time per Tot.time (d/c) # of Time per
Data File (s) (c) iter iter (s) (d) iter iter
p01-f40-v1t1 421.91 82 5.15 237.42 0.56 37 6.42
p01-f40-v1t2 2071.86 90 23.02 1189.89 0.57 38 31.31
p01-f40-v2t1 2403 97 24.77 1407.19 0.59 40 35.18
p01-f40-v2t2 - - - 7580.99 - 44 172.3
p01-l40-v1t1 305.09 84 3.63 185.2 0.61 40 4.63
p01-l40-v1t2 1633.48 108 15.12 1165.47 0.71 55 21.19
p01-l40-v2t1 2627.27 108 24.33 1386.55 0.53 44 31.51
p01-l40-v2t2 - - - 7560.21 - 48 157.5
p03-f40-v1t1 967.01 77 12.56 478.59 0.49 35 13.67
p03-f40-v1t2 7304.94 100 73.05 3020.07 0.41 45 67.11
p03-f40-v2t1 3665.82 124 29.56 1726.98 0.47 47 36.74
p03-f40-v2t2 - - - 1306.66 - 42 31.11
p03-m40-v1t1 372.55 89 4.19 192.03 0.52 37 5.19
p03-m40-v1t2 2102.19 117 17.97 976.05 0.46 44 22.18
p03-m40-v2t1 2080.48 101 20.6 879 0.42 45 19.53
p03-m40-v2t2 8504.13 112 75.93 638.1 0.08 38 16.79
p03-l40-v1t1 728.35 114 6.39 297.81 0.41 55 5.41
p03-l40-v1t2 2451.58 120 20.43 1622.2 0.66 64 25.35
p03-l40-v2t1 1898.16 92 20.63 929.9 0.49 43 21.63
p03-l40-v2t2 - - - 6838.52 - 44 155.42
p07-f40-v1t1 360.19 69 5.22 247.66 0.69 38 6.52
p07-f40-v1t2 1673.53 75 22.31 887 0.53 33 26.88
p07-f40-v2t1 2131.64 77 27.68 935.34 0.44 34 27.51
p07-f40-v2t2 11560.98 89 129.9 7571.9 0.65 45 168.26
p07-s40-v1t1 152.88 68 2.25 69.81 0.46 32 2.18
p07-s40-v1t2 580.9 76 7.64 220.19 0.38 28 7.86
p07-s40-v2t1 574.03 67 8.57 260.34 0.45 26 10.01
p07-s40-v2t2 3320.35 71 46.77 1877.25 0.57 34 55.21
p07-t40-v1t1 1026.73 77 13.33 534.15 0.52 36 14.84
p07-t40-v1t2 10006.02 80 125.08 928.37 0.09 38 24.43
p07-t40-v2t1 5200.24 76 68.42 3886.73 0.75 33 117.78
p07-t40-v2t2 - - - - - - -
Average 2819.46 90.37 30.91 1839.92 0.5 40.71 43.28

255
 Table A.8: 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 25-customer instances
1S-EBP with 2p-ESPPRC 1S-EBP with 2p-ESPPRC-LL + 2p-ESPPRC
Tot.time # of Avg. CG Tot.time # of Avg. CG (b/a)
Data File (s) (a) BP Nd. time (s) (s) (b) BP Nd. time (s)
p01-f25-v1t1 80.86 30 2.58 77.81 25 2.99 0.96
p01-f25-v1t2 14866.19 199 74.19 12488.94 297 41.32 0.84
p01-f25-v2t1 833.58 95 8.67 594.65 99 5.87 0.71
p01-f25-v2t2 1639.84 151 11.06 954.41 119 7.79 0.58
p01-l25-v1t1 36.64 24 1.48 35.45 24 1.43 0.97
p01-l25-v1t2 218.77 201 0.97 173.39 139 1.14 0.79
p01-l25-v2t1 310.93 141 2.08 341.98 157 2.03 1.1
p01-l25-v2t2 527.24 37 14.41 414.51 39 10.48 0.79
p03-f25-v1t1 168.4 282 0.42 182.13 236 0.6 1.08
p03-f25-v1t2 140.62 69 1.89 147.16 77 1.77 1.05
p03-f25-v2t1 334.41 489 0.43 278.07 393 0.5 0.83
p03-f25-v2t2 314.08 49 6.58 316.88 52 5.92 1.01
p03-m25-v1t1 288.67 519 0.18 314.12 509 0.22 1.09
p03-m25-v1t2 x x x x x x x

256
p03-m25-v2t1 165.69 59 2.7 175.58 59 2.87 1.06
p03-m25-v2t2 835.65 9 93.19 523.75 9 57.89 0.63
p03-l25-v1t1 35.35 17 2.02 36.21 17 2.07 1.02
p03-l25-v1t2 106.14 37 2.8 113.9 37 3 1.07
p03-l25-v2t1 168.53 41 4.05 152.39 41 3.66 0.9
p03-l25-v2t2 564.31 41 13.86 452.17 37 12.11 0.8
p07-f25-v1t1 79.43 49 1.58 81.28 55 1.44 1.02
p07-f25-v1t2 6964.47 67 103.82 7874.52 67 117.4 1.13
p07-f25-v2t1 478.86 165 2.84 529.59 229 2.23 1.11
p07-f25-v2t2 1193.31 135 8.99 907.33 197 4.39 0.76
p07-s25-v1t1 63.61 499 0.08 87.77 599 0.1 1.38
p07-s25-v1t2 196.13 115 1.54 176.64 123 1.28 0.9
p07-s25-v2t1 30.92 21 1.42 33.88 21 1.57 1.1
p07-s25-v2t2 209.55 102 2.13 217.97 100 2.1 1.04
p07-t25-v1t1 73.65 35 2.02 51.8 35 1.4 0.7
p07-t25-v1t2 1258.27 1151 0.39 1498.69 1345 0.37 1.19
p07-t25-v2t1 1396.96 1011 0.51 1129.81 777 0.81 0.81
p07-t25-v2t2 789.76 87 9.39 754.9 91 7.92 0.96
Average 1108.74 191.19 12.2 1003.8 193.71 9.83 0.95
 Table A.9: 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 25-customer instances, Cont.
2p-ESPPRC-CS, 2p-ESPPRC 2p-ESPPRC-LL, 2p-ESPPRC-CS, 2p-ESPPRC
Tot.time # of Avg. CG (c/a) Tot.time # of Avg. CG (d/a)
Data File (s) (c) BP Nd. time (s) (s) (d) BP Nd. time (s)
p01-f25-v1t1 78.91 25 3.01 0.98 54.54 31 1.68 0.67
p01-f25-v1t2 11159.51 221 49.85 0.75 16832.67 197 84.85 1.13
p01-f25-v2t1 1815.81 79 22.85 2.18 1462.03 97 14.9 1.75
p01-f25-v2t2 897.41 101 8.68 0.55 1224.16 71 17.02 0.75
p01-l25-v1t1 40.07 26 1.48 1.09 42.03 22 1.84 1.15
p01-l25-v1t2 242.72 171 1.26 1.11 200.56 121 1.54 0.92
p01-l25-v2t1 380.64 155 2.3 1.22 347.62 157 2.06 1.12
p01-l25-v2t2 546.99 37 14.63 1.04 513.4 39 13.01 0.97
p03-f25-v1t1 167.02 256 0.46 0.99 162.08 268 0.42 0.96
p03-f25-v1t2 144.84 69 1.93 1.03 134.75 75 1.64 0.96
p03-f25-v2t1 212.76 289 0.55 0.64 304.99 579 0.32 0.91
p03-f25-v2t2 365.63 50 7.1 1.16 403.99 49 8.05 1.29
p03-m25-v1t1 417.63 659 0.16 1.45 341.56 567 0.19 1.18
p03-m25-v1t2 x x x x x x x x

257
p03-m25-v2t1 216.57 63 3.31 1.31 213.6 59 3.49 1.29
p03-m25-v2t2 576.48 9 63.71 0.69 636.99 9 70.47 0.76
p03-l25-v1t1 34.43 17 1.97 0.97 36.6 17 2.08 1.04
p03-l25-v1t2 126.62 37 3.31 1.19 111.5 35 3.09 1.05
p03-l25-v2t1 185.95 47 3.89 1.1 170.48 43 3.91 1.01
p03-l25-v2t2 552.7 39 14.06 0.98 692.1 45 15.27 1.23
p07-f25-v1t1 90.86 59 1.48 1.14 95.54 57 1.63 1.2
p07-f25-v1t2 17030.53 71 239.67 2.45 7838.97 71 110.28 1.13
p07-f25-v2t1 485.99 205 2.3 1.01 571.31 263 2.1 1.19
p07-f25-v2t2 1076.36 219 4.71 0.9 1435.43 181 7.66 1.2
p07-s25-v1t1 49.33 291 0.13 0.78 65.99 441 0.11 1.04
p07-s25-v1t2 223.36 163 1.15 1.14 191.58 161 0.98 0.98
p07-s25-v2t1 36.5 21 1.68 1.18 32.91 21 1.51 1.06
p07-s25-v2t2 223.8 94 2.3 1.07 239.23 105 2.2 1.14
p07-t25-v1t1 69.15 35 1.89 0.94 76.94 39 1.88 1.04
p07-t25-v1t2 4061.29 2289 0.32 3.23 5506.6 2475 0.29 4.38
p07-t25-v2t1 1157.31 823 0.66 0.83 988.76 781 0.62 0.71
p07-t25-v2t2 1277.6 107 11.58 1.62 1244.91 149 7.91 1.58
Average 1417.57 217 15.24 1.18 1360.45 233.06 12.36 1.19
 Table A.10: Performance with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 40-customer instances
1S-EBP with 2p-ESPPRC 1S-EBP with 2p-ESPPRC-LL + 2p-ESPPRC
Tot.time # of Avg. CG Gap Tot.time # of Avg. CG Gap (b/a)
Data File (s) (a) BP Nd. time (s) % (s) (b) BP Nd. time (s) %
p01-f40-v1t1 19186.19 777 22.94 0 19060.3 935 18.17 0 0.99
p01-f40-v1t2 x x x 0.08 x x x 0.04 x
p01-f40-v2t1 12248.44 271 43.98 0 8910.97 262 32.77 0 0.73
p01-f40-v2t2 28808.63 12 2399.6 4.35 28800 * * 4.28 1
p01-l40-v1t1 29133.11 67 434.37 6.79 28800 * * 6.74 0.99
p01-l40-v1t2 10338.9 47 219.36 0 6701.66 47 142.07 0 0.65
p01-l40-v2t1 10006.77 91 109.51 0 8775.04 159 54.77 0 0.88
p01-l40-v2t2 28898.03 13 2221.55 7.9 18641.02 25 744.9 0 0.65
p03-f40-v1t1 28800 * * 0.01 26697.32 83 321.32 0 0.93
p03-f40-v1t2 23429.75 39 600.35 0 25008.13 40 624.79 0 1.07
p03-f40-v2t1 3048.36 3 1015.07 0 2176.3 3 724.45 0 0.71
p03-f40-v2t2 28876.5 2 14435.53 6.19 28800 * * 5.98 1
p03-m40-v1t1 18977.36 2423 2.33 0 16698.08 2407 1.44 0 0.88
p03-m40-v1t2 29199.38 311 90.15 0.06 28802.28 303 91.56 6.26 0.99

258
p03-m40-v2t1 28813.25 422 64.58 6.6 28800 * * 6.61 1
p03-m40-v2t2 29053.1 16 1814.3 9.12 28800 * * 8.38 0.99
p03-l40-v1t1 6419.83 203 30.93 0 5319.67 202 25.74 0 0.83
p03-l40-v1t2 28800.61 3646 5.06 0.03 28800 * * 0.08 1
p03-l40-v2t1 10177.31 163 61.72 0 10255.35 167 60.59 0 1.01
p03-l40-v2t2 29018.43 13 2230.68 5.63 28837.69 25 1152.22 5.5 0.99
p07-f40-v1t1 2293.58 103 21.94 0 1891.68 103 18.03 0 0.82
p07-f40-v1t2 16511 741 21.42 0 10654.8 303 34.39 0 0.65
p07-f40-v2t1 22830.17 443 50.58 0 12299.16 457 26.11 0 0.54
p07-f40-v2t2 28830.47 18 1600.46 4.67 28821.53 30 959.71 4.56 1
p07-s40-v1t1 469.4 77 5.91 0 440.66 94 4.53 0 0.94
p07-s40-v1t2 886.68 23 38.27 0 618.44 23 26.63 0 0.7
p07-s40-v2t1 14276.36 265 53.15 0 5420.89 239 21.94 0 0.38
p07-s40-v2t2 22036.13 279 78.44 0 15782.07 207 75.79 0 0.72
p07-t40-v1t1 28800 * * 0.16 28840.23 124 232.07 0.27 1
p07-t40-v1t2 28803.76 21 1370.22 2.71 28859.14 42 686.13 2.15 1
p07-t40-v2t1 28800 * * 4.07 28800 * * 4.07 1
p07-t40-v2t2 28000 * * ** 28800 * * 5.8 1.03
Average 20186.18 388.48 1075.64 1.88 18384.27 273.04 264.35 1.9 0.87
 Table A.11: 1S-EBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS: 40-customer instances, Cont.
2p-ESPPRC-CS, 2p-ESPPRC 2p-ESPPRC-LL, 2p-ESPPRC-CS, 2p-ESPPRC
Tot.time # of Avg. CG Gap (c/a) Tot.time # of Avg. CG Gap (d/a)
Data File (s) (c) BP Nd. time (s) % (s) (d) BP Nd. time (s) %
p01-f40-v1t1 26943.98 773 32.75 0 1.4 28836.2 1082 24.04 0.01 1.5
p01-f40-v1t2 - - - 0.07 - - - - 0.24 -
p01-f40-v2t1 11681.88 163 70.54 0 0.95 10523.39 241 42.28 0 0.86
p01-f40-v2t2 28800 12 * 4.35 1 28800 15 * 4.34 1
p01-l40-v1t1 28800 85 * 3.22 0.99 28858.46 160 179.34 6.32 0.99
p01-l40-v1t2 8577.94 44 194.35 0 0.83 7130.93 49 144.79 0 0.69
p01-l40-v2t1 12299.81 107 114.48 0 1.23 6978.62 77 90.16 0 0.7
p01-l40-v2t2 28800 16 * 7.88 1 22797.46 25 910.81 0 0.79
p03-f40-v1t1 28800 49 * 0.01 1 28800 66 * 0.01 1
p03-f40-v1t2 23798.86 39 609.81 0 1.02 26705.83 40 667.15 0 1.14
p03-f40-v2t1 2631.92 3 876.09 0 0.86 909.05 3 302.29 0 0.3
p03-f40-v2t2 28800 3 0 6.19 1 28842.86 6 4805.7 6.07 1
p03-m40-v1t1 24726.6 2499 3.44 0 1.3 20536.5 2395 2.09 0 1.08
p03-m40-v1t2 28800 301 * 4.27 0.99 28800 334 * 3.52 0.99

259
p03-m40-v2t1 28800 383 * 6.61 1 28800 451 * 6.59 1
p03-m40-v2t2 28800 20 * 8.96 0.99 28810.06 23 1251.14 8.57 0.99
p03-l40-v1t1 6309.74 198 31.22 0 0.98 5328.91 197 26.35 0 0.83
p03-l40-v1t2 28800 1735 * 0.08 1 - - - 0.08 0
p03-l40-v2t1 10110.99 165 60.63 0 0.99 8448.23 163 51.11 0 0.83
p03-l40-v2t2 28852.68 17 1696.11 5.58 0.99 28800 22 * 5.53 0.99
p07-f40-v1t1 2122.83 103 20.29 0 0.93 1940.71 103 18.51 0 0.85
p07-f40-v1t2 19084.75 497 37.56 0 1.16 7999.36 261 29.98 0 0.48
p07-f40-v2t1 17727.98 357 48.76 0 0.78 11036.22 419 25.58 0 0.48
p07-f40-v2t2 28871.14 27 1068.45 4.62 1 28800 32 * 4.55 1
p07-s40-v1t1 514.74 98 5.05 0 1.1 424.88 83 4.92 0 0.91
p07-s40-v1t2 924.37 23 39.87 0 1.04 544.78 23 23.41 0 0.61
p07-s40-v2t1 7630.35 255 29.18 0 0.53 4136.89 247 16.14 0 0.29
p07-s40-v2t2 18944.23 231 81.53 0 0.86 13794.52 269 50.83 0 0.63
p07-t40-v1t1 28800 152 * 0.16 1 29211.47 162 179.58 0.07 1.01
p07-t40-v1t2 28800 15 * 3.04 1 28832.01 53 543.16 1.84 1
p07-t40-v2t1 28800 18 * 4.07 1 28800 18 * 4.07 1
p07-t40-v2t2 28800 3 * 5.9 1.03 28800 5 * 5.8 1.03
Average 20198.54 270.68 264.22 2.03 1 18434.24 234.13 426.79 1.8 0.84
 Table A.12: HBP with 2p-ESPPRC-LL and 2p-ESSPRC-CS
2p-ESPPRC-LL 2p-ESPPRC-LL, 2p-ESPPRC-CS-12 2p-ESPPRC-LL, 2p-ESPPRC-CS-20
Best UB Gap time UB Gap time UB Gap time
Data File IP % (s) % (s) % (s)
p01-f40-v1t1 7131 7131 0 1594.47 7131 0 4635.03 7131 0 3191.8
p01-f40-v1t2 6983 6984 0.01 7201.8 6983 0 - 7170 2.68 7202.76
p01-f40-v2t1 6866 6866 0 203.18 6866 0 396.65 6866 0 293.74
p01-f40-v2t2 6818 6818 0 7202.26 6819 0.01 7205.48 6821 0.04 2123.59
p01-l40-v1t1 7159 7159 0 7202.5 7214 0.77 7207.93 7183 0.34 7206.03
p01-l40-v1t2 6934 6934 0 144.57 6934 0 292.24 6934 0 333.1
p01-l40-v2t1 6823 6823 0 51.64 6823 0 141.3 6823 0 145.77
p01-l40-v2t2 6612 6613 0.02 98.7 6621 0.14 339 6633 0.32 506.36
p03-f40-v1t1 8780 8780 0 270.95 8780 0 537.47 8780 0 574.53
p03-f40-v1t2 8753 8753 0 1509.47 8753 0 1559.74 8753 0 1679.57
p03-f40-v2t1 8440 8663 2.64 237.29 8663 2.64 446.09 8663 2.64 467.15
p03-f40-v2t2 8438 8441 0.04 134.27 8439 0.01 872.07 8438 0 690.57
p03-m40-v1t1 7253 7755 6.92 7204.8 7257 0.06 7201.92 7740 6.71 7200.88
p03-m40-v1t2 7075 7075 0 7238.97 7075 0 7205.53 7236 2.28 5037.24

260
p03-m40-v2t1 7003 7132 1.84 6900.02 7003 0 4211.66 7132 1.84 7202.01
p03-m40-v2t2 6947 6967 0.29 7202.03 6947 0 2436.43 6947 0 2234.61
p03-l40-v1t1 7167 7167 0 160.1 7167 0 223.06 7167 0 90.81
p03-l40-v1t2 6950 6950 0 7201.62 6950 0 7202.16 6950 0 7200.81
p03-l40-v2t1 6849 6849 0 130.85 6849 0 190.73 6849 0 232.33
p03-l40-v2t2 6639 6846 3.12 435.09 6846 3.12 535.26 6846 3.12 2341.28
p07-f40-v1t1 7167 7167 0 62.23 7167 0 172.17 7167 0 195.24
p07-f40-v1t2 7001 7001 0 506.83 7001 0 525.29 7001 0 662.4
p07-f40-v2t1 6881 6905 0.35 52.96 6881 0 485.49 6881 0 510.22
p07-f40-v2t2 6845 6850 0.07 762.36 6845 0 1235.29 6845 0 1218.72
p07-s40-v1t1 7343 7343 0 110.02 7343 0 46.67 7343 0 45.76
p07-s40-v1t2 7117 7117 0 37.54 7124 0.1 47.73 7124 0.1 107.18
p07-s40-v2t1 7000 7008 0.11 213.42 7000 0 62.46 7005 0.07 212.83
p07-s40-v2t2 6944 6944 0 444.3 6945 0.01 375.46 6944 0 341.39
p07-t40-v1t1 7012 7015 0.04 796.04 7012 0 504.59 7013 0.01 613.31
p07-t40-v1t2 6972 6972 0 7203.88 6972 0 7206.91 6972 0 4572.11
p07-t40-v2t1 6665 6858 2.9 1099.94 6858 2.9 2239.16 6857 2.88 4090.15
p07-t40-v2t2 6658 6658 0 3465.59 6660 0.03 822.08 6658 0 1566.81
Average 0.57 2408.74 0.31 2147.2 0.72 2190.35
 Table A.13: 2S-EBP with different designs of HBP
2S-EBP With HBP that uses
no 2p-ESPPRC-CS 2p-ESPPRC-CS-12 2p-ESPPRC-CS-20
Data Gap Step 2 (s) Tot (s) Gap Step 2 (s) Tot (s) Gap Step 2 (s) Tot (s)
p01-f40-v1t1 0.06 21600.52 23194.99 0 17884.51 22519.54 0 9002.07 12193.87
p01-f40-v1t2 0.33 21603.02 28804.82 0 - - 0.32 21600 28802.76
p01-f40-v2t1 0 11176.92 11380.1 0 10823.79 11220.44 0 6779.42 7073.16
p01-f40-v2t2 2.03 21780.84 28983.1 2.04 21767.73 28973.21 1.26 21640.84 23764.43
p01-l40-v1t1 3.02 21601.15 28803.65 3.71 21684.45 28892.38 3.22 21622.19 28828.22
p01-l40-v1t2 0 6330.69 6475.26 0 5239.58 5531.82 0 2880.06 3213.16
p01-l40-v2t1 0 9989.3 10040.94 0 9238.55 9379.85 0 6191.15 6336.92
p01-l40-v2t2 0.01 21691.61 21790.31 0 19555.27 19894.27 0 11253.68 11760.04
p03-f40-v1t1 0 9241.16 9512.11 0 21890.4 22427.87 0 5868.19 6442.72
p03-f40-v1t2 0 21572.76 23082.23 0 18963.12 20522.86 0 9276.91 10956.48
p03-f40-v2t1 0 991.78 1229.07 0 1016.06 1462.15 0 528.56 995.71
p03-f40-v2t2 0.49 21606.8 21741.07 0.38 21645.1 22517.17 0 8293.47 8984.04
p03-m40-v1t1 0.41 21610.1 28814.9 0.08 21600.67 28802.59 0 21604.08 28804.96
p03-m40-v1t2 0.08 21702.66 28941.63 0.08 21884.13 29089.66 2.95 2186.5 7223.74

261
p03-m40-v2t1 2.14 21626.98 28527 0.33 21604.12 25815.78 2.79 2163.77 9365.78
p03-m40-v2t2 3.01 21682.05 28884.08 2.59 22169.5 24605.93 4.18 2175.46 4410.07
p03-l40-v1t1 0 5515.75 5675.85 0 5640.92 5863.98 0 3334.91 3425.72
p03-l40-v1t2 0.05 21607.14 28808.76 0.04 21600.13 28802.29 0.05 21600 28800.81
p03-l40-v2t1 0 10199.23 10330.08 0 10642.46 10833.19 0 6828.42 7060.75
p03-l40-v2t2 3.14 21607.06 22042.15 3.2 21604.89 22140.15 0 17453.7 19794.98
p07-f40-v1t1 0 1927.39 1989.62 0 1824.84 1997.01 0 1325.01 1520.25
p07-f40-v1t2 0 7798.67 8305.5 0 10076.08 10601.37 0 4527.12 5189.52
p07-f40-v2t1 0.01 23958.09 24011.05 0 12047.83 12533.32 0 8941.95 9452.17
p07-f40-v2t2 2.53 21623.87 22386.23 2.46 21651.7 22886.99 2.34 21607.55 22826.27
p07-s40-v1t1 0 326.04 436.06 0 274.13 320.8 0 263.38 309.14
p07-s40-v1t2 0 516.54 554.08 0 662.59 710.32 0 390.17 497.35
p07-s40-v2t1 0 11463.4 11676.82 0 7868.75 7931.21 0 7716.19 7929.02
p07-s40-v2t2 0 16099.52 16543.82 0 14168.82 14544.28 0 12673.15 13014.54
p07-t40-v1t1 0.07 22408.12 23204.16 0.03 21614.97 22119.56 0 8870.59 9483.9
p07-t40-v1t2 0.94 21922.1 29125.98 1.21 21608.45 28815.36 0.35 21627.34 26199.45
p07-t40-v2t1 2.85 21700 22799.94 2.85 22600 24839.16 0.04 23464.62 27554.77
p07-t40-v2t2 1.72 22390.12 25855.71 1.76 21764.63 22586.71 1.64 21800.96 23367.77
Average 0.72 15839.73 18248.47 0.65 15245.75 17392.94 0.6 10484.11 12674.45
 Table A.14: Performance of I-Heur, BB, HBP: 25 customer instances
Upper Bound Values Time (s)
I-Heur BB in HBP HBP I-Heur BB in HBP HBP
Data HBP Rt W BB WO BB HBP Rt W BB WO BB
p01-f25-v1t1 4795 4722 4699 4699 0.02 10.5 18.32 9.26
p01-f25-v1t2 4795 4734 4511 4511 0.05 40.54 166.61 223.87
p01-f25-v2t1 4722 4701 4427 4425 0.02 27.7 31.09 19.49
p01-f25-v2t2 4722 4710 4425 4425 0.04 34.41 55.99 53.75
p01-l25-v1t1 5193 4963 4899 4947 0.04 2.59 10.56 12.87
p01-l25-v1t2 4968 4827 4815 4815 0.03 3.54 17.66 44.54
p01-l25-v2t1 4809 4785 4759 4759 0.04 0.87 13.82 21.26
p01-l25-v2t2 4762 4762 4650 4650 0.04 1.7 9.61 9.19
p03-f25-v1t1 5281 5082 5062 5062 0.06 6.06 67.16 71.24
p03-f25-v1t2 5139 5086 4837 4837 0.04 65.71 96.88 33.25
p03-f25-v2t1 5048 4785 4759 4759 0.04 1.54 73.3 621.59
p03-f25-v2t2 4831 4831 4742 4742 0.04 19.04 34.77 19.04
p03-m25-v1t1 4995 4853 4785 4785 0.04 28.28 101.27 208
p03-m25-v1t2 4869 4801 4785 4785 0.05 47.89 7200 7200

262
p03-m25-v2t1 4992 4726 4689 4689 0.04 4.56 21.79 18.63
p03-m25-v2t2 4767 4761 4548 4548 0.03 11.55 26.04 18.11
p03-l25-v1t1 5186 5073 4921 4921 0.04 5.31 10.34 8.7
p03-l25-v1t2 4871 4848 4803 4803 0.03 6.73 13.29 6.87
p03-l25-v2t1 4805 4805 4741 4741 0.02 1.28 6.65 6.34
p03-l25-v2t2 4805 4794 4734 4734 0.03 3.4 18.51 19.98
p07-f25-v1t1 4870 4768 4761 4761 0.03 1.51 9.34 8.67
p07-f25-v1t2 4823 4791 4761 4761 0.04 8.92 37.16 23.87
p07-f25-v2t1 4721 4721 4703 4703 0.02 1.96 72.08 40.57
p07-f25-v2t2 4721 4576 4482 4482 0.03 0.92 24.08 49.14
p07-s25-v1t1 5412 5103 5098 5098 0.04 1.34 17.34 52.67
p07-s25-v1t2 5238 5049 4889 4889 0.04 7.89 25.86 17.88
p07-s25-v2t1 5159 5030 4787 4787 0.04 8.57 11.8 3.88
p07-s25-v2t2 4934 4856 4781 4781 0.04 5.38 21.35 18.4
p07-t25-v1t1 5134 5128 4898 4898 0.03 8.1 15.01 7.63
p07-t25-v1t2 5162 5121 4886 4886 0.04 35.68 373.5 1690.54
p07-t25-v2t1 5025 4800 4771 4767 0.04 1.32 144.38 710.75
p07-t25-v2t2 4769 4769 4654 4607 0.03 3.4 14.76 13.25
Average 0.04 12.76 273.76 351.98
 Table A.15: Performance of I-Heur, BB, HBP: 40-customer instances
Upper Bound Values Time (s)
I-Heur BB in HBP HBP I-Heur BB in HBP HBP
Data HBP Rt W BB WO BB HBP Rt W BB WO BB
p01-f40-v1t1 7170 7170 7131 7131 0.05 10.7 3191.8 3128.83
p01-f40-v1t2 7170 7170 7170 7170 0.04 241.72 7202.76 7206.14
p01-f40-v2t1 7064 7064 6866 6866 0.06 20.32 293.74 286.94
p01-f40-v2t2 6984 6984 6821 6821 0.05 48.86 2123.59 2049.08
p01-l40-v1t1 7449 7339 7183 7179 0.07 218.09 7206.03 7206.02
p01-l40-v1t2 7236 7199 6934 6934 0.06 128.43 333.1 223.53
p01-l40-v2t1 7166 6970 6823 6823 0.04 9.21 145.77 149.28
p01-l40-v2t2 7166 7133 6633 6638 0.07 86.56 506.36 763.49
p03-f40-v1t1 9253 9013 8780 8780 0.04 45.13 574.53 722.31
p03-f40-v1t2 9028 8760 8753 8753 0.04 218.16 1679.57 1535.31
p03-f40-v2t1 8943 8922 8663 8663 0.06 98.67 467.15 353.33
p03-f40-v2t2 8943 8702 8438 8438 0.04 50.65 690.57 659.91
p03-m40-v1t1 7827 7740 7740 7827 0.06 2303.69 7200.88 7203.83
p03-m40-v1t2 7544 7503 7236 7236 0.1 3637.17 5037.24 2286.2

263
p03-m40-v2t1 7479 7479 7132 7132 0.06 449.2 7202.01 7200
p03-m40-v2t2 7430 7430 6947 6947 0.11 1339 2234.61 940.01
p03-l40-v1t1 7386 7193 7167 7167 0.05 10.42 90.81 95.16
p03-l40-v1t2 7212 7212 6950 6950 0.05 72.29 7200.81 7201.88
p03-l40-v2t1 7121 7112 6849 6849 0.05 16.84 232.33 209.03
p03-l40-v2t2 7018 7018 6846 6846 0.05 187.95 2341.28 2226.85
p07-f40-v1t1 7424 7424 7167 7167 0.04 86.14 195.24 117.71
p07-f40-v1t2 7340 7246 7001 7001 0.05 78.8 662.4 684.56
p07-f40-v2t1 7174 7173 6881 6881 0.05 25.72 510.22 677.89
p07-f40-v2t2 7000 6942 6845 6709 0.05 36.09 1218.72 1025.11
p07-s40-v1t1 7860 7367 7343 7343 0.05 4.92 45.76 46.34
p07-s40-v1t2 7656 7372 7124 7124 0.05 39.11 107.18 87.86
p07-s40-v2t1 7466 7441 7005 7005 0.04 137.35 212.83 87.12
p07-s40-v2t2 7241 7001 6944 6944 0.05 14.8 341.39 503.35
p07-t40-v1t1 7283 7283 7013 7012 0.05 12.78 613.31 555.03
p07-t40-v1t2 7058 7058 6972 6972 0.05 20.21 4572.11 4536.15
p07-t40-v2t1 6945 6945 6857 6857 0.04 15.51 4090.15 1401.86
p07-t40-v2t2 6945 6945 6658 6669 0.04 83.95 1566.81 2277.89
Average 0.05 304.64 2190.35 1989
 Table A.16: I-Heur, BB, HBPs Upper Bounds vs. LP Relaxations & Optimal (Best) IPs: 25-customer
%Gap btw UB & Optimal (or Best IP) %Gap btw UB & the LP Relax.
I-Heur BB in HBP HBP I-Heur BB in HBP HBP
Data HBP Rt W BB WO BB HBP Rt W BB WO BB
p01-f25-v1t1 2.04 0.49 0 0 6.45 4.83 4.32 4.32
p01-f25-v1t2 6.3 4.94 0 0 8.34 6.96 1.93 1.93
p01-f25-v2t1 6.71 6.24 0.05 0 8.15 7.67 1.39 1.35
p01-f25-v2t2 6.71 6.44 0 0 9.93 9.65 3.02 3.02
p01-l25-v1t1 6 1.31 0 0.98 7.08 2.33 1.01 2
p01-l25-v1t2 3.18 0.25 0 0 5.13 2.15 1.9 1.9
p01-l25-v2t1 1.05 0.55 0 0 3 2.49 1.93 1.93
p01-l25-v2t2 3.48 3.48 1.04 1.04 4.96 4.96 2.49 2.49
p03-f25-v1t1 4.33 0.4 0 0 9.79 5.66 5.24 5.24
p03-f25-v1t2 6.24 5.15 0 0 9.31 8.19 2.89 2.89
p03-f25-v2t1 6.09 0.57 0.02 0.02 9.59 3.88 3.32 3.32
p03-f25-v2t2 1.88 1.88 0 0 7.29 7.29 5.32 5.32
p03-m25-v1t1 4.39 1.42 0 0 6.34 3.31 1.87 1.87
p03-m25-v1t2 1.76 0.33 0 0 6.35 4.87 4.52 4.52

264
p03-m25-v2t1 6.46 0.79 0 0 11.28 5.35 4.53 4.53
p03-m25-v2t2 6.43 6.3 1.54 1.54 8.44 8.3 3.45 3.45
p03-l25-v1t1 7.1 4.77 1.63 1.63 8.15 5.79 2.62 2.62
p03-l25-v1t2 1.42 0.94 0 0 4.3 3.81 2.84 2.84
p03-l25-v2t1 1.35 1.35 0 0 3.98 3.98 2.6 2.6
p03-l25-v2t2 3.22 2.99 1.7 1.7 6.6 6.35 5.02 5.02
p07-f25-v1t1 2.29 0.15 0 0 4.9 2.7 2.55 2.55
p07-f25-v1t2 2.95 2.26 1.62 1.62 6.13 5.43 4.77 4.77
p07-f25-v2t1 0.38 0.38 0 0 5.25 5.25 4.85 4.85
p07-f25-v2t2 5.33 2.1 0 0 7.58 4.28 2.13 2.13
p07-s25-v1t1 6.16 0.1 0 0 8.24 2.06 1.96 1.96
p07-s25-v1t2 7.14 3.27 0 0 7.95 4.06 0.76 0.76
p07-s25-v2t1 7.77 5.08 0 0 8.85 6.13 1 1
p07-s25-v2t2 3.2 1.57 0 0 7.03 5.34 3.71 3.71
p07-t25-v1t1 4.82 4.7 0 0 6.15 6.03 1.27 1.27
p07-t25-v1t2 5.65 4.81 0 0 9.76 8.89 3.89 3.89
p07-t25-v2t1 5.41 0.69 0.08 0 8.15 3.31 2.69 2.6
p07-t25-v2t2 3.85 3.85 1.35 0.33 5.67 5.67 3.12 2.08
Average 4.41 2.48 0.28 0.28 7.19 5.22 2.97 2.96
 Table A.17: I-Heur, BB, HBPs Upper Bounds vs. LP Relaxations & Optimal (Best) IPs: 40-customer
%Gap btw UB & Optimal (or Best IP) %Gap btw UB & the LP Relax.
I-Heur BB in HBP HBP I-Heur BB in HBP HBP
Data HBP Rt W BB WO BB HBP Rt W BB WO BB
p01-f40-v1t1 0.55 0.55 0 0 2.78 2.78 2.22 2.22
p01-f40-v1t2 2.68 2.68 2.68 2.68 5.04 5.04 5.04 5.04
p01-f40-v2t1 2.88 2.88 0 0 5.43 5.43 2.47 2.47
p01-f40-v2t2 2.43 2.43 0.04 0.04 6.12 6.12 3.65 3.65
p01-l40-v1t1 4.05 2.51 0.34 0.28 7.96 6.37 4.11 4.05
p01-l40-v1t2 4.36 3.82 0 0 7.08 6.54 2.61 2.61
p01-l40-v2t1 5.03 2.15 0 0 8.3 5.34 3.12 3.12
p01-l40-v2t2 8.38 7.88 0.32 0.39 10.02 9.51 1.83 1.91
p03-f40-v1t1 5.39 2.65 0 0 6.39 3.63 0.95 0.95
p03-f40-v1t2 3.14 0.08 0 0 5.24 2.11 2.03 2.03
p03-f40-v2t1 5.96 5.71 2.64 2.64 5.97 5.72 2.65 2.65
p03-f40-v2t2 5.98 3.13 0 0 6.59 3.71 0.57 0.57
p03-m40-v1t1 7.91 6.71 6.71 7.91 11.01 9.77 9.77 11.01
p03-m40-v1t2 6.63 6.05 2.28 2.28 9.21 8.62 4.75 4.75

265
p03-m40-v2t1 6.8 6.8 1.84 1.84 10.08 10.08 4.98 4.98
p03-m40-v2t2 6.95 6.95 0 0 11.64 11.64 4.38 4.38
p03-l40-v1t1 3.06 0.36 0 0 6.73 3.94 3.56 3.56
p03-l40-v1t2 3.77 3.77 0 0 6.5 6.5 2.64 2.64
p03-l40-v2t1 3.97 3.84 0 0 6.88 6.74 2.79 2.79
p03-l40-v2t2 5.71 5.71 3.12 3.12 7.23 7.23 4.6 4.6
p07-f40-v1t1 3.59 3.59 0 0 5.26 5.26 1.62 1.62
p07-f40-v1t2 4.84 3.5 0 0 6.28 4.92 1.37 1.37
p07-f40-v2t1 4.26 4.24 0 0 6.35 6.34 2.01 2.01
p07-f40-v2t2 4.34 3.47 2.03 0 5.59 4.72 3.25 1.2
p07-s40-v1t1 7.04 0.33 0 0 8.32 1.53 1.19 1.19
p07-s40-v1t2 7.57 3.58 0.1 0.1 7.97 3.96 0.46 0.46
p07-s40-v2t1 6.66 6.3 0.07 0.07 7.82 7.46 1.16 1.16
p07-s40-v2t2 4.28 0.82 0 0 7.07 3.52 2.68 2.68
p07-t40-v1t1 3.86 3.86 0.01 0 4.93 4.93 1.04 1.03
p07-t40-v1t2 1.23 1.23 0 0 3.69 3.69 2.43 2.43
p07-t40-v2t1 4.2 4.2 2.88 2.88 4.44 4.44 3.12 3.12
p07-t40-v2t2 4.31 4.31 0 0.17 6.27 6.27 1.88 2.04
Average 4.74 3.63 0.78 0.76 6.88 5.75 2.84 2.82
 Table A.18: Performance of BB at Root Node of 1S-EBP: 25-customer

1S-EBP, BB at Root 1S-EBP WO BB Time

Data File BBIP-Opt Gap Nds Time Nds Time BB/WOBB
p01-f25-v1t1 0.19 25 95.11 25 77.81 1.22
p01-f25-v1t2 4.54 231 14091.47 297 12488.94 1.13
p01-f25-v2t1 0 91 697.91 99 594.65 1.17
p01-f25-v2t2 0.54 101 1233.77 119 954.41 1.29
p01-l25-v1t1 2.84 24 42.06 24 35.45 1.19
p01-l25-v1t2 0 123 191.48 139 173.39 1.1
p01-l25-v2t1 0.65 157 382.42 157 341.98 1.12
p01-l25-v2t2 1.3 39 491.26 39 414.51 1.19
p03-f25-v1t1 0.4 278 214.85 236 182.13 1.18
p03-f25-v1t2 0.74 81 187.65 77 147.16 1.28
p03-f25-v2t1 4.77 445 391.13 393 278.07 1.41
p03-f25-v2t2 0.72 48 393.62 52 316.88 1.24
p03-m25-v1t1 4.39 509 360.64 509 314.12 1.15
p03-m25-v1t2 0.4 - - - - -
p03-m25-v2t1 0.66 59 223.55 59 175.58 1.27
p03-m25-v2t2 6.43 9 620.86 9 523.75 1.19
p03-l25-v1t1 4.07 17 43.37 17 36.21 1.2
p03-l25-v1t2 0.5 37 118.88 37 113.9 1.04
p03-l25-v2t1 0.27 41 160.12 41 152.39 1.05
p03-l25-v2t2 2.34 37 519.68 37 452.17 1.15
p07-f25-v1t1 0 47 83.08 55 81.28 1.02
p07-f25-v1t2 1.66 67 10990.51 67 7874.52 1.4
p07-f25-v2t1 0.26 229 569.53 229 529.59 1.08
p07-f25-v2t2 0.29 185 1010.26 197 907.33 1.11
p07-s25-v1t1 0.1 501 79.39 599 87.77 0.9
p07-s25-v1t2 4.79 123 202.86 123 176.64 1.15
p07-s25-v2t1 0.08 21 35.61 21 33.88 1.05
p07-s25-v2t2 0 97 237.78 100 217.97 1.09
p07-t25-v1t1 0.1 41 66.95 35 51.8 1.29
p07-t25-v1t2 0.78 593 687.63 1345 1498.69 0.46
p07-t25-v2t1 0.08 957 1478.93 777 1129.81 1.31
p07-t25-v2t2 3.85 91 859.93 91 754.9 1.14
Average 1.49 171.1 1185.88 193.71 1003.8 1.15

266
 Table A.19: Performance of BB at Root Node of 1S-EBP: 40-customer
1S-EBP, BB at Root 1S-EBP WO BB Time
Data File BBIP-Opt Gap Nds Gap Time Nds Gap Time BB/WOBB
p01-f40-v1t1 0.55 935 0 19701.81 935 0 19060.3 1.03
p01-f40-v1t2 2.68 1208 0.04 - - 0.04 - -
p01-f40-v2t1 2.04 262 0 12093.16 262 0 8910.97 1.36
p01-f40-v2t2 1.38 14 3.34 28842.48 * 4.28 28900 1
p01-l40-v1t1 0.28 174 3.24 28881.47 * 6.74 28900 1
p01-l40-v1t2 0.48 43 0 7159.8 47 0 6701.66 1.07
p01-l40-v2t1 0.13 143 0 9707.42 159 0 8775.04 1.11
p01-l40-v2t2 3.64 25 0 22005.02 25 0 18641.02 1.18
p03-f40-v1t1 2.32 58 0.01 28900 83 0 26697.32 1.08
p03-f40-v1t2 0 39 0 27551.15 40 0 25008.13 1.1
p03-f40-v2t1 2.67 3 0 2435.82 3 0 2176.3 1.12
p03-f40-v2t2 0.05 6 0.51 28840.91 * 5.98 28900 1
p03-m40-v1t1 3.65 2521 0 21399.49 2407 0 16698.08 1.28
p03-m40-v1t2 7.31 231 6.9 28825.64 303 6.26 28802.28 1
p03-m40-v2t1 * 1 * 28900 * 6.61 28900 1

267
p03-m40-v2t2 0 26 2.16 28832.76 * 8.38 28900 1
p03-l40-v1t1 0.43 199 0 5869.16 202 0 5319.67 1.1
p03-l40-v1t2 0.65 3381 0.08 - * 0.08 28900 -
p03-l40-v2t1 3.59 167 0 12038.32 167 0 10255.35 1.17
p03-l40-v2t2 3.68 23 3.66 28812.64 25 5.5 28837.69 1
p07-f40-v1t1 1.14 101 0 1938.89 103 0 1891.68 1.02
p07-f40-v1t2 3.44 391 0 12329.6 303 0 10654.8 1.16
p07-f40-v2t1 3.07 529 0 16487.51 457 0 12299.16 1.34
p07-f40-v2t2 3.2 27 3.57 28831.65 30 4.56 28821.53 1
p07-s40-v1t1 2.89 94 0 535.39 94 0 440.66 1.21
p07-s40-v1t2 2.89 23 0 629.55 23 0 618.44 1.02
p07-s40-v2t1 2.81 247 0 8643.47 239 0 5420.89 1.59
p07-s40-v2t2 0.65 295 0 18052.95 207 0 15782.07 1.14
p07-t40-v1t1 3.14 156 0.07 28900 124 0.27 28840.23 1
p07-t40-v1t2 1.16 41 2.08 28854.13 42 2.15 28859.14 1
p07-t40-v2t1 3.66 18 3.57 28900 * 4.07 28900 1
p07-t40-v2t2 0.56 5 2.19 28823.02 * 5.8 28900 1
Average 2.07 1.05 19124.11 1.9 18410.08 1.1
 Table A.20: 1S-EBP vs 2S-EBP: Effect of HBP, 25-Customer Instances

Nds CPU Time (s) Time

Data File 1S-EBP 2S-EBP 1S-EBP 2S-EBP 1S-EBP/2S-EBP
p01-f25-v1t1 25 27 77.81 72.82 1.07
p01-f25-v1t2 297 125 12488.94 21876.01 0.57
p01-f25-v2t1 99 107 594.65 802.14 0.74
p01-f25-v2t2 119 159 954.41 1120.17 0.85
p01-l25-v1t1 24 25 35.45 34.65 1.02
p01-l25-v1t2 139 119 173.39 165.53 1.05
p01-l25-v2t1 157 145 341.98 329.88 1.04
p01-l25-v2t2 39 39 414.51 522.43 0.79
p03-f25-v1t1 236 235 182.13 241.56 0.75
p03-f25-v1t2 77 121 147.16 178.85 0.82
p03-f25-v2t1 393 743 278.07 603.42 0.46
p03-f25-v2t2 52 47 316.88 298.62 1.06
p03-m25-v1t1 509 357 314.12 340.02 0.92
p03-m25-v1t2 x x x x x
p03-m25-v2t1 59 59 175.58 180.23 0.97
p03-m25-v2t2 9 9 523.75 538.48 0.97
p03-l25-v1t1 17 17 36.21 36.8 0.98
p03-l25-v1t2 37 37 113.9 109.61 1.04
p03-l25-v2t1 41 51 152.39 161.1 0.95
p03-l25-v2t2 37 43 452.17 543.47 0.83
p07-f25-v1t1 55 37 81.28 79.15 1.03
p07-f25-v1t2 67 61 7874.52 13085.64 0.6
p07-f25-v2t1 229 181 529.59 466.23 1.14
p07-f25-v2t2 197 137 907.33 880.35 1.03
p07-s25-v1t1 599 305 87.77 65.79 1.33
p07-s25-v1t2 123 99 176.64 173.24 1.02
p07-s25-v2t1 21 23 33.88 41.57 0.82
p07-s25-v2t2 100 85 217.97 208.18 1.05
p07-t25-v1t1 35 47 51.8 96.07 0.54
p07-t25-v1t2 1345 3323 1498.69 10426.22 0.14
p07-t25-v2t1 777 1033 1129.81 2039.87 0.55
p07-t25-v2t2 91 127 754.9 1065.28 0.71
Average 1003.8 1831.72 0.87

268
 Table A.21: 1S-EBP vs 2S-EBP: Effect of HBP, 40-customer instances
Gap % Nds CPU Time (s) Time
Data File 1S-EBP 2S-EBP 1S-EBP 2S-EBP 1S-EBP 2S-EBP 2S-EBP/2S-EBP
p01-f40-v1t1 0 0 935 1047 19060.3 24008.84 0.79
p01-f40-v1t2 0.04 0.14 x 1259 x 28805.21 x
p01-f40-v2t1 0 0 262 185 8910.97 9143.19 0.97
p01-f40-v2t2 4.28 1.99 * 15 28900 24308.9 1.19
p01-l40-v1t1 6.74 3.29 * 179 28900 29026.46 1
p01-l40-v1t2 0 0 47 43 6701.66 5697.16 1.18
p01-l40-v2t1 0 0 159 97 8775.04 8983.17 0.98
p01-l40-v2t2 0 0 25 25 18641.02 17638.15 1.06
p03-f40-v1t1 0 0.01 83 58 26697.32 22503.4 1.19
p03-f40-v1t2 0 0 40 39 25008.13 21788.5 1.15
p03-f40-v2t1 0 0 3 3 2176.3 1776.48 1.23
p03-f40-v2t2 5.98 0.46 * 5 28900 22358.28 1.29
p03-m40-v1t1 0 0.41 2407 2191 16698.08 28812.76 0.58
p03-m40-v1t2 6.26 2.34 303 182 28802.28 26668.97 1.08
p03-m40-v2t1 6.61 2.56 * 205 28900 28812.92 1

269
p03-m40-v2t2 8.38 2.59 * 21 28900 24508.88 1.18
p03-l40-v1t1 0 0 202 199 5319.67 5650.57 0.94
p03-l40-v1t2 0.08 0.04 * 1471 28900 28801.69 1
p03-l40-v2t1 0 0 167 161 10255.35 10288.34 1
p03-l40-v2t2 5.5 3.08 25 24 28837.69 24050.68 1.2
p07-f40-v1t1 0 0 103 105 1891.68 2034.98 0.93
p07-f40-v1t2 0 0 303 427 10654.8 6579.34 1.62
p07-f40-v2t1 0 0 457 243 12299.16 11315.02 1.09
p07-f40-v2t2 4.56 2.46 30 26 28821.53 22903.7 1.26
p07-s40-v1t1 0 0 94 77 440.66 393.97 1.12
p07-s40-v1t2 0 0 23 23 618.44 687.58 0.9
p07-s40-v2t1 0 0 239 245 5420.89 7570.69 0.72
p07-s40-v2t2 0 0 207 259 15782.07 17281.98 0.91
p07-t40-v1t1 0.27 0.03 124 168 28840.23 23413.31 1.23
p07-t40-v1t2 2.15 1.06 42 42 28859.14 26690.69 1.08
p07-t40-v2t1 4.07 0.07 * 18 28900 26890.15 1.07
p07-t40-v2t2 5.8 1.73 * 3 28900 23295.55 1.24
Average 1.9 0.7 18410.08 17584.05 1.07
 Table A.22: Cover and GAP Inequalities for 25-customer Instances

1S-EBP-WO 1S-EBP with CUTGENG 1S-EBP with CUTGENK

Data file Nds Nds Time W/WO # cuts Nds Time W/WO # cuts
p01-f25-v1t1 28 26 1.01 1 24 1.11 5
p01-f25-v1t2 81 87 1.05 1 87 0.98 2
p01-f25-v2t1 25 25 0.96 0 33 1.05 1
p01-f25-v2t2 29 29 1 0 41 0.97 3
p01-l25-v1t1 29 29 0.95 0 29 0.95 0
p01-l25-v1t2 69 69 0.99 0 71 1.02 1
p01-l25-v2t1 127 127 0.98 0 132 0.94 5
p01-l25-v2t2 33 33 0.96 1 33 0.99 1
p03-f25-v1t1 241 233 1.04 9 189 0.88 32
p03-f25-v1t2 60 60 0.99 5 56 1.27 20
p03-f25-v2t1 151 193 1.03 2 153 1.08 19
p03-f25-v2t2 44 40 1.03 5 36 1.14 14
p03-m25-v1t1 205 273 1.34 5 205 1.15 15
p03-m25-v1t2 2655 x x x x 0 x
p03-m25-v2t1 53 49 1.01 5 43 1.14 10
p03-m25-v2t2 9 9 1.07 0 9 1.02 0
p03-l25-v1t1 15 15 0.99 0 15 1.1 1
p03-l25-v1t2 38 38 0.93 0 36 0.99 1
p03-l25-v2t1 40 40 0.97 0 41 0.98 1
p03-l25-v2t2 37 37 1.04 0 37 1.06 4
p07-f25-v1t1 69 69 0.98 0 44 0.95 1
p07-f25-v1t2 67 67 0.98 0 67 1.01 0
p07-f25-v2t1 158 158 1.02 0 158 0.97 0
p07-f25-v2t2 143 143 1.06 0 143 1.02 0
p07-s25-v1t1 263 253 0.98 1 193 1.03 5
p07-s25-v1t2 81 81 0.97 0 81 0.99 0
p07-s25-v2t1 20 20 1.05 0 20 1.01 1
p07-s25-v2t2 85 83 0.97 1 81 1 5
p07-t25-v1t1 43 43 1.02 0 41 1.08 1
p07-t25-v1t2 463 313 0.85 1 x x x
p07-t25-v2t1 165 165 1.03 0 149 1.1 1
p07-t25-v2t2 91 81 0.94 2 69 0.9 7
Average 1.01 1

270
 Table A.23: Cover and GAP Inequalities for 25-customer Instances: Cont.

Base Alg. Alg. with CUTGENK+G

Data file Nds Nds Time W/WO # gap cuts # cover cuts
p01-f25-v1t1 28 24 1.13 1 5
p01-f25-v1t2 81 99 1.25 1 2
p01-f25-v2t1 25 33 1.12 0 1
p01-f25-v2t2 29 41 0.9 0 3
p01-l25-v1t1 29 29 0.97 0 0
p01-l25-v1t2 69 71 1.06 0 1
p01-l25-v2t1 127 132 1.02 0 5
p01-l25-v2t2 33 33 0.99 1 1
p03-f25-v1t1 241 194 0.9 1 34
p03-f25-v1t2 60 56 1.3 1 19
p03-f25-v2t1 151 195 1.13 1 19
p03-f25-v2t2 44 36 1.14 0 14
p03-m25-v1t1 205 197 1.14 2 13
p03-m25-v1t2 2655 x x x x
p03-m25-v2t1 53 43 0.97 2 9
p03-m25-v2t2 9 9 1.03 0 0
p03-l25-v1t1 15 15 1.12 0 1
p03-l25-v1t2 38 36 0.98 0 1
p03-l25-v2t1 40 41 0.98 0 1
p03-l25-v2t2 37 37 1.05 0 4
p07-f25-v1t1 69 44 0.97 0 1
p07-f25-v1t2 67 67 1.05 0 0
p07-f25-v2t1 158 158 0.99 0 0
p07-f25-v2t2 143 143 1.01 0 0
p07-s25-v1t1 263 207 1.03 1 5
p07-s25-v1t2 81 81 1.01 0 0
p07-s25-v2t1 20 20 1.08 0 1
p07-s25-v2t2 85 81 0.99 0 5
p07-t25-v1t1 43 41 1.12 0 1
p07-t25-v1t2 463 x x x x
p07-t25-v2t1 165 149 1.16 0 1
p07-t25-v2t2 91 69 0.9 0 7
Average 1.05

271
Vita

Name: Zeliha Akça.

Place and year of birth: Cankiri, Turkey, 1979.

Education

• Ph.D. in Industrial and Systems Engineering, Lehigh University, Bethlehem, PA, December

2009.

• M.S. in Management Science, Lehigh University, Bethlehem, May 2005.

• B.S. in Industrial Engineering, Middle East Technical University, Ankara, Turkey, June

2002.

Professional Experience

• Research Assistant (Project Development), Lehigh University, Industrial and Systems Eng.

Dept. August 2008 August 2009.

• Research Assistant, Lehigh University, Industrial and Systems Eng. Dept. May 2003 Au-

gust 2008.

• Instructor, Software Tools Course, Lehigh University, Industrial and Systems Eng. Dept.

Spring 2003, Fall2005, Spring 2006, Spring 2007, Fall 2007.

272
 • Teaching Assistant, Lehigh University, Industrial and Systems Eng. Dept. August 2002 -

May 2003.

• Summer Intern, Eczacibasi Securities, Turkey, Summer 2001.

• Summer Intern, Vestel Electronics, Manisa, Turkey, Summer 2000.

Publications

• L. Snyder, Z. Akca, Pool Points for Peeps: A network Optimization Model for a Candy

Supply Chain, MSOM 2009 Supply Chain Management SIG Conference.

• Z. Akca, R. T. Berger, T. K. Ralphs, A Branch-and-Price Algorithm for Combined Loca-

tion and Routing Problems Under Capacity Restrictions, in the Proceedings of the Eleventh

INFORMS Computing Society Meeting, 2008.

• Z. Akca, R. T. Berger, T. K. Ralphs, Modeling and Solving Location Routing and Schedul-

ing Problems, COR@L Laboratory Technical Report, 2008 (submitted for publication).

Honors and Certification

• Awarded with 1st place in Student Poster Competition in Center For Value Chain Research,

Spring Symposium, May 2009.

• Certificates of Excellence & Completion for the Teacher Development Program, June 2008.

• Awarded with Graduate Student of the Year by Lehigh University, Industrial Eng. Dept.,

May 2006.

• Dean’s High Honor List in Middle East Technical University, Turkey, every semester from

Fall 1998 to Spring 2002.

• Ranked 63rd in Science and 57th in Mathematics among about 1.5 million high school

students in Nationwide University Entrance Exam in Turkey, August 1997.

273