Handbook of Test Problems in

Local and Global Optimization

Nonconvex Optimization and Its Applications
Volume 33

Managing Editors:
Panos Pardalos
University of Florida, U.S.A.

Reiner Horst
University ofTrier, Germany

Advisory Board:
Ding-Zhu Du
University of Minnesota, U.S.A.

C. A. Floudas
Princeton University, U.SA.

J. Mockus
Stanford University, U.S.A.

H. D. Sherali
Virginia Polytechnic Institute and State University, U.S.A.

The titfes pubfished in this se ries are Iisted at the end of this vofwlle.
Handbook of Test Problems
in Local and Global
Optimization

by

Christodoulos A. Floudas
Department of Chemical Engineering,
Princeton University,
Princeton, N.}., U.S.A.

Panos M. Pardalos
Department of Industrial and Systems Engineering,
University of Florida

and

Claire S. Adjiman
William R. Esposito
Zeynep H. Gm~
Stephen T. Harding
John L. Klepeis
Clifford A. Meyer
Carl A. Schweiger
Department of Chemical Engineering,
Princeton University.
Princeton, N.}., U.S.A.

Springer-Science+Business Media, B.V.

A C.I.P. Catalogue record for this book is available from the Library of Congress.

ISBN 978-1-4419-4812-0 ISBN 978-1-4757-3040-1 (eBook)

DOI 10.1007/978-1-4757-3040-1

Printed on acid-free paper

All Rights Reserved

1999 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 1999.
Softcover reprint ofthe hardcover 1st edition 1999
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical.
inc1uding photocopying. recording or by any information storage and
retrieval system. without written permission from the copyright owner
Preface

Significant research activities have taken place in the areas of local and global
optimization in the last two decades. Many new theoretical, computational,
algorithmic, and software contributions have resulted. It has been realized
that despite these numerous contributions, there does not exist a systematic
forum for thorough experimental computational testing and evaluation of the
proposed optimization algorithms and their implementations.
Well-designed nonconvex optimization test problems are of major impor-
tance for academic and industrial researchers interested in algorithmic and
software development. It is remarkable that eventhough nonconvex models
dominate all the important application areas in engineering and applied sci-
ences, there is only a limited dass of reported representative test problems.
This book reflects our long term efforts in designing a benchmark database
and it is motivated primarily from the need for nonconvex optimization test
problems. The present collection of benchmarks indudes test problems from
literature studies and a large dass of applications that arise in several branches
of engineering and applied science.

C.A. Floudas
P.M. Pardalos
C.S. Adjiman
W.R. Esposito
Z.H. Gm
S.T. Harding
J.L. Klepeis
C.A. Meyer
C.A. Schweiger

Princeton, New Jersey

February 1999

v
Contents

1 Introduction 1

2 Quadratic Programming Problems 5

2.1 Introduction .. 5
2.2 Test Problem 1 5
2.3 Test Problem 2 6
2.4 Test Problem 3 7
2.5 Test Problem 4 8
2.6 Test Problem 5 10
2.7 Test Problem 6 11
2.8 Test Problem 7 12
2.9 Test Problem 8 15
2.10 Test Problem 9 16
2.11 Test Problem 10 18

3 Quadratically Constrained Problems 21

3.1 Introduction .. 21
3.2 Test Problem 1 21
3.3 Test Problem 2 23
3.4 Test Problem 3 24
3.5 Test Problem 4 25

4 Univariate Polynomial Problems 27

4.1 Introduction .. 27
4.2 Test Problem 1 27
4.3 Test Problem 2 28
4.4 Test Problem 3 28
4.5 Test Problem 4 29
4.6 Test Problem 5 29
4.7 Test Problem 6 30
4.8 Test Problem 7 30
4.9 Test Problem 8 30
4.10 Test Problem 9 31
vii
viii CONTENTS

5 Bilinear problems 33
5.1 Introduction.. 33
5.2 Pooling Problems . . 34
5.2.1 Introduction 34
5.2.2 Haverly Pooling Problem 34
5.2.3 Ben-Tal et al. (1994) Problems: General Formulation 36
5.2.4 Ben-Tal et al. (1994) Problems: Test Problem 1 38
5.2.5 Ben-Tal et al. (1994) Problems: Test Problem 2 40
5.3 Distillation Column Sequencing Problems . . . . . . . . 43
5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . 43
5.3.2 Nonsharp separation of propane, isobutane and n-butane 44
5.3.3 Nonsharp separation of propane, isobutane, n-butane
and isopentane . . . . . . . . 46
5.4 Heat Exchanger Network Problems. 51
5.4.1 Introduction . 51
5.4.2 Test Problem 1 51
5.4.3 Test Problem 2 52
5.4.4 Test Problem 3 54

6 Biconvex and (D.C.) Problems 59

6.1 Introduction . . . . . . . . . . . . . . . . . . 59
6.2 Phase and Chemical Equilibrium Problems 59
6.2.1 Introduction . . . . . . . . 59
6.2.2 Mathematical Formulation 60
6.3 Biconvex Problems . . . 64
6.3.1 NRTL Equation 64
6.3.2 Test Problem 1 66
6.3.3 Test Problem 2 67
6.3.4 Test Problem 3 68
6.3.5 Test Problem 4 69
6.4 Difference of Convex Functions (D.C.) Problems 70
6.4.1 UNIQUAC Equation . 70
6.4.2 Test Problem 5 71
6.4.3 Test Problem 6 72
6.4.4 Test Problem 7 73
6.4.5 Test Problem 8 74
6.4.6 UNIFAC Equation 75
6.4.7 Test Problem 9 . 76
6.4.8 Test Problem 10 77
6.4.9 Test Problem 11 78
6.4.10 ASOG Equation 79
6.4.11 Test Problem 12 81
6.4.12 Test Problem 13 81
6.4.13 Modified Wilson Equation . 82
CONTENTS IX

6.4.14 Test Problem 14 . . . . . . . . 84

7 Generalized Geometrie Programming 85

7.1 Introduction. . . . . . . . . . . . . . . ....... 85
7.2 Literature Problems ......... . . . . . . . 86
7.2.1 Test Problem 1 : Alkylation process design 86
7.2.2 Test Problem 2 : CSTR Sequence Design 89
7.2.3 Test Problem 3 : Heat exchanger design . . 90
7.2.4 Test Problem 4 : Optimal Reactor Design . 91
7.2.5 Test Problem 5 : Colville's Test Problem 92
7.2.6 Test Problem 6 93
7.2.7 Test Problem 7 . 94
7.2.8 Test Problem 8 94
7.2.9 Test Problem 9 . 95
7.2.10 Test Problem 10 96
7.3 Robust Stability Analysis 97
7.3.1 Test Problem 11 98
7.3.2 Test Problem 12 99
7.3.3 Test Problem 13 100
7.3.4 Test Problem 14 101
7.3.5 Test Problem 15 102
7.3.6 Test Problem 16 103

8 Twiee Continuously Differentiable NLPs 107

8.1 Introduction . . . . . . 107
8.2 Literature Problems ......... 108
8.2.1 Test Problem 1 . . . . . . . . . 108
8.2.2 Test Problem 2: Pseudoethane 109
8.2.3 Test Problem 3: Goldstein and Price function . 110
8.2.4 Test problem 4: Three-hump camelback function 110
8.2.5 Test Problem 5: Six-hump Camelback Function . 111
8.2.6 Test Problem 6: Shekel Function 111
8.2.7 Test Problem 7 . . . . . . . . . . 112
8.2.8 Test Problem 8 . . . . . . . . . . 113
8.3 Batch Plant Design Under Uncertainty . 114
8.3.1 Introduction .......... 114
8.3.2 Single-Product Campaign Formulation . 115
8.3.3 Test Problem 1 118
8.3.4 Test Problem 2 . . . . . . . . . . . . . . 121
8.3.5 Test Problem 3 . . . . . . . . . . . . . . 123
8.3.6 Unlimited Intermediate Storage Formulation 124
8.3.7 Test Problem 4 . . . . . . . . . 126
8.3.8 Test Problem 5 . . . . . . . . . 127
8.4 Chemical Reactor Network Problems. 128
8.4.1 Introduction ......... 128
x CONTENTS

8.4.2 General Formulation . . 129

8.4.3 Specific Information . . 132
8.4.4 Problem Gharacteristics 133
8.4.5 Test Problems . . . . . 134
8.4.6 Test Problem 1 : Nonisothermal Van de Vusse Reaction
Gase I . . . . . . . . . . . . . . . . . . . . . . . . . . .. 134
8.4.7 Test Problem 2 : Isothermal Van de Vusse Reaction Gase 1139
8.4.8 Test Problem 3: Isothermal Van de Vusse Reaction Gase 11141
8.4.9 Test Problem 4: Isothermal Van de Vusse Reaction Gase
111 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 143
8.4.10 Test Problem 5: Isothermal Van de Vusse Reaction Gase
IV . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 145
8.4.11 Test Problem 6 : Isothermal Trambouze Reaction .. 147
8.4.12 Test Problem 7 : Isothermal Denbigh Reaction Gase I 149
8.4.13 Test Problem 8 : Isothermal Denbigh ReactionGase 11 151
8.4.14 Test Problem 9 : Isothermal Levenspiel Reaction . .. 153
8.4.15 Test Problem 10 : a-Pinene Reaction . . . . . . . . . 155
8.4.16 Test Problem 11 : Nonisothermal Van de Vusse Reaction
Gase 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.4.17 Test Problem 12 : Nonisothermal Naphthalene Reaction 159
8.4.18 Test Problem 13 : Nonisothermal Parallel Reactions 162
8.4.19 Test J>roble~ 14 : Sulfur Dioxide Oxidation. 164
8.5 Parameter Estimation problems. 166
8.5.1 Introduction . . . . . . . . . . 166
8.5.2 General formulation . . . . . . 167
8.5.3 Test Problem 1 : Linear Model 167
8.5.4 Test Problem 2 : Polynomial Model 169
8.5.5 Test Problem 3 : Non-linear Model. 169
8.5.6 Test Problem 4: Respiratory Mechanical Model . 171
8.5.7 Test Problem 5: Kowalik Problem . . . . 172
8.5.8 Test Problem 6: Pharmacokinetic Model. . . . . 173
8.5.9 Test Problem 7: Steady-State GSTR . . . . . . . 174
8.5.10 Test Problem 8: Vapor-Liquid Equilibrium Model 176
8.6 Phase and Ghemical Equilibrium Problems . . . . . . . . 178
8.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . 178
8.6.2 General formulation - Tangent Plane Distance Minimiza-
tion . . . . . . . . . . . . 179
8.6.3 Van der Waals Equation . 180
8.6.4 Test Problem 1 180
8.6.5 Test Problem 2 181
8.6.6 SRK Equation 182
8.6.7 Test Problem 3 182
8.6.8 Test Problem 4 183
8.6.9 Peng-Robinson Equation. 184
CONTENTS Xl

8.6.10 Test Problem 5 . . . . . . 185

8.6.11 Test Problem 6 . . . . . . 186
8.7 Clusters of Atoms and Molecules 186
8.7.1 Introduction ....... 186
8.7.2 General Formulation . . . 187
8.7.3 Lennard-Jones Potential. 188
8.7.4 Morse Potential. . 193
8.7.5 Tersoff Potential ..... 197
8.7.6 Brenner Potential .... 199
8.7.7 Bolding-Andersen Potential 202

9 Bilevel Programming Problems 205

9.1 Introduction . . . . . . . . . . . 205
9.1.1 Terminology and Properties . 206
9.1.2 Solution Techniques ..... 206
9.2 Bilevel Linear Programming Problems 207
9.2.1 Karush-Kuhn-Tucker Approach. 208
9.2.2 Test Problem 1 208
9.2.3 Test Problem 2 210
9.2.4 Test Problem 3 211
9.2.5 Test Problem 4 212
9.2.6 Test Problem 5 213
9.2.7 Test Problem 6 214
9.2.8 Test Problem 7 215
9.2.9 Test Problem 8 216
9.2.10 Test Problem 9 218
9.2.11 Test Problem 10 219
9.3 Bilevel Quadratic Programming Problems 220
9.3.1 Introduction 220
9.3.2 Test Problem 1 221
9.3.3 Test Problem 2 222
9.3.4 Test Problem 3 223
9.3.5 Test Problem 4 225
9.3.6 Test Problem 5 226
9.3.7 Test Problem 6 227
9.3.8 Test Problem 7 228
9.3.9 Test Problem 8 229
9.3.10 Test Problem 9 230

10 Complementarity Problems 233

10.1 Introduction . . . . . . . . . 233
10.2 Linear Complementarity Problems 234
10.2.1 Test Problem 1 . 234
10.2.2 Test Problem 2 . 235
10.2.3 Test Problem 3 . 236
xii CONTENTS

10.2.4 Bimatrix Games .......... 236

10.2.5 Test Problem 4 . . . . . . . . . . . 237
10.2.6 Equilibrium Transportation Model 238
10.2.7 Test Problem 5 . . . . . . . . 239
10.2.8 Trafik Equilibrium Problem. . 240
10.2.9 Test Problem 6 . . . . . . . . . 241
10.3 Nonlinear Complementarity Problems 242
10.3.1 Test Problem 1 243
10.3.2 Test Problem 2 . . 243
10.3.3 Test Problem 3 . . 244
10.3.4 Nash Equilibrium. 245
10.3.5 Test Problem 4 .. 246
10.3.6 Test Problem 5 . . 246
10.3.7 Invariant Capital Stock Problem 247
10.3.8 Test Problem 6 . . . . . . . . . 248

11 Semidefinite Programming Problems 251

11.1 Introduction . . . . . . . . . . . . . . . 251
11.1.1 Problem formulation . . . . . . 251
11.1.2 Semidefinite Programming Applications 252
11.2 Educational Testing Problem 254
11.2.1 General formulation 255
11.3 Maximum Cut Problem .. 258
11.3.1 General formulation 259

12 Mixed-Integer Nonlinear Problems 263

12.1 Introduction . . . . . . 263
12.2 Literature Problems . 264
12.2.1 Test Problem 1 264
12.2.2 Test Problem 2 265
12.2.3 Test Problem 3 266
12.2.4 Test Problem 4 267
12.2.5 Test Problem 5 267
12.2.6 Test Problem 6 268
12.3 Heat Exchanger Network Synthesis 269
12.3.1 Introduction . . . . 269
12.3.2 General Formulation . . . . 270
12.3.3 Test Problem 1 . . . . . . . 273
12.4 Heat Exchanger Networks: Arithmetic Mean 278
12.4.1 Introduction . . . . 278
12.4.2 General Formulation 278
12.4.3 Test Problem 1 . 280
12.4.4 Test Problem 2 . 282
12.5 Pump network synthesis 285
12.5.1 Introduction .. 285
CONTENTS Xlll

12.5.2 Test Problem 1 . 286

12.6 Trim Loss Minimization 289
12.6.1 Introduction .. 289
12.6.2 General Formulation 290
12.6.3 Test Problem 1 291
12.6.4 Test Problem 2 294
12.6.5 Test Problem 3 296
12.6.6 Test Problem 4 299

13 Combinatorial Optimization Problems 303

13.1 Modeling with Integer Programming 303
13.1.1 Test problems in the Internet 304
13.2 Quadratic Integer Programming . . . 304
13.2.1 Quadratic 0-1 Test problems 304
13.3 Satisfiability Problems . . . . . . 306
13.3.1 SAT Test Problems . . . 307
13.4 The Traveling Salesman Problem 308
13.4.1 TSP test Problems. 309
13.5 Assignment Problems . . . 309
13.5.1 QAP Test Problems 310
13.6 Graph Coloring . . . . . . . 313
13.6.1 Test Problems . . . 314
13.7 Maximum Clique Problem. 314
13.7.1 Maximum Clique: Coding Theory Test Problems 317
13.7.2 Maximum Clique: Keller Graphs 317
13.8 Steiner Problems in Networks (SPN) 318
13.8.1 Test Problems . . . . . 319

14 Nonlinear Systems of Equations 325

14.1 Literature problems . . . . . . . . . . . . . . . . 326
14.1.1 Test Problem 1: Himmelblau function . . 326
14.1.2 Test Problem 2: Equilibrium Combustion 327
14.1.3 Test Problem 3 328
14.1.4 Test Problem 4 328
14.1.5 Test Problem 5 329
14.1.6 Test Problem 6 329
14.1.7 Test Problem 7 331
14.1.8 Test Problem 8 332
14.1.9 Test Problem 9 333
14.2 Enclosing All Homogeneous Azeotropes 334
14.2.1 Introduction . . . . . . . . . . . 334
14.2.2 General Formulation - Activity Coefficient Equations . 335
14.2.3 Wilson Equation 337
14.2.4 Test Problem 1 337
14.2.5 Test Problem 2 . 338
xiv CONTENTS

14.2.6 Test Problem 3 . 339

14.2.7 NRTL Equation 340
14.2.8 Test Problem 4 . 341
14.2.9 Test Problem 5 . 341
14.2.lOUNIQUAC Equation . 342
14.2.11 Test Problem 6 .. 343
14.2.12 Test Problem 7 .. 344
14.2.13 UNIFAC Equation 345
14.2.14 Test Problem 8 346
14.2.15 Test Problem 9 .. 347

15 Dynamic Optimization Problems 351

15.1 Introduction . . . . . . . . . . . . 351
15.1.1 General Formulation .. . 351
15.1.2 Solution Techniques (Local) 352
15.1.3 Control Parameterization .. 352
15.1.4 Solution of NLP /DAE . . . . 354
15.2 Chemical Reactor Network Problems . 354
15.2.1 Introduction . . . . 354
15.2.2 General Formulation . . 355
15.2.3 Specific Information . . 363
15.2.4 Problem Characteristics 364
15.2.5 Test Problems . . . . . 364
15.2.6 Test Problem 1 : Nonisothermal Van de Vusse Reaction
Case I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
15.2.7 Test Problem 2 : Isothermal Van de Vusse Reaction Case 1369
15.2.8 Test Problem 3 : Isothermal Van de Vusse Reaction Case Ir371
15.2.9 Test Problem 4 : Isothermal Van de Vusse Reaction Case
111 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
15.2.10 Test Problem 5 : Isothermal Van de Vusse Reaction Case
IV . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 374
15.2.11 Test Problem 6 : Isothermal Trambouze Reaction . . . 376
15.2.12 Test Problem 7 : Isothermal Denbigh Reaction Case I . 377
15.2.13 Test Problem 8 : Isothermal Denbigh Reaction Case Ir. 379
15.2.14 Test Problem 9 : Isothermal Levenspiel Reaction . . . . 381
15.2.15 Test Problem 10 : a-Pinene Reaction . . . . . . . . . . 383
15.2.16 Test Problem 11 : Nonisothermal Van de Vusse Reaction
Case Ir . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
15.2.17 Test Problem 12 : Nonisothermal Naphthalene Reaction 387
15.2.18 Test Problem 13 : Nonisothermal Parallel Reactions 389
15.2.19 Test Problem 14 : Sulfur Dioxide Oxidation. 391
15.3 Parameter Estimation Problems 393
15.3.1 Introduction . . . . 393
15.3.2 General Formulation . . . 394
CONTENTS xv

15.3.3 Test Problem 1 . . . . . . . . . . . . . . . . . . . 394

15.3.4 Test Problem 2 . . . . . . . . . . . . . . . . . . . 396
15.3.5 Test Problem 3 : Catalytic Cracking of Gas on . 398
15.3.6 Test Problem 4 : Bellman's Problem . . . . . . . 399
15.3.7 Test Problem 5 : Methanol-to-Hydrocarbons Process . 401
15.3.8 Test Problem 6 : Lotka-Volterra Problem 403
15.4 Optimal Control Problems. 404
15.4.1 Introduction . . . . . . . . . . . . . . . . 404
15.4.2 Test Problem 1 . . . . . . . . . . . . . . . 405
15.4.3 Test Problem 2 : Singular Control Problem 406
15.4.4 Test Problem 3 : CSTR Problem . . . . . . 407
15.4.5 Test Problem 4 : Oil Shale Pyrolysis . . . . 408
15.4.6 Test Problem 5 : Bifunctional Catalyst Blend Problem. 410
Chapter 1

Introduction

During the last two decades, a significant growth has taken place in algorithmic
and software development of local and global optimization methods for a va-
riety of classes of nonlinear, discrete, and dynamic mathematical problems.
These problems include (i) multi-quadratic programming, (ii) bi linear and
biconvex, (iii) generalized geometrie programming, (iv) general constrained
nonlinear optimization, (v) bilevel optimization, (vi) complementarity, (vii)
semidefinite programming, (viii) mixed-integer nonlinear optimization, (ix)
combinatorial optimization, and (x) optimal control problems. Relative to
these advances there have been very limited efforts in establishing a system-
atic benchmark framework for the evaluation of the algoritms and their im-
plementations (Hock and Schittkowski (1981); Floudas and Pardalos (1990);
Bongartz et al. (1995)). A well-designed experimental computational testing
framework is of primary importance in identifying the merits of each algorithm
and implementation.
The principal objective of this book is to present a collection of challeng-
ing test problems arising in literat ure studies and a wide spectrum of appli-
cations. These applications include : poolingjblending operations, heat ex-
changer network synthesis, phase and chemical reaction equilibrium, robust
stability analysis, batch plant design under uncertainty, chemie al reactor net-
work synthesis, parameter estimation and data reconcilliation, conformational
problems in clusters of atoms and molecules, pump network synthesis, trim
loss minimization, homogeneous azeotropic separation system, dynamic opti-
mization problems in parameter estimation and in reactor network synthesis,
and optimal control problems.
This book reflects our long term efforts in establishing a benchmark database
of nonconvex optimization problems. These efforts started a decade ago with
the book by Floudas and Pardalos (1990) which introduced the first collection
of nonconvex test problems for constrained global optimization algorithms.
Several approaches have been proposed to address the challenging task
of testing and benchmarking local and global optimization algorithms and
software. In regard to the test problems presented in this book, the following

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
2 CHAPTER 1. INTRODUCTION

approaehes have been eonsidered.

Collections of randomly generated test problems with known solution.

Problem instanees with eertain eharacteristies that have been used to test
some aspeets of speeifie algorithms.

Colleetion of real-world problems, that is, problems that model a variety of

praetical applieations.

Furthermore, there exist different types of specifie problem instanees that

are used to test some aspects of an algorithm. Such types of test problems
include :

Worst ease test problems : For example aglobai optimization test problem
with an exponential number of loeal minima ean be used to check the
eflicieney of an algorithm based on loeal searehes or simulated annealing
methods.

Standard test problems: A test problem beeomes standard if it is used fre-

quently. Standard test problems are usually small dimension problems
published in papers to illustrate the main steps of a partieular algorithm.
Most of the referenees listed at the end of this book eontain standard test
problems.
This book eontains many noneonvex optimization test problems that model
a diverse range of praetical applications. The main eriteria in selecting such
test problems have been (a) the size (ranging from small to medium to large),
(b) the mathematical properties (exhibiting different types of noneonvexities),
and (e) the degree of diflieulty (resulting from the wide range of applieations).
Chapter 2 presents test problems in quadratic programming. Chapter 3
diseusses multi-quadratic programming problems, and Chapter 4 presents uni-
variate polynomial test problems. Chapter 5 deals with bilinear problems and
their applieations in poolingjblending operations, distillation eolumn sequene-
ing, and heat exehanger networks. Chapter 6 addresses bieonvex and (D.C.)
optimization problems with applieations in phase and ehemieal reaetion equi-
librium. Chapter 7 introduces generalized geometrie programming problems
with applieations in design and robust stability analysis. Chapter 8 foeuses on
the large class of twice-eontinuously differentiable NLPs and presents literature
test problems and applieations in the areas of bateh design under uneertainty,
ehemical reaetor networks, parameter estimation, phase and ehemieal equilib-
rim with equations of state, and clusters of atoms and moleeules. Chapter
9 presents test problems for bilevel optimization and foeuses on bilevel lin-
ear and bilevel quadratic programming problems. Chapter 10 diseusses test
problems for linear and nonlinear eomplementarity. Chapter 11 introduces
test problems in the area of semidefinite programming. Chapter 12 addresses
 3

mixed-integer nonlinear optimization test problems and presents literat ure test
problems and applications in heat exchanger network synthesis, pump network
synthesis, and trim loss minimization problems. Chapter 13 intro duces several
combinatorial optimization problems. Chapter 14 presents test problems of
nonlinear constrained systems of equations with applications in homogeneous
azeotropic systems. Finally, Chapter 15 intro duces dynamic optimization test
problems with applications in reactor networks, parameter estimation, and
optimal control.
The input files of the algebraic test problems presented in this book are
available in the GAMS modeling language format, while the input files of
the differential-algebraic test problems are available in the MINOPT modeling
language format. Both types of input files of the aforementioned test problems
are at the following internet address :

http://titan.princeton.edu/TestProblems

These input files can also be accessed via anonymous ftp as follows :

ftp titan.princeton.edu <enter>

name: anonymous <enter>
passvord: (your e-mail address) <enter>
cd pub/TestProblems <enter>
get filename.gms
quit
Chapter 2

Quadratic Programming
Problems

2.1 Introduction
In this chapter nonconvex quadratic programming test problems are consid-
ered. These test problems have a quadratic objective function and linear con-
straints. Quadratic programming has numerous applications (Pardalos and
Rosen (1987), Floudas and Visweswaran (1995)) and plays an important role
in many nonlinear programming methods. Recent methods of generating chal-
lenging quadratic programming test problems and disjointly constrained bi-
linear programming test problems can be found in the work of Vicente et al.
(1992) and Calamai et al. (1993). Furthermore, a very broad class of difficult
combinatorial optimization problems such as integer programming, quadratic
assignment, and the maximum clique problem can be formulated as nonconvex
quadratic programming problems.

2.2 Test Problem 1

Objective function

min I(z) = cT Z - 0.5z T Qz

Z

Constraints

20Xl + 12x2 + llX3 + 7X4 + 4X5 < 40

Variable bounds

O~z~l

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
6 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

Data

c = {42,
(42, 44, 45, 47, 47.5)T

Q = 1001

where 1 is the identity matrix.

Problem Statistics
No. of continuous variables 5
No. of linear inequalities 1
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: -17

Continuous variables

a: = {I, 1,0,1, of.

(1,1,0,1,

2.3 Test Problem 2

Objective function

min /(a:, y)
a:,y
= eTa: - 0.5a: T Qa: + dy

Constraints

6Xl + 3X2 + 3X3 + 2X4 + X5 ::;< 6.5

lOXl + lOx3 + Y
y < 20

Variable bounds

0::;a:::;1
O::;y
2.4. TEST PROBLEM 3 7

Data

e = (-10.5, -7.5, -3.5, -2.5, -1.5f

d=-l0

Q = 1001
where 1 is the identity matrix.

Problem Statistics
No. of continuous variables 6
No. of linear inequalities 2
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: -413
Continuous variables

x = (0,1,0,1, l)T

y= 20.

2.4 Test Problem 3

Objective function
min f(x, y) = e T x - 0.5x T Qx
x,y
+ dT Y

Constraints
2Xl + 2X2 + Y6 + Y7 < 10
2Xl + 2X3 + Y6 + Ys < 10
2X2 + 2X3 + Y7 + Ys < 10
-8Xl + Y6 < 0
-8X2 +Y7 < 0
- 8X 3 + Ys < 0
- 2X4 - Yl + Y6 < 0
-2Y2 - Y3 + Y7 < 0
-2Y4-Y5+YS < 0
8 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

Variable bounds

0 < a:::;1
0 < Yi ::; 1 i = 1,2,3,4,5,9
0 < Yi ::; 3 i = 6,7,8

Data

c = (5,5,5,5f

d = (-1, -1, -1, -1, -1, -1, -1, -1, -1f

Q = 1001

where 1 is the identity matrix.

Problem Statistics
No. of continuous variables 13
No. of linear inequalities 9
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Continuous variables

a: = (1,1,1, 1)T

y = (1,1,1,1,1,3,3,3,1)T.

2.5 Test Problem 4

Objective function

min f(x, y) = 6.5x - 0.5x 2 - Yl - 2Y2 - 3Y3 - 2Y4 - Y5

a:,y
2.5. TEST PROBLEM 4 9

Constraints

Az < b
Z (x,y)T

Variable bounds

o:::;a: < 1
Yi < 1 i = 3,4
Y5 < 2

Data
A is the following (5x6) matrix:

-~ 1
2 8 1 3
-8-4 -2 2 4
( 1
A= 2 0.5 0.2 -3 -1 -4
0.2 2 0.1 -4 2 2
-0.1 -0.5 2 5 -5 3

b = (16, -1,24,12, 3)T

Problem Statistics
No. of continuous variables 6
No. of linear inequalities 5
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: -11.005

Continuous variables

x=o

y = (6,0,1, 1,of.
10 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

2.6 Test Problem 5

Objective function

min I(x, y) = cT X - 0.5xT Qx + dT Y

x,y

Constraints

Az < b
z (x,yf

Variable bounds

O~z~1

Data
A is the following (l1xlO) matrix:

-2 -6 -1 -3 -3 -2 -6 ':"'2 -2
6 -5 8 -3
1 3 8 9 -3
-5
9
-8
6
5
5
0
7 -4
3
-9
8
1
-5 -9
-8
-8
9 2
3 -9
1 -7 -1

-9 -3
-9

3 -2
A= -7 -5 -2
-6 -6 -7 -6 7 7

4
1 -3 -3 -4 -1 -4 1 6
1 -2 6 9 -7 9 -9 -6
-4 6 7 2 2 6 6 -7 4
1 1 1 1 1 1 1 1 1 1
-1 -1 -1 -1 -1 -1 -1 -1 -1 -1

b = (-4,22, -6, -23, -12, -3, 1, 12, 15, 9, - l f

d = (10,10, lOf

c = (-20, -80, -20, -50, -60, -90, O)T

Q= 101

where 1 is the identity matrix.

2.7. TEST PROBLEM 6 11

Problem Statistics
No. of continuous variables 10
No. of linear inequalities 11
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: -268.01463

Continuous variables

x = (1,0.90755,0,1,0.71509,1, O)T

y = (0.91698,1, 1f.

2.7 Test Problem 6

Objective function

Constraints

Ax < b
x E !RIO

Variable bounds

0:::;x:::;1

Data
A is the following (5xlO) matrix:

( -2
-6 -1 0 -3 -3 -2 -6 -2

A~ -~
-5
6
8 -3
5 3
0
8
1
-8
3
9
8
2
9 -3
0 -9
-2)
5 0 -9 1 -8 3 -9 -9 -3
-8 7 -4 -5 -9 1 -7 -1 3 -2
12 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

b = (-4 , 22 , -6 , -23 , -12)T

c= (48,42,48,45,44, 41, 47, 42, 45, 46)T

Q = 1001
where 1 is the identity matrix.

Problem Statistics
No. of continuous variables 10
No. of linear inequalities 5
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: -39

Continuous variables

x = (I,O,O,I,I,I,O,I,I,I)T.

2.8 Test Problem 7

This dass of the test problems belongs to separable concave quadratic pro-
gramming problems.
Objective function

nMn f(x) = -0.5 L Ai(Xi - (Yi)2

i

Constraints

Ax < b
x E lR20

Variable bounds
2.8. TEST PROBLEM 7 13

Data

b = (-5,2, -1, -3,5,4, -1,0,9, 40)T

The sets of parameters involve the }.'s which correspond to a set of non-
negative eigenvalues, and the a's. These parameter values will be provided for
five cases along with the corresponding best known optimum solutions in the
following sections.
Ais a (lOx20) matrix and AT is the following (20xlO) matrix:

-3 7 0 -5 1 1 0 2 -1 1
7 0 -5 1 1 0 2 -1 -1 1
0 -5 1 1 0 2 -1 -1 -9 1
-5 1 1 0 2 -1 -1 -9 3 1
1 1 0 2 -1 -1 -9 3 5 1
1 0 2 -1 -1 -9 3 5 0 1
0 2 -1 -1 -9 3 5 0 0 1
2 -1 -1 -9 3 5 0 0 1 1
-1 -1 -9 3 5 0 0 1 7 1
-1 -9 3 5 0 0 1 7 -7 1
A= -9 3 5 0 0 1 7 -7 -4 1
3 5 0 0 1 7 -7 -4 -6 1
5 0 0 1 7 -7 -4 -6 -3 1
0 0 1 7 -7 -4 -6 -3 7 1
0 1 7 -7 -4 -6 -3 7 0 1
1 7 -7 -4 -6 -3 7 0 -5 1
7 -7 -4 -6 -3 7 0 -5 1 1
-7 -4 -6 -3 7 0 -5 1 1 1
-4 -6 -3 7 0 -5 1 1 0 1
-6 -3 7 0 -5 1 1 0 2 1

Problem Statistics
No. of continuous variables 20
No. of linear inequalities 10
No. of convex inequalities
No. of nonlinear equalities
0

Global Solution

Case 1 : }.i = 1, ai = 2 for i = 1, ... , 20

Objective function: -394.7506
14 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

Continuous variables

x = (0,0,28.8024,0,0,4.1792,0,0,0,0,0,0,0,0,
0.6188,4.0933,0,2.3064,0,0)T

Case 2 : Ai = 1, ai = -5 for i = 1, ... ,20

Objective function: -884.75058

Continuous variables

x = (0,0,28.8024,0,0,4.1792,0,0,0,0,0,0,0,0,
0.6188,4.0933,0,2.3064,0,of

Case 3 : Ai = 20, ai
= for i = 1, ... ,20

Objective function: -8695.01193

Continuous variables

x = (0,0,28.8024,0,0,4.1792,0,0,0,0,0,0,0,0,
0.6188,4.0933,0,2.3064,0,of

Case 4 : Ai = 1, ai = 8 for i = 1, ... , 20

Objective function: -754.75062

Continuous variables

x = (0,0,28.8024,0,0,4.1792,0,0,0,0,0,0,0,0,
0.6188,4.0933,0,2.3064,0,of

Case 5 : Ai = i, ai = 2 for i = 1, ... ,20

Objective function: -4150.4101

Continuous variables

x = (0,0,1.04289,0,0,0,0,0,0,0,1.74674,0,0.43147,0,0,
4.43305,0,15.85893,0, 16.4889)T.
2.9. TEST PROBLEM 8 15

2.9 Test Problem 8

This dass of global optimization problems belongs to minimum concave cost
trasportation problems.
Objective function
m n
m~n f(x) = L L(CijXij + dijx;j)
i=l j=l

Constraints
m
L Xij = bj j = 1, ... , n
i=l
n
L Xij ai i = 1, ... , m
j=l

Variable bounds

Data
The parameters dij , ai, and bj satisfy the following conditions :

dij < 0 i = 1, ... ,m, j = 1, ... ,n

m n
Lai Lbj i = 1, ... ,m, j = 1, ... ,n
i=l j=l

n=4

m=6

a = (8,24,20,24,16, 12f

b = (29,41,13, 21f

C = (Cij) is the following (6x4) matrix:

16 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

300 270 460 800

740 600 540 380
300 490 380 760
C= 430 250 390 600
210 830 470 680
360 290 400 310

D = (dij) is the following (6x4) matrix:

-7 -4 -6 -8
-12 -9 -14 -7
-13 -12 -8 -4
D= -7 -9 -16 -8
-4 -10 -21 -13
-17 -9 -8 -4

Problem Statistics
No. of continuous variables 24
No. of linear equalities 10
No. of convex inequalities
No. of nonlinear equalities

The resulting model contains (n + m) equality constraints and (n x m) vari-

ables. There is exactly one redundant equality constraint. When any one of
the constraints is dropped the remaining constraints form a linearly indepen-
dent system of constraints, which results in a basic vector for the trasportation
problem consisting of (n + m - 1) basic variables. Furthermore, if all the ai's,
and bj 's are positive integers, then every basic solution is an integer solution.

Global Solution
Objective function: 15639

Continuous variables

Xn =6 X12 =2 X22 =3 X24 = 21 X31 = 20

X41 = 24 X51 = 3 X53 = 13 X62 = 12.

2.10 Test Problem 9

Given an undirected graph G(V, E) where V is a set of vertices and E is
a set of edges, a clique is defined to be a set of vertices that is completely
2.10. TEST PROBLEM 9 17

interconnected. The maximum clique problem consists of determining a clique

of maximum cardinality.
The maximum clique problem can be stated as a nonconvex quadratic
programming problem over the unit simplex (Pardalos and Phillips (1990))
and its general formulation is :

nMn l(x) = I: XiXj

(i,j)EE

iEV
Xi > 0 i = 1, ... , IVI

If k is the size of the maximum clique of G, then l(x*) = 0.5(1 - (ljk)).

The global maximum x* is defined by xi = (ljk) ifthe vertex i belongs to the
clique and zero otherwise.
In the following, a specific instance of this general formulation is presented.
Objective function
n-l n-2
nMn l(x) = I: XiXi+l + I: Xi X i+2 + X1X9 + X1XlQ + X2 X lQ + X1X5 + X4 X 7
i=l i=l

Constraints

Variable bounds

o~ x i = 1, ... ,n

Data

n= 10

Problem Statistics
No. of continuous variables 10
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities
18 CHAPTER 2. QUADRATIC PROGRAMMING PROBLEMS

Global Solution
Objective function: 0.375

Continuous variables

::r: = (0,0,0,0.25,0.25,0.25,0.25,0, of.

2.11 Test Problem 10

This test problem has a separable quadratic objective function where there
exists a convex part (Le., the summation multiplied by +0.5) and a concave
part (i.e., the summation multiplied by -0.5).
Objective function

Constraints

A l ::r:+A 2 y < b
::r: E !R 1O
Y E !R10

Variable bounds

Data

A = (63, 15,44, 91,45, 50, 89, 58, 86, 82)T

~ = (42,98,48,91, 11,63,61,61,38,26)T

a = (-19, -27, -23, -53, -42,26, -33, -23,41, 19f

= (-52, -3,81,30, -85,68,27, -81, 97, -73f

Al, which corresponds to the concave variables, is the (10xl0) matrix:
2.11. TEST PROBLEM 10 19

3 5 5 6 4 4 5 6 4 4
5 4 5 4 1 4 4 2 5 2
1 5 2 4 7 3 1 5 7 6
3 2 6 3 2 1 6 1 7 3
6 6 6 4 5 2 2 4 3 2
A1 =
5 5 2 1 3 5 5 7 4 3
3 6 6 3 1 6 1 6 7 1
1 2 1 7 8 7 6 5 8 7
8 5 2 5 3 8 1 3 3 5
1 1 1 1 1 1 1 1 1 1

A 2 , that corresponds to the convex variables, is the following (10x10) matrix:

8 4 2 1 1 1 2 1 7 3
3 6 1 7 7 5 8 7 2 1
1 7 2 4 7 5 3 4 1 2
7 7 8 2 3 4 5 8 1 2
7 5 3 6 7 5 8 4 6 3
A2 =
4 1 7 3 8 3 1 6 2 8
4 3 1 4 3 6 4 6 5 4
2 3 5 5 4 5 4 2 2 8
4 5 5 6 1 7 1 2 2 4
1 1 1 1 1 1 1 1 1 1

b= (380,415, 385,405,470,415, 400,460, 400, 200)T

Problem Statistics
No. of continuous variables 20
No. of linear inequalities 10
No. of convex inequalities
No. of nonlinear equalities

Global Solution
Objective function: 49318

Continuous variables

x = (0,0,0,0,0,4.348,0,0,0, of
y = (0,0,0,62.609,0,0,0,0,0,0)T.
 Chapter 3

Quadratically Constrained
Problems

3.1 Introduction
In this chapter, we discuss nonconvex quadratically constrained test problems.
These include problems with separable quadratic constraints, complementarity
type constraints, and integer type constraints. Notice that every simple binary
constraint of the form :

x E {0,1}

can be written as :

_x 2 +x < 0
O::;x < 1

or

In the following sections, a number of standard test problems that are

quadratically constrained are presented.

3.2 Test Problem 1

This test problem is taken from Hock and Schittkowski (1981) (Problem 106).
Its objective function is linear while it features three linear inequalities and
21

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
22 CHAPTER 3. QUADRATICALLY CONSTRAINED PROBLEMS

three nonconvex inequalities.

Objective function

Constraints

-1 + 0.0025(X4 + X6) < 0

-1 + 0.0025( -X4 + X5 + X7) < 0
-1 + 0.01( -X5 + xs) < 0
100Xl - X1X6 + 8333.33252x4 < 0
- 83333.333
X2X4 - X2X7 - 1250x4 + 1250x5 < 0

X3X5 - X3XS - 2500X5 + 1250000 < 0

Variable bounds

100 ~ Xl ~ 10000
1000 ~ X2 ~ 10000
1000 ~ X3 ~ 10000
10 ~ X4 ~ 1000
10 ::; X 5 ::; 1000
10 ~ X6 ~ 1000
10 ~ X7 ~ 1000
10 ~ Xs ~ 1000

Problem Statistics
No. of continuous variables 8
No. of linear inequalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 3

Global Solution
Objective function: 7049.25

Continuous variables

:z: = (579.19,1360.13,5109.92,182.01,295.60,217.99, 286.40, 395.60f.

3.3. TEST PROBLEM 2 23

3.3 Test Problem 2

This problem consists of a nonconvex quadratic objective function subject to
six nonconvex quadratic inequality constraints, and it is taken from Colville
(1970).
Objective function

min f(ro) = 37.293239xl

ro
+ 0.8356891xlX5 + 5.3578547x~ - 40792.141

Constraints

-0.0022053x3X5 + 0.0056858x2X5 + 0.0006262xlX4 - 6.665593 < 0

0.0022053x3X5 - O.0056858x2X5 - 0.0006262xlX4 - 85.334407 < 0
0.0071317x2X5 + 0.0021813x~ + 0.0029955xlX2 - 29A8751 < 0
-0.0071317x2x5 - 0.0021813x~ - 0.0029955xlX2 + 9.48751 < 0
0.0047026x3X5 + 0.0019085x3X4 + 0.0012547xlX3 - 15.699039 < 0
-0.0047026x3X5 - 0.0019085x3x4 - 0.0012547xlX3 + 10.699039 < 0

Variable bounds

78 ~ Xl ~ 102
33 ~ X2 ~ 45
27 ~ X3 ~ 45
27 ~ X4 ~ 45
27 ~ X5 ~ 45

Problem Statistics
No. of continuous variables 5
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 6

Global Solution
Objective function: -30665.5387

Continuous variables

ro = (78,33,29.9953,45,36. 7758f.
24 CHAPTER 3. QUADRATICALLY CONSTRAINED PROBLEMS

3.4 Test Problem 3

This test problem is taken from Hesse (1973), and it features a eoneave quadratie
objeetive function subjeet to linear and quadratie eonstraints. There exist 18
loeal minima, and this problem ean be deeomposed into three smaller inde-
pendent problems.
Objective function

min f(a;)
a;
= -25(Xl - 2)2 - (X2 - 2)2 - (X3 - 1)2
-(X4 - 4)2 - (X5 - 1)2 - (X6 - 4)2

Constraints

(X3 - 3)2 + X4 > 4

(X5 - 3)2 + X6 > 4
Xl - 3X2 < 2
-Xl + X2 < 2
Xl +X2 < 6
Xl +X2 > 2

Variable bounds

o ::;Xl
o ::; X2
1 ::; X3 ::; 5
o ::; X4 ::; 6
1 ::; X5 ::; 5
o ::; X6 ::; 10
Problem Statistics
No. of eontinuous variables 6
No. of linear inequalities 4
No. of eonvex inequalities
No. of nonlinear equalities
No. of noneonvex inequalities 2

Global Solution
Objective funetion: -310
3.5. TEST PROBLEM 4 25
Continuous variables

x = (5,1,5,0,5,lOf.

3.5 Test Problem 4

This test problem is taken from Ben-Saad (1989), and it features a linear
objective function subject to a single reverse convex inequality and three linear
inequalities.
Objective function

min f(x) =
X
-2Xl + X2 - X3

Constraints

x T AT Ax - 2y T Ax+ Y 11 2 -0.2511 b- z >

11 2

11

+ X2 + X3 - 4 <

Xl

3X2 + X3 - 6 <
x E ~3

Variable bounds

~
~ Xl ~2
x2
~ x3 ~ 3

Data
Ais the following (3x3) matrix :

b = (3,0, _4)T

y = (1.5, -0.5, _5)T

z = (0, -1, _6)T

26 CHAPTER 3. QUADRATICALLY CONSTRAINED PROBLEMS

Problem Statistics
No. of continuous variables 1
No. of linear inequalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 1

Global Solution
Objective function: -4

Continuous variables

x = (O.5,O,3f.
 Chapter 4

U nivariate Polynomial
Problems

4.1 Introduction
In this chapter, we describe both unconstrained and constrained nonconvex
univariate polynomial problems. The reported global solutions are from the
studies of Visweswaran and Floudas (1992).

4.2 Test Problem 1

This test problem is taken from Wingo (1985).
Objective function

52 5 39 4 71 3 79 2 1
min x - - x + - x + - x - - x - x + -10
6
x 25 80 10 20

Variable bounds

-2 ~ x ~ 11

Global Solution
Objective function: -29763.233

Continuous variables

x = 10.

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
28 CHAPTER 4. UNIVARIATE POLYNOMIAL PROBLEMS

4.3 Test Problem 2

This example concerns the minimization of a 50 th degree polynomial, and it is
taken from Moore (1979).
Objective function
50
min L
i=l
ai xi

where
a = (-500.0,2.5,1.666666666,1.25,1.0,0.8333333, 0.714285714,
0.625,0.555555555,1.0, -43.6363636,0.41666666,0.384615384,
0.357142857,0.3333333,0.3125,0.294117647, 0.277777777, 0.263157894,
0.25,0.238095238,0.227272727,0.217391304, 0.208333333, 0.2,
0.192307692,0.185185185,0.178571428, 0.344827586, 0.6666666,
-15.48387097,0.15625,0.1515151,0.14705882,0.14285712,
0.138888888,0.135135135,0.131578947,0.128205128,0.125,
0.121951219,0.119047619,0.116279069,0.113636363,0.1111 111,
0.108695652,0.106382978,0.208333333,0.408163265,0.8).

Variable bounds

Global Solution
Objective function: -663.5
Continuous variables
x = 1.0911.

4.4 Test Problem 3

This example is taken from Wilkinson (1963).
Objective function

min 0.000089248x - 0.0218343x 2

x
+ 0.998266x3 - 1.6995x4 + 0.2x 5

Variable bounds
0:::; x:::; 10
4.5. TEST PROBLEM 4 29

Global Solution
This problem has loeal minima at x = 6.325, f = -443.67, x = 0.4573, f =
-0.02062, x = 0.01256, f = 0.0 and x = 0.00246, f = 0 amongst others.

4.5 Test Problem 4

This example is taken from Dixon and Szeg (1975).
Objeetive function

Variable bounds

-5 ~ x ~ 5

Global Solution
This problem has two global minima at x = 0, f = 0, and x = 2, f = O. There
is a loeal maximum at x = 1.

4.6 Test Problem 5

This example is taken from Dixon and Szeg (1975) and it is denoted as the
three-hump eamel-baek funetion.
Objective function

~n f(x) = 2x~ - 1.05xf + ~x~ - X1X2 + x~

Variable bounds

Global Solution
Objeetive funetion: 0

Continuous variables

Xl 0
X2 O.
30 CHAPTER 4. UNIVARIATE POLYNOMIAL PROBLEMS

4.7 Test Problem 6

This example is taken from Goldstein and Price (1971).
Objective function
min x 6
x
- 15x 4 + 27x 2 + 250
Variable bounds
-5 ::; x ::; 5

Global Solution
This function has a local minima at (0,250), and two global minima at (3,7)
and (-3,7).

4.8 Test Problem 7

This example is taken from Dixon (1990).
Objective function
min x 4
x
- 3x 3 - 1.5x 2 + lOx
Variable bounds
-5 ::; x ::; 5

Global Solution
This function has a global solution of -7.5 at x = -1.

4.9 Test Problem 8

This example is taken from Soland (1971).
Objective function
min -12xl -
x
7X2 + x~
Constraints
-2xi + 2 - X2 =0

Variable bounds
o ::; Xl ::; 2
o ::; x2 ::; 3
4.10. TEST PROBLEM 9 31

Global Solution
Objective function: -16.73889

Continuous variables

Xl = 0.7175
X2 1.47.

4.10 Test Problem 9

This is a test example that has a feasible region consisting of two disconnected
sub-regions.
Objective function

Constraints

X2 < 2 + 2xi - 8x~ + 8x~

X2 < 4xi - 32x~ + 88x~ - 96xI + 36

Variable bounds

o ::; Xl ::; 3
o ::; x2 ::; 4
Global Solution
Objective function: -5.50796

Continuous variables

Xl = 2.3295
X2 3.17846.
 Chapter 5

Bilinear problems

5.1 Introduction
Bilinear problems are an important subdass of nonconvex quadratic program-
ming problems whose applications encompass pooling and blending, separation
sequencing, heat exchanger network design and multicommodity network flow
problems. The general form of a bilinear problem is given by

min x T AoY + c x + d~ y
er x +
x,y
x T Aiy + = 1, ... ,p
erx +
B.t. dry::; bi, i
x T Aiy + dry = bi, i =P+ 1, ... ,p + q
x E Rn
yERm

where x and y are n- and rn-dimensional vectors respectively, Ai, i = 1, . .. ,p+

q, are n x rn matrices, Ci, i = 1, ... ,p + q, are n-dimensional real vectors,
di, i = 0, ... ,p+q, are rn-dimensional real vectors and bis a (p+q)-dimensional
real vector.
Since the specific structure of this dass leads to almost linear problems, it
has attracted the attention of researchers in global optimization early on. The
first methods to tackle these problems were proposed by Mangasarian (1964)
and Mangasarian and Stone (1974). Since then, a large number of solution
techniques have been presented in the literat ure (Altmann, 1968; Falk and
Soland, 1969; Falk, 1973; Konno, 1976; McCormick, 1976; Vaish and Shetty,
1976; Gallo and Ulcul, 1977; Sherali and Shetty, 1980; Thieu, 1980; AI-
Khayyal and Falk, 1983; Baker and Lasdon, 1985; Floudas and Visweswaran,
1990; Visweswaran and Floudas, 1991; Floudas and Visweswaran, 1993). In
particular, McCormick (1976) derived the convex envelope of bilinear terms,
providing the tightest possible convex relaxation. Reviews of techniques that
address quadratic and bilinear problems can be found in Horst and Tuy (1993),
Floudas and Visweswaran (1995) and in the forthcoming book by Floudas
(2000).
33

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
34 CHAPTER 5. BILINEAR PROBLEMS

Several examples of bilinear problems are listed in this chapter. They

focus on three types of engineering design applications: pooling, separation
sequencing and heat exchanger network design.

5.2 Pooling Problems

5.2.1 Introduction
Pooling and blending problems are very common in the chemical and petro-
chemical industries. Given several feeds with different properties and costs,
the aim is to determine the optimum flowrates that yield intermediate and
product streams of satisfactory quality. These problems have been studied by
Haverly (1978), Lasdon et al. (1979), Baker and Lasdon (1985), Visweswaran
and Floudas (1990), Floudas and Aggarwal (1990), Ben-Tal et al. (1994),
Visweswaran and Floudas (1996b) and Androulakis et al. (1996).

5.2.2 Haverly Pooling Problem

Haverly (1978) formulated aseries of pooling problems based on the flowsheet
shown in Figure 5.1. Three feeds, A, Band C, with sulfur contents of 3%,
1% and 2% respectively, are combined to form two products streams x and y.
The maximum sulfur content content for these two streams is 2.5% and 1.5%
respectively. The bilinearities in the problem arise from the fact that A and B
must be pooled together.

A -------1[>-
x
B
y

c--------t
splitter

Figure 5.1: Schematic of the Haverly pooling problem.

Formulation
This problem is formulated as a profit maximization.
Objective function

max 9x + 15y - 6A - ciB -lO(Gx + Gy)

x,y,A,B ,p,c""cy ,P", ,Py
5.2. POOLING PROBLEMS 35

Constraints

Px +Py - A- B 0
x-Px -Cx 0
y - Py - Cy 0
PPx + 2Cx - 2.5x < 0
pPy + 2Cy - 1.5y < 0
PPx + pPy - 3A - B 0

Variable bounds

o :::;
x :::; C2
o :::;
Y :::; 200
o< A,B,Cx,Cy,P,Px,Py :::; 500

Variable definitions
A and Bare the flowrates for the two feeds to the pool; C is the flowrate of
the feed to the splitter; p is the sulfur content of the streams coming out of
the pool; Px and Py are the flowrates of the streams from the pool to mixers
Mx and My, respectively; C x and C y are the flowrates of the streams from the
splitter to mixers Mx and My, respectively; x and y are the flowrates of the
two product streams. Three cases have been defined for different values of the
parameters Cl and C2.

Data
Case 1: (Cl, C2) = (16,100) .

Case 2: (Cl, C2) = (16,600).

Problem Statistics
No. of continuous variables 9
No. of linear equalities 3
No. of nonlinear equalities 1
No. of nonconvex inequalities 2
No. of known solutions - Case 1 1
No. of known solutions - Case 2 2
No. of known solutions - Case 3 1
36 CHAPTER 5. BILINEAR PROBLEMS

Global Solution

Objective function

Case1 Case2 Case3

400 600 750

Continuous variables

Case1
x =0.0 A = 0.0 C x = 0.0 Px = 0.0 p = 1.0
Y=
Y = 200.0 B =
= 100.0 Cy = = 100.0 Py =
= 100.0
Case 2
x =
= 600.0 A =
= 300.0 C x = 300.0 Px =
= 300.0 = 3.0
pp=
Yy= 0.0
=0.0 B =0.0
B=O.O C y = 0.0 Py =
= 0.0
Case3
= 0.0
x =0.0 A =
= 50.0 Cx = = 0.0 Px =
= 0.0 p = 1.5
Y
Y =
= 200.0 B =
= 150.0 Cl!.y =
C 0.0
=0.0 Py =
= 200.0

5.2.3 Ben-Tal et al. (1994) Problems: General Formulation

Ben-Tal et al. (1994) proposed a more general formulation for pooling prob-
lems which can accomodate any number of feed streams, pools and products,
and in which any feed stream may reach any pool and product. A graphical
representation of a system involving three feeds, one pool and two products
is shown in Figure 5.2. The x, y and z vectors represent the flows between
The:l:,
different units. Ben-Tal et al. (1994) suggested the substitution offlowrate
of flowrate Xii,
denoting the flow from feed i to pooll, with a fractional flowrate, qi/,
qil, denoting
the fraction of flow to pooll coming from feed i.
Objective function

lfgxz L L(dj - L C;qit}Ylj + L L(dj - C;)Zij

" jE.:T lEe iEI iEI jE.:T
5.2. POOLING PROBLEMS 37

Feeds Products

Figure 5.2: Schematic of a general pooling problem.

Constraints

L L qilYlj +L Zij < Ai, Vi EI

lEe jE:! jE:!

LYlj < St,Vl E.c

JE:!

LYlj + LZij ::; D j , Vj E.J

lEe iEI

L ( L Cikqil - Pjk) Ylj + L(Cik - Pjk)Zij < O,Vj E J,Vk E K-

lEe iEI iEI

Lqil 1, Vl E .c
iEI

Variable bounds

o ~ qil ~ 1, Vi E I, Vi E .c
o ~ Ylj ~ Dj, Vl E .c, Vj E J
o ~ Zij ~ Dj , Vi EI,Vj E J

Variable definitions
The sets are defined as folIows. I is the set of feeds, J is the set of products,
38 CHAPTER 5. BILINEAR PROBLEMS

C is the set of pools and /C is the set of components whose quality is being
monitored.
The variables are qi/, the fractional flow from feed i to pool I ; Ylj, the flow
from pool I to product j; and Zij, the flow from feed i directly to product j.
The parameters are Ai, the maximum available flow of feed i; D j , the
maximum demand for product j; SI, the size of pool I; Cik, the percentage of
component k in feed i; Pjk, the maximum allowable percentage of component
kin product j; Ci, the unit price of feed i; and dj , the unit price of product j.

5.2.4 Ben-Tal et al. (1994) Problems: Test Problem 1

Ben-Tal et al. (1994) proposed a first instance of their general formulation
consisting of four feeds, one pool and two products. The quality of only one
component is monitored. The following restrictions are imposed on the prob-
lem: feeds 1, 2 and 4 can be sent to the pool and feed 3 cannot. Thus, there
are only three q variables, qn, q21 and q41, and two Z variables, Z31 and Z32.
The corresponding flowsheet is shown in Figure 5.3.

Feeds Products

pool]

Figure 5.3: First instance of the Ben-Tal et al. (1994) pooling problems
5.2. POOLING PROBLEMS 39

Formulation

Objective function

max (9 - 6qll - 16q21 - 15q41)Yll

q,y,z

Constraints

q41Yll + q41Y12 < 50

Yll + Z31 < 100
Y12 + Z32 < 200
(3qll + q21 + q41 - 2.5)Yll - 0.5Z31 < 0
(3qll + q21 + q41 - 1.5)Y12 + 0.5Z32 < 0
qll + q21 + q41 1

Variable bounds

0 < qll < 1

0 < q21 < 1
0 < q41 < 1
0 < Yll < 100
0 < Y12 < 200
0 < Z31 < 100
0 < Z32 < 200

Data
A = (00, 00, 00, 50)T D = (100,200)T
C = (3,1,2, l)T P = (2.5, 1.5)T
c = (6,16,10, 15)T d = (9, 15)T

Problem Statistics
No. of continuous variables 7
No. of linear equalities 1
No. of convex inequalities 2
No. of nonconvex inequalities 3
No. of known solutions 3

Global Solution
Objective function: 450

Continuous variables
40 CHAPTER 5. BILINEAR PROBLEMS

qll = 0 q21 = 0.5 q41 = 0.5

Yll = 0 Y12 = 100
Z31 =0 Z32 = 100

5.2.5 Ben-Tal et al. (1994) Problems: Test Problem 2

The second case suggested by Ben-Tal et al. (1994) consists offive feeds, three
pools and five products. The quality of two components is monitored. The
following restrietions are imposed on the problem: feeds 1, 2, 4 and 5 can be
sent to the pool and feed 3 cannot. Thus, there are twelve q variables, and
three z variables. The corresponding flowsheet is shown in Figure 5.4.

Feeds Products

Figure 5.4: Second instance of the Ben-Tal et al. (1994) pooling problems
5.2. POOLING PROBLEMS 41

Formulation

Objective function

max (18 - 6qn - 16q21 - 15q41 - 12q5t}Yll

q,y,z
+(18 - 6q12 - 16q22 - 15q42 - 12q52)Y21
+(18 - 6q13 - 16q23 - 15q43 - 12q53)Y31
+(15 - 6qn - 16q21 - 15q41 - 12q5t}Y12
+(15 - 6q12 - 16q22 - 15q42 - 12q52)Y22
+(15 - 6q13 - 16q23 - 15q43 - 12q53)Y32
+(19 - 6qn - 16q21 - 15q41 - 12q5t}Y13
+(19 - 6q12 - 16q22 - 15q42 - 12q52)Y23
+(19 - 6q13 - 16q23 - 15q43 - 12q53)Y33
+(16 - 6qn - 16q21 - 15q41 - 12q5t}Y14
+(16 - 6q12 - 16q22 - 15q42 - 12q52)Y24
+(16 - 6q13 - 16q23 - 15q43 - 12q53)Y34
+(14 - 6qn - 16q21 - 15q41 - 12q5t}Y15
+(14 - 6q12 - 16q22 - 15q42 - 12q52)Y25
+(14 - 6q13 - 16q23 - 15q43 - 12q53)Y35
+8Z31 + 5Z32 + 9Z33 + 6Z34 + 4Z35

Constraints

q4IYn + q41Y12 + q41Y13 + q41Y14 + q41Y15 + q42Y21 + q42Y22 + q42Y23

+q42Y24 + q42Y25 + q43Y31 + q43Y32 + q43Y33 + q43Y34 + q43Y35 ~ 50
Yll + Y21 + Y31 + Z31 ~ 100
Y12 + Y22 + Y32 + Z32 ~ 200
Y13 + Y23 + Y33 + Z33 ~ 100
Y14 + Y24 + Y34 + Z34 ~ 100
Y15 + Y25 + Y35 + Z35 ~ 100
(3qll + q21 + q41 + 1.5q51 - 2.5)Yll
+(3q12 + q22 + q42 + 1.5q52 - 2.5)Y21
+(3q13 + q23 + q43 + 1.5q53 - 2.5)Y31 - 0.5Z31 ~ 0
(qn + 3q21 + 2.5q41 + 2.5q51 - 2)Yn
+(q12 + 3q22 + 2.5q42 + 2.5q52 - 2)Y21
+(q13 + 3q23 + 2.5q43 + 2.5q53 - 2)Y31 + 0.5Z31 ~ 0
(3qn + q21 + q41 + 1.5q51 - 1.5)Y12
+(3q12 + q22 + q42 + 1.5q52 - 1.5)Y22
+(3q13 + q23 + q43 + 1.5q53 - 1.5)Y32 + 0.5Z32 ~ 0
(qn + 3q21 + 2.5q41 + 2.5q51 - 2.5)Y12
+(q12 + 3q22 + 2.5q42 + 2.5q52 - 2.5)Y22
+(q13 + 3q23 + 2.5q43 + 2.5q53 - 2.5)Y32 ~ 0
(3qll + q21 + q41 + 1.5q51 - 2)Y13
+(3q12 + q22 + q42 + 1.5q52 - 2)Y23
+(3q13 + q23 + q43 + 1.5q53 - 2)Y33 ~ 0
42 CHAPTER 5. BILINEAR PROBLEMS

(qll + 3q21 + 2.5q41 + 2.5q51 - 2.6)Y13

+(q12 + 3q22 + 2.5q42 + 2.5q52 - 2.6)Y23
+(q13 + 3q23 + 2.5q43 + 2.5q53 - 2.6)Y33 - 0.lz33 ~ 0
(3qll + q21 + q41 + 1.5q51 - 2)Y14
+(3q12 + q22 + q42 + 1.5q52 - 2)Y24
+(3q13 + q23 + q43 + 1.5q53 - 2)Y34 ~ 0
(qll + 3q21 + 2.5q41 + 2.5q51 - 2)Y14
+(q12 + 3q22 + 2.5q42 + 2.5q52 - 2)Y24
+(q13 + 3q23 + 2.5q43 + 2.5q53 - 2)Y34 + 0.5Z34 ~ 0
(3qn + q21 + q41 + 1.5q51 - 2)Y15
+(3q12 + q22 + q42 + 1.5q52 - 2)Y25
+(3q13 + q23 + q43 + 1.5q53 - 2)Y35 ~ 0
(qn + 3q21 + 2.5q41 + 2.5q51 - 2)Y15
+(q12 + 3q22 + 2.5q42 + 2.5q52 - 2)Y25
+(q13 + 3q23 + 2.5q43 + 2.5q53 - 2)Y35 + 0.5Z35 ~ 0
qn + q21 + q41 + q51 =1
q12 + q22 + q42 + q52 =1
q13 + q23 + q43 + q53 = 1

Variable bounds

o ~ qn ~ 1 o ~ q12 ~ 1 o ~ q13 ~ 1
o ~ q21 ~ 1 0~q22~1 o ~ q23 ~ 1
o ~ q41 ~ 1 o ~ q42 ~ 1 o ~ q43 ~ 1
o ~ q51 ~ 1 0:::; q52 :::; 1 0:::; q53 :::; 1
o ~ Yn ~ 100 o ~ Y21 ~ 100 o ~ Y31 ~ 100
o ~ Y12 ~ 200 o ~ Y22 ~ 200 o ~ Y32 ~ 200
o ~ Y13 ~ 100 o ~ Y23 ~ 100 o ~ Y33 :::; 100
o ~ Y14 ~ 100 o ~ Y24 ~ 100 o ~ Y34 ~ 100
o ~ Y15 ~ 100 o ~ Y25 ~ 100 o ~ Y35 ~ 100
o ~ Z31 ~ 100 o ~ Z32 ~ 200 o ~ Z33 ~ 100
o ~ Z34 ~ 100 o ~ Z35 ~ 100

Data
A = (00, 00, 00, 50,oof D = (100,200,100,100, 100)T S = (00, 00, oo)T
c= (6, 16, 10, 15, 12)T d= (18, 15, 19, 16, 14)T

C =
3. 10
0 .
( 1.0 3.0 ) P =
2.5
15
( 2:0
2.0
25
2:6
1
2.0 2.5 20 20
1.5 2.5 2:0 2:0
5.3. DISTILLATION COLUMN SEQUENCING PROBLEMS 43

Problem Statistics
No. of continuous variables 32
No. of linear equalities 3
No. of convex inequalities 5
No. of nonconvex inequalities 11
No. of known solutions 40

Global Solution
Objective function: 3500

Continuous variables: the variable values at the solution with the greatest
objective function are reported here. The next best solution is very elose,
with a relative difference of only 1.5 x 10-8 .

qn =1 q21 =0 q41 =0 q51 =0

q12 =0 q22 =0 q42 =0 q52 =1
q13 = 0.2757 q23 = 0 q43 =0 q53 = 0.7243
Yn = 51.5527 Y12 = 0 Y13 =0 Y14 = 7.9609 Y15 = 15.7866
Y21 = 0 Y22 = 200 Y23 =0 Y24 = 0 Y25 = 20.5622
Y31 = 17.9504 Y32 = 0 Y33 =0 Y34 = 92.0391 Y35 = 63.6512
Z31 = 30.4969 Z32 = 0 Z33 = 100 Z34 = 0 Z35 = 0

5.3 Distillation Column Sequencing Problems

5.3.1 Introduction
The separation of products from waste and recyclable material is a central
part of chemical processes, in terms of both the technical challenges it presents
and the capital and operating costs it entails. A common means of achieving
separation is through the use of aseries of distillation columns. The sequence
in which different components in the initial feed are separated has a significant
impact on the economics of the process. It is therefore beneficial to optimize
this sequence in order to obtain the minimum annualized cost for the separation
section of the plant. Research efforts have been directed towards this problem
in the past three deeades (Glinos et al. , 1985; Westerberg, 1985; Floudas,
1987; Wehe and Westerberg, 1987; Floudas and Anastasiadis, 1988; Aggarwal
and Floudas, 1990, 1992). While distillation eolumns are very eomplex, their
operation ean be approximated through bilinear models for the purpose of
sequenee design. Several examples of such models are given in this section.
A partieular diffieulty of these problems, from the point of view of global
optimization, is the presenee of numerous equality constraints whieh, when
relaxed, lead to a mueh enlarged feasible region and henee poor lower bounds
on the global solution.
44 CHAPTER 5. BILINEAR PROBLEMS

.....----.-t>1 e
~ 1
PI
N
1
A 13
9
B
C 2
P2

Figure 5.5: Distillation sequence superstructure for the separation of a propane,

isobutane and n-butane mixt ure

5.3.2 Nonsharp separation of propane, isobutane and n-butane

The goal of this problem, taken from Aggarwal and Floudas (1990) anel Ad-
jiman et al. (1998b), is to identify the optimum design fr the separation of
a mixt ure of propane, isobutalle and n-butane into two products: PI, with a
total flow of 110 kgmoljh consisting of 30 kgmoljh of propane, 50 kgmoljh
of isobutane and 30 kgmoljh of n-butane, and P2, with a total flow of 190
kgmoljh consisting of 70 kgmoljh of propane, 50 kgmoljh of iso butane and 70
kgmoljh of n-butane. Two distillation columns alld a number of mixers and
splitters can be used, as ShOWll in Figure 5.5.

Formulation

Objective function

min 0.9979 + 0.00432FI + 0.00432F13 + 0.01517F 2 + 0.01517 F g

F,x

Constraints

F I + F2 + F3 + F4 = 300
F5 - F6 - F7 = 0
F 8 - F g - F lO - F ll = 0
F I2 - F I3 - F I4 - F I5 = 0
F I6 - F I7 - F I8 = 0
F 13 XA,12 - F 5 + 0.333 F I = 0
F 13 XB,12 - F8 xB,8 + 0.333 F I = 0
5.3. DISTILLATION COLUMN SEQUENCING PROBLEMS 45

- F8XC,8 + 0.333FI = 0
-FI2XA,12 - 0.333 F 2 = 0
F 9 XB,8 - F 12 XB,12 + 0.333 F 2 = 0
F9XC,8 - F I6 + 0.333 F 2 = 0
F14XA,12 + 0.333 F3 + F6 = 30
FlOXB,8 + F14XB,12 + 0.333 F3 = 50
F lO xc,8 + 0.333 F 3 + F I7 = 30
XA,8 = 0
XB,8 + XC,8 = 1
XA,12 + XB,12 = 1
XC,12 =0

Variable bounds

0::; Fi::; 300, i = 1, ... ,18

0::; Xj,i ::; 1, jE {A, B, Cl, i E {8, 12}

Variable definitions
A, Band C denote propane, isobutane and n-butane respectively. Fi denotes
the flowrate, in kgmol/h, for stream i as indicated on Figure 5.5. Xj,i denotes
the mole fraction of component j in stream i.

Problem Statistics
No. of continuous variables 24
No. of linear equalities 9
No. of nonlinear equalities 9
No. of known solutions 1

Global Solution
Several of the streams in Figure 5.5 do not exist at the optimum solution. The
corresponding global optimum network (Adjiman et al. , 1998b) is therefore
shown in Figure 5.6.

Objective function: 1.8642

Continuous variables

FI = 60.0601 Fs = 20 F g =40 F I3 =0 F17 =0

F2 =0 F6 =0 FlO = 0 F I4 = 20 F I8 = 20
F 3 = 90.0901 F 7 = 20 F ll = 0 F 1S = 0
F 4 = 149.8498 F 8 =40 F I2 = 20 F 16 = 20
XA,8 = 0 XB,8 = 0.5 XC,8 = 0.5
XA,12 = 0 XB,12 = 1 XB,12 = 0
46 CHAPTER 5. BILINEAR PROBLEMS

C
0 PI
L
U 1
M
1 N
1
A 8
C
B 0
L
9
C U
M
N P2
16

Figure 5.6: Optimal configuration for the separation of a mixt ure of propane,
isobutane and n-butane

5.3.3 Nonsharp separation of propane, isobutane, n-butane

and isopentane
In this problem taken from Aggarwal and Floudas (1990) a mixt ure of propane,
isobutane, n-butane and isopentane is considered. It must be separated into
two products: PI, with a total flow of 315 kgmoljh consisting of 75 kgmoljh
of propane, 100 kgmoljh of iso-butane, 40 kgmoljh of n-butane and 100 kg-
moljh of isopentane, and P2, with a total flow of 285 kgmoljh consisting of 75
kgmoljh of propane, 100 kgmoljh of iso-butane, 60 kgmoljh of n-butane and
50 kgmoljh of isopentane. Three distillation columns and a number of mixers
and splitters can be used, as shown in Figure 5.7.

Formulation

Objective function

min aO,1 + (al,l + 0.85a2,1 + a3,1 + bA,IXA,6 + b B ,IXB,6 + bC,IXC,6) F 6

F,x
+ aO,2 + (al,2 + 0. 85a 2,2 + 0. 85a 3,2 + bA,2 X A,15 + bB ,2 X B,15 + bC,2XC,15) F l5
+ aO,3 + (al,3 + a2,3 + 0. 85a 3,3 + bA ,3 X A,24 + b B ,3 X B,24 + bC,3XC,24) F24

Constraints
FI + F2 + F3 + F4 + F 5 600
F6 - F I - F 17 - F26 0
F l5 - F 2 - Fn - F27 0
F 24 - F 3 - F l2 - F 21 0
5.3. DISTILLATION COLUMN SEQUENCING PROBLEMS 47

Pt

A
B
C
D
P2

I0 splitter I

Figure 5.7: Distillation sequence superstructure for the separation of propane,

isobutane, n-butane and isopentane

F 7 - F8 - F 9 0
F lO - F u - FI2 - F 13 - F I4 0
F I6 - FI7 - F I8 - F I9 = 0
F 20 - F 2I - F 22 - F 23 0
F 25 - F 26 - F 27 - F 28 - F 29 0
F30 - F 3I - F 32 0
F 7xA,7 - O.85F6xA,6 0
F16XB,16 - O.85FI5 XB,15 0
F 25 XC,25 - F 24 XC,24 = 0
FlOXB,lO - F 6xB,6 0
F 2oxC,20 - O.85FI5 XC,15 = 0
F 3o XD,30 - O.85F24 XD,24 0
F 6xA,6 - F 7xA,7 - FlOxA,lO 0
F 6xB,6 - F 7xB,7 - HOXB,lO 0
48 CHAPTER 5. BILINEAR PROBLEMS

F6Xe,6 - FlOXe,lO 0
F 6 xD,6 - FlOxD,IO 0
F 15 XA,15 - F 16 xA,16 0
F 15 XB,15 - F 16 XB,16 - F20XB,20 0
F15Xe,15 - F16Xe,16 - F2oXe,2o 0
F 15 XD,15 - F20XD,20 0
F 24 XA,24 - F25XA,25 0
F 24 XB,24 - F 25 XB,25 0
F24Xe,24 - F25Xe,25 - F30Xe,3o 0
F24XD,24 - F 25 XD,25 - F30XD,30 0
0.250FI + F 17XA,16 + F 26 XA,25 - F 6 xA,6 0
0.333F I + F 17 xB,16 + F 26 XB,25 - F 6 xB,6 0
0.167FI + F17Xe,16 + F26Xe,25 - F6Xe,6 0
0.250FI + F26XD,25 - F 6xD,6 0
0.250F2 + F l1 xA,lO + F 27 XA,25 - F 15 XA,15 0
0.333F 2 + FUXB,lO + F 27 XB,25 - F 15 XB,15 0
0.167F2 + FuXe,lo + F27Xe,25 - F15Xe,15 0
0.250F2 + FUXD,IO + F 27 XD,25 - F 15 XD,15 0
0.250F3 + F I2 XA,1O - F 24 XA,24 0
0. 333F3 + F I2 XB,1O + F 21 XB,20 - F24XB,24 0
0.167F3 + F12Xe,lo + F 21 XC,20 - F 24 XC,24 0
0.250F3 + FI2XD,1O + F 21 XD,20 - F24XD,24 0
0.250F4 + FSXA,7 + FI3XA,1O + F IS XA,16 + F 2S XA,25 50
0.222F4 + F SXB,7 + F I 3 X B,1O + F IsXB,16 + F 22 XB,20 + F2SXB,25 100
0. 167F4 + F 13 X e,lo + F Is Xe,16 + F22Xe,2o + F2sXe,25 + F31Xe,3o 40
0.250F4 + FI3XD,IO + F22XD,20 + F 2S XD,25 + F 31 XD,30 100
XA,6 + XB,6 + Xe,6 + XD,6 1
XA,7 + XB,7 1
XA,lO + XB,lO + Xe,lO + XD,lO 1
XA,15 + XB,15 + Xe,15 + XD,15 1
XA,16 + XB,16 + Xe,16 1
XB,20 + Xe,2o + XD,20 1
XA,24 + XB,24 + Xe,24 + XD,24 1
XA,25 + XB,25 + Xe,25 + XD,25 1
Xe,3o + XD,30 1
5.3. DISTILLATION COLUMN SEQUENCING PROBLEMS 49

Variable bounds

0::; Fi ::; 600, i = 1, ... ,32

0::; X .
J,t ::; 1, jE {A, B, C, D}, i E {6, 10, 15, 24, 25}
0::; Xj,7 ::; 1, jE {A,B}
0::; Xj,16 ::; 1, jE {A,B,C}
0::; Xj,20 ::; 1, jE {B,C,D}
0::; Xj,30 ::; 1, jE {C,D}

Variable definitions
A, B, C and D denote propane, iso-butane, n-butane and isopentane respec-
tively. Fi denotes the flowrate, in kgmol/h, for stream i as indicated on Fig-
ure 5.7. Xj,i denotes the mole fraction of component j in stream i.

Data

0.31569 0.96926 0.40281 )

_ ( -0.0112812 -0.0413393 -0.0119785
a - 0.0072698 0.0228203 0.0082055
0.0064241 0.0257035 0.009819

0.0016446 0.0015625 -0.001748)

b = ( 0.0018611 0.0091604 -0.0002583
0.001262 0.0076758 -0.0004691

Problem Statistics
No. of continuous variables 62
No. of linear equalities 19
No. of nonlinear equalities 34
No. of known solutions 1

Best known solution

Several of the streams in Figure 5.7 do not exist at the optimum solution. The
corresponding network is therefore shown in Figure 5.8.

Objective function: 3.2340

Continuous variables
50 CHAPTER 5. BILINEAR PROBLEMS

Pt

A
B
C
D
P2

1 0 splitter I

Figure 5.8: Optimal configuration for the separation of a mixt ure of propane,
isobutane, n-butane and isopentane

FI = 161.99 F9 = 34.42 FI7 = 0 F25 = 100.27

F 2 = 115.49 FlO = 127.57 FIS = 64.46 F 26 = 0
F 3 = 93.68 Fu =0 FI9 = 0 F 27 =0
F 4 = 60.20 F I2 =0 F20 = 51.04 F2S = 0
F 5 = 168.62 F I3 = 127.57 F 21 = 51.04 F29 = 100.27
F 6 = 161.99 F14 =0 F22 =0 F30 = 44.45
F 7 = 34.42 FI5 = 115.49 F 23= 0 F 3I = 44.45
F s =0 FI6 = 64.46 F24 = 144.72 F 32 = 0
XA,6 = 0.250 XB,6 = 0.333 XC,6 = 0.167 XD,6 = 0.250
XA,7 = 1.000 XB,7 = 0.000
XA,10 = 0.048 XB,10 = 0.423 XC,10 = 0.212 XD,lO = 0.317
XA,15 = 0.250 XB,15 = 0.333 XC,15 = 0.167 XD,15 = 0.250
XA,16 = 0.448 XB,16 = 0.507 XC,16 = 0.045
XB,20 = 0.113 XC,20 = 0.321 XD,20 = 0.566
XA,24 = 0.162 XB,24 = 0.256 XC,24 = 0.221 XD,24 = 0.361
X A,25 = 0.234 XB,25 = 0.369 XC,25 = 0.319 XD,25 = 0.078
XC,30 = 0.000 XD,30 = 1.000
5.4. HEAT EXCHANGER NETWORK PROBLEMS 51

5.4 Heat Exchanger Network Problems

5.4.1 Introduction
Heat integration brings about significant savings when the right balance be-
tween capital investment and operating costs is found. Thus, heat exchanger
network problems have attracted attention for many years. Initially, these
problems were formulated as bilinear optimization problems.

5.4.2 Test Problem 1

The following problem, from Avriel and Williams (1971), addresses the design
of the heat exchanger network shown in Figure 5.9. One cold stream must
be heated from 1000 F to 500 0 F using three hot streams with different inlet
temperatures. The goal is to minimize the overall heat exchange area.

THJ ,in=300 TH2 ,in=400

~
TC';'=IOO~ A, 1----1> TC,out=500

THl,out

Figure 5.9: Heat exchanger network design problem 1, Avriel and Williams
(1971).

Formulation

Objective function
min Al +A2+A3
A,Tc,TH,out

Constraints
TC,1 + THI,out - TC,in - THI,in < 0
-TC,1 + TC,2 + T H2 ,O'Ut - THl,in < 0
TH3,out - TC,2 - TH3,in + Tc,out < 0
FCp FCp
Al - AITHl,O'Ut + ---u;TC,1 - ---u;Tc,in < 0
FCp FC
A 2T c I - A 2TH2 out - ---rr-T.c I
, , U2'
+ --T.c
U2
p
'
2 < 0

A FCp FCp
3T C,2 - A 3TH3,O'Ut - U3 TC,2 + U3 Tc,out < 0
52 CHAPTER 5. BILINEAR PROBLEMS

Variable bounds

100::; Al ::; 10000

1000 ::; A2,A3 ::; 10000
10::; Tc, I ::; 1000
10::; TC,2 ::; 1000
10::; THI,out ::; 1000
10::; TH2,out ::; 1000
10::; TH3,out ::; 1000

Data

FCp = 105 [Btu/(hr F)]

U = (120,80,40)T[Btu/(hr ft2 F)]

Problem Statistics
No. of continuous variables 8
No. of convex inequalities 3
No. of nonconvex inequalities 3
No. of known solutions 1

Best Known Solution

Objective function: 7512.23 ft 2

Continuous variables:

A = (102.69,1000.00, 5485.28)T

Tc = (265.06, 280.59)T

TH,out = (134.49,284.47, 380.59)T

5.4.3 Test Problem 2

This problem is taken from Visweswaran and Floudas (1996b). It involves two
hot streams and one cold stream. While the objective function is convex, a11
the nonconvexities in the problem stern from bilinear constraints. The heat
exchanger network is shown in Figure 5.10.
5.4. HEAT EXCHANGER NETWORK PROBLEMS 53

fHl,in = 10
THJ,in =500

fl,J h,2
~ f1,4
H1-C
T1,I. T1,0

9 f1,3
THJ ,out=250
TC,in =150 ..J Tc,out =310
I...
TH2,in=350
fe,in= 10 fe,out = 10
~fH2,in= 10 h,3
12,i 12,0
H2-C
h,l 12,2 12,4
9
TH2out =200

Figure 5.10: Heat exchanger network design problem 2, Visweswaran and

Floudas (1996b).
Formulation

Objective function

Constraints

11,1 + 12,1 10
11,1 + 12,3 - 11,2 0
12,1+11,3-12,2 0
11,4 + 11,3 - 11,2 0
12,4+12,3-12,2 = 0
15011,1 + T2,oh3 - T1,ih,2 0
15012,1 + T1,oh,3 - T2,ih2 0
54 CHAPTER 5. BILINEAR PROBLEMS

h,2(TI ,o - TI,i) 1000

h,2(T2,o - T2,i) 600
.!lTn + TI,o 500
.!lT12 + TI,i 250
.!lT21 +T2,o 350
.!lT22 + T2,i 200

Variable bounds
1O~ .!lTI,1 ~ 350
10 ~ .!lTI,2 ~ 350
1O~ .!lT2,1 ~ 200
1O~ .!lT2,2 ~ 200
O~ h,j ~ 10, j = 1, ... ,4
O~ h,j ~ 10, j = 1, ... ,4
150 ~ T-.
J,t ~ 310, j = 1, ... ,2
150 ~ Tj,o ~ 310, j = 1, ... ,2

Problem Statistics
No. of continuous variables 16
No. of linear equalities 9
No. of nonlinear equalities 4
No. of known solutions 9

Global Solution
Objective function: 4845
Continuous variables:
AT = (190 40)
140 50

0 10 0 0)
j= ( 10 10 10 0

TI,i = 210 TI,o = 150 T2,i = 310 T2,o = 210

5.4.4 Test Problem 3
This problem is taken from Floudas and Ciric (1988). It involves three hot
streams and one cold stream. The formulation has a convex objective function
and bilinear and linear constraints. The heat exchanger network is shown in
Figure 5.11.
5.4. HEAT EXCHANGER NETWORK PROBLEMS 55

Formulation

Objective function

Constraints

h,1 +12,1 +/3,1 45

h,1 + 12,3 + /3,4 - h,2 0
12,1 + h,3 + /3,3 - 12,2 0
/3,1 + h,4 + 12,4 - /3,2 0
h,5 + h,3 + h,4 - h,2 0
12,5 + 12,3 + 12,4 - 12,2 0
/3,5 + /3,3 + /3,4 - /3,2 0
100h,1 + T2,oh3 + T3,o/3,4 - T1,d1,2 0
10012,1 + T1,oh,3 + T3,o/3,3 - T2,d2,2 0
100/3,1 + T1,oh,4 + T2,oh4 - T3,d3,2 = 0
h,2(T1,o - T1,i) 2000
h2(T2,o - T2,i) 1000
/3,2(T3,o - T3,i) 1500
oTll + T1,o 210
oT12 +T1,i 130
oT21 + T2,o = 210
oT22 + T2,i = 160
oT31 +T3,o 210
oT32 + T3,i = 180
56 CHAPTER 5. BILINEAR PROBLEMS

fm,in= 25
THl ,in=21O

f1,1 f1,2
~ iJ,5
..----.-.,...----t>lH1-c I-----r------i>
Ti,i T1,o

~
THl ,out=130

TH2,in=21O
iJ 4 f1,3 ~ fH2,in = 20 h,3
TC,in =100 T.C 2 T2 Tc.out=200
________~r--r+~-;~'>H2-C~--o--~----~~--~>
fe,in= 45 12.1 \. h,i 1 h,5 fe,out= 45
13,3 V h,4
f 3,4 r
" TH2,out=160
~~------------~

TH3 ,in=21O
~fH3,in= 50
T3,2 T3,o
H3-C
13,1 h,i 13,5
~
TH3 ,out=180

Figure 5.11: Heat exchanger network design problem 3, Floudas and Ciric
(1988)
5.4. HEAT EXCHANGER NETWORK PROBLEMS 57

Variable bounds

1O~ T1,1 ~ 110

1O~ T1,2 ~ 110
1O~ T2,1 ~ 110
1O~ T2,2 ~ 110
1O~ T3,1 ~ 110
1O~ T3,2 ~ 110
O~ h,j ~ 45, j = 1,. " ,5
O~ !2,j ~ 45, j = 1, . " ,5
O~ h,j ~ 45, j = 1, ... ,5
100~ T
J,t ~ 200, j = 1, ... ,3
100~ Tj,o ~ 200, j = 1, ... ,3

Problem Statistics
No. of continuous variables 27
No. of linear equalities 13
No. of nonlinear equalities 6
No. of known solutions 9

Global Solution
Objective function: 10077.8
Continuous variables:

52.86 30.00)
aT = ( 10.00 60.00
10.00 22.86

Tl,i = 100 T1,o = 157.14

T2 ,i = 100 T2 ,o= 200
T 3,i = 157.14 T3,o = 200
 Chapter 6

Biconvex and Difference of

Convex Functions (D.C.)
Problems

6.1 Introduction
Biconvex and Difference of Convex functions (D.C.) problems are subclasses
of the general C 2 nonlinear programming problems. However, they possess
special structure and therefore are treated separately in this chapter.
Biconvex problems possess the following special characteristic: the vari-
ables can be partitioned into two disjoint sets (X,Y) in such a way that when
all of the variables in set X are fixed, the problem is convex in the remaining
variables Y, and when all of the variables in set Y are fixed, the problem is
convex in the set X.
A (D.C.) function /(x) is of the form:

where C 1 (x) and C2 (x) are both convex functions. Therefore the function
contains a convex part and a reverse-convex part.

6.2 Phase and Chemical Equilibrium Problems

6.2.1 Introduction
The phase and chemical equilibrium problem is extremely important for pre-
dicting fluid phase behavior for most separation process applications. Process
simulators must be able to reliably and efficiently predict the correct number
and type of phases that exist at equilibrium and the distribution of compo-
nents within those phases. The Gibbs free energy is the thermodynamic func-
tion most often used for equilibrium calculations because it can be applied at
59

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
60 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

conditions of constant temperature and pressure. A global minimum of the

Gibbs free energy corresponds to the true equilibrium solution. For many sys-
tems, the Gibbs free energy surface is nonconvex, therefore local optimization
methods can provide no guarantees that the correct equilibrium solution has
been located.
White et al. (1958) were among the first to use optimization methods for
minimizing the Gibbs free energy with their RAND algorithm, developed for
ideal systems. Castillo and Grossmann (1981) proposed a method whereby
one problem is solved to determine the maximum number of phases present
in the system and then a second problem is solved to minimize the Gibbs free
energy. Paules and Floudas (1989) used the Global Optimum Search method
of Floudas et al. (1989) to attempt to avoid converging to local extrema of the
Gibbs function.
The GLOPEQ algorithm developed by McDonald and Floudas (1997) can
provide a theoretical guarantee of convergence to the true equilibrium solution.
The GLOPEQ algorithm is a two-step procedure in which the first step is to
postulate the number of phases at equilibrium and a solution to the Gibbs
free energy minimization problem is obtained. In the second step, the Gibbs
tangent plane stability criterion is used to determine whether the candidate
solution is the true equilibrium solution. This is done by solving the Gibbs
tangent plane distance minimization problem to global optimality. If necessary,
the proposed phase configuration is updated and the process is repeated.
A presentation of the theoretical and algorithmic issues in global optimiza-
tion approaches for the phase and chemical reaction equilibrium problems can
be found in the book by Floudas (2000).

6.2.2 Mathematical Formulation

At constant temperature and pressure, the condition of equilibrium is that
the Gibbs free energy function attains its global minimum. The condition is
subject only to molar balances in the case of non-reacting systems, or elemen-
tal balances for reacting systems. For the following formulations, the Gibbs
free energy function is expressed in terms of the activity coefficient for liquid
phases, and vapor phases are assumed to be ideal. Depending upon the activ-
ity coefficient expression used, this may lead to either a biconvex formulation
or a formulation that is the difference of two convex functions.
For nonideal systems, there may be multiple solutions to the Gibbs energy
minimization problem. The tangent plane criterion can be used to determine
whether or not a candidate solution is the true equilibrium solution. The tan-
gent plane distance function is the difference between the Gibbs energy surface
for a potential new phase and the tangent plane to the Gibbs energy surface
constructed at the candidate solution point. If the tangent plane distance is
non-negative over the whole composition space, then the candidate solution is
the true equilibrium solution. The tangent plane distance function may itself
be nonconvex, therefore this problem must be solved to global optimality in
6.2. PHASE AND CHEMICAL EQUILIBRIUM PROBLEMS 61

order to guarantee that the true equilibrium solution has been obtained.

Gibbs Free Energy Minimization

In the Gibbs energy minimization problem we are given a system of C com-
ponents at a fixed temperature and pressure. P phases are postulated to be
present in the equilibrium state. In the case of a reacting system, the compo-
nents are composed of E elements.
Standard constraints
Elemental Balances - Reacting Systems

L L aein~ = be VeEE
iEC kEP

Molar Balances - Non-reacting Systems

Ln~ =nf Vi E C
kEP

Variable bounds

Variable definitions

nf - the number of moles of component i in phase k.

Parameter definitions

aei - number of elements e in component i.

bi - total number of elements e in mixt ure.

nf - total number of moles of component i.

The objective function is the Gibbs free energy function which is given by
the equation: G = L: L: n~ J.t~. The chemical potential of component i in
kEPiEC
phase k is denoted as J.t~. When the vapor phase is taken as ideal and the
set of liquid phases pL is represented by an activity coefficient equation, the
Gibbs free energy minimization problem is formulated as:
62 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

min
n
Jlr =

subject to

where t1G~'/ is the Gibbs free energy of formation of component i in phase k

at the system temperature. For a liquid phase,
'k
fi _ 'kxk
fik,o - 'Yi i

and for a vapor phase,

The symbol' denoted that the property is calculated for a component in a

mixt ure. In the case of an ideal vapor phase, the fugacity coefficient, ~i, is equal
to one. If the correct number of phases have been postulated, then the global
minimum of the Gibbs free energy will correspond to the true equilibrium state
of the system.
In many cases, the formulation shown above can be simplified. In the
first case, when reaction does not occur and both vapor and liquid phases
are present, the relation t1Gf'/ = t1GY'/ + RTln~sat can be used to remove
the need to calculate Gibbs energies of formation. In the second case, when
reaction does not occur and only liquid phases are present, the linear term in-
volving the Gibbs energy of formation becomes a constant and can be removed
from the objective function. Taking these simplifications into consideration, we
define three different formulations of the minimum Gibbs free energy problem.
Objective I - Reacting System

G { t1G~'/ jk }
mJn RT = LLn~ R~ +ln ~,o
kEPiEC fi
Objective II - Non-reacting System, both vapor and liquid phases present

'k
min.!l...-
n RT
"n~z In psat
= "L...J L...J z
+ " L...J n z In
L...J "
k 1L
k,o
kEpL iEC kEP iEC fi
Note that the first summation is only over the liquid phases, pL.
6.2. PHASE AND CHEMICAL EQUILIBRlUM PROBLEMS 63

Objective III - Non-reacting System, only liquid phases present

G
min- = ~~n~ln-'-
jk
n RT L.J L.J' fk,o
kEPiEC i

Tangent Plane Distance Minimization

In the tangent plane distance minimization problem, we are given a candidate
equilibrium solution, z, for a system of C components at fixed temperature and
pressure. The tangent plane distance minimization problem must be solved to
global optimality to determine if the candidate solution is stable or unstable
with respect to an incipient phase with composition x. If the global minimum
is greater than or equal to zero, then the candidate phase is stable, otherwise
it is unstable.
Elemental Balances - Reacting Systems

L
aeiXi = be V e E E
iEC
Molar Balances - Non-reacting Systems

Variable bounds

Variable definitions
Xi - The mole fraction of component i in the incipient phase.
The tangent plane distance function is the difference between the Gibbs free
energy of aphase with composition x, given by E XiJ-Li(X), and the tangent
iEC
plane to the Gibbs free energy surface constructed at the candidate phase
composition Z, given by E XiJ-Li(Z). Using these definitions, the tangent plane
iEC
distance minimization problems is formulated as follows:

minF=
z
E Xi {AG{ + In Xi + lni'i - J-Li(Z)}
iEC
subject to E aeiXi = be V e E E
iEC
~ Xi = 1
iEC
0:::; Xi:::; 1
64 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Tbe tangent plane distance formulation shown above can also be simplified,
as discussed in the previous section for the Gibbs energy minimization problem.
Objective I - Reacting System

Objective II - Non-reacting System, both vapor and liquid phases present

minF=
z
L Xi { RT In
iEC
pr t + In Xi + In 1'i - ILi (z)}

Objective III - Non-reacting System, only liquid phases present

minF=
z
LXi {lnxi + In1'i - ILi(Z)}
iEC

6.3 Biconvex Problems

6.3.1 NRTL Equation
Renon and Prausnitz (1968) derived the NRTL equation from Scott's two-
liquid theory using the assumption of non-randomness. The NRTL equation is
capable of representing liquid-liquid immiscibility for multicomponent systems
using only binary parameters. The analysis of McDonald and Floudas (1995c)
shows that by introducing a substitution variable, IJIf, the problem can be
written in a biconvex form.

Gibbs Free Energy Minimization

Objective function

minG=
n,>lI

where,

Additional Constraints
6.3. BICONVEX PROBLEMS 65

Transformation constraint

Variable definition

n~ - the number of moles of component i in phase k.

w~ - an additional variable created by substitution in order to give the

problem a bieonvex form.

Parameter definition

Tij, Qij - non-symmetrie binary interaction parameters.

~G~,J - Gibbs energy of formation of component i in phase k.

PL - denotes the subset of phases that are liquid.

Tangent Plane Distance Minimization

Objective function

Additional Constraints
Transformation constraint

Wi {L
lEG
QijXj } - Xi = 0 Vi E C

Variable definition

Xi - the mole fraction of component i in the incipient phase.

Wi - an additional variable created by substitution in order to give the

problem a bieonvex form.
66 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Parameter definition

Tij, gij - non-symmetrie binary interaction parameters.

t:.G{ - Gibbs energy of formation of eomponent i.

J1-i (z) - ehemical potential of the eandidate solution z.

6.3.2 Test Problem 1

The system of n-Butyl-Aeetate and Water is a ehallenging two eomponent,
two phase example. This example was studied by Heidemann and Mandhane
(1973) to demonstrate the potential eomplexities of the NRTL equation. In
this example the Gibbs free energy is minimized. The reported global solution
is from the work of MeDonald and Floudas (1997). Sinee only liquid phases
are present, this problem uses objeetive III for the Gibbs energy minimization.

Explicit Formulation

minG=
n,>It
1
n11 [ In::TT-r
n
n 1 n2
] + n21 [ In::iT-:"r
n
n
n2
1
1
]

+n12 [ In -::i? + n22 [ In $n n 2n 2 ]

2
n ]
n 1 n2 1

+n} [g11 T11IJ1} + g12T121J1~]

+n~ [g21T211J1~ + g22T221J1~]
+n~ [g11 T11IJ1~ + g12T121J1~]
+n~ [g21 T211J1~ + g22T221J1~]

subjeet to
g11lJ1}n} + g121J1}n~ - n} 0
g 21 1JI12n 11 + g 22 1JI12n 21 - n 21 0
g 11 1J121n12 + g 12 1J121n22 - nl2 0
g211J1~n~ + g221J1~n~ - n~ 0
n 11 +n 21 = n T1
n 2f +n22 nT2
1 122
nl' n2' n 1, n2 > 0
6.3. BICONVEX PROBLEMS 67

Data

P 1.0 atm
T = 298 K
nT = (0.50,0.50f
( 0.0 0.391965 )
a
0.391965 0.0
( 0.0 3.00498 )
7' = 4.69071 0.0

Problem Statistics
No. of continuous variables 8
No. of linear equalities 2
No. of convex inequalities
No. of nonlinear equalities 4
No. of nonconvex inequalities

Global Solution
Objective function: -0.02020
Continuous variables

= (0.4993 0.0007)
n 0.3441 0.1559

6.3.3 Test Problem 2

This is the same system as Test Problem 1. In this case, the tangent plane
distance is minimized for a given candidate solution. The reported global
solution is from the work of McDonald and Floudas (1995a). This problem
uses objective III for the tangent plane distance minimization.

Data

p 1.0 atm
T 298 K
z = (0.50,0.50)T
IJ.? (-0.06391,0.02875f

a ( 0.0 0.391965 )
=
0.391965 0.0

7'
( 0.0 3.00498 )
4.69071 0.0
68 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 2
No. of nonconvex inequalities

Global Solution
Objective function: -0.03247

Continuous variables
x = (0.00421,0.99579)T

6.3.4 Test Problem 3

This is a three component system of Toluene - Water - Aniline that was inves-
tigated by Castillo and Grossmann (1981). Two liquid phases are postulated
and there is no reaction. In this example, the Gibbs free energy is minimized
and objective III is used because only liquid phases are present. The reported
global solution is from the work of McDonald and Floudas (1995c).

Data

P = 1.0 atm
T 298 K
nT (0.2995,0.1998,0.4994)T
0.2485 0.30 )
( 00
0 0.2485 0.0 0.3412
0.30 0.3412 0.0

( 0.0 4.93035 1.59806 )

T 7.77063 0.0 4.18462
0.03509 1.59806 0.0

Problem Statistics
No. of continuous variables 12
No. of linear equalities 3
No. of convex inequalities
No. of nonlinear equalities 6
No. of nonconvex inequalities
6.3. BICONVEX PROBLEMS 69

Global Solution
Objective function: -0.3574

Continuous variables

0.29949 0.00001)
n = ( 0.06551 0.13429
0.49873 0.00067

6.3.5 Test Problem 4

This is the same three component system as in Test Problem 3, but in this
case the tangent plane distance is minimized, using objective IH. The reported
global solution is from the work of McDonald and Floudas (1995c).

Data

p = 1.0 atm
T 298 K
z (0.29989,0.20006,0.50005f
I-'? = (-0.28809,0.29158, -0.59336f
0.0 0.2485 0.30 )
a = ( 0.2485 0.0 0.3412
0.30 0.3412 0.0
0.0 4.93035 1.59806)
T ( 7.77063 0.0 4.18462
0.03509 1.27932 0.0

Problem Statistics
No. of continuous variables 6
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 3
No. of nonconvex inequalities -

Global Solution
Objective function: -0.29454

Continuous variables
z = (0.00007,0.99686, 0.00307f
70 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

6.4 Difference ofConvex Functions (D.C.) Problems

6.4.1 UNIQUAC Equation
The UNIQUAC equation was originally proposed by Abrams and Prausnitz
(1975) and postulates that the excess Gibbs energy is composed of a combina-
torial contribution due to differences in the sizes of molecules, and a residual
contribution due to energetic interactions between molecules. A modified form
of the equation proposed by Anderson {tnd Prausnitz (1978a), (1978b) is used
for better results with systems containing water and alcohols.

Gibbs Free Energy Minimization

Objective function

minG(n)
n
= ~ Ck - ~ ~ tPinf lnnf
kEP kEPiEG

Ck (n) is convex as is the term, ~

where Ck L ~L 1/Jinf nf and,
tPinf In nf
kEPiEG
kEPiEC

t:.Ck,f }
Ck(n) = ~nf { Rf -zfrilnri+~qilnqi
iEG

+zA .. ~ rinf In (~
lEG JEG
r n1) + .~ z[lrinf In ( E n~nk)
j
lEG jEC J

+~ ~ qinf In (~)
'EG E qJn
+ q~nf In (~ q'.nk) j 'EG J J
I k ( j E :k ) J
+ .~ qi'ni In E i
q'.Tj;n k
lEG jEC J J

tPi = q~ + rizA 'V i E C

and,

zll R = .!.q;-l
~qi-l
Z
1
= ~
ri
zA
zA L [zf
z~ + ~ [zr -- z~]
iEG
iEC
z
I ~
jf:.i
jf.i
[zf -
L [z~ z~] 'V i E C
J

z~ ,
m.inzf
m.inzr
I

Tangent Plane Distance Minimization

Objective function
6.4. DIFFERENCE OF CONVEX FUNCTIONS (D.C.) PROBLEMS 71

minF = CU (z) - ~ tPiXi lnxi

Z iEG

where CU (z) is convex as is the term, ~ tPiXi

'ifJixi In Xi
Xi and,
iEG
iEC

CU(z) = ~Xi{6.G{-J.L?(z)-zfrilnri}
iEG

+zA .. ~ riXi In (~ riXi) + .~ Zfrixi In ( EX;jXj)

tEG JEG tEG jEC

+~ .~ qiXi In ( 5!i:;Xj) + q~xi In (~ qjXi)

tEG jEC JEG

+ .~ iXi In ( E ;~jiXj)
tEG jEC J

tPi
'ifJi = q~ + ri [zr
ri [zf + zf]
zf] Vi E
\;j C

6.4.2 Test Problem 5

This ternary example, SBA - DSBE - Water, arises from an azeotropic distil-
lation problem. The data for the problem and the reported global solution are
from McDonald and Floudas (1997). Because of the sensitivity of this example
to the parameters, it represents achallenging test for aglobai optimization
method. In this example, the Gibbs free energy is minimized for a postulated
LLV system, using objective II.
H.

Data

P 1.16996 atm
T = 363.19909 K
nT (40.30707,5.14979,54.54314)T
,
RTln~sat
RTlnp~at = (-0.3658348, -0.9825555, -0.3663657f
q = (3.6640,5.1680, 1.4000)T
(3.6640,5.1680,1.4000)T
q' (4.0643,5.7409,1.6741)T
rr (3.9235,6.0909,0.9200)T
0.0 -193.141 424.025)
T = ( 415.855 0.0 315.312
103.810 3922.5 0.0
72 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Problem Statistics
No. of continuous variables 9
No. of linear equalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities -

Global Solution

Objective function: -70.75208

Continuous variables

31.459 0.902 7.946)

n = ( 3.103 0.0 2.046
26.167 15.014 13.362

6.4.3 Test Problem 6

This is the same system as presented in Test Problem 5, but in this case the
tangent plane distance function is minimized for a given candidate solution.
Since both vapor and liquid phases are possible, objective II is used. The
reported global minimum solution is from McDonald and Floudas (1995a).

Data

p 1.16996 atm
T 363.19909 K
z (0.51802, 0.05110, OA3088f
,
RTlnpfJat (-0.3658348, -0.9825555, -0.3663657)T
JI.? = (-0.92115, -2.27779, -OAOI39)T
q = (3.6640,5.1680,1.4000)T
q' = (4.0643,5.7409, 1.6741)T
r (3.9235, 6.0909, 0.9200f
0.0 -193.141 424.025)
T = ( 415.855 0.0 315.312
103.810 3922.5 0.0
6.4. DIFFERENCE OF CONVEX FUNCTIONS (D.C.) PROBLEMS 73

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities

Global Solution
Objective function: 0.0

Continuous variables
:z: = (0.51802,0.05110, 0.43088)T

6.4.4 Test Problem 7

The system of Ethylene Glycol - Lauryl Alcohol - Nitromethane presents an
interesting and challenging example because it may form three liquid phases.
The data for this example and the reported global solution are given in McDon-
ald and Floudas (1997). The Gibbs free energy is minimized for a postulated
LLL system, using objective IH.

Data

P 1.0 atm
T = 295 K
nT = (0.4,0.1,0.5f
q (2.2480,7.3720,1.8680)T
q' (2.2480,7.3720,1.8680)T
r (2.4088,8.8495,2.0086)T

( 0.0
247.2 54.701 )
T 69.69 0.0 305.52
467.88 133.19 0.0

Problem Statistics
No. of continuous variables 9
No. of linear equalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
74 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Global Solution
Objective function: -0.16085

Continuous variables

0.00881 0.33595 0.05525)

n = ( 0.00065 0.00193 0.09741
0.30803 0.14703 0.04537

6.4.5 Test Problem 8

The system for this example is the same ternary system as Test Problem 7,
but in this case the tangent plane distance is minimized using objective III for
a given candidate solution. The data for this example and the reported global
solution are given in McDonald and Floudas (1995a).

Data

P 1.0 atm
T = 295 K
z (O.29672,0.46950,0.23378)T
p.? = (-0.20102, -0.56343, -0.02272f
q = (2.2480,7.3720,1.8680)T
q' (2.2480,7.3720,1.8680)T
r = (2.4088,8.8495,2.0086f
247.2 54.701 )
( 0.0
T 69.69 0.0 305.52
467.88 133.19 0.0

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities -

Global Solution
Objective function: -0.0270

Continuous variables
x = (0.71540,0.00336, 0.28124)T
6.4. DIFFERENGE OF GONVEX FUNGTIONS (D.G.) PROBLEMS 75

6.4.6 UNIFAC Equation

The UNIFAC equation was developed by Fredenslund et al. (1975) and is based
on the UNIQUAC equation. The UNIFAC equation is a group contribution
method and therefore can give estimates of activity coefficients in the absence
ofreliable experimental data. Many ofthe same parameters as the UNIQUAC
equation are used, and it leads to a formulation that has the same character-
istics as the UNIQUAC equation.

Gibbs Free Energy Minimization

Objective function

minG(n)
n
= L Ck - L L 1Pin f In nf
kEP kEPiEG

where, Ck
Ck (n) is convex as is the term, l: L
L 'l/Jinf In nf
l: 1Pinf nf and,
kEPiEC
kEPiEG

Ck(n) L nf { t:.Gk,f
Jd-. - z[lri In ri + ~qi In qi - . }
v()
iEG

+zA . .L rinf In (.L rjnj ) + .L zfrinf In ( L n!ink)

z ",,'EG kI ( ~:~ ) :EIG ( "" k)iEG
+2 LJ qini n L q;nk + qini n LJ qjnj
J

.EG k JEG (J nk ) lEG

+L ni L VliQlln L ~kv.
iEG lEG JEG J lJ

1Pi = qi + rizA Vi E G

and,
76 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

qi = L V/iQI
lEG
ri = L V/iRI
lEG
Vii = L QmVmiWml
mEG
V(i) = L vliAV)
lEG
In A(i) = Q I {I -ln Pli
q. - L vmi9m'ltlm}
I mEG
v .
m.
~qi-l
z!l
t = Ti
zA = Z~ +
L [zf - Z~]
iEG
z!1
t = L [z!l- z~] Vi E C
j#i J
zR
M = m~nzf
t

Tangent Plane Distance Minimization

Objective function

where, CU (:z:) is convex as is the term, L tPiXi In Xi and,

iEG

CU(:z:) = L Xi {~ - JL?(Z) - zfrilnri + ~qilnqi - V(i)}

iEG
+zA. L rixi In (L rjXj)
iEG jEG
+L
iEG
Zfrixi In (~)
jEC J J

.L qiXi In ( r:XJjXj) + qiXi In (.L qjXj)

+~ tEG jEC JEG

+ LXi
iEG
L vliQ/ln
lEG
(r:Xjv .)
jEC J IJ

6.4.7 Test Problem 9

The system of n-Butyl-Acetate and Water is achallenging two component, two
liquid phase example. This example was studied by Heidemann and Mandhane
(1973) to demonstrate the potential complexities of the NRTL equation. In
this example the Gibbs free energy is minimized using the UNIFAC equation
with objective III. The reported global solution is from the work of McDonald
and Floudas (1997).
6.4. DIFFERENCE OF CONVEX FUNCTIONS (D.C.) PROBLEMS 77

Data

P = 1.0 atm
T 298 K
nT (0.50,0.50)T
Q (0.540,0.848,1.728,1.400f
R = (0.6744,0.9011,1.9031,0.9200)T

vT (~ 1 1 0)
0 o 1
0.0 972.4 1300.0 )
( 00
0.0 0.0 972.4 1300.0
T
-320.1 -320.1 0.0 385.9
342.4 342.4 -6.32 0.0

Problem Statistics
No. of continuous variables 4
No. of linear equalities 2
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities

Global Solution
Objective function: -0.03407

Continuous variables

_ (0.4998 0.0002)
n - 0.0451 0.4549

6.4.8 Test Problem 10

The Ethanol - Benzene - Water system has been extensively studied due to the
importance of the azeotropic distillation process used to separate ethanol and
water using benzene as an entrainer. In this example the Gibbs free energy is
minimized using the UNIFAC equation and two liquid phases are postulated,
so objective III is used. The binary interaction parameters are obtained from
Magnussen et al. (1981). The reported global solution is from the work of
McDonald and Floudas (1997).
78 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

Data

P 1.0 atm
T 298 K
nT = (0.20,0.40,0.40f
Q = (1.972,0.400,1.400)T

n
R (2.1055,0.5313,0.9200f

U
0
v 6
0
( 0.0
89.60 353.5 )
T = 636.1 0.0 903.8
-229.1 362.3 0.0

Problem Statistics
No. of continuous variables 6
No. of linear equalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities -

Global Solution

Objective function: -3.02954

Continuous variables

0.0818 0.1181)
n = ( 0.3910 0.0090
0.0136 0.3864

6.4.9 Test Problem 11

This is the same ternary system as in Test Problem 10, but in this case the
tangent plane distance is minimized. The candidate solution is a three phase
liquid-liquid-vapor solution, so objective 11 must be used. The example and
the reported global solution appear in McDonald (1995).
6.4. DIFFERENGE OF GONVEX FUNGTIONS (D.G.) PROBLEMS 79

Data

P = 1.0 atm
T 338 K
z (0.27208,0.00663, 0.72129)T
,
RTlnpflat (-2.562164, -2.084538, -3.482138)T
J.'? (-1.59634, -0.50349, -1.64535f
Q (1.972,0.400,1.400)T

C0 0)
R (2.1055,0.5313,0.9200f

v 060
001

T =
( 636.1
0.0 89.60 353.5 )
0.0 903.8
-229.1 362.3 0.0

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities

Global Solution
Objective function: 0.0
Continuous variables
x = (0.00565,0.99054, 0.00381)T

6.4.10 ASOG Equation

Like the UNIFAC equation, the ASOG equation is based on group contribution
methods. The method is described in the monograph of Kojima and Tochigi
(1979). The activity coefficient expression is proposed to be made up of two
contributions, one associated with the size of the moleeules and a second asso-
ciated with the groups that make up the molecules. The ASOG equation also
leads to a D.C. formulation for the Gibbs energy and tangent plane distance
minimization problems.

Gibbs Free Energy Minimization

Objective function
80 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

minG(n) = ECk - E E tPinfinnf

tPinflnnf
n kEP kEPiEG

where, Ckk(n), is convex as is the term, E E tPinf

tPinf In nf, and,
kEPiEG

E nf { =m.-
Gk,! (. }}
k
Ck(n) = G .!
+ In Vi - v')
V')
iEG

+.E.E nfnflnIn ((Lntnk)

+ +(.E nfvr)
L n!jnk) + nfvr) In (E njvf ) njv;)
lEG k (J n ')EG
J(J jEC JEG k

+E nn i E Vii In L ~ki)
iEG lEG jEC J /J.
J /J
.

tPi vr Vi E C
and,

Vii E Qmvmialm
mEG
v-?t E Vii
lEG
V(i) E VliAI
lEG
AI =
In AI 1 - In ~
L..J
X malm - ~
L..J EXma m/
v -
lEG mEG nEG
L XiV/i
iEC
XI E LE XiVmi
L
iECmEG
Inal m mlm+~

Tangent Plane Distance Minimization

Objective function

minF == CU (x) - E tPiXi Inxi

lnxi
Z iEG
where, CU(x), is convex as is the term, E tPixiInxi,
tPixi1nxi, and,
iEG

CU(x) E Xi {~
iEG
- p.?(z) + In Vi - v(i)}
+ .E
'EG
Xi In (~) + (.E
jECtEG
Xivr) In (E JEG
XjV;)

+E Xi E Vii In ( E X;i) . )
iEG lEG jEC J /J

tPi Vr Vi E C
6.4. DIFFERENCE OF CONVEX FUNCTIONS (D.C.) PROBLEMS 81

6.4.11 Test Problem 12

The system of n-Butyl-Acetate and Water is achallenging two component, two
liquid phase example. This example was studied by Heidemann and Mandhane
(1973) to demonstrate the potential complexities of the NRTL equation. In
this example the Gibbs free energy is minimized using the ASOG activity
coefficient equation and objective IH. The reported global solution is from the
work of McDonald and Floudas (1997).

Data

p 1.0 atm
T 298 K
nT = {0.50,0.50)T
II {8,lf
vT (5 3
o 0 1~6 )
( 0.0
-15.2623 -0.2727 )
m = -0.3699 0.0 -2.5548
0.5045 -2.4686 0.0

( 0.0
515.0 -277.3 )
n 162.6 0.0 659.9
-2382.3 565.7 0.0

Problem Statistics
No. of continuous variables 4
No. of linear equalities 2
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities -

Global Solution
Objective function: 0.28919
Continuous variables

_ (0.4994 0.0006)
n - 0.1179 0.3821

6.4.12 Test Problem 13

The ternary system Ethanol - Ethyl Acetate - Water was studied by Walraven
and van Rompay (1988) and by McDonald and Floudas (1997), whose global
82 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

solution is reported. For this problem the Gibbs free energy is minimized using
the ASOG activity coefficient equation. Two liquid phases are postulated for
the solution, so objective III is used.

Data

p = 1.0 atm
T = 343 K
nT = (0.08,0.30,0.62)T
v (3,6,1)T

0
1 0
v 0 1
0 0 1~6 )
-41.2503 -15.2623 -0.2727 )
( 00
4.7125 0.0 0.0583 -5.8341
m
-0.3699 -0.0296 0.0 -2.5548
0.5045 1.4318 -2.4686 0.0
7686.4 515.0 -277.3 )
( -3060.0
00 0.0 -455.3 1582.5
n = 162.6 2.6 0.0 659.9
-2382.3 -280.2 565.7 0.0

Problem Statistics
No. of continuous variables 6
No. of linear equalities 3
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities -

Global Solution
Objective function: -0.25457

Continuous variables

0.0480 0.0320 )
n = ( 0.0527 0.2473
0.5884 0.0316

6.4.13 Modified Wilson Equation

Wilson (1964) proposed an expression for the excess Gibbs energy that con-
tained a single nonsymmetrie binary interaction parameter. The Wilson equa-
6.4. DIFFERENCE OF CONVEX FUNCTIONS (D.C.) PROBLEMS 83

tion successfully predicts the equilibrium state for a large number of vapor-
liquid systems. However, the Gibbs function that is obtained from the Wilson
equation is convex, therefore the Wilson equation cannot predict liquid-liquid
immiscibility. Tsuboka and Katayama (1975) proposed a modification of the
Wilson equation that can overcome this limitation. The T-K-Wilson equation
leads to a D.C. formulation for both the Gibbs free energy and tangent plane
distance minimization problems.

Gibbs Free Energy Minimization

Objective function

minG(n) =
n
2: Ckk -- 2: 2: 1fJin~ In n~
kEP kEPiEC
kEPiEG

where, Ck ( n), is convex as is the term, 2: 2: 1fJin~ In n~, and,

kEPiEC
kEPiEG

t!.c k .!
+ In E n;kAji }
k k
Ck(n) ,2: n i
k {
~
n
+ In E1nk
~EG jEG J jEG J

+2: { 2: Pjinj } In {,2: Pjinj } + ,2: 2:,pji n j In E nLnk

~EG lEG lEG ~EG rf.~ lEG I

1fJi 1 + 2: Pji Vi E C
jf-i

Tangent Plane Distance Minimization

Objective function

min
minFF = CU (z) - 2: 1fJixi In
lnxi
Xi
Z iEG
iEC

where, CU(z), is convex as is the term, 2: 1fJixilnxi, and,

I
iEG
iEC

CU(z) ,2: Xi t!.c! - ILi0)

7ij!- (z + In E X i}
AjiXj
~EG jEG

+ 2: 2: Pji Xj } In {,2: Pji Xj } + ,2: 2:Pji Xj In E x;liXI

iEG JEG lEG ~EG lf.l lEG

1fJi 2: Pji Vi E C
jf-i
84 CHAPTER 6. BICONVEX AND (D.C.) PROBLEMS

6.4.14 Test Problem 14

The binary system Methanol - Cyclohexane is considered for the modified
Wilson equation. The system was studied by Tsuboka and Katayama (1975)
and McDonald and Floudas (1995b). The Gibbs free energy is minimized with
two liquid phases postulated using objective IH.

Data

p = 1.0 atm
T = 298 K
nT (0.50,0.50)T
( 1.0 0.30384 )
A = 0.095173 1.0
( 1.0 0.374 )
P 2.6738 1.0

Problem Statistics
No. of continuous variables 4
No. of linear equalities 2
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities

Global Solution
Objective function: -0.07439

Continuous variables

= (0.0583 0.4417)
n 0.4080 0.0920
 Chapter 7

Generalized Geometrie
Programming Problems

In this ehapter, we will diseuss test problems that arise from generalized ge-
ometrie programming applieations. For a thorough theoretieal and algorith-
mie exposition of global optimization approaehes for generalized geometrie
programming problems, the reader is direeted to the artide of Maranas and
Floudas (1997), and the book by Floudas (2000).

7.1 Introduction
Generalized geometrie or signomial programming (GGP) is the dass of opti-
mization problems where the objective function and eonstraints are the difIer-
enee of two posynomials. A posynomial G(z) is simply the sum of a number
of posynomial terms or monomials 9k(Z), k == 1, ... ,K multiplied by some
positive real eonstants Ck, k = 1, ... ,K.

G(z) = CI91(Z) + C292(Z) + ... + cKgK(Z)

Note that Ck E ~+, k = 1, ... ,K. Each monomial 9k(Z) gk(Z) is in turn the
produet of a number of positive variables raised to some real power,

gk (Z) dl k
X2d2 'k ... XdN k k = 1, ... ,K
Xl ' n "

where dl,k,
d1,k, d2,k, . .. ,dN,k E ~ and are not neeessarily integer. The term
geometrie programming was adopted beeause of the key role that the well
known arithmetie-geometrie inequality played in the initial developments.
By grouping together monomials with identieal sign, the generalized geo-
metrie (GGP) problem ean be formulated as the following nonlinear optimiza-
tion problem:
85

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
86 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

min Go(t) Gt(t) Gij(t)

t
subjeet to Gj(t) Gj(t) < 0,
Gj(t) :::; j = 1, ... ,M
ti ~ 0, i = 1, ... ,N
N
where Gj(t) = L Cjk II t~ijk, j=o, ...
j=O, ... ,M
kEKt i=l

Gj(t) = L Cjk
C'
Jk rr
N
IItt Qijk
i=l
Q""k
-= 0,, ... ,M
i 'J ', Jj'-0
t

kEKj

where t == (tl , ... ,tN) is the positive variable vectorj Gj, Gj, j == 0, ... ,M
are positive posynomial funetions in tj CXijk are arbitrary real constant expo-
nentsj whereas Cjk are given positive coefficients. Finally, sets Kt, Kj- count
how many positively jnegatively signed monomials form posynomials Gj, Gj
respectively. In general, formulation (GGP) corresponds to a highly nonlinear
optimization problem with nonconvex objective funetion andjor constraint set
and possibly disjoint feasible region.
A review of the previous work for posynomials and signomials along with
global optimization methods for generalized geometrie programming problem
can be found in the book by Floudas (2000).

7.2 Literature Problems

In this seetion, chemieal engineering example problems that are modelled as
generalized geometrie programming problems will be presented. The following
examples correspond to an alkylation process design, a CSTR sequence design,
a heat exchanger design, an optimal reactor design, and several problems taken
from the literature. The reported global solutions are from the work of Maranas
and Floudas (1997) and Adjiman et al. (1998b).

7.2.1 Test Problem 1 : Alkylation process design

This example involves the design of an alkylation unit as modeled by Dembo
(1976). The objective is to improve the oetane number of some olefin feed by
reaeting it with iso butane in the presence of acid.
Objeetive Funetion

min C1Xl + C2X1X6 + C3X3 + C4X2 + C5 - C6X3X5

7.2. LITERATURE PROBLEMS 87

Constraints
2
C7 x 6 + -1
csx 1 X3 - CgX6 < 1
-1
ClOX1X3 + -1
Cl1X1X3 X6 - C12X1X3 X6
-1 2
< 1
2
C13X6+ C14 X 5 - C15 X 4 - C16 X 6 < 1
+ C1S x 5-1 X6 + C19X4X5-1 - C20 x 5-1 X62
-1
C17 x 5 < 1
C21 X 7 + C22X2X3 x 4 - C23 X 2 X 3
-1 -1 -1
< 1
C24X7 + C25 X 2X 3 x 7 - C26 X 2 X 3 x 4 X 7
-1 -1 -1 -1 -1 -1
< 1
-1 + -1 < 1
C27 X 5 C2Sx5 X7

C29 X 5 - C30 X 7 < 1

C31 X3 - C32 X l < 1
C33X1X3
-1
+ -1
C34X3 < 1
-1 -1 -1
C35 X 2X 3 x4 - C36X2X3 < 1
C37X4 + -1
C3Sx2 X3 X 4 < 1
C39XIX6 +
C40 X l - C41 X 3 < 1
-1
C42Xl X3 + -1
C43 X l - C44 X 6 < 1

Variable Bounds

1500 < Xl < 2000

1 < X2 < 120
3000 < X3 < 3500
85 < X4 < 93
90 < X5 < 95
3 < X6 < 12
145 < X7 < 162

Data
The parameters Ci, i = 1, ... ,44 are shown in Table 7.1.

Problem Statistics
No. of continuous variables 7
No. of linear inequalities 2
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 12

Global Solution
Objective Function : 1227.23
88 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

i Ci i Ci i Ci

1 1.715 16 0.19120592 &1 31 0.00061000

2 0.035 17 0.56850750 E+2 32 0.0005
3 4.0565 18 1.08702000 33 0.81967200
4 10.000 19 0.32175000 34 0.81967200
5 3000.0 20 0.03762000 35 24500.0
6 0.063 21 0.00619800 36 250.0
7 0.59553571 &2 22 0.24623121 E+4 37 0.10204082 E-1
8 0.88392857 23 0.25125634 E+2 38 0.12244898 E-4
9 0.11756250 24 0.16118996 E+3 39 0.00006250
10 1.10880000 25 5000.0 40 0.00006250
11 0.13035330 26 0.48951000 E+6 41 0.00007625
12 0.00660330 27 0.44333333 E+2 42 1.22
13 0.66173269 &3 28 0.33000000 43 1.0
14 0.17239878 &1 29 0.02255600 44 1.0
15 0.56595559 &2 30 0.00759500

Table 7.1: Coefficients for alkylation example

7.2. LITERATURE PROBLEMS 89

Continuous Variables
xi = 1698.18
x2 == 53.66
xj == 3031.30
x4 == 90.11
Xs = 95
x6 = 10.50
x7 = 153.53
7.2.2 Test Problem 2 : CSTR Sequence Design
This example was originally posed by Manousiouthakis and Sourlas (1992) and
has been studied by Maranas and Floudas (1997), Ryoo and Sahinidis (1995)
and Adjiman et al. (1998b). It involves the design of a sequence of two CSTR
reactors where the consecutive reaction A --+ B --+ C takes place. This problem
is known to have caused difficulties for other global optimization methods.
Objective Function
min- X4

Constraints
Xl+ klXIX5 1
X2 - Xl+ k2X2X6 = 0
X3 + Xl + k3X3X5 = 1
X4 - X3 + X2 - Xl + k4X4X6 0
xg 5 + x~5 < 4
kl 0.09755988
k2 O.99k 1
k3 0.0391908
k4 = 0.9

Variable Bounds

(0,0,0,10- 5 ,10- 5 ) S (Xl, X2, X3, X4, X5) S (1,1,1,1,16,16)

Problem Statistics
No. of continuous variables 6
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities 4
No. of nonconvex inequalities 1
90 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Global Solution
A number of close loeal minima exist, as shown below.
obj Xl X2 X3 X4 X5 X6
Global -0.38881 0.772 0.517 0.204 0.388 3.036 5.097
Loeal1 -0.38808 1.0 0.393 0.0 0.388 10- 5 15.975
Loeal2 -0.37461 0.391 0.391 0.375 0.375 15.975 10- 5

7.2.3 Test Problem 3 : Heat exchanger design

This standard heat exehanger design problem was first formulated as a gener-
alized geometrie problem by Avriel and Williams (1971), as an alternative to
the formulation of Test Problem 1 of Seetion 5.4. It ean be written as folIows:
Objeetive Function
min Xl + X2 + X3
Constraints
833.33252xllX4Xiil + 100.0xii l - 83333.333x 1lx ii l
< 1
1250.0x21X5X7l + 1.0x4x7l - 1250.0X2lx4X7l < 1
1250000.0x3lxi!1 + 1.0x5xi!1 - 2500.0x3 l x5xi!1 < 1
0.0025x4 + 0.0025x6 < 1
-0.0025x4 + 0.0025x5 + 0.0025x7 < 1
O.Olxs - 0.01x5 < 1

Variable Bounds
100 ::; Xl < 10000
1000 < X2 < 10000
1000 < X3 < 10000
10 < X4 < 1000
10 < X5 < 1000
10 < X6 < 1000
10 < X7 < 1000
10 ::; Xs < 1000

Problem Statisties
No. of eontinuous variables 8
No. of linear inequalities 3
No. of eonvex inequalities
No. of nonlinear equalities
No. of noneonvex inequalities 3
7.2. LITERATURE PROBLEMS 91

Global Solution
Objective Function : 7049.25

Continuous Variable

xi
x*1 =
= 579.31
x;2
x* 1359.97
x; 5109.97
x4
x4* =
= 182.01
Xs
x*5 295.60
x*6
xij 217.98
x7 = = 286.42
Xg
x*s == 395.60

7.2.4 Test Problem 4 : Optimal Reactor Design

This problem has been proposed by Dembo (1976) and it addresses the optimal
design of a reactor.
Objective Function

min 0.4x~67 X7o. 67 + 0.4x~67 x SO. 67 + 10.0 - Xl - X2

Constraints

0.0588x5X7 + 0. lXlI
O. 1X < 1
0.0588x6XS + 0.1Xl
O.lXl + 0. 1X2
lx 2 < 1
4X3X5"1 + 2x3".71 x5"1 + 0.0588x3"1.3 x7 < 1
4X4Xijl + 2xio.71x61
4X4X61 2x4".71xij l + 0.0588xi1.3x8
0.0588x4"1.3 x8 < 1

Variable Bounds

0.1 < < 10,

Xi ::; i = 1, ... ,8

Problem Statistics
No. of continuous variables 8
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 4

Global Solution
Objective Function : 3.9511
92 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Continuous Variables

xi 6.4747
xi = 2.2340
x*3 = 0.6671
x*4 0.5957
x5 = 5.9310
x*6 5.5271
x*7 = 1.0108
x*8 0.4004

7.2.5 Test Problem 5 : Colville's Test Problem

This problem is taken from the compilation of test problems by Rijckaert and
Martens (1978b) and it appears to have troubled the local solvers on it was
tested.
Objective Function

mm 5.3578t~ + 0.8357tlt5 + 37.2392tl

Constraints

0.00002584t3t5 0- 0.00006663t2t5 - 0.0000734h t4 < 1

0.000853007t2t5 + 0.00009395tl t4 - 0.00033085t3t5 < 1
1330.3294t2" t 5 1 - 0.42tlt51 - 0.30586t2"lt~t51
l < 1
0.00024186t2t5 + O.00010159tlt2 + O.00007379t~ < 1
2275.1327t31t 5 1 - 0.2668tlt51 - 0.40584t4t51 < 1
0.00029955t3t5 + 0.00007992tl t3 + 0.00012157t3t4 < 1

Variable Bounds

78 < h < 102

33 < t2 < 45
27 < t3 < 45
27 < t4 < 45
27 < t5 < 45

Problem Statistics
No. of continuous variables 5
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 6
7.2. LITERATURE PROBLEMS 93

Global Solution
Objective Function : 1.1436

Continuous Variables

ti
ti 78
t2 33
t;
t; 29.998
t 4 45
t s 36.7673

7.2.6 Test Problem 6

This is problem 10 of Rijckaert and Martens (1978b).
Objective Function

min 0.5tlti1 - tt - 5ti1

Constraints

0.Olt2t3"1 + 0.01tt + 0.0005ttt3 < 1

Variable Bounds

1 < tl,t2,t3 < 100

Problem Statistics
No. of continuous variables 3
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 1

Global Solution
Objective Function : -83.254

Continuous Variables

ti =
ti 88.2890
t2 7.7737
t; =
t; 1.3120
94 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

7.2.7 Test Problem 7

This is problem 11 from Rijckaert and Martens (1978b).
Objective Function

Constraints
0.05882t3t4 + 0.1tt < 1
4t2t4"1 + 2tio.71t4"1 + 0.05882ti1.3t3 :s; 1

Variable Bounds

Problem Statistics
No. of continuous variables 4
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 2

Global Solution
. Objective Function : -5.7398
Continuous Variables
t*I 8.1267
t*2 0.6154
t*3 = 0.5650
t*4 = 5.6368

7.2.8 Test Problem 8

This is Problem 12 from Rijckaert and Martens (1978b).
Objective Function

min -tl - t5 + 0.4t~67 t 30. 67 + 0.4t~67 fi.67

Constraints

+ Olt l
0.05882t3t4 :s; 1
0.05882t7tS + Olt l + 0 lt5 < 1
4t2t4"1 + 2tio.71t4"1 + 0.05882ti1.3t3 < 1
4t6tsl + 2t6".71 ts l + 0.05882t6"1.3 t7 < 1
7.2. LITERATURE PROBLEMS 95

Variable Bounds
0.01 < ti < 10, i = 1, ... ,8

Problem Statistics
No. of continuous variables 8
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 4

Global Solution
Objective Function : -6.0482
Continuous Variables
ti = 6.4225
t*2 0.6686
t*3 1.0239
t*4 5.9399
t*5 2.2673
t*6 0.5960
t*7 = 0.4029
t*s 5.5288

7.2.9 Test Problem 9

This is Problem 14 from Rijckaert and Martens (1978b).
Objective Function

Constraints
tllti1.5t3tilt,51 + 5tlltilt3t2 < 1
0.05t3 + 0.05t2 < 1
lOt3"l - M3"l < 1
tillt7"1.5tstgltlOl + 5tillt7"ltstt2 < 1
ti1t7 + ti1ts < 1
t1ts1 - t6ts1 < 1
tlO < 0.1

Variable Bounds
0.01 < ti < 15, i = 1, ... ,10
96 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Problem Statistics
No. of continuous variables 10
No. of linear inequalities 1
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 6

Best Known Solution

Objective Function: 1.1436

Continuous Variables

t*1 2.0953 t*6 = 0.4548

t2 12.0953 t7 10.4548
t*3 = 7.9047 t*8 = 1.6405
t*4 0.4594 t*9 1.1975
t*5 = 0.3579 tio = 0.1000

7.2.10 Test Problem 10

This is Problem 17 from Rijckaert and Martens (1978b).
Objective Function

min

Constraints

Olt lO + t7t lO < 1

10tl t4 + lOt 1 t4t~ < 1
ti 1 - 100t7t lO < 1
tlOt1l- lOt8 < 1
t11t2t5 + tllt2t5t~ < 1
t s
1 - lOt 1 1t8t l1 < 1
lOtl1 - lOt9 < 1
t21t3t6 + t21t3t6t~ < 1
t 61 - t 2 1t9 < 1

Variable Bounds

0.01 < ti < 10, i = 1, ... ,11

7.3. ROBUST STABILITY ANALYSIS 97

Problem Statistics
No. of continuous variables 11
No. of linear inequalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 9

Best Known Solution

Objective Function: 1.1406

Continuous Variables

t*1 7.004
t*7 = 0.3820
ti = 7.646
t*8 0.3580
t*3 7.112
t*9 0.3530
t*4 0.0125
tio 2.0770
t*5 0.8120
th 0.4530
t*6 0.9558

7.3 Robust Stability Analysis

Robust stability analysis of linear systems involves the identification of the
largest possible region in the uncertain model parameter space for which the
controller manages to stabilize any disturbances in the system. The stability of
a feedback structure is determined by the roots of the closed loop characteristic
equation:

det (I + P(s, q)C(s, q)) = 0

where q is the vector of the uncertain model parameters, and P(s), C(s) are
the transfer functions of the plant and controller respectively. After expanding
the determinant we have:

where the coefficients ai(q), i = 0, ... ,n are typically multivariable polyno-

mial functions.
Astability margin km can then be defined as follows:

km(jw) = inf {k : P (jw, q(k)) = 0, \f q E Q}

Robust stability for this model is then guaranteed if and only if:
98 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Note that, real parameter uncertainty is typically expressed as positive and

negative deviations of the real parameters from some nominal values. Check-
ing the stability of a particular system with characteristic equation P(jw, q)
involves the solution of the following nonconvex optimization problem.

min k (8)
qi,k 2:0,w 2:0

subject to Re [P(jw, q)] 0

Im [P(jw, q)] 0

where qN is a stable nominal point for the uncertain parameters and q+, q-
are the estimated bounds. Note that it is important to be able to always locate
the global minimum of (8), otherwise the stability margin might be overesti-
mated. This overestimation can sometimes lead to the erroneous conclusion
that a system is stable when it is not.
The reported global solutions are from the work of Maranas and Floudas
(1997) and Adjiman et al. (1998b).

7.3.1 Test Problem 11

This example was first proposed by de Gaston and Safonov (1988). The sta-
bility margin formulation is:
Objective Function

min k

Constraints

lOq~qg + lOq~q~ + 200q~q~ + 100q~q3

+100q2qg + qlq2q~ + qlq~q3 + 1000q2q~
+8qlq~ + 1000q~q3 + 8qlq~ + 6qlq2q3
+60ql q3 + 60ql q2 - q? - 200ql ::::; 0

800 - 800k < ql < 800 + 800k

4 - 2k < q2 < 4 + 2k
6 - 3k < q3 < 6 + 3k
7.3. ROBUST STABILITY ANALYSIS 99

Problem Statistics
No. of continuous variables 4
No. of linear inequalities 6
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 1

Global Solution
Objective Function : 0.3417

Continuous Variables

qi 1073.4
q2 3.318
q3 4.975

7.3.2 Test Problem 12

This example examines the Zoo stability margin for the closed-Ioop system
described in Sideris and Sanchez (1988).
Objective Function

mm k

Constraints

1.4 - 0.25k < ql < 1.4 + 0.25k

1.5 - 0.20k < q2 < 1.5 + 0.20k
0.8 - 0.20k < q3 < 0.8 + 0.20k

Problem Statistics
No. of continuous variables 4
No. of linear inequalities 6
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 1

Global Solution
Objective Function : 1.089
100 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Continuous Variables

qi 1.1275
q2 1.2820
q; 1.0179

7.3.3 Test Problem 13

This example involves affine coefficient functions and it was first proposed by
Ackermann et al. (1991).
Objective Function

min k

Constraints

q3 + 9.625qlw + 16q2w + 16w2 + 12 - 4ql - q2 - 78w 0

16qlw + 44 - 19q1 - 8q2 - q3 - 24w 0

2.25 - 0.25k < ql < 2.25 + 0.25k

1.5 - 0.50k < q2 < 1.5 + 0.50k
1.5 - 1.50k < q3 < 1.5 + 1.50k

Problem Statistics
No. of continuous variables 5
No. of linear inequalities 6
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 2

Global Solution
This solution is valid for 0 :::; w :::; 10.

Objective Function : 0.8175

Continuous Variables

qi 2.4544
q2 1.9088
q; 2.7263
w* = 1.3510
7.3. ROBUST STABILITY ANALYSIS 101

7.3.4 Test Problem 14

This example studies the stability of the mechanical system introduced by
Vicino et al. (1990).
Objective Function

mm k

Constraints

a4(q)w 4 - a2(q)w 2 + ao(q) 0

a3(q)w 2 - al(q) 0

10.0 - LOk < ql < 10.0 + LOk

1.0 - O.lk < q2 < 1.0 + O.lk
1.0 - O.lk < q3 < 1.0 + O.lk
0.2 - O.Olk < q4 < 0.2 + O.Olk
0.05 - 0.005k < q5 < 0.05 + 0.005k

where a4(q) q~q2 (4q2 + 7qt)

a3(q) 7q4q~q2 - 64.918q~q2 + 380.067q3q2 + 3q5q2 + 3q5ql
a2(q) 3 (-9.81q3q~ - 9.81q3qlq2 - 4.312q~q2 + 264.896q3q2)
+ 3 (q4q5 - 9.274q5)
1
al (q) 5 (-147.15q4q3q2 + 1364.67q3q2 - 27.72q5)
ao(q) 54.387q3q2

Problem Statistics
No. of continuous variables 7
No. of linear inequalities 10
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 2

Global Solution
This solution is valid for 0 ::; w ::; 10.

Objective Function : 6.2746

102 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

Continuous Variables
qi 16.2746
q2 1.6275
qj = 1.6275
q: = 0.1373
q; = 0.0186
w* 0.9864

7.3.5 Test Problem 15

This example addresses the stability of a wire guided Daimler Benz 0305 bus
and was studied by Ackermann et al. (1991).
Objective Function
mm k
Constraints
as( q)w s - a6(q)w 6 + a4(q)w 4 - a2(q)w 2 + ao( q) 0
a7(q)w6 - a5(q)w 4 + a3(q)w 2 - al(q) 0
17.5 - 14.5k ~ ql ~ 17.5 + 14.5k
20.0 - 15.0k ~ q2 ~ 20.0 + 15.0k

where aO(ql, q2) 453106q~

al(ql, q2) 528 106 q~ + 3640 106 ql
a2(ql,q2) = 5.72106q~q2 + 113106q~ + 4250 106 ql
a3(ql,q2) = 6.93106q~q2 + 911106 ql + 4220 106
a4(ql,q2) 1.45106q~q2 + 16.8106qlq2 + 33810 6
a5(ql, q2) 15.6103q~q~ + 840q~q2 + 1.35106 qlq2 + 13.5106
a6(ql,q2) 1.25103q~q~ + 16.8q~q2 + 53.9103 qlq2 + 270103
a7(ql, q2) 50q~ q~ + 1080ql q2
as(ql, q2) = q~q~

Problem Statistics
No. of continuous variables 4
No. of linear inequalities 4
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 2

Global Solution
The system was found to be stable, km > 1.0, for the given range of uncertain
parameters and 0 ~ wleqlO.
7.3. ROBUST STABILITY ANALYSIS 103

7.3.6 Test Problem 16

In this example, an analysis of the stability margin of the spark ignition engine
Fiat Dedra (Barmish (1994), Abate et al. (1994)) is carried out for a continuous
range of operating conditions involving seven uncertain parameters.
Objective Function
mm k

Constraints
- a6(q)w 6 + a4(q)w 4 + -a2(q)w 2 + ao(q) 0
a7(q)w6 - as(q)w 4 + a3(q)w 2 - al(q) 0

3.4329 - 1.2721k < ql < 3.4329

0.1627 - 0.06k < q2 < 0.1627
0.1139 - 0.0782k < q3 < 0.1139
0.2539 < q4 < 0.2539 + 0.3068k
0.0208 - 0.0108k < qs < 0.0208
2.0247 < q6 < 2.0247 + 2.4715k
1.0000 < q7 < 1.0000 + 9.0000k

a7(q) q?

as(q) = 0.0189477qlq? + 0.11104q2q? + 0.1826qsq6 + 0.1104qsq7

+0.0237398q6q7 + q~q? + 0.1586ql q2q? + 0.0872ql q4q7
+O.0215658ql q6q7 + O.3652q2q6q7 + 2q3q4q7 - 0.0848q3q6q7
+q~ + 7.61760 1O- 4q? + 0.3172qlqSq7 + 4q2qSq7

a4(q) 0.1586qlq~+ 4.021411O- 4qlq? + 2q2q~ + 0.00152352q2q?

+0.0237398qsq6 + 0.00152352qsq7 + 5.16120 1O- 4q6q7
+0.0552q~q? + 0.01898477q1q2Q? + 0.0872qlq4qs
+0.034862qlq4q7 + 0.0215658q1qsq6 + 0.00287416qlq6q7
+0.0474795q2q6q7 + 2q3q4qS + 0.1826q3q4q6 + 0.1104q3q4q7
-0.0848q3qSq6 - 0.00234048q3q6q7 + 2q~qSq7 + 0.1826q~Q6q7
+ 0.0872ql q2q4q7 + 0.3172ql q2qsq7 + 0.0215658ql q2q6q7 + 0.1586ql q3q4q7
+ 2q2q3q4q7 - 0.0848q2q3q6q7 + 0.0552q~
+0.3652q2qSq6 + 0.0378954qlqsq7 + 0.2208q2qsq7
104 CHAPTER 7. GENERALIZED GEOMETRIC PROGRAMMING

a3(q) 0.0189477qlq~ + 0.1104q2q~ + 5.16120 1O- 4q5q6 + q~q~

+7.61760 1O-4q~q? + q~q~ + 0.1586qlq2q~ + 4.021411O- 4qlq2q?
+0.0872ql q3q~ + 0.034862ql q4q5 + 0.00336706ql q4q7
+ 0.00287416qlq5q6 + 6.28987 1O- 5qlq6q7 + 0.00103224q2q6q7
+ 0.1104q3q4q5 + 0.0237398q3q4q6 + 0.00152352q3q4q7
- 0.00234048q3q5q6 + 0.1826q~q5q6 + 0.1104q~q5q7
+0.0237398q~q6q7 - 0.0848q~q4q6 + 0.0872qlq2q4q5
+ 0.034862ql q2q4q7 + 0.0215658ql q2q5q6 + 0.0378954ql q2q5q7
+ 0.00287416ql q2q6q7 + 0.1586ql q3q4q5 + 0.0215658ql q3q4q6
+ 0.0189477qlq3q4q7 + 2q2q3q4q5 + 0.1826q2q3q4q6 + 0.1104q2q3q4q7
- O.0848q2q3q5q6 - 0.00234048q2q3q6q7 + 7.61760 1O-4q~
+ 0.0474795q2q5q6 + 8.04282 1O- 4qlq5q7 + 0.00304704q2q5q7

a2(q) 4.021411O-4qlq~ + 0.00152352q2q~ + 0.0552q~q~ + 0.0552q~q~

+0.0189477ql q2qg + 0.034862ql q3q~ + 0.00336706ql q4q5
+6.82079 1O- 5qlq4q7 + 6.28987 1O- 5qlq5q6 + 0.00152352q3q4q5
+5.16120 1O- 4q3q4q6 - 0.00234048q~q4q6 + 0.034862qlq2q4q5
+0.0237398q~q5q6 + 0.00152352q~q5q7 + 5.16120 1O-4q~q6q7
+0.00336706qlq2q4q7 + 0.00287416qlq2q5q6 + 8.04282 1O- 4qlq2q5q7
+6.28987 10-5 ql q2q6q7 + 0.0189477ql q3q4q5 + 0.00287416ql q3q4q6
+4.02141 1O- 4qlq3q4q7 + 0.1104q2q3q4q5 + 0.0237398Q2Q3Q4Q6
+0.00152352q2q3q4q7 - 0.00234048q2q3q5q6 + 0.00103224q2Q5Q6

al(q) 7.61760 1O-4q~qg + 7.61760 1O-4q~q~ + 4.021411O- 4qlq2qg

+0.00336706qlq3q~ + 6.82079 1O- 5qlq4q5 + 5.16120 1O-4q~q5q6
+0.00336706ql q2q4q5 + 6.82079 10-5 ql q2Q4q7 + 6.28987 10- 5 Ql Q2Q5Q6
+4.02141 1O- 4qlQ3q4q5 + 6.28987 1O- 5qlq3q4q6 + 0.00152352q2q3q4q5
+5.16120 1O- 4q2q3q4q6

Problem Statistics
No. of continuous variables 9
No. of linear inequalities 14
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities 2
7.3. ROBUST STABILITY ANALYSIS 105

Global Solution
The system was found to be stable km > 1.0, for the given range of uncertain
parameters and 0 ~ wleqlO.
 Chapter 8

Twice Continuously
Differentiable NLP Problems

8.1 Introduction
Twiee eontinuously differentiable NLPs represent a very broad class of prob-
lems with diverse applieations in the fields of engineering, seienee, finanee
and eeonomies. Specifie problems include phase equilibrium eharaeterization,
minimum potential energy eonformation of clusters and moleeules, distillation
sequeneing, reaetor network design, bateh proeess design, VLSI ehip design,
pro tein folding, and portfolio optimization.
The general form of the optimization problems to be solved is
mm l(x)
x
s.t. g(x)::; 0
h(x) = 0
xE [xL,x U]

where x is a veetor of n eontinuous variables, l(x) is the objeetive funetion,

g(x) is a veetor of inequality eonstraints, and h(x) is a veetor of equality
eonstraints. xL and x U denote the lower and upper bounds on x respectively.
All the functions are twiee eontinuously differentiable.
Many loeal optimization teehniques are available to solve this type of
problem. The most popular are redueed gradient teehniques and sueeessive
quadratic programming. Commercial implementations of algorithms based on
these approaehes have been available for many years. When the objective fune-
tion and the inequality eonstraints are eonvex and the equality eonstraints are
linear, there is a unique solution which ean be identified with a loeal solver.
In many applieations, NLPs have several loeal solutions beeause of the
noneonvexity of the objeetive function and/or feasible region. The global op-
timization of these problems has attraeted growing attention in the last two
deeades, either through stoehastie or deterministie methods. The stoehas-
tie algorithms fit within the following classes: genetie algorithms (Goldberg,
107

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
108 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

1989), taboo search (Glover et al., 1993; Cvijovic and Klinowski, 1995), and
simulated annealing (Kirkpatrick et al., 1983). The deterministic approaches
include primal-dual methods (Stephanopoulos and Westerberg, 1975; Floudas
and Visweswaran, 1990, 1993, 1996a, 1996b), interval methods (Neumaier,
1990; Hansen, 1992; Vaidyanathan and El-Halwagi, 1994) and branch-and-
bound methods (Horst and Tuy, 1993; Ryoo and Sahinidis, 1995; Smith and
Pantelides, 1996; Adjiman et al., 1998a,b). The difficulty of these problems
arises from the wide variety of functional forms that can be encountered. As
a result, the techniques developed must be applicable to and effective for very
broad classes of problems.
The reader is referred to Floudas and Grossmann (1995), Horst and Parda-
los (1995), Floudas and Pardalos (1996), Grossmann (1996), and Floudas
(1997) for thorough descriptions of these methods and their applications to
process synthesis and design, process control and computational chemistry.
The forthcoming book of Floudas (2000) discusses the theoretical, algorithmic
and application-oriented issues of deterministic global optimization approaches
applicable to twice continuously differentiable NLPs.
This chapter presents a variety of test problems that demonstrate the
breadth of global optimization applications of twice continuously differentiable
NLPs. Section 8.2 discusses literat ure test problems. Section 8.3 presents
applications from the area of batch design under uncertainty. Section 8.4 in-
troduces test problems from the area of chemical reactOF network synthesis.
Section 8.5 discusses applications that arise in parameter estimation and data
reconcilliation. Section 8.6 describes test problems from the area of phase and
chemical reaction> equilibrium using equations of state. Finally, section 8.7
discusses test problems that arise in atomic and molecular clusters.

8.2 Literature Problems

8.2.1 Test Problem 1
This unconstrained problem and the reported global solution are taken from
Adjiman et al. (1998a).

Formulation

Objective function

Variable bounds
-1:::; Xl :::; 2
-1:::; X2 :::; 1
8.2. LITERATURE PROBLEMS 109

Problem Statistics
No. of continuous variables 2
No. of known solutions 3

Global Solution
Objective function: -2.02181.

Continuous variables

Xl =2 X2 = 0.10578

8.2.2 Test Problem 2: Pseudoethane

The goal of this example is to determine the molecular conformation of pseudo-
ethane, an ethane moleeule in which aB the hydrogen atoms have been replaced
by C, N or 0 atoms. It is taken from Maranas and Floudas (1994), where the
global minimum potential energy conformation of smaB moleeules is studied.
The Lennard-Jones potential is expressed in terms of a single dihedral angle.
The contribution from 1-2 and 1-3 atom interactions have not been ineluded
in the formulation since they are constant. This problem has three very elose
minima.

Formulation

Objective function

588600
min
t (3r - 4 cosOr - 2 (sin2 Ocos(t -2n - 0) r)6cos 2
1079.1
(3rZ - 4 COSOTZ - 2 (sin Ocos(t - 2n - cos 0) TZ)3
2 2

600800
+ 6
( 3 TZ - 4 cos OTZ - 2 (sin2 0 cos (t) - cos 2 0) TZ)
1071.5
(3 TZ - 4 COSOTZ - 2 (sin2 Ocos(t) - cos 2 0) TZ)3
481300
+
(3TZ - 4 cosOT5 - 2 (sin2 Ocos(t +2n - cos 0) T5) 2
6

1064.6
(3T5 - 4 cosOT5 - 2 (sin Ocos(t + 2n - cos 0) T5)3
2 2

Variable bounds

o ~ t ~ 211"
110 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Variable definitions
ro is the covalent bond length (1.54 A), () is the covalent bond angle
(109.5), and t is the dihedral angle.

Problem Statistics
No. of continuous variables 1
No. of known solutions 3

Global Solution
Objective function: -1.0711 kcal/mol.
Continuous variable: t = 183.45.

8.2.3 Test Problem 3: Goldstein and Price function

This problem is taken from Goldstein and Price (1971).

Formulation

Objective function

mln
x,y
[1 + (x + y + 1f (19 - 14x + 3x 2 - 14y + 6xy + 3y 2) ]
x [30 + (2x - 3y)2 (18 - 32x + 12x 2 + 48y - 36xy + 27y2)]

Problem Statistics
No. of continuous variables 2
No. of known solutions 4

Global Solution
Objective function: 3
Continuous variables: x = 0, y = -1.

8.2.4 Test problem 4: Three-hump camelback function

This problem is taken from Dixon and Szeg (1975).

Formulation

Objective function
min 12x 2 - 6.3x 4
x,y
+ x6 - 6xy + 6y 2
8.2. LITERATURE PROBLEMS 111

Problem Statistics
No. of continuous variables 2
No. of known solutions 3

Global Solution
Objective function: 0

Continuous variables: x = 0, y = o.

8.2.5 Test Problem 5: Six-hump Camelback Function

This problem is taken from Dixon and Szeg (1975).

Formulation

Objective function

min
x,y

Problem Statistics
No. of continuous variables 2
No. of known solutions 6

Global Solution
The six solutions represent only three objective function values. For each pair
with the same objective function, the variable values are opposite in sign.

Objective function: -1.03163

Continuous variables: x = 0.08984, Y = -0.71266.

8.2.6 Test Problem 6: Shekel Function
This problem is taken from Dixon and Szeg (1975).

Formulation

Objective function

min 1 1
x,y - (x-4)2+(y-4)2+0.1 - (x-l)2+(Y-l)2+0.2
1
112 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of continuous variables 2
No. of known solutions 3

Global Solution
Objective function: -10.0860
Continuous variables: x = 4, Y = 4.
8.2.7 Test Problem 7
This is a small nonconvex problem from Murtagh and Saunders (1993). The
reported global solution is from the work of Adjiman et al. (1998b).

Formulation

Objective function

Constraints

3V2+2
2V2-2
2

Variable bounds
xE [-5,5]5

Problem Statistics
No. of continuous variables 5
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 3
No. of nonconvex inequalities -
No. of known solutions 5

Global Solution
Objective function: 0.0293
Continuous variables
x = (1.1166,1.2204,1.5378,1.9728, 1.7911)T
8.2. LITERATURE PROBLEMS 113

A A
B B
C C
A

Figure 8.1: Test problem 8.

8.2.8 Test Problem 8

The following example, taken from Ryoo and Sahinidis (1995), is a reactor
network dp.sign problem, describing the system shown in Figure 8.1. The goal
is to maximize the concentration of product B in the exit stream, GB2. The
reported global solution is from the work of Adjiman et al. (1998b). This
problem has three solutions with very dose objective function values. However,
these solutions correspond to markedly different network configurations.

Formulation

Objective function

min-GB2

Constraints

GAI+ kleAl VI 1
GA2 - GAI + k2GA2V2
GBI + GAI + k 3 GBI VI 1
GB2 - GBI + GA2 - GAI + k 4 GB2V2
vt 5 + V20. 5 < 4

Variable bounds

(0,0,0,0,1O- 5 ,1O- 5 ):s (GAI,GA2,GBI,GB2,V1 ,Vid:S (1,1,1,1,16,16)

114 GHAPTER 8. TWIGE GONTINUOUSLY DIFFERENTIABLE NLPS

Data
k 1 = 0.09755988 8 k2 = 0.09658428 8
k 3 = 0.0391908 8- 1 k 4 = 0.9 8- 1

Problem Statistics
No. of continuous variables 6
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 4
No. of nonconvex inequalities 1
No. of known solutions 3

Global Solution
Objective function: -0.38881.

Continuous variables

GA1 = 0.772 GA2 = 0.517 GEl = 0.204 GB2 = 0.388

VI = 3.036 V2 = 5.097

8.3 Batch Plant Design Under Uncertainty

8.3.1 Introduction
The design of batch processes has received much attention in the literat ure
over the past two decades. Batch processing is a popular method for manu-
facturing products in low volume or which require several complicated steps
in the synthesis procedure. Batch plant designs are generally grouped into
two categories: (1) multiproduct batch plants, where all products follow the
same sequence of processing steps, and (2) multipurpose batch plants, where
equipment can be used for more than one task and products may have different
routes through the plant. In addition, batch plants can be operated ac cord-
ing to different scheduling policies. In a single-product campaign (SPC), aB
batches of one product are manufactured in order, followed by aB batches of
another product. Mixed-product campaigns do not have this restriction, and
therefore allow for more efficient operation of the plant.
The final design of the batch plant is dependent upon the values of several
parameters. These include the demand for the products and the processing
times for each product in every stage. A batch plant design based on fixed
values for these parameters can lead to poor performance if the parameters
change (Le., the demand for the products increases). Batch design under un-
certainty models the range of possible parameter values and determines the
best trade-off between efficiency and flexibility. Among the first to address the
problem of batch design under uncertainty were Marketos (1975) and Johns
8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 115

et al. (1978). Reinhart and Rippin (1986), (1987) formulated the multiproduct
batch plant design problem with uncertainties in the demand for the products
and in the processing times and size factors for each product in each stage.
Subrahmanyam et al. (1994) addressed the problem by using a discrete
probability distribution for the demands. They formulated the problem us-
ing scenarios for different demand realizations. Ierapetritou and Pistikopoulos
(1995), (1996) used a continuous probability distribution for the demands and
used a stochastic programming formulation. In addition, they allowed for un-
certainty in the size factors and processing times by using scenarios. Their
formulation results in a single large-scale nonconvex optimization problem.
Harding and Floudas (1997) studied the mathematical structure via eigen-
value analysis and identified exact lower bounds. Based on these properties,
they proposed a global optimization approach based on the aBB method (An-
droulakis et al. (1995); Adjiman and Floudas (1996)) to solve this problem for
large examples in reasonable CPU time. They also extended the formulation
to apply to multipurpose batch plant design under uncertainty. For a detailed
exposition to the global optimization approach for the design of batch pro-
cesses under uncertainty, the reader is directed to the book by Floudas (2000)
and the article of Harding and Floudas (1997).

8.3.2 Single-Product Campaign Formulation

In the design of multiproduct batch plants, a scheduling policy is given (e.g.,

single product campaign or mixed product campaign), and the task is to de-
termine the optimal equipment sizes in each processing step, and the optimal
batch sizes for each product. When uncertainty is introduced in the processing
times, tij, and size parameters, Sij, the design must be feasible over the whole
range of possible parameter values. The range of parameter values are repre-
sented using a number of scenarios, P. When uncertainty is introduced in the
product demands, then the amount of each product to be produced becomes
a variable to be optimized. The joint probability distribution for the product
demands is discretized using Gaussian quadrat ure with a total of Q quadrat ure
points. The objective of batch plant design under uncertainty is to minimize
equipment costs minus the expected revenues from the sale of the products.
The chosen scheduling policy must be enforced, and production must be com-
pleted by some time, H, for all scenarios, P. Finally, the batch sizes for all of
the products must not be too large for the equipment, for all scenarios, P.
Objective function
The objective is to minimize the equipment cost less the expected revenues
plus a penalty term that reHects the cost of unfilled orders.
116 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

M
mln 8 L: ajNj exp (jVj)
bi,Vj ,Q1 P j=1
P 1 Q N
- L: wP L: wqJq L: PiQ1P
p=1 q=1 i=1
P {N
+ 'Y ~ ~p ~ wqJq ?: p/Jl - i~PiQrp}
Q

p-1 q-1 1=1

Constraints
Batch Size Constraints
The batch size fr
for each product must be small enough to fit in the equip-
ment in each stage.

Vj - bi ~ In(S~) ViEN VjEM VpEP

Planning Horizon Constraint

All products must be manufactured in H hours or less.

N
L Qi p exp(~i - bii ) < H Vq E Q V pEP
i=1
i=l

where the limiting time step is defined as,

~i =m;x In ~ { t!:.}
Variable bunds
bounds

(}L
1
<
-
Q~P
1
<
-
()?
1
Vi E N V q E Q VpEP
ln(l/~l) $ Vj $ In(V~n V jEM
V
vL
?-) $ bi $ ~in ln( ?-) Vi E N
v,U
~in ln(
hP ij J~ ij

Variable definition

Q1
QiP - the amount of product i made given process parameter scenario p
and market demand corresponding to Gaussian quadrature grid point q .

Vj - Naturallogarithm of the equipment volume in stage j.

8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 117

bi - Naturallogarithm of the batch size of product i.

Parameter definition
Sij - size factor for product i in stage j for scenario p.
tfj
tfj -- processing time for product i in stage j for scenario p.
N j -- number of parallel units in stage j.
Nj

H - planning horizon time.

0i()l - the demand realization for product i at quadrat ure point q.
O~ = of (1 + vi) + Of(l - vi)
~ 2
OrJ
~
Oi!'
~
+ 40<7~

OL
~
O~ -40~
~ ~

()f, Of
Of, ()f - the mean and standard deviation of the probability distribution
of the demand for product i.
wqJq -- the weighted probability density function for the demands.
wqJq

L (99-91'
Jq _
_ exp(_iEN y ) 2
2)
.rr v'2/ffO<7~
~EN

vi, wq - Gaussian quadrat ure parameters that depend on the number

of quadrat ure points used. These can be obtained from handbooks of
mathematical tables. The values for a five-point quadrat ure, which is
used in all of the test problems in this handbook, are provided in the
table below.
Quadrature Point vq wq
1 -0.9061798459 0.23692885
2 -0.5384693101 0.4786286705
3 0.0 0.56888889
4 0.5384693101 0.4786286705
5 0.9061798459 0.23692885

w P - the weight for scenario p.

Pi - the price of product i.

'Y - penalty parameter.
"f

8,aj ,j - cost function parameters.

J,aj
118 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

8.3.3 Test Problem 1

The first example originally appeared in Grossmann and Sargent (1979). Hard-
ing and Floudas (1997) proposed aglobaI optimization approach for the formu-
lation given above. This example is for a batch plant with (N = 2) products
in (M = 3) stages with only (P = 1) scenario for the processing parameters.
For each product, five quadrat ure points are used to approximate the demand
distribution, giving a total of (Q = 25) quadrat ure points for the problem.
The reported global solution is from the work of Harding and Floudas (1997).

Explicit Formulation

Objective function

min <l [alNI exp(lvt} + a2N2 exp(!hv2) + a3N3 exp(3v3) + a4N4 exp(4v4)]
+uh-w(l,l) J(l,l) [(1 -1')pI Q~l,l),l + (1 _1')p2Q~I,1),1 + 1'P10[1,1) + 1'P20~1,1)]

+uh- w (1,2) J(I,2) [(1 -1')PI Q~I,2),1 + (1 _1')P2Q~I,2),1 + 1'P10[1,2) + 1'P20~1,2)]

+uh- w (1,3)J(I,3) [(1-1')PIQ~I,3),1 + (1 _1')p2Q~1,3),1 + 1'P10P,3) + 1'P20~1,3)]

+uh- w (I,4)J(1,4) [(1_1')PIQ~1,4),1 + (1 _1')P2Q~1,4),1 + 1'P10P,4) + 1'P20~1,4)]

+uh- w (I,5) J(1,5) [(1 - 1')PI Qil ,5),1 + (1 _1')P2Q~1,5),1 + 1'P10[1,5) + 1'P20~1,5)]

+uh- w (2,I)J(2,1) [(1_1')P1Q~2,1),1 + (1 _1')P2Q~2,1),1 + 1'P10F,I) + 1'P20~2,1)]

+uh- w (2,2) J(2,2) [(1 - 1')Pl Q~2,2),1 + (1 - 1')P2Q~2,2),1 + 1'P 10[2,2) + 1'P20~2,2)]

+uh-w(2,3) J(2,3) [(1 - l' )PI Q~2,3),1 + (1 - l' )p2Q~2,3),1 + 1'Pl OF,3) + 1'P20~2,3)]

+uh- w (2,4) J(2,4) [(1 - 1')PI Q~2,4),1 + (1 _1')P2Q~2,4),1 + 1'P 10[2,4) + 1'P20~2,4)]

+uh- w (2,5) J(2,5) [(1 - 1')PI Q~2,5),1 + (1 _1')p2Q~2,5),1 + 1'Pl 0[2,5) + 1'P20~2,5)]

+uh- w (3,I)J(3,1) [(1 -1')PIQi3,1),1 + (1_1')p2Q~3,1),1 + 1'P10P,1) + 1'P20~3,1)]

+uh- w (3,2)J(3,2) [(1-1')P1Qi3,2),1 + (1_1')P2Q~3,2),1 + 1'P10P,2) + 1'P20~3,2)]

+uh- w (3,3)J(3,3) [(1-1')P1Qi3,3),1 + (1_1')p2Q~3,3),1 + 1'P10P,3) + 1'P20~3,3)]

+uh- w (3,4)J(3,4) [(1-1')P1Qi3,4),1 + (1 _1')p2Q~3,4),1 + 1'P10P,4) + 1'P20~3,4)]

+uh- w (3,5)J(3,5) [(1-1')PIQi3,5),1 + (1 _1')P2Q~3,5),1 + 1'P10[3,5) + 1'P20~3,5)]

+uh- w (4,I)J(4,1) [(1 -1')PIQi4,1),1 + (1 _1')p2Q~4,1),1 + 1'P10[4,1) + 1'P20~4,1)]

+uh-w(4,2)J(4,2) [(1 -1')P1Qi4,2),1 + (1 _1')p2Q~4,2),1 + 1'P 10[4,2) + 1'P20~4,2)]

8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 119

+~W(4,3) J(4,3)[(1- 'Y)Pl Ql4,3),1 + (1 - 'Y)P2Q~4,3),1 + 'YPl(J~4,3) + 'YP2(J~4,3)]

+~w(4,4)J(4,4) [(1- 'Y)PlQi4,4),1 + (1 - 'Y)P2Q~4,4),1 + 'YPI(J~4,4) + 'YP2(J~4,4)]
+~w(4,5)J(4,5) [(1 - 'Y) PlQl4,5),1 + (1 - 'Y)P2Q~4,5),1 + 'YPl(J~4,5) + 'YP2(J~4,5)]

+~w(5,1) J(5,1) [(1 - 'Y)Pl Qi5,1),1 + (1 - 'Y )P2Q~5,1),1 + 'YPI (Ji5,1) + 'YP2(J~5,1)]
+~w(5,2)J(5,2) [(1- 'Y)PIQi5,2),1 + (1 - 'Y)P2Q~5,2),1 + 'YPI(J~5,2) + 'YP2(J~5,2)]
+~w(5,3)J(5,3) [(1- 'Y)PlQi5,3),1 + (1 - 'Y)P2Q~5,3),1 + 'YPl(J~5,3) + 'YP2(J~5,3)]
+~w(5,4) J(5,4) [(1 - 'Y)Pl Ql5,4),1 + (1 - 'Y)P2Q~5,4),1 + 'YPl(J~5,4) + 'YP2(J~5,4)]
+~w(5,5)J(5,5) [(1 - 'Y) PIQl5,5),1 + (1 - 'Y)P2Q~5,5),1 + 'YPl(J~5,5) + 'YP2(J~5,5)]

Constraints

VI - bl > InStl
V2 - bl > InSt2
V3 - bl > InSt3
Vl-~ > InS~l
V2 -~ > InS~2
V3- b2 ~ InS~3
Q(l,l),l
I (tl
exp LI - I b ) + Q(l,l),l
2 exp (tlL2- ~) < H
Q(1,2),1
1 exp (tlLI - b1) + Q(I,2),1
2 exp (tlL2 - b2) < H
Qil ,3),1 exp(tb -"- bJ) + Q~1,3),1 exp(tb - b2) < H
Q(1,4),1
I exp (tlLI - bI ) + Q(1,4),1
2 exp (tlL2- ~) < H
Qll,5),1 exp(tb - bl ) + Q~I,5),1 exp(tb - b2) < H
Q(2,1),1
I exp (tlLI - bI ) + Q(2,1),1
2 exp (tlL2 - b2) < H
Q(2,2),1
1 exp (tlLI - b1) + Q(2,2),1
2 exp (tlL2 - b2) < H
Ql2,3),1 exp(tb - bJ) + Q~2,3),1 exp(tb - b2) ::; H
Q(2,4),1
I exp (tlLI - bI ) + Q(2,4),1
2 exp (tlL2 - b2) < H
Qi2,5),1 exp(tlI - bJ) + Q~2,5),1 exp(tb - ~) < H
Q(3,1),1
I exp (tlLI - bI ) + Q(3,1),1
2 exp (tlL2- ~) < H
Q(3,2),1
I exp (tlLI - bI ) + Q(3,2),1
2 exp (tlL2 - b2) ::; H
Q(3,3),1
I exp (tlLI - b1) + Q(3,3),1
2 exp (tlL2- ~) < H
Q(3,4),1
1 exp (tlLI - bI ) + Q(3,4),1
2 exp (tlL2 - b2) < H
120 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Ql3,5),1 exp{ti1 - bd + Q~3,5),1 exp{tb - ~) < H

Q(4,1),1
1 exp (tlL1 - b1) + Q(4,1),1
2 exp (tlL2- ~) < H
Q(4,2),1
1 exp (tlL1 - b1) + Q(4,2),1
2 exp (tlL2- ~) < H
Q(4,3),1
1 exp (tlL1 - b1) + Q(4,3),1
2 exp (tlL2- ~) < H
Q(4,4),1
1 exp (tlL1 - b1) + Q(4,4),1
2 exp (tlL2- ~) < H
Q(4,5),1
1 exp (tlL1 - b1) + Q(4,5),1
2 exp (tlL2- ~) < H
Q(5,1),1
1 exp (tlL1 - b1) + Q(5,1),1
2 exp (tlL2- ~) < H
Q(5,2),1
1 exp (tlL1 - b1) + Q(5,2),1
2 exp (tlL2- ~) < H
Ql5,3),1 eXp(ti1 - bd + Q~5,3),1 exp(tb - ~) < H
Q(5,4),1
1 exp (tlL1 - b1) + Q(5,4),1
2 exp (tlL2 - b2 ) < H
Ql5,5),1 exp{tL - b1 ) + Q~5,5),1 exp{tb - ~) < H

Data

(h(/-L = 200, a = 10) (}2(/-L = 100, a = 10)

s = ( 2.0 3.0 4.0 )

4.0 6.0 3.0
20.0 8.0 )
T exp{t) = ( 186~0 4.0 4.0
a = {1O.0, 10.0, 10.0)T
= (0.6,0.6,0.6)T
N (1,1,1f
p = {5.5,7.of
'Y = 0.0
8 0.3
H = 8.0
w = 1.0
VL = (500,500,500f
VU = (4500, 4500, 4500)T

Problem Statistics
No. of continuous variables 55
No. of linear equalities 6
No. of nonconvex inequalities 25
8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 121

Global Solution

Objective function: -979.186

Continuous variables

v = exp(v) = (1800,2700, 3600)T

B = exp(b) = (900, 450)T
( 163.753, 163.753, 163.753, 163.753,
178.461, 178.461, 178.461, 178.461, 163.753)
178.461
Ql = 200.000, 200.000, 200.000, 200.000, 200.000
221.539, 221.539, 221.539, 221.539, 221.539
236.247, 236.247, 236.247, 236.247, 236.247

( 63.753, 78.461, 100.000, 121.539, 122.654)

63.753, 113.461, 113.461
78.461, 100.000,
Q2 = 63.753, 78.461, 100.000, 100.000, 100.000
63.753, 78.461, 86.539, 86.539, 86.539
63.753, 77.346, 77.346, 77.346, 77.346

8.3.4 Test Problem 2

This example is for a batch plant with (N = 4) products in (M = 6) stages

with (P = 3) scenarios for the processing parameters. For each product, five
quadrature points are used to approximate the demand distribution, giving a
total of (Q = 625) quadrature points for the problem. Due to its size in terms
ofvariables (7510) and number ofnonconvex inequality constraints (1875), this
is achallenging problem. This example and the reported global solution are
taken from Harding and Floudas (1997).

Data

fh ('" = 150, u = 10) 82 (", = 150, u = 8)

83 (", = 180, u = 9) 84('" = 160, u = 10)
122 CHAPTER8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

SI = CO 2.0 5.2 4.9 6.1

0.7 0.8 0.9 3.8 2.1 42
0.7 2.6 1.6 3.4 3.2 2.9
4.7 2.3 1.6 2.7 1.2 2.5
2.5 )

S2 C5 2.5 5.7 5.4 6.6

1.2 1.3 1.4 4.3 2.6 47
1.2 3.1 2.1 3.9 3.7 3.4
5.2 2.8 2.1 3.2 1.7 3.0
3.0 )

S3 =
C 5 1.5 4.7 4.4 5.6
0.2 0.3 0.4 3.3 1.6 3.7
0.2 2.1 1.1 2.9 2.7 2.4
4.2 1.8 1.1 2.2 0.7 2.0
2.0 )

Tl = CO 8.3 6.0 7.0 6.5

6.8 5.0 6.0 4.8 5.5 80
4.0 5.9 5.0 6.0 5.5 4.5
2.4 3.0 3.5 2.5 3.0 2.8
5.8 )

T2 Co 7.3 5.0 6.0 5.5

5.8 4.0 5.0 3.8 4.5 4.8
3.0 4.9 4.0 5.0 4.5 3.5
1.4 2.0 2.5 1.5 2.0 1.8
70 )

T3 = Co 9.3 7.0 8.0 7.5

7.8 6.0 7.0 5.8 6.5 6.8
5.0 6.9 6.0 7.0 6.5 5.5
3.4 4.0 4.5 3.5 4.0 3.8
90 )

Ct= (10.0, 10.0,10.0, 10.0, 10.0, 1O.0)T

= (0.6,0.6,0.6,0.6,0.6,0.6) T
N = (l,l,l,l,l,l)T
P = (3.5,4.0, 3.0, 2.0)T
W (3.0,3.0,3.0)T
'Y 0.0
0 = 0.3
H 8.0
VL = (500, 500,500,500, 500,500)T
VU = (4500,4500,4500, 4500, 4500, 4500)T

Problem Statistics
No. of continuous variables 7510
No. of linear equalities 72
No. of nonconvex inequalities 1875
8.3. RATCR PLANT DESIGN UNDER UNCERTAINTY 123

Global Solution

Objective function: -552.665

Continuous variables

V exp(v) = (3036,1726,2036,2714,2357, 1894)T

B exp(b) = (357,631,557, 584f

8.3.5 Test Problem 3

This example is for a batch plant with (N = 5) products in (M = 6) stages

with (P = 1) scenario for the processing parameters. For each product, five
quadrat ure points are used to approximate the demand distribution, giving a
total of (Q = 3125) quadrat ure points for the problem. Due to its size in terms
of variables (15636) and number of nonconvex inequality constraints (3125),
this is achallenging problem. This example and the reported global solution
are taken from Harding and Floudas (1997).

Data

(it (JL = 250, a = 10)

fh (JL = 150,a = 8)
(h (JL = 180,a = 9)
04 (JL = 160, a = 6)
Os (JL = 120, a = 3)
124 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

( 0.7
2.0 5.2 4.9 6.1
7.9 0.8 0.9 3.4 2.1 2.5 4,2)
S 0.7 2.6 1.6 3.6 3.2 2.9
4.7 2.3 1.6 2.7 1.2 2.5
1.2 3.6 2.4 4.5 1.6 2.1

T
( 6.8
4.7 8.3 3.9 2.1
6.4 6.4 6.5 4.4 2.3 3.2
1.0 6.3 5.4 11.9 5.7 6.2
12)
3.2 3.0 3.5 3.3 2.8 3.4
2.1 2.5 4.2 3.6 3.7 2.2
a (0.25,0.25,0.25,0.25, 0.25, 0.25)T
= (0.6, 0.6, 0.6, 0.6, 0.6, 0.6f
N (3,2,3,2, 1, 2f
p (3.5,4.0, 3.0,2.0, 4.5)T
'Y = 0.0
6 0.3
H 6.0
w = 1.0
VL = (500, 500, 500,500, 500,500)T
VU = (4500,4500,4500, 4500, 4500, 4500)T

Problem Statistics
No. of continuous variables 15636
No. of linear equalities 30
No. of nonconvex inequalities 3125

Global Solution
Objective function: -3731.079

Continuous variables

v = exp(v) = (2789,1901,1836,2460,2187, 1982f

B exp(b) = (353,724,683,593, 528)T

8.3.6 Unlimited Intermediate Storage Formulation

The single-product campaign formulation uses the maximum processing time
for each product, tLi = . max
J=l, ... ,M
{*}. Therefore, it overestimates the time
8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 125

required to process each product. This can result in an overdesign of the equip-
ment sizes. A mixed-product campaign allows for products to be processed in
any order that is desired. In general, mixed-product campaigns require less
time to process the required amount of products.
In order to allow mixed-product campaigns, the horizon constraint must be
replaced by an expression that takes into account the processing times in each
stage, rather than the maximum processing time. This can be done according
to the analysis of Birewar and Grossmann. The new horizon constraint has
the form:

N
L Qi p exp(tfj < H
- bi) :-:; V Q
Vq E Q V pEP
V
i=l

Using this analysis, the unlimited intermediate storage problem is formulated

as folIows.
Objective function
The objective is to minimize the equipment cost less the expected revenues
plus a penalty term that reflects the cost of unfilled orders.

M
min 8L cxjNj exp (jVj)
bi ,Vj,Q1P j=l
P Q N
- L ~P L wqJq LPiQiP
p=l q=l i=l

P {N
i~ PiQi
Q
+ 'Y P"f1 ~p qE wqJq i~ Pier -
P
}

Constraints
Batch Size Constraints

Vj ;::: In(S~) + bi ViEN VjEM VpEP

Planning Horizon Constraint

N
L Qi p exp(tfj < H
- bi ) :-:; Vq E Q
V V pEP
V
i=l

Variable bounds
126 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

(J!-t <
-
Q~P
t
<
-
(J~
t "liEN 'v q E Q 'VpEP
ln("~/) :5 Vj :5 In(v~F) 'v jE M
v.L VU
~inln(#) :5 bi :5 ~inln(#) "liEN
J~ ij hP ij

Variable and Parameter definitions - For a description of the variables and

parameters that are used in this formulation, refer to section 8.3.2.

8.3.7 Test Problem 4

This example is for a batch plant with (N = 2) products in (M = 3) stages
with (P = 1) scenario for the processing parameters. For each product, five
quadrat ure points are used to approximate the demand distribution, giving a
total of (Q == 25) quadrat ure points for the problem. This example and the
reported global solution are taken from Harding and Floudas (1997).

Data
(h (p,
(Jl (I' = 200, a = 10) fh(p,
(J2(p, = 100, a = 10)

3.0 4.0)
S = (2.0
( 2.0 3.0 4.0 )
4.0 6.0 3.0

T ( 8.0 20.0 8.0)

8.0 )
16.0 4.0 4.0
a 1O.of
(10.0, 10.0, 1O.of

= (0.6, 0.6, 0.6)T
N = (1,1,I)T
P (5.5,7.0)T
"I 0.0
8 = 0.3
H 8.0
w 1.0
VL
V (500,500,500)T
VU
V = (4500, 4500, 4500)T

Problem Statistics
No. of continuous variables 55
No. of linear equalities 6
No. of nonconvex inequalities 75
8.3. BATCH PLANT DESIGN UNDER UNCERTAINTY 127

Global Solution
Objective function: -1197.132

Continuous variables

Y = exp(v) (1200, 1800, 2400)T

B = exp(b) (600,300)T

8.3.8 Test Problem 5

This example is for a batch plant with (N = 4) products in (M = 6) stages
with (P = 1) scenario for the processing parameters. For each product, five
quadrat ure points are used to approximate the demand distribution, giving a
total of (Q = 625) quadrat ure points for the problem. Due to its size in terms
ofvariables (2510) and number ofnonconvex inequality constraints (3750), this
is achalienging problem.

Data
fh (J.l = 150, er = 10) fh(J.l = 150, er = 8) fh(J.l = 180, er = 9) ()4(J.l = 160, er = 10)

SI = CO 2.0
0.7 0.8
0.7 2.6
4.7 2.3
5.2
0.9
1.6
1.6
4.9
3.8
3.4
2.7
6.1
2.1
3.2
1.2
42
2.5 )
2.9
2.5

Tl CO 8.3 6.0 7.0

6.8 5.0 6.0 4.8
4.0 5.9 5.0 6.0
2.4 3.0 3.5 2.5
6.5
5.5
5.5
3.0
80
5.8 )
4.5
2.8
a (1O.0,1O.0,1O.0,1O.0,1O.0,10.0)T
(0.6,0.6,0.6,0.6,0.6,0.6)T
N (l,l,l,l,l,l)T
P (3.5,4.0, 3.0, 2.0)T
'Y 0.0
0 0.3
H 8.0
w 1.0
VL (500, 500, 500,500, 500, 500)T
yU (4500, 4500, 4500, 4500, 4500, 4500)T
128 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of continuous variables 2510
No. of linear equalities 24
No. of nonconvex inequalities 3750

Global Solution
Objective function: -830.338

Continuous variables

V exp(v) = (2703,1323, 1757,2045,2061, 1475f

B = exp(b) = (338,538,509, 575f

8.4 Chemical Reactor Network Problems

8.4.1 Introduction
Reactor network synthesis has been a widely studied problem in the area of
process synthesis. It deals with determining the types, sizes, and interconnec-
tions of reactors which optimize a desired performance objective.
Previous work in the area of reactor network synthesis can be catego-
rized as either superstructure-based approaches or targeting approaches. The
superstructure-based approaches propose a network of reactors which contain
all the possible design alternatives of interest. The solution strategy then de-
termines the optimal structure of the flowsheet. The early developments of this
approach was done by Horn and Tsai (1967) and Jackson (1968). The approach
was furt her developed by Achenie and Biegler (1986), Kokossis and Floudas
(1990), Kokossis and Floudas (1994), and Schweiger and Floudas (1998b).
The targeting approaches determine a target or bound on the performance
index of the reactor network regardless of the reactor types and configura-
tion. These approaches utilize the idea of the attainable region, which was
first defined by Horn (1964), as the set of all possible conditions that can be
achieved through reaction and mixing. Further work in this area has been done
by Glasser et al. (1987), Hildebrandt et al. (1990), Hildebrandt and Glasser
(1990), Achenie and Biegler (1988), Balakrishna and Biegler (1992a), Balakr-
ishna and Biegler (1992b), and Feinberg and Hildebrandt (1997). The work of
Lakshmanan and Biegler (1996) combined the ideas of the targeting approaches
with the superstructure approaches.
Problem Statement
In reactor network synthesis, the goal is to determine the reactor network that
transforms the given raw materials into the desired products. The following
information is assumed to be given in the problem definition:
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 129

the reaction mechanism and stoichiometry

the chemical kinetic data

the rate laws

the energetic data

the inlet stream(s) conditions

the performance objective (output).

The synthesis problem is to determine:

the type, size, and interconnection of reactor units

the stream Howrates, compositions, and temperatures

the composition and temperatures within the reactors

the heating and cooling requirements

the optimal performance objective

There are many different reactor types that can be considered for the reac-
tor network. Two extreme reactor types are often considered: the Continuous
Stirred Tank Reactor (CSTR) and the Plug Flow Reactor (PFR). The CSTR
is assumed to be perfectly mixed such that there are no spatial variations in
concentration, temperature, and reaction rate within the reactor. The PFR is
a tubular reactor where there are no radial variations in concentration, temper-
ature, and reaction rate. Since the mathematical modeling of the PFR results
in differential equations and leads to a dynamic model, they will be omitted
from the present model. The network that will be considered here consists
only of CSTRs. Models and test problems which include the dynamic models
are discussed in Chapter 15, Section 15.2.
The reactor network consists of mixers, splitters, and CSTRs. Splitters are
used to split the feed and the out let of each reactor. Mixers are situated at
the inlet to each reactor to mix the streams which come from the feed splitter
and the splitters after each reactor. A final mixer is used to mix the streams
which go from the splitter after each reactor unit to the product stream. A
fiowsheet which has two CSTRs is shown in Figure 8.2

8.4.2 General Formulation

In the general formulation, two assumptions will be made. First, constant
density is assumed which implies that volume is conserved. The second as-
sumption is that for nonisothermal systems, the heat capacity is assumed to
be constant in temperature and the same for all the components.
130 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

CSTR Fg II

vt

c-A-T '1

CSTR F g1'
vi"

Figure 8.2: Reactor network consisting of two CSTRs

The general formulation described below has been taken from Schweiger
and Floudas (1998b). It has been derived for any number of CSTR reactor
units, any number of reaction components, any number of reaction paths, any
number of feed streams, and any number of product streams.
In the general formulation, the following sets will be used:
Set Description
I Components
J Reactions
L CSTR units
R Feeds
P Products

Constraints
Feed splitter

Fa
r = '"
L..J Fad
r,l Vr E R
IEL

CSTR inlet mixer total balance

Fl = 2: F:,1 + 2: FI~1 Vl E L
rER l'EL

CSTR inlet mixer component balance

cl,i Fl = 2: c~,iF:,1 + 2: df"iFI~~ Vi E I Vl E L

rER l'EL

CSTR total balance

Fjg = Fl Vl E L
CSTR component balance

c'/,iFjg = cl,i Fl + Vi m 2: Vi,jriJ Vi E I Vl E L

JEJ
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 131

CSTR reaction rates

riJ = /J(cl,i, 1'r) Vj E J Vl E L

CSTR outlet splitter

Fjg = LF,~~ + LF,~; Vl

Fj9 Vi E L
I'EL pEP

Product mixer total balance

F.
ph
p
h = '""" F,gh
L..J
L...J p,9 h
I,p Vp E P
IEL

Product mixer component balance

c~,iF:
cp,i p = L d!,iFC
h F.h _ ' " JI r;.gh
- L...J C'f,iL'l,p Vi I
vZ E I
\.I.
Vp
\.I P
vp E P
IEL

Objective Function
The objective function varies from problem to problem. In many cases it is
the maximization of the yield of the desired product. It may be based upon
economic criteria reflecting the value of the products, cost of reactants, cost of
the utilities, and the cost of the reactors.
Variables
132 CHAPTER8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Flowrates
Fad VrER Vl E L Volumetrie flowrates
T,I
from feed splitters to CSTR mixers
Fd Vl E L Volumetrie flowrates
I
into CSTRs
F9 Vl E L Volumetrie flowrates
I
out of CSTRs
F:g d Vl E L Vl' E L Volumetrie flowrates
,
11'
from CSTR outlets to CSTR inlets
F:g h Vl E L VpEP Volumetrie flowrates
I,p
from CSTR outlets to produet mixer
F ph VpEP Volumetrie flowrates
of the produet streams
Coneentrations
d Vl E L Vi E I Coneentrations of species
cI,i
in the CSTR inlet streams
ej.,t Vl E L Vi E I Coneentrations of species
in the CSTR outlet streams
c;,i VpEP Vi E I Coneentrations of species
in the produet streams
Temperatures
T,m Vl E L Temperatures
I
in the CSTRs
Reaetor Volumes
Vjm Vl E L Volume of the CSTRs
Reaetion Rates
r l,7 Vl E L Vj E J Rate of reaetion j in reactor 1

Variable Bounds
The variable bounds vary from one problem to the next. The flowrates, eon-
eentrations, temperatures, volumes and reaetion rates have a lower bound of
zero due to physieal restrietions. The upper bounds on the flow rates are
set arbitrarily to help the solution algorithm. The upper bounds on the eon-
eentrations are set to maximum possible eoneentration that any speeies eould
aehieve. The temperature ranges for the nonisothermal problems are usually
speeified for eaeh problem. The upper bounds on the reaetor volumes and
reaction rates are set arbitrarily to help the solution algorithm.
Sinee the reeycle around a CSTR unit is not neeessary, the upper and lower
bounds for these flowrates are set to zero.

8.4.3 Specific Information

Speeifie information is provided for eaeh problem. This information includes
the reaetion meehanism, the reaetion parameters, the rate expressions, and
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 133

any additional information required for the problem.

Reaction Mechanism
For each problem, areaction mechanism indicating the species and reaction
steps involved is provided.
Parameters
The parameters for the problems are the stoichiometric coefficient matrix, the
kinetic constants, and the feed conditions. The stoichiometric coefficient ma-
trix, Vi,j indicates the relative number of moles of products and reactants that
participate in a given reaction.
Reaction constants
Vi E I Vj E J Stoichiometric coefficient matrix
Vj E J Kinetic rate constant
Vj E J Pre-exponential factor
Vj E J Activation Energy
Vj E J Heat of reaction
Feed conditions
F: Vr E R Flowrate of feed streams
c~!i Vr E R Vi E I Concentrations of species in feed streams

Rate Expressions
The rate expressions (JJ{Ci, T)) are specified for each problem. These expres-
sions often involve the products of the concentrations of the species. The rate
expressions have the form

fJ{Ci,T) = kj{T) II c~i,j

I

iEI

where vf.j correspond to the positive coefficients in the stoichiometric matrix

representing the forward reactions. (For equilibrium reactins, both the for-
ward and reverse reactions must appear separately and the rate constants for
both must be provided.)
Fr isothermal reactions, the values of k j are fixed. For most nonisothermal
problems, Arrhenius temperature dependence will be assumed:

kj{T) = kj exp (~i)

Additional Information
Some problems have additional restrictions on the sizes of the reactors and the
amount of conversion.

8.4.4 Problem Characteristics

This problem has bilinearities which result from the mass balances of the mix-
ers. The rate expressions usually involve the products of the concentrations
134 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

and thus lead to squared or bilinear terms for first order kinetics, trilinear
terms for second order kinetics, and so on. For nonisothermal problems, the
temperature dependence is an exponential which is multiplied by the product
of the concentrations. Because of these terms in the formulation, the problem
is nonconvex.
For each problem, the number of CSTRs in the network must be given.
Existence of the CSTRs in the network could be handled with binary variables,
however, this would lead to an MINLP.

8.4.5 Test Problems

For all of the example problems, 5 CSTRs are included in the reactor network.
The test problems described here have been used as examples in many of the
references listed above. These problems have been formulated in both the
MINOPT and GAMS modeling languages and the solutions provided have
been obtained using MINOPT (Schweiger and Floudas, 1998c, 1997). These
problems have multiple local optima which have been found by using MINOPT
to solve the problem with random starting points.

8.4.6 Test Problem 1 : Nonisothermal Van de Vusse Reaction

Case I
Problem Information

Reaction Mechanism

k} k k2 k
A I " B ----.
A----- 2 C

k3
2A
2A-----
3 .. D

Objective:
Maximize the yield of B

h
maxcl,B

Parameters

Vi,j =
[-1~ 0~ -2]~
1 -1 0
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 135

Rate Constants
reaction k E t:;.H
pGp
1 x 1099 h- l1
5.4 X 15.84 kcaljmol 84 K Ljmol
2 1.6 x10 12 Lj(mol h) 23.76 kcaljmol 108 K Ljmol
3 3.6 x10 5 h- l1 7.92 kcaljmol 60 K Ljmol
Feed Conditions
F: 100 Ljs
c~,i 1.0 moljL A, 0 moljL B, 0 moljL C

Rate Expressions

1r = kle ~
A

RT CA
I 2r = k2 e ~
A

RT CB

la = k eA

3
~
RT C~

Additional Information
The temperatures in the reactors are bounded between 300 K and 810 K.

Explicit Formulation

max = -d;B
FaI1 =
Fa - Fad'ad' + Fad
+ Fad + Fad
+ Fad + Fad
+ Fad + Fad
+ Fad
- F 1,1 1,2 1,3 1,4 1,5
ad Fgd Fgd p,gd F.g d p,gdF.g d Fgd Fg d p,gd
Fd
Fd
I1
F
= 1,1
F ad
11+ + 1,111+ + 2,121+ + 3,131+ + 4,1 41+ + p,gd 51
5,1
F.d2 = F1,2dd
ad F9d
Fgd F.9
p,gdd F.9
p,gdd F9
Fgd d p,9
p,gd d
p,d F
12+ + 1,2 + 2,2 + 3,2 + 4,2 + 5,2
12+ 22+ 32+ 42+ 52
F3d = FF1,3dd + F9d F.9 d p,gdF.9 d Fgd F9 d p,gdF.9 d
1,3 + 2,3 + 3,3 + 4,3 +
p,d ad Fgd p,gd
+ + + + + 5,3
Fd ad F9d F.g d F.g d Fg d p,gd
Fd
4 = F1,4
= F ad
14+ + 1,4
F9d
14+ + 2,4
p,gd
24+ + 3,4
p,gd
34+ + 4,4
Fgd
44+ + 5,4
p,gd
54
p,d F dd Fgd F9d p,gd p,9d p,gd F.9d Fg F9dd F,gd p,9d
p,d
5 = 1,5 +
F ad
+ 1,5 ++ 2,5 ++ 3,5 +
+ 4,5 +
+ 5,5
dd FdFd aa Fad
Fad _0 Fgb Fgb n9 p,gb p,gb _0 p,gbF.gb _0 Fgb Fgb _0 p,gb F.gb
C1,A
Cl A I1 = C1 ,A 1,1 + Gi,A 1,1 +
+ Cl,A + "2,A 2,1 + + C3,A 3,1 + + C4,A 4,1 + + C5,A 5,1
d Fd
d' Fd aa Fad
Fad Fgb
Fgb F.gb
p,gb F.gb
p,gb Fgb
Fgb p,gb
p,gb
= C1 ,B 1,1 + Gi,B 1,1 + + C2,B 2,1 + + C3,B 3,1 + + C4,B 4,1 +
_0 _0 _0 31 _0
C1,B I1
CI,B + Cl,B + C5,B 5,1
dd Fd Fd aa FadFad _0 Fg Fg bb _0 p,gb F.g b _0 p,gb F.gb _0 Fg Fgbb p,gb
F,gb
,e 1,1 + Gi,e 1,1 + C2,e 2,1 + C3,e 3,1 + C4,e 4,1 +
_0
C1 ,e 1
C1,G = C1,G
= + Cl,G + C2,G + C3,G + C4,G C5,e 5,1
+ C5,G
dd Fd Fd aa Fad
Fad Fgb
Fgb p,gb
p,gb F.gb
p,gb Fgb
Fgb F.gb
p,gb
= C1,D 1,1 + Gi,D 1,1 + + C2,D 21 + + GJ,D 3,1 + + C4,D 4,1 +
_0 _0 _0 _0 _0
C1,D 1 = + Cl,D + C5,D 5,1
dd p,dF.d aa FadFad _0 Fgb Fgb _0 p,gb F.g b _0 p,gb F.gb Fgb
_0 Fgb p,gb
p,gb
= C2,A 1,2 + Gi,A 1,2 + + C2,A 2,2 + + C3,A 3,2 + + C4,A 4,2 +
_0
_Q
C2,A 2 = + Cl,A + C5,A 5,2
dd p,d p,d
C2,B 2 c~,BFt~
a Fad
= C2 ,B 1,2 + df,BFf,t
_0
+ Cl,B
Fgb
1,2 + 4,B
_0
+ C2,B F i,t +
p,gb
2,2 4,B
_0
+ C3,B F j,t +
p,gb
3,2 q,B
_0
+ C4,B F l,t +
Fgb
4,2 S,BFl,t
_0
+ C5,B
p,gb
5,2
C2
dd p,d
C2,G ,e F.d
2
a Fad
c~,eFt~
= C2 ,G 1,2 + df,eFf,t
_0
+ Cl,G
Fgb
1,2 +
_0
+ C2,G F
4,e 2,2 i,t +
p,gb
4,eF!,~
_0
+ GJ,G
p,gb
3,2 +
_0
+ C4,G
Fgb
q,eFl,~
4,2 +
_0
+ C5,G
p,gb
S,eFl,~ 5,2
rE
-;t.,D F.2
p,dd
~,DFt~
a Fad
= C2,D 1,2 +
_0
+ Cl,D
Fg b
df,DFf,~
1,2 + 4,DFi,~
_0
+ C2,D
p,gb
2,2 + 4,DF!,~
_0
+ GJ,D
p,gb
3,2 + q,DFt,~
_0
+ C4,D
Fgb
4,2 + S,DFg,~
_0
+ C5,D
p,gb
5,2
136 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

a Fad Fgb p,gb

F.gb pgb
Fgb Fgb
Fgb p,gb
Dgb
= C3,A 1,3 + Ci.,A 1,3 + "2,A 2,3 + ca,A 3,3 + C4,A 4,3 + C5,A
d pd ~q ~ _q
-q ~q ~q
C3,A 3 C5,AL'5,3
5,3
d pd a Fad ~q Fgb ~ p,gb
F.gb ~q pgb
Fgb Fg b
~q Fgb ~q p,gb
Dgb
C3,B 3 = C3,B 1,3 + Ci.,B 1,3 + "2,B 2,3 + ca,B 3,3 + C4,B 4,3 + C5,B
C5,BL'5,35,3
a Fad ~q FgbFg b F.g b
~ p,gb F.g b
~q p,gb Fg b
~q Fgb ~q p,gb
pgb
= C3,C 1,3 + Ci.,c
Ci.,c 1,3 + "2,C 2,3 + ca,c
ca,c 3,3 + C4,c
C4,c 4,3 + C5,c
d pd
C3 ,C 3 C5,c 5,3
a Fad ~q Fgb p,gb
F.g b p,gb
F.gb ~q Fgb ~q p,gb
Dgb
C3,D 1,3 + Ci.,D 1133 + C2,D 2233 + ca,D 3,3 + C4,D 4,3 + C5,D
d pd _0 _0
C3,D 3 = C3,D C5,DL'5,3
5,3
a Fad ~O Fg b ~ p,gb F.g b ~q p,gb
F.g b ~O Fgb
Fg b p,gb
Dgb
4,A 1,4 + Ci.,A 1,4 + "2,A 2,4 + ca,A 3,4 + C4,A 4,4 + C5,A
d Fd _0
C4,A 4 =Cc4,A C5,AL'5,4
5,4
a Fad ~q FgbFg b F.g b
~ p,gb F.g b
~q p,gb Fg b
~O Fgb p,gb
Dgb
d Fd
= C4,B 1,4 + Ci.,B 1,4 + "2,B 2,4 + ca,B 3,4 + C4,B 4,4 + C5,B
_q
C4 ,B 4 C5,BL'5,45,4
a Fad ~fI Fg b F.g b
~ p,gb F.g b
~fI p,gb Fg b
~fI Fgb ~fI p,gb
Dgb
4,C 1,4 + Ci.,c
Ci.,c 1,t + "2,C, 2,t + ca,c
ca,c 3,4 + C4,c
C4,c 4,4 + C5,c
d Fd
C4 ,C 4 =Cc4,C C5,CL'5,45,4
d Fd
C4 ,D 4 =~D~~+4D~4+~D~4+~D~:+~D~:+~Dn:
a' Fad ~q' Fg b F.g b
~q' p,gb ~fI ' F.g b
pgb ~q , Fg b
Fgb p,gb'
F.gb'
= C5,A 1,5 + Ci.,A 1,5 + C2,A 2,5 + ca,A 3,5 + C4,A 4,5 + C5,A 5,5
d p,d _0'
C5,A 5
a Fad ~q Fgb ~q p,gb
F.gb ~q pgb
F.gb ~O Fgb
Fgb ~q p,gb
Dgb
= C5,B 1,5 + Ci.,B 1,5 + C2,B 2,5 + ca,B 3,5 + C4,B 4,5 + C5,B
d p,d
C5,B 5 C5,BL'5,5
5,5
d p,d a Fad ~q Fgb F.g b
~q p,gb ~q p,gb
F.gb Fgb
Fgb ~q p,gb
Dgb
= C5 ,C 1,5 + Ci.,c
Ci.,c 1,5 + C2,c
C2,c 2,5 + ca,c
ca,c 3,5 + C4,c
C4,c 4,5 + C5,c
_0
C5 ,C 5 C5,CL'5,5
5,5
d p,d a Fad ~q FgbFg b ~fI F.g b
p,gb ~q F.g b
p,gb ~O Fgb p,gb
pgb
C5,D 1,5 + Ci.,D 1,5 + C2,D 2,5 + ca,D 3,5 + C4,D 4,5 + C5,D 5,5
JI
C5,D 5 = C5,D
= 5,5

Fg
1 =Ff
=Ff
P,g =Ft
2
P,g =Ff
3
Fg =Ft
4
P,g -
--
p,d
F.5d
5
4A F f = cf,AFf + V1m (lIA,l r i':1 + lIA,2 r i':2 + lIA,3 r i':3)
ef'BFf = C1,B Ff + Vr(VB,1 r ;'\ + VB,2 r i':'2 + lIB,3
= VB,3 r i':'3)
4'c F f = cf,cFf
cf,cFf + Vr(lIC,1 r i':1 + lIC,2 r i':2 + lIC,3 r i':3)
4'D F f = cf,DFf
= cf,DFf + Vr(lID,l r i':l + lID,2 r i':2 + lID,3 r i':3)
~'AFi = c~,AFt + V2m (lIA,l r ;r:1 + lIA,2 r ;r:2 + lIA,3 r ;r:3)
4B Fi = c~,BFt + V2m (lIB,l r ;r:1 + lIB,2 r ;r:2 + lIB,3 r ;r:3)
4c F i = c~,cFt + V2m (lIC,1
(lIC,l r ;r:1 + lIC,2 r ;r:2 + lIC,3 r ;r:3)
4D Fi = c~,DFt + V2m (lID,l r ;r:1 + lID,2 r ;r:2 + lID,3 r ;r:3)
~'AFi = cg,AFf + Vr(lIA,l r T,l + lIA,2 r T,2 + lIA,3 r T,3)
4B Fl = cg,BFf + V3m (lIB,1
(lIB,l r T,1 + lIB,2 r T,2 + lIB,3 r T,3)
4c F l = cg,cFf + Vr(lIC,l r T,l + lIC,2 r T,2 + lIC,3 r T,3)
~DFI = cg,DFf + V3m (lID,l r T,1 + lID,2 r T,2 + lID,3 r T,3)
~'AFI = C~,AFt + v4m(lIA,l r T,1 + lIA,2 r T,2 + lIA,3 r T,3)
~'BFI = c~,BFt + V4m (lIB,1
(lIB,l r T,1 + lIB,2 r T,2 + lIB,3 r T,3)
~'cFl = c~,cFt + V4 (lIC,l r T,1 + lIC,2 r T,2 + lIC,3 r T,3)
m
4D Fl = c~,DFt + V4m (lID,l r T,1 + lID,2 r T,2 + lID,3 r T,3)
~',AFt = cg,AFf + v5m(lIA,l r r,1 + lIA,2 r r,2 + lIA,3 r r,3)
~BFt = cg,BFf + V5m (lIB,l r r,1 + lIB,2 r r,2 + lIB,3 r r,3)
~'cFt = cg,cFf + v5m(lIC,1 r r,1 + lIC,2 r r,2 + lIC,3 r r,3)
s'DFt
, = cg,DFf + Vr(lID,1 r r,1 + lID,2 r r,2 + lID,3 r r,3)
8.4. CHEMICAL REACTOR NETWORK PROBLEMS
B.4. 137

rr1 = 5.4 x 10ge-7971.82/Tj" qA

cf B
rr2 = 1.6 x 1012e-1l957.7/Tj"

rr3 = 3.6 x 105e-3985.91/Tj" q ~q A

5.4 x 10ge-7971.82/Tr' q'A '
rr,l =
rr,2 = 1.6 x 1012C1l957.7/Tr' ~ B
rr,3 = 3.6 x 105e-3985.91/Tr' q ~q A
rr,l = 5.4 x 10ge-7971.82/T3' C(A '
rr,2 = 1.6 x 1012e-1l957.7/T3' 4B
rr,3 = 3.6 x 105e-3985.91/T3' 4,~4,A
rT,l = 5.4 x 10 g e- 7971.82/TJ" A q
rT,2 = 1.6 x 1012e-1l957.7 /TJ" ~ B

rT,3 = 3.6 x 105e-3985.91/TJ" q ~q A

rg:1 = 5.4 x 10ge-7971.82/T5' ~'A '

rg:2 = 1.6 x 1012e-1l957.7/T5' ~,B

rg:3 = 3.6 x 105e-3985.91/T5' ~ A~ A
9d pgd pgd pgd p' gd pgh
Ff = P11+ 11+ 11+ 11+ 11+ 11
Fi 9d p9d pgd
= P21+ 21+ 21+
p9d p9d p9h
21+ 21+ 21
pg P 9d p9d p9d p9d p9d p9h
3 = 31+ 31+ 31+ 31+ 31+ 31
9d F9d Fgd Fgd Fgd Fgh
Fg = F4,1
4 + 4,1 + 4,1 + 4,1 + 4,1 + 4,1
9d pgd pgd vgd vgd vgh
Fl = 51 + 51 + 5,1 +r51 +r51 +r51
P
'h 'h h 'h 'h '
Ff = Ff,l + Fi,l + F!.l + Fl,l + F!,l
ct,AFf = q,AFf,~ + q,AFi,~ + 4,AF!.~ + q,AF!,~ + ~,AFl,~
ct,BFf = q,BFf,~ + q,BFi,~ + 4,BF1,~ + dl,BF!,~ + ~,BF!,~
ct,cFf = q,cFf,~ + q,cFi,~ + 4,cF!.~ + dl,cFl,~ + ~,cF!,~
ct,DFf = q,DFf,~ + q,DFi,~ + 4,DF!.~ + dl,DFl,~ + ~,DFl,~
Problem Statistics
No. of continuous variables 115
No. of linear equalities 17
No. of nonlinear equalities 59

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series .
Objective function: 0.8189924 mol/L

Variables

F:1, = (100.00,0.00,0.00,0.00, O.OO)T

F/d = (100.00,100.00,100.00,100.00, 100.00f

138 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Fj9 = (100.00,100.00,100.00,100.00, 100.00f

0.00 100.00 0.00 0.00 0.00)

0.00 0.00 100.00 0.00 0.00
Fj~~ = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fj~: = (0.00,0.00,0.00,0.00, 100.00f

F; = 100.00
0.000000 0.0000000 0.0000000 )
( 1.000000
0.831745 0.150840 0.0086813 0.0043669
Cl,i = 0.589306 0.355891 0.0257788 0.0145124
0.309826 0.581538 0.0495779 0.0295292
0.022384 0.805048 0.0779508 0.0473088

0.831745 0.150840 0.0086813 0.0043669)

0.589306 0.355891 0.0257788 0.0145124
c'/.,i = ( 0.309826 0.581538 0.0495779 0.0295292
0.022384 0.805048 0.0779508 0.0473088
0.003923 0.818992 0.0809024 0.0480911

C~,i = (0.003923,0.818992,0.0809024,0.0480911 f

Vjm = (0.00346267,0.0377339,0.506607,1865.11, 10000.00f

11m = (578.081,514.215,459.703,347.888, 311.952)T

4606.885 250.7101 252.2287 )

588.7233 45.31080 53.77399
rrJ =
(
49.23861 4.697744 5.928367
0.0135050 0.00152125 0.00190655
0.000168959 2.95161 X 10-5 1.56466 X 10- 5
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 139

8.4.7 Test Problem 2 : Isothermal Van de Vusse Reaction ease

I
Problem Information

Reaction Mechanism
k1 k2
A - -....a B ---=---....- C
k3
2 A - -....- D

Objective:
Maximize the yield of B:

h
maxc1,B

Parameters

Rate Constants
k1 10 s 1 (first order)
k2 1 S-l (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
Fra 100 L/s
c~,i 0.58 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

J[ = k 1cA
12 = k 2CB
Ir = k3C~
Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49
140 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series .
Objective function: 0.41233 mol/L
Variables
F:1, = (100.00,0.00,0.00,0.00, o.oof
Fjd = (100.00,100.00,100.00,100.00, 100.00f

Fr = (100.00,100.00,100.00,100.00, 100.00f

0.00 100.00 0.00 0.00 0.00)

0.00 0.00 100.00 0.00 0.00
Fl~~ = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fl~: = (0.00,0.00,0.00,0.00, 100.00)T

F; = 100.00
0.000000 0.0000000
( 0058~
0.355250 0.204543 0.0124966
000~000
0.00385517
)
Cl,i = 0.227037 0.312646 0.0297604 0.00527831
0.147296 0.371414 0.0495755 0.00585706
0.096150 0.400948 0.0707007 0.00610060

0.204543 0.0124966
( 0355250 0000385517)
0.227037 0.312646 0.0297604 0.00527831
c'/,i = 0.147296 0.371414 0.0495755 0.00585706
0.096150 0.400948 0.0707007 0.00610060
0.062929 0.412330 0.0923321 0.00620448

C;,i = (0.062929,0.412330,0.0923321, 0.00620448f

Vi m = (6.10951,5.52155,5.33504,5.26882, 5.24612f

3.55250 0.204543 0.0631012)

2.27037 0.312646 0.0257729
rIJ = ( 1.47296 0.371414 0.0108481
0.96150 0.400948 0.0046224
0.62929 0.412330 0.0019800
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 141

8.4.8 Test Problem 3 : Isothermal Van de Vusse Reaction Case

11
Problem Information

Reaction Mechanism

k1 k2
A .. B .. C
k3
2A .. D

Objective:
Maximize the yield of B:

h
maxcl,B

Parameters

[-~
0

Vi,j =
-1
1
0
-~ 1
Rate Constants
kilOs 1 (first order)
k2 1 s-1 (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
Far 100 L/s
c~,i 0.58 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

f[ = k l cA
1'2 = k 2cB
13 = k3C~

Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49
142 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series.
Objective function: 0.4166031 mol/L
Variables
F:1, = (100.00,0.00,0.00,0.00, O.OO)T
Eid = (100.00,100.00,100.00,100.00, 100.00f

Fjg = (100.00,100.00,100.00,100.00, 100.00f

0.00 100.00 0.00 0.00 0.00)

0.00 0.00 100.00 0.00 0.00
Fi~~ = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fl~; = (0.00,0.00,0.00,0.00, 100.00)T

F; = 100.00
( 0.580000
0.000000 0.0000000 0.00000000 )
0.365118 0.199499 0.0115306 0.00385254
C~i = 0.234447 0.311496 0.0286909 0.00536656
0.151716 0.373396 0.0488990 0.00598941
0.098500 0.404599 0.0706506 0.00625021

( 0.365118
0.199499 0.0115306 0.00385254 )
0.234447 0.311496 0.0286909 0.00536656
d/,i = 0.151716 0.373396 0.0488990 0.00598941
0.098500 0.404599 0.0706506 0.00625021
0.064040 0.416603 0.0929969 0.00636020

C~,i = (0.064040,0.416603,0.0929969, 0.00636020)T

Vi m = (5.77978,5.50900,5.41197,5.37610, 5.36393f

3.65118 0.199499 0.0666554)

2.34447 0.311496 0.0274827
(
rIj = 1.51716 0.373396 0.0115088
0.98500 0.404599 0.0048512
0.64040 0.416603 0.0020505
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 143

8.4.9 Test Problem 4 : Isothermal Van de Vusse Reaction Case

111
Problem Information

Reaction Mechanism

k. k2
A ~B -c
k3
2A - D

Objective:
Maximize the yield of B:

h
maxc1,B

Parameters

[-~ -n
0
-1
Vi,j = 1
0

Rate Constants
k1 10 S-1 (first order)
k2 1 s-l (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
pa 100 L/s
r
c~,i 5.8 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

f[ = k 1cA
12 = k 2CB
Ir = k3C~
Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49
144 CHAPTER B. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series .
Objective function: 3.579982 mol/L
Variables
F:,1 = (100.00,0.00,0.00,0.00, O.OO)T

Fl = (100.00,100.00,100.00,100.00, 100.00)T

Fj9 = (100.00,100.00,100.00,100.00, 100.00f

0.00 100.00 0.00 0.00 0,00)

0.00 0.00 100.00 0.00 0.00
Fl~~ = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fj~: = (0.00,0.00,0.00,0.00, 100.00f

F; = 100.00
0.00000 0.000000
( 5.80~
2.18714 2.61064 0.353851
O.~OO
0.324187
)
Cl,i = 1.42109 3.13347 0.501745 0.371845
0.98303 3.39459 0.639475 0.391450
0.69584 3.52700 0.775577 0.400792

2.61064 0.353851
( 2.18714 0,324187)
1.42109 3.13347 0.501745 0.371845
d!,i = 0.98303 3.39459 0.639475 0.391450
0.69584 3.52700 0.775577 0.400792
0.49779 3.57998 0.911252 0.405487

C~,i = (0.49779,3.57998,0.911252, 0.405487)T

ltlm = (13.55421,4.71981,4.05734,3.85885,3. 78984)T

21.87137 2.61064 2,39178)

14.21093 3.13347 1.00975
riJ = ( 9.83035 3.39459 0.48318
6.95837 3.52700 0.24209
4.97791 3.57998 0.12390
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 145

8.4.10 Test Problem 5 : Isothermal Van de Vusse Reaction

Case IV
Problem Information

Reaction Mechanism
k1 k2
A ----=------ B ----=------ C
k3
2A ------=-- D
Objective:
Maximize the yield of B:

h
maxcl,B

Parameters

Rate Constants
k1 1 s-1 (first order)
k2 2 s-1 (first order)
k3 10 L / (mol s) (second order)
Feed Conditions
F; 100 L/s
c~,i 1 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

![ = k l cA
J; = k2CB
Ir = k St3

Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49
146 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series .
Objective function: 0.06911967 mol/L
Variables

, = (100.00,0.00,0.00,0.00, O.OO)T
F:r

Fl = (100.00,100.00,100.00,100.00, 100.00f

F/g = (100.00,100.00,100.00,100.00, 100.00)T

0.00 100.00 0.00 0.00 0,00)

0.00 0.00 100.00 0.00 0.00
F/~1 = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

F/~: = (0.00,0.00,0.00,0.00, 100.00f

F; = 100.00
0.0000000 0.0000000
L~OO
( 0.307501 0.0594212 0.0374313
O.OO~O
0.297823
)
Cl,i = 0.242938 0.0645822 0.0432904 0.324595
0.204348 0.0671803 0.0482783 0.340097
0.176733 0.0685462 0.0530021 0.350859

( 0.307502 0.0594212 0.0374313 0.297823 )

0.242938 0.0645822 0.0432904 0.324595
df,i = 0.204348 0.0671803 0.0482783 0.340097
0.176733 0.0685462 0.0530021 0.350859
0.155418 0.0691197 0.0576168 0.358923

C;,i = (0.155418,0.0691197,0.0576168, 0.358923f

"{Ir = (31.49660,4.53616,3.71232,3.44572, 3.33816f

0.307502 0.118842 0,945572)

0.242938 0.129164 0.590188
TrJ =
( 0.204348 0.134361 0.417581
0.176733 0.137092 0.312346
0.155418 0.138239 0.241549
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 147

8.4.11 Test Problem 6 : Isothermal Trambouze Reaction

Problem Information

Reaction Mechanism

Objective:
Maximize the selectivity of C to A
h
cl C
max '
1- chA
1,

Parameters

Rate Constants
kl 0.025 gmol/(L min) (first order)
k2 0.2 min- l (first order)
k3 0.4 L/(mol min) (second order)
Feed Conditions
Fra 100 L/min pure A
c~,i 1 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

Ir = k l
12 = k 2CA
Ir = k 3 St
Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49
148 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Best Known Solution

Objective function: 0.5
Variables

F:r. = (100.00,0.00,0.00,0.00, o.oof

Fl = (100.00,0.00,0.00,0.00, O.OO)T

Fl = (100.00,0.00,0.00,0.00, O.OO)T
0.00 0.00 0.00 0.00 0.00)
0.00 0.00 0.00 0.00 0.00
FI~~ = ( 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00

Fl~; = (100.00,0.00,0.00,0.00, o.oof

F; = 100.00
0.00 0.00
0.00
( 100 0.00 0.00 000
0.00 )
C~i = 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00

0.250011 0.187489 0.374994

0.187506)
0.00 0.00 0.00 0.00
(
df i = 0.00 0.00 0.00 0.00
. 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00

C~.i = (0.2500113,0.1874887,0.3749944, 0.1875056)T

vr = (749.9550,0.00,0.00,0.00, o.oof
0.02500000 0.05000225
0.02500225)
0.02500000 0.000000 0.000000
rrJ =
(
0.02500000 0.000000 0.000000
0.02500000 0.000000 0.000000
0.02500000 0.000000 0.000000
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 149

8.4.12 Test Problem 7 : Isothermal Denbigh Reaction Case I

Problem Information

Reaction Mechanism

D E

Objective:
Maximize the selectivity of B to D:
h
cl B
max-'
cf.D
,

Parameters

Rate Constants
kILO L/(mol s) (second order)
k2 0.6 S-l (first order)
k3 0.6 s-l (first order)
k4 0.1 L/(mol s) (second order)
Feed Conditions
Fa
r 100 L/s
c~,i 6.0 mol/L A, 0 mol/L B, 0 mol/L C, 0.6 mol/L D, 0 mOl/L E

Rate Expressions

Ir = k1C~
12 = k 2cB
1'3 = k3cA
14 = k4C~
150 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of continuous variables 126
No. of linear equalities 27
No. of nonlinear equalities 65

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series.

Objective function: 1.232619

Variables

F:1, = (100.00,0.00,0.00,0.00, O.OO)T

Fl = (100.00,100.00,100.00,100.00, 100.00)T

Fr = (100.00,100.00,100.00,100.00, 100.00f
0.00 100.00 0.00 0.00 0.00)
0.00 0.00 100.00 0.00 0.00
Fl~t = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fj~: = (0.00,0.00,0.00,0.00, 100.00)T

F; = 100.00
0.000000 0.0000000 0.600000 O.OOOOO~ )
[ 600000
5.07454 0.405458 0.0078186 0.697855 0.00052835
Ct,i = 4.31316 0.722212 0.0233875 0.790834 0.00240236
3.68278 0.966340 0.0465604 0.879148 0.00613451
3.15761 1.150582 0.0771169 0.963005 0.01199413

0.405458 0.0078186 0.697855 0.00052835 )

[ 5.0m4
4.31316 0.722212 0.0233875 0.790834 0.00240236
d!,i = 3.68278 0.966340 0.0465604 0.879148 0.00613451
3.15761 1.150582 0.0771169 0.963005 0.01199413
2.71737 1.285165 0.1147744 1.042629 0.02006015
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 151

c;,i = (2.71737,1.285165,0.1147744,1.042629, 0.02006015)T

vr = (3.21391,3.59287,3.99668,4.42623,4.88362)
. . T

25.75090
18.60339
0.243275
0.433327
3.04472
2.58790
0.016440
0.052159
1
rij = ( 13.56289 0.579804 2.20967 0.093381
9.970488 0.690349 1.89457 0.132384
7.384113 0.771099 1.63042 0.165165

8.4.13 Test Problem 8 : Isothermal Denbigh ReactionCase 11

Problem Information

Reaction Mechanism

Objective:
Maximize the production of C subject to 95% conversion of A

maxd;,c

Parameters

-~ I
0 -1
[ -I -1 0
viJ ~ ~ 1
0
0
1
0 0

Rate Constants
k1 1.0 L/(mol s) (second order)
k2 0.6 s-1 (first order)
k3 0.6 S-1 (first order)
k4 0.1 L/(mol s) (second order)
Feed Conditions
F; 100 L/s pure A
c~,i 6.0 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D, 0 mol/L E
152 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Rate Expressions

Ir = kl~
12 = k2 CB
13 =k3cA
Ir = k4~
Additional Information
There is an additional constraint due to the requirement of 95% conversion of
A:

cf,1 = 0.3mol/L
Problem Statistics
No. of continuous variables 126
No. of linear equalities 28
No. of nonlinear equalities 65

Best Known Solution

Objective function: 3.256119 mol/L

Variables

F:,1 = (100.00,0.00,0.00,0.00, o.oof

Fjd = (100.00,94.2474,94.2474,94.2474,94.2474, 94.2474)T

Fl = (100.00,94.2474,94.2474,94.2474,94.2474, 94.2474)T

0.00 94.2474
0.00 0.00
0.00
94.2474
0.00
0.00
0.00
0.00
1
Fl~f, = ( 0.00 0.00 0.00 94.2474 0.00
0.00 0.00 0.00 0.00 94.2474
0.00 0.00 0.00 0.00 0.00

Fl~: = (5.75258,0.00,0.00,0.00, 94.2474)T

F; = 100.00
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 153

0.00000 0.00000 0.00000

O.~OOO
c~
)
5.12504 0.76777 0.01374 0.09170 0.001758
Cf,i = 2.96703 2.19583 0.28238 0.45469 0.100073
1.67993 2.44890 0.77615 0.79341 0.301607
0.35851 0.81306 2.65204 1.62058 0.555808

0.76777 0.01374 0.09170

( 5.125M 0,001758)
2.96703 2.19583 0.28238 0.45469 0.100073
cf,i = 1.67993 2.44890 0.77615 0.79341 0.301607
0.35851 0.81306 2.65204 1.62058 0.555808
0.00549 0.01260 3.45402 1.97039 0.557492

C~,i = (0.30000,0.056039,3.25612,1.86232, 0.525523f

vr = (2.98204,19.2174,31.6718,362.412, 10000.0)T

(26,2660 0.46066 3.075023 0,058947)

8.80325 1.31750 1.780217 0.482166
rlj = 2.82215 1.46934 1.007955 0.599713
0.12853 0.48784 0.215108 0.066107
0.00003 0.00756 0.003297 0.000016

8.4.14 Test Problem 9 : Isothermal Levenspiel Reaction

Problem Information

Reaction Mechanism
k
A+ B - - - -.. 2B
Objective:
Maximize the production of B
h
maxcl,B

Parameters

Rate Constants
k1 1.0 L/(mol s) (second order)
Rate Constants
F: 100 L/s
c~!i 0.45 mol/L A, 0.55 mol/L B
154 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Rate Expression

Problem Statistics
No. of continuous variables 78
No. of linear equalities 18
No. of nonlinear equalities 27

Best Known Solution

Objective function: 0.763002 mol/L
Variables

F:1, = (100.00,0.00,0.00,0.00, o.oof

Fl = (100.00,0.00,0.00,0.00, O.OO)T
Fl = (100.00,0.00,0.00,0.00, O.OO)T

0.00 100.00 0.00 0.00 0.00)

0.00 0.00 100.00 0.00 0.00
(
FI~t = 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

FI~; = (0.00,0.00,0.00,0.00, 100.00)T

F; = 100.00
0.450000 0.550000 )
0.388501 0.611499
(
Cl,i = 0.340633 0.659367
0.300790 0.699210
0.266667 0.733333

0.388501 0.611499 )
0.340633 0.659367
(
cY,i = 0.300790 0.699210
0.266667 0.733333
0.236998 0.763002
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 155

C~,i = (0.236998,0.763002)T

v;m = (25.8871,21.3123,18.9442,17.4497, 16.4067)T

rf,j = (0.237568,0.224602,0.210315,0.195555, 0.180830)T

8.4.15 Test Problem 10 : a-Pinene Reaction

Problem Information

Reaction Mechanism

Objective:
Maximize the selectivity of C to D

ctc
max-'
ctD,
Parameters

-~ I
-1 0 0 -2 0
[ -1 0 1 0 0 -1
vij ~ ~ 0 0 1
1 -1 -2
1
0
0
1
0 0 0 0 0

Rate Constants
k1 0.33384 min 1 (first order)
k2 0.26687 min- 1 (first order)
k3 0.14940 min- 1 (second order)
k4 0.18957 LJ(mol min) (second order)
k5 0.009598 LJ(mol min) (second order)
k6 0.29425 min- 1 (second order)
k7 0.011932 min- 1 (second order)
Feed Conditions
F: 100 LJmin pure A
c~.i 1.00 molJL A, 0 molJL B, 0 molJL C, 0 molJL D, 0 molJL E
156 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Rate Expressions

Ir = k1cA
12 =k2cA
13 = k3 CD
Ir = k4 bC
I~ = k 5 St
1[' = k6 C B
17 = k 7 Cc

Problem Statistics
No. of continuous variables 78
No. of linear equalities 17
No. of linear inequalities 1
No. of nonlinear equalities 27

Best Known Solution

Objective function: 1.546076
Variables
F:,1 = (89.74737,10.25263,0.00,0.00, o.oof

Fjd = (89.74737,10.25263,10.25263,10.25263, 1O.25263f

Fj9 = (89.74737,10.25263,10.25263,10.25263, 1O.25263f

0.00
0.00
( 0.00
0.00 0.00
0.00 10.25263
0.00
0.00
0.00
0.00
1
Fj~~ = 0.00 0.00 10.25263 0.00
0.00 0.00 0.00 0.00 10.25263
0.00 0.00 0.00 0.00 0.00

Fl~: = (89.74737,0.00,0.00,0.00, 10.25263f

F: = 100.00

C~i =
[1.~
1.00000
0.000000
0.000000
0.48804 0.029814
0.000000
0.000000
0.013609
0.000000
0.000000
0.174779
o.ooo~
0.000000
0.280150
1
0.00441 0.056105 0.139568 0.111465 0.548884
0.00004 0.050143 0.149973 0.098560 0.551315
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 157

d!,i =
(
1.00000
0.48804
0.00441
0.000000
0.029814
0.056105
0.000000
0.013609
0.139568
0.000000
0.174779
0.111465
0.000000
0.280150
0.548884
1
0.00004 0.050143 0.149973 0.098560 0.551315
0.00000 0.049532 0.150798 0.097536 0.551335

C;,i = (0.8974737,0.005078374,0.01546076, 0.01000000, 0.05652636f

Vjm = (0.000000,17.62920,1871.154,2051.577, 2059.640)T

T
0.33384 0.16293 0.001472 0.00001 0.00000
0.26687 0.13024 0.001177 0.00001 0.00000
0.00000 0.02611 0.016653 0.01472 0.01457
r/,= ,)
0.00000 0.00579 0.002355 0.00184 0.00180
0.00960 0.00229 0.000000 0.00000 0.00000
0.00000 0.00877 0.016509 0.01475 0.01457
0.00000 0.00016 0.001665 0.00179 0.00180

8.4.16 Test Problem 11 : Nonisothermal Van de Vusse Reac-

tion Case 11
Problem Information

Reaction Mechanism

A---- B
k1
---=-----.
k 2
C
k3
2A---- D

Objective:
Maximize the yield of B

h
maxcl,B

Parameters

I/' ' -
I,) -
[-~ -~ 1
0
2
1 -00
1
o 0
158 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Rate Constants
reaction k E AH
pCp
1 5.4x109
5.4 x 109 h-
h 1 15.84 kcal/mol 84 K L/mol
2 3.6x105 L/(mol h) 7.92 kcal/mol 108 K L/mol
3 1.6x1012 h- 1 23.76 kcal/mol 60 K L/mol
Feed Conditions
F: 100 L/s
c~.i
c~!i 1.0 mol/L A, 0 mol/L B, 0 mol/L C

Rate Expressions

I 1r
= k1e
~
RT CA

f 2r =
~
k 2 e RT CB
I 3r
= k3 e
::.
RT C~

Additional Information
The temperatures in the reactors are bounded between 450 K and 810 K.

Problem Statistics
No. of continuous variables 115
No. of linear equalities 17
No. of nonlinear equalities 59

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series.

Objective function: 0.8004376 mol/L

Variables

F:1, = (100.00,0.00,0.00,0.00, o.oof

Fl = (100.00,100.00,100.00,100.00, 100.00)T

Fr = (100.00,100.00,100.00,100.00, 100.00f

Fg d =
Fl~t
,' -
II
000
100.00
(0.00 100.00
( 0.00
0.00
0.00
0.00
0.00
0.00 0.00
0.00
100.00 0.00 0.00
0.00
0.00 100.00 0.00
1
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 159

Fl~; = (0.00,0.00,0.00,0.00, 100.00)T

F; = 100.00

0.000000 0.0000000
(L~
0.751524 0.217939 0.0142688
O.OO~O
0.0081340
)
Cl,i = 0.432529 0.480538 0.0397326 0.0236002
0.062313 0.769965 0.0765880 0.0455668
0.029024 0.793461 0.0844164 0.0465491

0.217939 0.0142688
( 0.751524 0.OO813~ )
0.432529 0.480538 0.0397326 0.0236002
d!,i = 0.062313 0.769965 0.0765880 0.0455668
0.029024 0.793461 0.0844164 0.0465491
0.014464 0.800438 0.0915585 0.0467697

C;,i = (0.014464,0.800438,0.0915585, 0.0467697f

vr = (0.0578041,0.0175675,0.00182321, 0.00037580, 0.000339863f

Tzm = (494.276,562.614,810.000,810.000, 810.000f

401.72 24.685 28,143)

1639.75 144.949 176.077
rr.J =
( 17896.06 2021.458 2409.663
8335.56 2083.146 522.770
4154.12 2101.461 129.837

8.4.17 Test Problem 12 : Nonisothermal Naphthalene Reac-

tion
Problem Information

Reaction Mechanism
160 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Objective:
Maximize the yield of B
h
maxc1,B

Parameters

[-1 -1
-1 -1 0

-n-fl
[ 1 0
0-1-1
=
~o ~1 ~1
lIi,j
ViJ =
o 0 0
Rate Constants
.
reaction kk E
E AH
pC"
1 x 1013 h-
2.0 X h 11 kcaljmol
38.00 kcal/mol 364 K Ljmol
L/mol
2 2.0 x 10 13
13 13 h- 1 kcaljmol
1 38.00 kcal/mol 129 K Ljmol
L/mol
3 8.15 X10 17
17 h- 1 1 kcaljmol
50.00 kcal/mol 108 K
K Ljmol
L/mol
4 2.1 x 105 h- 1 kcaljmol
20.00 kcal/mol 222 K
K Ljmol
L/mol
Feed Conditions
F:
Fa r Ljh
100 L/h
c~!i
c~.i moljL A, 0 mol/L
1.0 mol/L moljL B, 0 mol/L
moljL C

Rate Expressions

I 1r
= k 1e
~

12 = k. e
RT CA
~
2 RT CA

I 3r = k 3 e =
RT CB

I 4r = k 4 e ~
RT Cc

Additional Information
The temperatures in the reactors are bounded between 900 K and 1500 K

Problem Statistics
No. of continuous variables 110
No. of linear equalities 27
No. of nonlinear equalities 49

Best Known Solution

The best known solution is a network of the 5 CSTRs connected in series.

moljL
Objective function: 0.9993058 mol/L
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 161

Variables
F:,1 = (100.00,0.00,0.00,0.00, O.OO)T

Fl = (100.00,100.00,100.00,100.00, 100.00)T

F l9 = (100.00,100.00,100.00,100.00, 100.00)T

0.00 100.00 0.00 0.00 0,00]

0.00 0.00 100.00 0.00 0.00
Fj~~ = ( 0.00 0.00 0.00 100.00 0.00
0.00 0.00 0.00 0.00 100.00
0.00 0.00 0.00 0.00 0.00

Fj~: = (0.00,0.00,0.00,0.00, 100.00)T

F: = 100.00

0.00000000 0.000000

c~
0.000000000 ]
0.236839 0.00352521 0.759441 0.000194604
Cf.i = 0.061809 0.00082200 0.937006 0.000362509
0.013153 0.00011647 0.986242 0.000488288
0.001184 0.00000439 0.998228 0.000583951

( 0.236839 0.00352521 0.759441 0,000194604]

0.061809 0.00082200 0.937006 0.000362509
d!,i = 0.013153 0.00011647 0.986242 0.000488288
0.001184 0.00000439 0.998228 0.000583951
0.000045 0.00000007 0.999306 0.000648867

C:,i = (4.526830 X 10-5 ,6.976425 X 10-8 ,0.9993058,0.0006488673) T

v;m = (0.0053859,0.0029219,0.0010590,0.00018950,3.03094 x 1O- 5 )T

T{" = (941.126,964.010,1030.617,1208.119, 1461.435f
7084.837 7084.837 7019.384 3,613234]
2995.179 2995.179 3087.696 5.746510
(
rlj = 2297.241 2297.241 2363.862 11.87698
3158.110 3158.110 3217.256 50.48148
1877.826 1877.826 1892.068 214.1787
162 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

8.4.18 Test Problem 13 : Nonisothermal Parallel Reactions

Problem Information

Reaction Mechanism

c
y
A+B~
k~D

Objective:
Maximize the yield of C while minimizing the volume of the reactor

max 100c~,c - Lvi

IEL

Parameters

[
-1
-1 -1
-1 1
Vi,j = ~ ~

Rate Constants
reaction k E ~H
pC"
1 7
5.4 Xx 10 h-
7 h 1 19.138 kcal/mol 10 K L/mol
2 3.6 x10 5 L/(mol h) 9.569 kcal/mol 20 K L/mol
Feed Conditions
Ff 50 L/s pure A
F~ 50 L/s pure B
cti
c~ ,i 1.0 mol/L A, 0 mol/L B
~.i
~Ji 0 mol/L A, 1.0 mol/L B

Rate Expressions

~
f[ = k1e RT CAC~3
A

Pi = k2e RT c~5c18
~ A

Additional Information
The temperatures in the reactors are bounded between 450 K and 800 K.
B.4. CHEMICAL REACTOR NETWORK PROBLEMS 163

Problem Statistics
No. of continuous variables 115
No. of linear equalities 18
No. of nonlinear equalities 54

Best Known Solution

Objective function: 43.08948

Variables
Fad = (50.00000 0.000000 0.000000 0.000000 0.000000)
r,1 23.38640 13.09379 7.499338 4.505853 1.514625

Eid = (73.38640,86.48018,93.97952,98.48538, 100.0000)T

Fl = (73.38640,86.48018,93.97952,98.48538, 100.0000)T

0.000000 73.38640 0.000000 0.000000 0,000000)

0.000000 0.000000 86.48018 0.000000 0.000000
Fj~t = ( 0.000000 0.000000 0.000000 93.97952 0.000000
0.000000 0.000000 0.000000 0.000000 98.48538
0.000000 0.000000 0.000000 0.000000 0.000000

FI~; = (0.000000,0.000000,0.000000,0.000000, 100.0000f

( 0.681325
0.318264
0.318675
0.161930
0.000000
0.258306
o.ooooomo)
0.00159743
Ct,i = 0.158249 0.094187 0.370593 0.00318901
0.079973 0.064593 0.422854 0.00486290
0.040627 0.040627 0.452554 0.00681863

0.3750494 0.0123990 0.304393 0,00188245)

0.1719714 0.0156372 0.402730 0.00346555
(
c'/,i = 0.0838071 0.0197455 0.443128 0.00509605
0.0412520 0.0258728 0.459514 0.00692349
0.0212265 0.0212265 0.470887 0.00788614

C;,i = (0.0212265,0.0212265,0.470887, 0.00788614f

164 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

vr = (0.697019,0.792777,0.827854,0.821420, 0.860194f
Tjm = (800.0000,800.0000,800.0000,800.0000, 800.0000f

32.04841
15.75456
0.1981960
0.2037839
1
TrJ = [ 8.234240 0.2164909
4.395426 0.2470580
2.131292 0.1241009

8.4.19 Test Problem 14 : Sulfur Dioxide Oxidation

Problem Information

Reaction Mechanism

-=~. 80 3
80 2 + 21 0 2 -.-

8pecies 1 is 80 2, species 2 is 02, species 3 is 803, and species 4 is N2.

Objective:
Maximize the yield of 803

h
maxcl,3

Parameters

___
vt,J - [
-1 0.51
-0.5
1-1
1
o 0

Rate Constants
reaction k E
1 11
6.284 x 10 molj(hr kgcat) 15.500 kcaljmol 96.5 K kgjmol
2 2.732 x10 16 molj(hr kgcat) 26.7995 kcaljmol 96.5 K kgjmol
Feed Conditions
Fra 7731 kgjh
C~,i 2.5 moljkg 80 2, 3.46 moljkg 02,
o moljkg 80 3 , 26.05 moljkg N 2
8.4. CHEMICAL REACTOR NETWORK PROBLEMS 165

Rate Expressions
The rates are given in terms of kg mol 803 per kg catalyst.

Tl = k' le =Ei
T . (,1 )/1\
(cf )O.Sd,l

i~cf,i

Additional Information
The temperatures in the reactors are bounded between 300 K and 1200 K.

Problem Statistics
No. of continuous variables 110
No. of linear equalities 17
No. of nonlinear equalities 54

Best Known Solution

Objective function: 2.498196 mol/kg
moljkg

Variables

F:r = (7731.000,0.0000,0.0000,0.0000, O.OOOO)T

, =

Ft =
= (7731.000,7731.000,7731.000,7731.000, 7731.000)T

Fj9
F/9 = (7731.000,7731.000,7731.000,7731.000, 7731.000f
= 7731.000)T

0.0000 7731.000 0.0000 0.0000

0.0000 0.0000 7731.000 0.0000
( 0.0000 0.0000
0.0000
0.0000
1
Fj~l~ = 0.0000 7731.000 0.0000
0.0000 0.0000 0.0000 0.0000 7731.000
0.0000 0.0000 0.0000 0.0000 0.0000

Fj~:
F/~: = (0.0000,0.0000,0.0000,0.0000, 773l.000f
7731.000f

F: == 7731.000
 1
166 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

3.460000 0.000000
( 2.500~ 26.05000
0.150151 2.285075 2.349849 26.05000
Cl,i = 0.024384 2.222192 2.475616 26.05000
0.007785 2.213892 2.492215 26.05000
0.003335 2.211668 2.496665 26.05000

(
0.150151
0.024384
2.285075
2.222192
2.349849
2.475616
26.05000
26.05000
1
d/,i = 0.007785 2.213892 2.492215 26.05000
0.003335 2.211668 2.496665 26.05000
0.001804 2.210902 2.498196 26.05000

C~,i = (0.001804,2.210902,2.498196, 26.05000f

Vjm = (16783.9,30239.2,34328.8,45139.1, 49155.4)T

Tr = (739.297,658.340,617.072,589.701, 571.340f
2 549641
0.07608014
1.467253
0.04392624
1
riJ = ( 0.008873133 0.005134993
0.001808408 0.001046388
0.0005712887 0.0003304261

8.5 Parameter Estimation problems

8.5.1 Introduction
The estimation of parameters in semi-empirical models is a common problem
in the science and engineering fields. One statistical approach which has been
widely studied from a computational point of view, is the maximum likelihood
method. The method seeks to maximize the likelihood of the set of observed
values occurring given a set of parameters for the model and the statistical
distribution of the observations. In general, the error, i.e., the difference be-
tween the observed values and the unknown true values, is assumed to be
normally distributed with zero mean and known diagonal covariance matrix.
This assumption results in an optimization problem which seeks to minimize
the weighted squared error between the observed data and the unknown true
values obtained from the given model. A fuH treatment of the statistical as-
pects of the approach can be found in the book by Bard (1974).
Computational difficulties arise when the models used are nonlinear in
nature because the optimization problem is then nonconvex. The resulting
8.5. PARAMETER ESTIMATION PROBLEMS 167

formulation may eontain multiple minima in the area of interest and there-
fore result in inaeeurate parameter estimates. The first global optimization
approach for parameter estimation and data reeoneiliation problems was pro-
posed reeently by Esposito and Floudas (1998a,b). For detailed exposition to
the theoretical and algorithmic aspects, the reader is direeted to the book by
Floudas (2000) and the articles by Esposito and Floudas (1998a,b).
First a general formulation of the problem will be presented. Seeond, a
group of test problems will be given. The first three represent rat her simple
single equation models with only a few known loeal minima. The next three are
also only single equation models, but add an extra layer of diffieulty due to a
larger number of known loeal minima. The last two problems are real examples
of eommonly eneountered problems: a kinetie estimation from CSTR data,
and a nonideal vapor-liquid equilibrium model. These two problems involve
multiple equation models and a high degree of nonlinearity.

8.5.2 General formulation

A general formulation is presented below. Most of the examples in this section
follow this general formulation. Exeeptions to this will be noted.
Objective Function
r m
!flin ~ ~ (ZJ.I,i - Z .)2
%1',8 L...J L...J J.I,t
J.I=l i=l al
Constraints

f(zJ.l,8) = 0 p, = 1, ... ,m

Variable Bounds
Zp, -
ZJ.I 30'
30' ::;
::; zJ.l
zp, ::;
::; zp,
zJ.I + 30'
30'
8(JLL (J ::; 8
::; 8
::; u
(Ju

Variable definitions
fJ is a system of 1 algebraic funetions which represent the nonlinear model, 8(J is
a veetor of p parameters, and zJ.lzp, is a vector of i fit ted data variables at the p,th
zp, is a veetor of i experimentally observed values at the p,th data
data point. zJ.I
point, and 0'0' is a veetor of i values which represent the experimental standard
deviation of these observations.

8.5.3 Test Problem 1 : Linear Model

This example appears in Tjoa and Biegler (1992). The model represents the
fitting of a straight line to a set of data, but due to the optimization domain,
168 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

the model is in fact nonlinear. The reported global solution is from the work
of Esposito and Floudas (1998a).

Formulation

Objective Function

!,l1in
%,,, (J
L L2 (2
10 . - z""I
11,1
".2
l
"'= l '1= 1 vI

Constraints

Variable Bounds
Z", - 0.5 ::::; z'" ::; z", + 0.5
(0, -2) ::::; (J ::::; (10,2)

Data
J.I. 1 2 3 4 5 6 7 8 9 10
Z",,1 0.0 0.9 1.8 2.6 3.3 4.4 5.2 6.1 6.5 7.4
Z",,2 5.9 5.4 4.4 4.6 3.5 3.7 2.8 2.8 2.4 1.5
(7 = (1,1)

Problem Statistics
No. of variables 12
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 10
No. of nonconvex inequalities

Global Solution
Objective function: 0.61857
Parameters
(J = (5.7084, -0.54556f

Fitted Data Variables

1 2 3 4 5 6 7 8 9 10
2",,1 -0.049 0.86 1.97 2.50 3.50 4.27 5.26 5.96 6.43 7.50
2",,2 5.81 5.32 4.71 4.42 3.87 3.46 2.91 2.53 2.27 1.69
8.5. PARAMETER ESTIMATION PROBLEMS 169

8.5.4 Test Problem 2 : Polynomial Model

This example also appears in Tjoa and Biegler (1992). The model represents
the fitting of a third order polynomial to a set of data points. The reported
global solution is from the work of Esposito and Floudas (1998a).

Formulation

Variable Bounds

Z/J - 0.5 ~ z/J ~ z/J+0.5

(0, -2, -2, -2) ~ (J ~ (10,2,2,2)

Data
The data is the same as in the previous example

Problem Statistics
No. of variables 14
No. of linear equalities
No. of eonvex inequalities
No. of nonlinear equalities 10
No. of noneonvex inequalities
No. of known solutions 2

Global Solution
Objeetive funetion: 0.485152

Parameters
(J = (6.0153, -0.9998,0.15247, -0.01324)T

Fitted Data Variables

J.L 1 2 3 4 5 6 7 8 9 10
Z/J,1 0.057 0.82 1.90 2.45 3.44 4.31 5.31 6.00 6.46 7.45
Z/J,2 5.95 5.29 4.58 4.29 3.84 3.48 3.02 2.64 2.35 1.55

8.5.5 Test Problem 3 : Non-linear Model

This example appears in Tjoa and Biegler (1992) and Rod and HaneH (1980).
The reported global solution is from the work of Esposito and Floudas (1998a).
170 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Formulation

1
(h + - -
Z1 - (h

Variable Bounds

Zp. - 0.05 ::; zp. ::; zp. + 0.05

(1,1) ::; (J ::; (10,10)

Data
I-" 1 2 3 4 5 6 7 8 9
Zp.,1 0.113 0.126 0.172 0.155 0.219 0.245 0.274 0.264 0.312
Zp.,2 1.851 1.854 1.849 1.815 1.828 1.847 1.804 1.832 1.838
I-" 10 11 12 13 14 15 16 17 18
Zp.,1 0.324 0.333 0.399 0.417 0.419 0.439 0.475 0.506 0.538
Zp.,2 1.817 1.820 1.845 1.829 1.832 1.820 1.820 1.799 1.838
I-" 19 20 21 22 23 24 25
Zp.,1 0.538 0.591 0.578 0.626 0.659 0.668 0.687
Zp.,2 1.835 1.811 1.794 1.825 1.801 1.810 1.802

(j (1,1)

Problem Statistics
No. of variables 52
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 25
No. of nonconvex inequalities

Global Solution
Objective function: 4.64972 x 10-3

Parameters
(J = (2.06533,4.51079f

Fitted Data Variables

I-" 123456789
Zp.,1 0.112 0.125 0.171 0.156 0.219 0.244 0.276 0.264 0.312
Zp.,2 1.841 1.840 1.838 1.839 1.836 1.834 1.832 1.833 1.830
8.5. PARAMETER ESTIMATION PROBLEMS 171

J-t
J..t 10 11 12 13 14 15 16 17 18
ZJL,l 0.325 0.334 0.398 0.417 0.418 0.439 0.475 0.507 0.537
ZJL,2 1.830 1.829 1.825 1.824 1.824 1.823 1.821 1.819 1.816

J-t
J..t 19 20 21 22 23 24 25
ZJL,l 0.537 0.591 0.579 0.625 0.660 0.668 0.687
ZJL,2 1.817 1.813 1.814 1.811 1.809 1.808 1.807

8.5.6 Test Problem 4: Respiratory Mechanical Model

This example appears in Csendes and Ratz (1995). The model is complex in
nature, therefore the real and imaginary parts must be treated separately when
applying a standard NLP algorithm. The reported global solution is from the
work of Esposito and Floudas (1998a).

Formulation

Model

ZJL (BI + ~) + J ( WJL 04 - ~)

WJL = % J = A

Variable Bounds

ZJL - (1 + J) ::; zJL ::; zJL + (1 + J)

(0,0,1.1,0,0) ::; () ::; (1.0,1.0,1.3,1.0, 1.0)

Data
J-t11 2 3 4 5 6
ZyI5 - 5J 3 - 2J 2-J 1.5 - 0.5J 1.2 - 0.2J 1.1 - O.lJ

a = (1 + J)

Problem Statistics
No. of variables 17
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 12
No. of nonconvex inequalities
172 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Global Solution
Objective function: 0.212460

Parameters
8 == (0.60630,0.55676,1.1318,0.75020, 0.62190)T

Fitted Data Variables

~
2 3
415.13 _1 4 .93J
4.93J 2.67 -_22.07J
2.07J 1.91 - 1.lOJ

5 6
411.54 : 0.58J 1.34 -_50.23J
0.23J 1.20:
1.20 + 0.04J

8.5.7 Test Problem 5: Kowalik Problem

This example appears in Moore et al. (1992). In the model only ZI is assumed
to contain error, Zz is therefore treated as a constant and does not appear
in the objective function. The reported global solution is from the work of
Esposito and Floudas (1998a).

Formulation

Model

() zi + Zz (}z
ZI
1 zi + Zz (}3 + (}4

Variable Bounds

Zp.,1 - 0.02 ~ Zp.,1 ~ Zp.,1 + 0.02

-0.2892 ~ 8 ~ 0.2893

Data
J.L 1 2 3 4 5 6
Zp.,1 0.1957 0.1947 0.1735 0.1600 0.0844 0.0627
1/zp.,z 0.25 0.5 1 2 4 6
J.L 7 8 9 10 11
Zp.,1 0.0456 0.0342 0.0323 0.0235 0.0246
1/zp.,z 8 10 12 14 16

a 1
8.5. PARAMETER ESTIMATION PROBLEMS 173

Problem Statistics
No. of variables 15
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 11
No. of nonconvex inequalities
No. of known solutions 7

Global Solution
Objective function: 3.075 x 10- 4
Parameters
8 = (0.19283,0.19088,0.12314, 0.13578)T
Fitted Data Variables
J.L 1 2 3 4 5 6
Z/L,1 0.1944 0.1928 0.1824 0.1489 0.0928 0.0624
J.L 7 8 9 10 11
Z/L,1 0.0456 0.0355 0.0288 0.0241 0.0207

8.5.8 Test Problem 6: Pharmacokinetic Model

The model represents the fitting of kinetic data to a sum of three exponential
terms. In the model only Z2 is assumed to contain error, Z1 is therefore treated
as a constant and does not appear in the objective function. This example and
the reported global solution is from the work ofEsposito and Floudas (1998b).

Formulation

3
Z2 L fh,j exp [-02,j zIJ
j=1

Variable Bounds
o :s Z/L,2 :s 1
(-10,0) :s 8j :s (10,0.5)
Data
J.L 1 2 3 4 5 6 7 8
Z/L,1 4 8 12 24 48 72 94 118
Z/L,2 0.1622 0.6791 0.6790 0.3875 0.1822 0.1249 0.0857 0.0616
(72 = Z2
174 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of variables 14
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 8
No. of nonconvex inequalities
No. of known solutions 4

Global Solution
Objective function: 1.14 x 10- 3

Parameters
8 1 == (0.3554,2.007, -4.575f
8 2 = (0.0149,0.1102, 0.2847)T

Fitted Data Variables

p
P, i 1 2 3 4 5 6 7 8
Zp.,2 I 0.1622 0.6776 0.6817 0.3860 0.1838 0.1222 0.0875 0.0612

8.5.9 Test Problem 7: Steady-State CSTR

This model represents a steady state adiabatic CSTR with an irreversible first
order reaction (A ~ B). This model is presented by Kim et al. (1990). The
reported global solution is from the work of Esposito and Floudas (1998a).

Formulation

Model
Mass Balances
1.
r
(ZI -
1
Z2) k 1 Z2 =0
- T Z3 + k 1 Z2 =0
Energy Balance
1 -!.1H
- (Z4 - Z5) + __ (k 1 Z2)
T = 0
T pCp

Kinetic Expression

k = 0 exp [-0 (~:

1 1 2 - 1) ]

Variable Bounds

Zp. - 30' ::; zp. ::; zp. + 30'

(0.0001,5) ::; 8 ::; (0.1,15)
8.5. PARAMETER ESTIMATION PROBLEMS 175

Variable Definitions
Zl and Z2 are the inlet and outlet concentrations respectively of component A,
Z3 is the outlet concentration of component B, and Z4 and Z5 are the inlet and
out let temperatures (K) respectively. Tr is a reference temperature (800K),
H r is the heat ofreaction (-4180 Jjmol), T is the residence time of the reactor
(100s), p is the density of the reaction mixture (1.0 gjl), and Cp is the heat
capacity of the reaction mixture (4.18 Jjg K). The parameters ()l and ()2 rep-
resent transformations of the Arrhenius parameters, with ()l = Cl exp R~: and
()2 -- .!lL
lIT
r

Data

Jl. Zp.,l Zp.,2 Zp.,3 Zp.,4 Zp.,5

1 0.9871 0.8906 0.1157 547.47 663.48
2 1.0003 0.8350 0.1380 531.77 676.04
3 1.0039 0.8255 0.1850 512.21 684.81
4 0.9760 0.8020 0.2005 490.59 695.47
5 1.0129 0.7520 0.2420 464.67 703.69
6 1.0083 0.7193 0.2739 438.47 714.90
7 1.0075 0.6861 0.3215 408.04 726.09
8 0.9994 0.6388 0.3741 375.56 735.44
9 1.0007 0.5970 0.3926 340.26 745.70
10 0.9973 0.5580 0.4703 306.55 753.94
u = (0.01,0.01,0.01,1.0,1.0)

Problem Statistics
No. of variables 52
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 40
No. of nonconvex inequalities

Global Solution

Objective function: 29.0473

Parameters
() = (0.0168, 12.4332f

Fitted Data Variables

176 CHAPTER8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

J.L Z/1,l Z/1,2 Z/1,3 Z/1,4 Z/1,5

1 0.9985 0.8826 0.1159 547.84 663.78
2 0.9878 0.8437 0.1441 531.49 675.63
3 1.0011 0.8282 0.1729 512.06 684.95
4 0.9920 0.7874 0.2046 490.84 695.44
5 1.0058 0.7660 0.2398 464.80 704.62
6 1.0005 0.7242 0.2763 438.34 714.62
7 0.9995 0.6822 0.3173 407.66 724.94
8 0.9997 0.6377 0.3600 375.35 735.32
9 1.0013 0.5962 0.4052 340.39 745.55
10 1.0071 0.5587 0.4484 306.56 754.96

8.5.10 Test Problem 8: Vapor-Liquid Equilibrium Model

A two parameter Van Laar equation is used to model binary vapor-liquid equi-
librium data which consists of four measured qualtities (Pressure, Tempera-
ture, liquid and vapor mole fractions). The data is for the system consisting of
methanol (component 1) and 1,2-dichloroethane (component 2). This model
was studied by Kim et al. (1990) and Esposito and Floudas (1998a). The
reported global solution is from the work of Esposito and Floudas (1998a).

Formulation

Model
Equilibrium Expressions
')'1 Zl pr(T) - Z2 Z4 = 0
')'2 (1 Zl) p~(T) - (1 - Z2) Z4 = 0

Antoine Equation

p,/(T) = exp [Ci ,l - T, Ci,2 C ] i = 1,2

r Z3 - i,3

Two Parameter Van Laar Equation

In')'l = !!.t
Z3
(1 + !!.t ~)-2
82 1 - Zl

In')'2 =~
Z3
(1 + ~~)-2
81 Zl

Variable Bounds

Z/1 - 30- ~ z/1 ~ z/1 + 30-

1 ~ 0 ~ 2

Variable Definitions
Zland Z2 are the liquid and vapor mole fractions respectively of component 1
8.5. PARAMETER ESTIMATION PROBLEMS 177

(methanol), Z3 is the scaled system temperature, (scaled by Tr = 323.15 K),

and Z4 is the system pressure (mmHg). The parameters fh and fh are scaled
Van Laar coefficients, with fh = Ar
and (h = ~r. p~(T) is the vapor press ure
of the pure component i, described by the Antoine Equation with coefficients,
Ci,j. "/i is the activity coefficient of component i.

Data

Antoine Coefficients

1 18.5875 3626.55 34.29

2 16.1764 2927.17 50.22

Observations
J.L Z/"l Z/,,2 Z/,,3 Z/,,4
1 0.30 0.591 1.00 483.80
2 0.40 0.602 1.00 493.20
3 0.50 0.612 1.00 499.90
4 0.70 0.657 1.00 501.40
5 0.90 0.814 1.00 469.70

(7' (0.005,0.015,3.09 x 10-4, 0.75)

Problem Statistics
No. of variables 42
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities 30
No. of nonconvex inequalities

Global Solution
Objective function: 3.32185
Parameters
(J = (1.9177, 1.6082f

Fitted Data Variables

J.L Z/"l Z/,,2 Z/,,3 Z/,,4

1 0.299 0.596 1.00 483.99
2 0.400 0.612 1.00 493.28
3 0.500 0.624 1.00 499.67
4 0.699 0.667 1.00 501.29
5 0.901 0.810 1.00 469.71
178 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

8.6 Phase and Chemical Equilibrium Problems - Equa-

tions of State

8.6.1 Introduction

The phase and chemical equilibrium problem is extremely important for pre-
dicting fluid phase behavior for most separation process applications. Process
simulators must be able to reliably and efficiently predict the correct number
and type of phases that exist at equilibrium and the distribution of compo-
nents within those phases. The Gibbs free energy is the thermodynamic func-
tion most often used for equilibrium calculations because it can be applied at
conditions of constant temperature and pressure. A global minimum of the
Gibbs free energy corresponds to the true equilibrium solution. For many sys-
tems, the Gibbs free energy surface is nonconvex, therefore local optimization
methods can provide no guarantees that the correct equilibrium solution has
been located.
Despite the importance of the phase and chemical equilibrium problem,
there has been relatively little work published on methods for solving the
problem using equations of state. The use of equations of state results in more
complicated expressions for thermodynamic quantities compared to activity
coefficient equations, see Chapter 6. Most of the methods using equations of
state that have appeared in the literat ure are confined to verifying the stability
of a solution by use of the tangent plane stability criterion. Michelsen (1982)
presented a local solution method that uses multiple, intelligently chosen initial
guesses. Eubank et al. (1992) developed an "area" method. This method can
be reliable, but provides no mathematical guarantee that the global minimum
solution is found. Sun and Seider (1995) applied a homotopy-continuation
method, and Hua et al. (1998) presented an interval Newton/generalized bi-
section method for finding all stationary points.
In this section, the phase and chemical equilibrium problem using equations
of state is presented. Equations of state have a number of advantages over
activity coefficient equations. First is that they are applicable over a wide
range of pressures, while activity coefficient equations can be used only at low
press ure. Second is that many equations of state can be used to represent the
behavior of both liquid and vapor states.
When formulated using equations of state, the Gibbs free energy and tan-
gent plane distance problems do not contain special structures and therefore
they fall into the general dass of twice-continuously differentiable nonlinear
programming problems.
8.6. PHASE AND CHEMICAL EQUILIBRIUM PROBLEMS 179

8.6.2 General formulation - Tangent Plane Distance Minimiza-

tion
For a discussion of the general formulation of the tangent plane distance min-
imization problem, refer to Chapter 6.
Objective function
The objective is to minimize the distance between the Gibbs energy sur-
face for an incipient phase and the tangent plane to the Gibbs energy surface
constructed at the feed point (xF,zF). If a negative solution exists, then the
feed composition corresponds to an unstable solution .

."
~l~ L.." Xi In Xi
, iEC
"A
+ L.." Xi In CPi (Xi, z) -
iEC
"
L.." Xi In xiF'F
iEC
CPi

Constraints
Equation of State

E(x,z) = 0

Mole balance

LXi 1
iEC

Variable bounds

o~ Xi ~ 1 Vi E C
O~ z

Variable definition

Xi - the mole fraction of component i in the incipient phase.

z - the compressibility of the incipient phase.

~i - the fugacity coefficient of component i in the incipient phase.

180 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

8.6.3 Van der Waals Equation

One of the earliest and most famous equations of state developed for nonideal
systems is that of van der Waals (1873). Despite its relative simplicity, the van
der Waals equation is remarkably aceurate for many systems. The equation of
state in terms of the eompressibility factor, and the fugaeity eoeffieient term
derived from it, are given below.

I: Xi ln~i(xi' z)
B 2A
- - -ln(z-B) - -
z-B z
iEC
E(x, z) Z3 - (B + l)z2 + Az - AB

where,

A I: I: aijXiXj
iEC jEC

B I:biXi
iEC

8.6.4 Test Problem 1

Green et al. (1993) used this ternary system to demonstrate the failure of
Newton's method to eonverge to the true equilibrium solution. For the feed
eomposition given in this example, the van der Waals equation prediets three
loeal extrema. The global minimum is negative, therefore the feed is an un-
stable solution.

Explicit Formulation

min X1lnx1 + x2lnx2

+..l!...
z-B
-ln(z - B) _ 2A
z
F"F F"F
-X1 1nx1 <PI - x2lnx2 <P2
subject to z3 - (B + l)z2 + Az - AB = 0
A- al1xI - 2a12X1X2 - a22X~ 0
B - b1X1 - b2X2 = 0
Xl +X2 = 1
X1,X2,Z 2: 0
8.6. PHASE AND CHEMICAL EQUILIBRlUM PROBLEMS 181

Data

p = 80 atm
T = 400 K
R = 82.06 cm3 (atm)mol- 1 K- 1
zF = 0.55716
a: F = (0.83,0.085,0.085 f
b = (O.14998,0.14998,0.14998f
lnt/J[ = (-0.244654, -1.33572, -0.457869)T
( 0.37943 0.75885 0.48991 )
a 0.75885 0.88360 0.23612
0.48991 0.23612 0.63263

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities -
No. of known solutions 3

Global Solution
Objective function: -0.00988

Continuous variables

z = 0.26249
a: (0.6967,0.2090, 0.0943f

8.6.5 Test Problem 2

This is the same ternary system as in Example 1. In this case the feed com-
position given is a stable solution.

Data

ZF 0.331438
a: F (0.69,0.155, 0.155f
lnt/J[ (-0.214553, -1.67364, -0.319125)T
182 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities
No. of known solutions 3

Global Solution
Objective function: 0.0

Continuous variables

z = 0.284913
x = (0.69,0.155, 0.155)T

8.6.6 SRK Equation

The Redlick-Kwong equation has been used extensively since its introduction.
It has a form similar to that of the van der Waals equation but the results are
often a significant improvement. Soave (1972) proposed a modification of the
Redlich-Kwong equation. The modified equation of state (SRK) is given below
in terms of the compressibility factor.

L Xi ln~i(xi, z)
iEG
z - l-ln(z - B) - ~ In (1 + ~)
E(x, z) z3 - z2 + (A - B 2 - B)z - AB

where,

A LLaijXiXj
iEG JEG

B = LbiXi
iEG

8.6.7 Test Problem 3

The binary system of hydrogen sulfide and methane has appeared frequently
in the literature of phase equilibria. Michelsen (1982) has demonstrated that
for certain feed composition, like the one given in this example, this problem
can be diflicult to solve because it contains multiple local minima.
8.6. PHASE AND CHEMICAL EQUILIBRlUM PROBLEMS 183

Data

P = 40.53 bar
T = 190 K
R = 83.14 cm3 (bar)mol- 1K- 1
zF 0.531868
xF (0.0187,0.9813)T
In ,pr (-1.0672, -0.3480)T
b = (0.0771517,0.0765784f

a ( 1.04633 0.579822 )
0.579822 0.379615

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities -
No. of known solutions 5

Global Solution

Objective function: -0.004

Continuous variables

z 0.16436
x (0.0767,0.9233f

8.6.8 Test Problem 4

The binary system methane - propane was shown by Hua et al. (1998) to
contain three stationary points for the feed condition given below.
184 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Data

P = 100 bar
T 277.6 K
R 83.14 cm3 (bar)mol- 1 K- 1
zF 0.439853
zF = (0.68,0.32f
lnc/Jf (0.0234027, -2.13584)T
b (0.12932,0.271567)T
( 0.352565 0.844083 )
a = 0.844083 2.14335

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities
No. of known solutions 3

Global Solution
Objective function: -0.00033
Continuous variables

z = 0.54594
z (0.7721, 0.2279f

8.6.9 Peng-Robinson Equation

The Peng-Robinson cubic equation of state was developed in order to predict
the behavior of the liquid phase more accurately than previous equations like
van der Waals and SRK. The Peng-Robinson equation is slightly more com-
plicated than its predecessors, due to the increased dependence on the mixing
terms A and B. Both the equation of state and the fugacity coefficient expres-
sion derived from it are given below.

~::>i In ~i(Xi, z) = z -l-ln(z _ B) _ ~

21.5B
In (z + (1 + J2)B)
z + (1- J2)B
iEC
E(z, z) z3 - (1 - B)z2 + (A - 3B 2 - 2B)z - AB + B 3 + B 2
8.6. PHASE AND CHEMICAL EQUILIBRIUM PROBLEMS 185

where,

A = 2: 2: ijXiXj
iEC jEC

B = 2:biXi
iEC

8.6.10 Test Problem 5

The binary system, carbon dioxide - methane, was shown by Hua et al. (1998)
to contain three stationary points for the feed condition given in this example
problem. The global minimum tangent plane distance is negative, therefore
the given feed condition is unstable.

Data

p = 60.8 bar
T = 220 K
R 83.14 cm3 (bar)mol- 1 K- 1
zF 0.445886
re F = (0.20,0.80)T
In 1/>[ (-0.965195, -0.323254)T
b (0.0885973,0.0890893)T
( 0.884831 0.555442 )
a = 0.555442 0.427888

Problem Statistics
No. of continuous variables 3
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities
No. of known solutions 3

Global Solution
Objective function: -0.007
Continuous variables

z 0.15949
re (0.4972,0.5028)T
186 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

8.6.11 Test Problem 6

This example is the ternary system of nitrogen - methane - ethane. For the
feed conditiongiven, the Peng-Robinson equation has three stationary points,
as reported by Hua et al. (1998). The global minimum tangent plane distance
is negative, therefore the feed composition is an unstable solution.

Data

P 76 bar
T 270 K
R = 83.14 cm3 (bar)mol- 1 K- 1
zF 0.448135
zF = (0.15,0.30,0.55f
lnq,f (0.468354, -0.067012, -1.0288f
b = (0.0815247,0.0907391,0.13705f
( 0.142724 0.206577 0.342119 )
a = 0.206577 0.323084 0.547748
0.342119 0.547748 0.968906

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of convex inequalities
No. of nonlinear equalities 1
No. of nonconvex inequalities
No. of known solutions 3

Global Solution
Objective function: -0.0012
Continuous variables

z 0.30572
z (0.097,0.245, 0.658f

8.7 Clusters of Atoms and Moleeules

8.7.1 Introduction
The theoretical prediction of the physical properties of materials is an im-
portant tool for understanding and interpreting experimental results at the
8.7. CLUSTERS OF ATOMS AND MOLECULES 187

atomistic level. Such properties are governed by the interatomic and inter-
molecular forces of the system, which can be studied through computational
simulations of atomic and molecular clusters. For this reason, clusters have
been the focus of many recent theoretical, as weIl as experimental, studies.
One important result of such studies will be the identification of the ap-
propriate form of interatomic potentials for the prediction of bulk material
properties. That is, the reliability of theoretical predictions is dependent on
the choice of potential model. Accurate results can be obtained by solving the
Schrdinger equation exactly, although the large computational effort associ-
ated with this problem greatly limits the size of system that can be considered.
Therefore, empirical interatomic potentials have been developed in order to
effectively address larger systems. As a first order approximation, these poten-
tials can been modeled as pairwise interaction (two-body) terms (e.g., this is
appropriate for rare gas clusters). However, when considering atoms without
closed-shell structures, the use of more detailed potentials that incorporate
many body terms (i.e., usually three-body) is required.
Once a potential has been chosen, a method for examining the potential en-
ergy surface must also be selected. The ability of these procedures to traverse
complicated energy hypersurfaces must be measured, since this can greatly
affect the quality of theoretical predictions. Particularly important is the effi-
ciency of these methods in locating the global minimum energy configuration
of the cluster. TypicaIly, the identification of these global minimum energy
structures is used to set benchmarks for global optimization algorithms.
A database of best known solutions for several classes of cluster problems
(including Lennard-Jones and Morse potentials) can also be found at the fol-
lowing web address :

http://brian.ch.cam.ac.uk/CCD.html

8.7.2 General Formulation

A general formulation is presented below. This formulation represents aglobai
minimization problem for a cluster system containing N particles. For com-
pleteness, the formulation includes up to N-body interaction terms, although
most examples consider only two- and three-body interaction terms. The one-
body term, which describes external forces on the system, is generally disre-
garded so that the expansion begins with the pair interaction terms.
Objective function

LU2 (ri,rj) + L U3 (ri,rj,rk)+ ... +UN(ri,rj, ... ,rN)

i<j i<j<k
188 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Constraints

Xl = Yl = Zl = Y2 = Z2 = Z3 = 0

Variable Bounds

Xf ::; Xi ::; xf, = 2, ... ,N

yf ::; Yi ::; yf, = 3, ... ,N
xf ::; Xi ::; xf, = 4, ... ,N

Variable definitions
c:P represents the potential to be minimized, which is a function of the Carte-
sian coordinate vectors for the N individual atoms. The convention used for
indexing the summations (which will be used in the following test problems) is
to include all combinations of indices for which the imposed condition applies.
For example, L:: implies that all combinations of i and j for which i < j should
i<j
be included in the summation. The Cartesian coordinate vector for atom i is
given by ri = (Xi, Yi, Zi)T. U2, U3 and UN represent the two-, three- and
N-body terms, respectively. In order to set the translational and rotational
degrees of freedom, six constraints are added. These set the first atom at the
origin, the second along the x-axis and the third in the x-y plane. Finally, xf,
yf, zf, xf, yfand zf denote the lower and upper bounds for the x, y, and z
Cartesian coordinates, respectively.

8.7.3 Lennard-Jones Potential

The global optimization of Lennard-Jones clusters is a two-body problem used
to simulate heavy atom rare gas clusters such as argon, xenon and krypton.
Computer simulations have been successful in predicting transitions from non-
crystalline geometries to bulk face-centered cubic (fee) crystalline forms (Leary
(1997)). The problems are highly non-convex; that is, the number of distinct
local minima is believed to grow exponentially as the number of atoms, N,
increases (Hoare and McInnes (1983)).
The objective function is measured in units of energy, and interatomic
distances are usually given in angstroms [A]. Both values are easily scaled for
homogeneous clusters of atoms. A plot of the scaled Lennard-Jones potential
for a single atomic pair is given in Figure 8.3. The potential represents the
difference between a strong, short-range repulsive term (r- 12 ) and a longer-
range attractive term (2r- 6 ).
8.7. CLUSTERS OF ATOMS AND MOLECULES 189
2~r-----,-------~--------r-------,-------~-------.
Scaled Lennard-Jones - -

1.5

~
CI)
c:
CI)

~
CI) 0.5
8.
i
<ii
dl
o

-0.5

-1 L-____ ~~~ ____ ~ ________L __ _ _ _ _ _ ~ ______ ~ ______ ~

0.8 12 1A 1 1.8 2
Scaled interatomic distance

Figure 8.3: Scaled Lennard-Jones potential

Formulation

Objective function

min
rl, ... ,rN
4E2:[(;.)12 _(;.)6]
i<j tJ tJ

Constraints

Xl = Y1 = zl = Y2 = z2 = Z3 = 0

Variable definitions
E and 21/ 6 a are the pair equilibrium weH depth and separation, respectively.
TypicaHy, for homogeneous atom clusters, reduced units are used; that is E =
a = 1. After global minimum energy geometries have been found, the
distances can be rescaled for various microcluster systems. It should also be
noted that rij refers to the Euclidean distance between atoms i and j (i.e.,
rij = [(Xi - Xj)2 + (Yi - Yj)2 + (Zi - zj)21~).
190 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Problem Statistics
No. of continuous variables 3N-6
No. of linear equalities 6
No. of known solutions Grows exponentially with N

Best Known Solutions

The following table provides the best known solutions for the scaled Lennard-
Jones (homogeneous) microcluster problem. From left to right the columns
indicate: number of atoms in the cluster, scaled energy, geometry of the clus-
ter, and the earliest reference in which the solution was reported. The geome-
try classification codes include polytetrahedral structures (PT), face-centered
cubic lattice structures (FCC), and two multilayer icosahedrallattice confor-
mations : sublattices that comprise Mackay icosahedron (IC) and sublattices
consisting of tetrahedrally bonded face-centered sites (FC).

N Energy Geometry Reference

4 -6.000000 PT Hoare and Pal (1971)
5 -9.103852 PT Hoare and Pal (1971)
6 -12.712062 FCC Hoare and Pal (1971)
7 -16.505384 PT Hoare and Pal (1971)
8 -19.821489 PT Hoare and Pal (1971)
9 -24.113360 PT Hoare and Pal (1971)
10 -28.422532 PT Hoare and Pal (1971)
11 -32.765970 PT Hoare and Pal (1971)
12 -37.967600 PT Hoare and Pal (1971)
13 -44.326801 IC Hoare and Pal (1971)
14 -47.845157 FC Hoare and Pal (1971)
15 -52.322627 FC Hoare and Pal (1971)
16 -56.815742 FC Hoare and Pal (1971)
17 -61.317995 FC Freeman and Doll (1984)
18 -66.530949 FC Hoare and Pal (1971)
19 -72.659782 FC Hoare and Pal (1971)
20 -77.177043 FC Hoare and Pal (1971)
21 -81.684571 FC Hoare and Pal (1971)
22 -86.809782 FC Northby (1987)
23 -92.844472 FC Farges et al. (1985)
24 -97.348815 FC Wille (1987)
25 -102.372663 FC Hoare and Pal (1971)
26 -108.315616 FC Hoare and Pal (1971)
27 -112.873584 FC Northby (1987)
28 -117.822402 FC Northby (1987)
29 -123.587371 FC Hoare and Pal (1971)
30 -128.286571 FC Northby (1987)
8.7. CLUSTERS OF ATOMS AND MOLECULES 191

N Energy Geometry Reference

31 -133.586422 IC Northby (1987)
32 -139.635524 IC Northby (1987)
33 -144.842719 IC Northby (1987)
34 -150.044528 IC Northby (1987)
35 -155.756643 IC Northby (1987)
36 -161.825363 IC Northby (1987)
37 -167.033672 IC Northby (1987)
38 -173.928427 IC Pillardy and Piela (1995),
Doye et al. (1995)
39 -180.033185 IC Northby (1987)
40 -185.249839 IC Northby (1987)
41 -190.536277 IC Northby (1987)
42 -196.277534 IC Northby (1987)
43 -202.364664 IC Northby (1987)
44 -207.688728 IC Northby (1987)
45 -213.784862 IC Northby (1987)
46 -220.680330 IC Northby (1987)
47 -226.012256 IC Northby (1987)
48 -232.199529 IC Northby (1987)
49 -239.091864 IC Northby (1987)
50 -244.549926 IC Northby (1987)
51 -251.253964 IC Northby (1987)
52 -258.229991 IC Northby (1987)
53 -265.203016 IC Northby (1987)
54 -272.208631 IC Northby (1987)
55 -279.248470 IC Hoare and Pal (1972)
56 -283.643105 IC Northby (1987)
57 -288.342625 FC Northby (1987)
58 -294.378148 FC Northby (1987)
59 -299.738070 FC Northby (1987)
60 -305.875476 FC Northby (1987)
61 -312.008896 FC Northby (1987)
62 -317.353901 FC Northby (1987)
63 -323.489734 FC Northby (1987)
64 -329.620147 FC Northby (1987)
65 -334.971532 FC Xue (1994)
66 -341.110599 FC Coleman et al. (1994),
Xue (1994)
67 -347.252007 FC Northby (1987)
68 -353.394542 FC Northby (1987)
69 -359.882566 FC Wales and Doye (1997)
192 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

N Energy Geometry Reference

70 -366.892251 FC Northby (1987)
71 -373.349661 FC Northby (1987)
72 -378.637253 FC Coleman et al. (1994)
73 -384.789377 FC Northby (1987)
74 -390.908500 FC Northby (1987)
75 -397.492331 FC Doye et al. (1995)
76 -402.894866 FC Doye et al. (1995)
77 -409.083517 FC Doye et al. (1995)
78 -414.794401 FC Wales and Doye (1997)
79 -421.810897 FC Northby (1987)
80 -428.083564 FC Northby (1987)
81 -434.343643 FC Northby (1987)
82 -440.550425 JC Northby (1987)
83 -446.924094 JC Northby (1987)
84 -452.657214 JC Northby (1987)
85 -459.055799 FC Northby (1987)
86 -465.384493 JC Northby (1987)
87 -472.098165 JC Northby (1987)
88 -479.032630 JC Deaven et al. (1996)
89 -486.053911 IC Northby (1987)
90 -492.433908 JC Northby (1987)
91 -498.811060 JC Northby (1987)
92 -505.185309 JC Northby (1987)
93 -510.877688 JC Northby (1987)
94 -517.264131 Ie Northby (1987)
95 -523.640211 JC Northby (1987)
96 -529.879146 JC Northby (1987)
97 -536.681383 JC Northby (1987)
98 -543.642957 JC Deaven et al. (1996)
99 -550.666526 IC Northby (1987)
100 -557.039820 JC Northby (1987)
101 -563.411308 JC Northby (1987)
102 -569.363652 JC Doye and Wales (1995)
103 -575.766131 JC Doye and Wales (1995)
104 -582.086642 JC Doye and Wales (1995)
105 -588.266501 JC Northby (1987)
106 -595.061072 JC Northby (1987)
107 -602.007110 JC Wales and Doye (1997)
108 -609.033011 JC Northby (1987)
109 -615.411166 JC Northby (1987)
110 -621.788224 JC Northby (1987)
8.7. CLUSTERS OF ATOMS AND MOLECULES 193

N Energy Geometry Reference

111 -628.068416 Ie Northby (1987)
112 -634.874626 IC Northby (1987)
113 -641. 794704 Ie Barron et al. (1997)
114 -648.833100 Ie Northby (1987)
115 -655.756307 IC Barron et al. (1997)
116 -662.809353 Ie Northby (1987)
117 -668.282701 Ie Northby (1987)
118 -674.769635 Ie Northby (1987)
119 -681.419158 Ie Northby (1987)
120 -687.021982 Ie Northby (1987)
121 -693.819577 IC Northby (1987)
122 -700.939379 Ie Northby (1987)
123 -707.802109 Ie Northby (1987)
124 -714.920896 Ie Northby (1987)
125 -721.303235 Ie Northby (1987)
126 -727.349853 IC Northby (1987)
127 -734.4 79629 IC Northby (1987)
128 -741.332100 Ie Northby (1987)
129 -748.460647 IC Northby (1987)
130 -755.271073 Ie Northby (1987)
131 -762.441558 Ie Northby (1987)
132 -768.042203 Ie Northby (1987)
133 -775.023203 IC Northby (1987)
134 -782.206157 Ie Xue (1994)
135 -790.278120 Ie Northby (1987)
136 -797.453259 Ie Northby (1987)
137 -804.631473 IC Northby (1987)
138 -811.812780 Ie Northby (1987)
139 -818.993848 IC Northby (1987)
140 -826.174676 Ie Northby (1987)
141 -833.358586 Ie Northby (1987)
142 -840.538610 Ie Northby (1987)
143 -847.721698 Ie Northby (1987)
144 -854.904499 IC Northby (1987)
145 -862.087012 Ie Northby (1987)
146 -869.272573 Ie Northby (1987)
147 -876.461207 Ie Northby (1987)

8.7.4 Morse Potential

The Morse (1929) potential, like the Lennard-Jones potential, is used to model
clusters by considering only two-body interaction terms. However, in this
case a single variable parameter can be adjusted to shift the range of the
194 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

potential. This diversity offers an increased probability of accurately predicting

geometries of various metal clusters. This increased variability coupled with a
rapid increase in the number of distinct local minima as cluster size increases,
makes the Morse potential a rigorous test problem for global optimization
algorithms. As with the Lennard-Jones potential, the objective function is
measured in units of energy. A plot of the scaled Morse potential for a single
atomic pair (for several values of its variable parameter) is given in Figure 8.4.

Scaled Morse p=3 -

Scaled Morse p=6 ----.
Scaled Morse p=14
Scaled Lennard-Jones
1.5

0.5

-0.5

0.8 1.2 1.4 1.6 1.B 2

Scaled interatomic distance

Figure 8.4: Scaled Morse potential

Formulation

Objective function

mm
rl,,rN

Constraints

Xl = Yl = Zl = Y2 = Z2 = Z3 = 0

Variable definitions
8.7. CLUSTERS OF ATOMS AND MOLECULES 195

and 1'0 are the pair equilibrium weIl depth and separation, respectively. For
homogeneous atom clusters, reduced units can be used (i.e., f = a = 1),
and then later rescaled. p is an adjustable parameter that broadens (small
p) or narrows (large p) the range of the potential. It should also be noted
that 1'ij refers to the Euclidean distance between atoms i and j (i.e., 7'ij
[(Xi - Xj)2 + (Yi - Yj)2 + (Zi - Zj)2l~).

Problem Statistics
No. of continuous variables 3N-6
No. of linear equalities 6
No. of known solutions Grows exponentially with N

Best Known Solutions

The following table provides the best known solutions for the reduced Morse
potential microcluster problem. Results for p values of 3.0 and 6.0 are pre-
sented. This information was compiled from results presented in Doye and
Wales (1997) and Doye et al. (1995).

N Energy (p = 3.0) Energy (p = 6.0)

5 -9.299500 -9.044930
6 -13.544229 -12.487810
7 -17.552961 -16.207580
8 -22.042901 -19.327420
9 -26.778449 -23.417190
10 -31.888630 -27.473283
11 -37.930817 -31.521880
12 -44.097880 -36.400278
13 -51.737046 -42.439863
14 -56.754744 -45.619277
15 -63.162119 -49.748409
16 -69.140648 -53.845835
17 -75.662417 -57.941386
18 -82.579266 -62.689245
19 -90.647461 -68.492285
20 -97.417393 -72.507782
21 -104.336946 -76.529139
22 -112.041223 -81.136735
23 -120.786879 -86.735494
24 -127.884549 -90.685398
25 -136.072704 -95.127899
26 -145.322134 -100.549598
27 -152.513867 -104.745275
28 -160.773356 -108.997831
196 CHAPTERB. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

N Energy (p = 3.0) Energy (p = 6.0)

29 -170.115560 -114.145949
30 -177.578647 -118.432844
31 -185.984248 -122.857743
32 -195.468461 -127.771395
33 -204.208737 -132.287431
34 -214.068392 -136.797544
35 -221. 771452 -141.957188
36 -230.508264 -147.381965
37 -240.008130 -151.891203
38 -249.159174 -157.477108
39 -258.944962 -163.481990
40 -268.394773 -167.993097
41 -278.405573 -172.526828
42 -288.335415 -177.680222
43 -298.172449 -183.092699
44 -308.277011 -187.626292
45 -318.660653 -192.954739
46 -327.033118 -199.177751
47 -336.666189 -203.704178
48 -346.662788 -209.044000
49 -356.412817 -215.253702
50 -366.635589 -219.820229
51 -376.673413 -225.391240
52 -387.587332 -231.615013
53 -398.783184 -237.834976
54 -407.966010 -244.058174
55 -417.918562 -250.286609
56 -428.611289 -253.922955
57 -439.960320 -258.041717
58 -449.432282 -263.410755
59 -459.509280 -267.945226
60 -470.448485 -273.341243
61 -482.025765 -278.726626
62 -491.052378 -283.183002
63 -501. 731893 -288.560948
64 -512.831683 -293.931716
65 -523.446427 -298.392345
66 -534.464040 -303.763297
67 -544.754257 -309.130322
68 -555.582496 -314.374880
69 -566.364140 -319.819905
8.7. CLUSTERS OF ATOMS AND MOLECULES 197

N Energy (p = 3.0) Energy (p = 6.0)

70 -577.739782 -325.887749
71 -588.396687 -331.588748
72 -599.690310 -336.121753
73 -610.936684 -341.266253
74 -622.679870 -346.610834
75 -633.513370 -351.472365
76 -644.951602 -356.372708
77 -656.079789 -361.727086
78 -667.576295 -366.761455
79 -678.940231 -372.832290
80 -690.577890 -378.333471

8.7.5 Tersoff Potential

The Tersoff (1988) potential combines both two and three-body interaction
terms together into one functional form. That is, for each pair of atoms i
and j, a multi-body term, which depends on the positions and the neighbors
of atom i, is also included into the pairwise potential formulation. As in the
Morse potential, exponential functions are chosen to model the attractive and
repulsive forces. This empirical many-body potential was originally employed
to model the bulk properties of silicon (Si), but has since been appIied to
various metal cluster systems.

Formulation

Objective function

min
rl,,rN
L fc(rij) [VR(rij) + bijVA(r;j)]
i<j

where

c2 ~
g((}) = 1 + J2 - J2 + (h - cos(}) 2 ,

1' r < R-D,

{
fc(r) = ~ - ~sin[1T(;;R)], R - D < r < R + D,
0, r > R+D.
198 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Constraints

Xl = Y1 = Zl = Y2 = Z2 = Z3 = 0

Variable definitions
VR is a repulsive term, VA is an attractive term, and bij is the multi-body
term that depends on the positioning of atoms i and j, as weIl as the atoms
neighboring atom i. The !c(rij) term acts as a switching function so that
the potential falls smoothly to zero at short distances. (jijk is the bond angle
between bonds ij and ik. rij and rik refers to the Euclidean distance between
atoms i and j, and i and k, respectively (Le., rij = [(Xi - Xj)2 + (Yi _ Yj)2 +
(Zi - zj)21~ and rik = [(Xi - Xk)2 + (Yi - Yk)2 + (Zi - zk)21~)

Data
Data are provided for two models of silicon.
Si(B):

A = 3264.7 eV B = 95.373 eV >'1 = 3.23941- 1 >'2 = 1.32581- 1

'Y = 0.33675 n = 22.956 d = 2.0417 h = 0.0000
c = 4.8381 R = 3.01 D = 0.21

Si(C):

A = 1830.8 eV B = 471.18 eV >'1 = 2.47991- 1 >'2= 1. 73221 -1

'Y = 1.0999 X 10- 6 n = 0.78734 d = 16.218 h = -0.59826
C = 1.0039 X 10 5 R = 2.851 D = 0.151

Problem Statistics
No. of continuous variables 3N-6
No. of linear equalities 6
No. of known solutions Grows exponentially with N

Best Known Solutions

The following table provides the best known solutions for small clusters of
silicon modeled by the Tersoff potential. The second column provides energy
values for the Si(B) clusters and the third column provides energy values for
the Si(C) clusters. Results are compiled from Ali and Smith (1993) and Ali
et al. (1997).
8.7. CLUSTERS OF ATOMS AND MOLECULES 199

N Si(B) Energy (eV) Si(C) Energy (eV)

3 -7.87 -5.33
4 -15.71 -8.97
5 -20.80 -12.46
6 -26.82 -15.88

8.7.6 Brenner Potential

This potential was developed by Brenner (1990) for carbon clusters and is very
similar in form to the Tersoff equation. Three-body terms are included by
considering the neighboring atoms of i when considering pairwise interactions
with atom j. Originally the potential energy function was applied to the
calculation of properties for diamond and graphite. It has also been applied
to the prediction of small carbon cluster geometries.

Formulation

Objective function

min
rl,,rN
L fc(rij) [VR(r;j) + bijVA(rij)]
i<j

where

bij = (1 +Zij)-n, Zij = L fC(rik)g(()ijk)em(rij-rikl,

k-:pi,j

c2 2 ]
g(()) = a [ 1 + rf2 - rf2 + (h + cosO)2 '

I, < R-D,
r
fc(r) = { !- !sin[11"(;;RlJ, R - D < r < R + D,
0, r > R+D.

Constraints

Xl = YI = Zl = Y2 = Z2 = Z3 = 0

Variable definitions
200 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

VR is a repulsive term, VA is an attractive term, and bij is the multi-body

term that depends on the positioning of atoms i and j, as well as the atoms
neighboring atom i. The fc(rij) term acts as a switching function so that
the potential falls smoothly to zero at short distances. ()ijk is the bond angle
between bonds ij and ik. rij and rik refers to the Euclidean distance between
atoms i and j, and i and k, respectively (Le., rij = [(Xi - Xj)2 + (Yi _ Yj)2 +
(Zi - zj)2l~ and rik = [(Xi - Xk)2 + (Yi - Yk)2 + (Zi - Zk)2l~).

Data
D e = 6.325 eV r e = 1.315A = 1.5A- 1 S = 1.29
n 0.8047 a = 0.0113 C = 19.0 d = 2.5
h = 1.0 m = 2.25A- 1 R = 1.85A D = 0.15A
Problem Statistics
No. of continuous variables 3N-6
No. of linear equalities 6
No. of known solutions Grows exponentially with N

Best Known Solutions

The following table provides the best known solution for small carbon cluster
problems modeled by the Brenner potential. From left to right the columns
indicate: number of atoms in the cluster, energy, and the geometry of the clus-
ter. The geometries include, linear, mono- and poly-cyclic ring and fullerene
conformations. The data is compiled from results presented in Hobday and
Smith (1997).

N Energy (eV) Geometry

3 -12.40 linear
4 -18.58 linear
5 -26.33 m-cyclic ring
6 -33.96 m-cyclic ring
7 -41.10 m-cyclic ring
8 -47.91 m-cyclic ring
9 -54.50 m-cyclic ring
10 -60.95 m-cyclic ring
11 -67.32 m-cyclic ring
12 -73.61 m-cyclic ring
13 -79.90 m-cyclic ring
14 -86.11 m-cyclic ring
15 -92.34 m-cyclic ring
16 -98.61 m-cyclic ring
17 -140.80 m-cyclic ring
18 -111.38 p-cyclic ring
8.7. CLUSTERS OF ATOMS AND MOLECULES 201

N Energy (eV) Geometry

19 -117.40 p-cyclic ring
20 -128.39 fullerene
21 -131.70 p-cyclic ring
22 -138.36 p-cyclic ring
23 -147.39 fullerene
24 -157.16 fullerene
25 -163.13 fullerene
26 -171.98 fullerene
27 -178.01 fullerene
28 -186.88 fullerene
29 -192.67 fullerene
30 -200.44 fullerene
31 -207.00 fullerene
32 -216.66 fullerene
33 -221.78 fullerene
34 -230.30 fullerene
35 -236.38 fullerene
36 -245.45 fullerene
37 -251.23 fullerene
38 -259.97 fullerene
39 -265.59 fullerene
40 -274.64 fullerene
41 -280.38 fullerene
42 -289.34 fullerene
43 -295.12 fullerene
44 -304.09 fullerene
45 -309.02 fullerene
46 -318.54 fullerene
47 -324.06 fullerene
48 -333.38 fullerene
49 -339.03 fullerene
50 -348.37 fullerene
51 -353.88 fullerene
52 -362.83 fullerene
53 -367.92 fullerene
54 -377.31 fullerene
55 -383.08 fullerene
56 -392.12 fullerene
57 -397.73 fullerene
58 -407.25 fullerene
59 -412.74 fullerene
202 CHAPTER 8. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

N Energy (eV) Geometry

60 -422.55 fullerene

8.7.7 Bolding-Andersen Potential

The Bolding and Andersen (1990) potential intrinsically contains many multi-
body effects in its formulation. Although it is similar to the Tersoff and Brenner
potentials in that it contains only pairwise terms, complicated expressions are
used to model up to four-body interaction energies. These express ions were
developed by fitting parameters to a variety of data on silicon, including the
structure and energy of small clusters and surface properties.

Formulation

Objective function

min
rl,,TN
<Jl(r1, ... ,rN) L !c(rij)
fc(rij) [VR(rij) + IljV (rij) + IijVa(rij)]
7r

i<j
where

VR(r) = !(Vbond(r) + vanti(r)), V7r (r) = -a1tanh[a2(r - a3)]- 1,

Va(r) = ~(Vbond(r) - vanti(r)) - V (r), 7r

2

6
Vbond(r) = -D e (1 + L bipi)e-blP,
i=l
i=1

vanti(r) = (b s + bgp + blOp2)e-b7P, p = r - re , rtJ = 1..1..

h [2'

It == 1 + Ztj6 (L a4 Sik
S ik S jk + L
a4 S ik S il Sjk
a5 Sik
a5 S jk SjlSkl) ,
ki-i,j k<li-i,j

h == 1 + Ztj7 (L [SikP(Bjik) + SjkP(Bijk )]) ,

ki-i,j

6
P(B) = L: dncosn(B), Zij = L: (Sik + Sjk),
n=O ki-i,j

r < re ,
{ 1, )
Sik = S(rik) -- S( r ) = fc(r) VR(Te)+V,,(T
VR(r)+V,,(r
e) ' r 2 re,
8.7. CLUSTERS OF ATOMS AND MOLECULES 203

Iij
I~.
Q
S(zj)S(Z~)e-I3
= S(Z{)S(Zj)e-
, J
I3 ,
'
133 =
1 E F3 + E F4 + E F5 ,
kf.i,j k,lf.i,j k,l,mf,i,j
k,l,mf.i,j

I, Z <2,
S(Z) = { Cl + C2Z + C3Z2 + C4Z3, 2 ~ Z ~ 2.5, zl = E Sik,
kf.i,j
0, Z> 2.5,

F 3 = aSSikSjk +1 :9 Zij [Sik(l - Sfk) + Sjk(l - sfA:)]

F4 = alO(SikSi/SjkSjl)
+ l~~$. [SikSjl(l - Sil)(l - Sjk) + SilSjk(l - Sik)(l - Sjl)]
'~
+ al2l1i~kl (S;kS~1 + S~S~k + ~ + S~kSn
T ij TikT jl TilT jk TikT il TjkT jl

F5 = al3(SikSjkSilSjmSklSkm + SikSjkSimSjlSklSkm)
aI3(SikSjkSi/SjmSkISkm
+ aI3(Si/SjISikSjmSkISlm + SimSjmSi/SjkSkmSlm)
al3(SilSjlSikSjmSklSlm SimSjmSilSjkSkmSlm)
+ al3 (Si/SjISimSjkSklSlmSimSjmSikSjISkmSlm)
al3(SilSjISimSjkSkISlmSimSjmSikSjISkmSlm)

Vijkl = rij . rjl x rik

I, r<R-D,
fc(r) = { ! - !sin[1r(;JjR)]
!sin[1r(;D )] , R - D < r < R + D,
R ,
0, r > R+D.

Constraints

Xl = Yl
YI = Zl = Y2 = Z2 = Z3 = 0

Variable definitions
VR is a repulsive term. The attractive term is broken up into two contributions,
Vu and V1r , which represent favorable interactions due to bonding of a and 7r
orbitals, respectively. These terms are modulated by the interference functions
Ir; and Iij. There are two switching function, fc(r) and S(Z), which are used
Iij
to smooth the potential to zero. In addition, the bonding function S(R) is used
to measure the extent to which a pair of atoms are bonded. (}ijk(Jijk is the bond
angle between bonds ij and ik. rij refers to the Euclidean distance between
atoms i and j (i.e., rij = [(Xi - Xj)2 + (Yi - Yj)2 + (Zi - Zj)2]!).
204 CHAPTER B. TWICE CONTINUOUSLY DIFFERENTIABLE NLPS

Data
a1 15.211 kcal
mol a2 = 4.090011 a3 2.365011
a4 0.0090 a5 = 0.0100 a6 1.0000
a7 0.3333 as = 0.4521 ag -0.8763
alO = 1.8387 an = 75.3470 a12 = 0.3425
a13 = 5.0000 b 1 = 2.385011- 1 b2 = 0.729611- 2
b3 -2.296511- 3 b4 = -1.367711- 4 b5 = 2.048511- 5
b6 = -0.474511- 6 b7 = 3.483511- 1 b = 165.16 kcal
s mol
bg = 96.70 kcal 11- 1
mol
b
10
= 495.73 kcal
mol 11-2 do = 0.005387
d1 = 0.007894 d2 = -0.081615 d3 = -0.083367
d4 = 0.380043 d5 = 0.374569 d6 = 0.054232
Cl = -175.0 C2 = 240.0 C3 = -108.0
C4 = 16.0 Te = 2.18411 D = 75.47 kcal
e mol

Problem Statistics
No. of continuous variables 3N-6
No. of linear equalities 6
No. of known solutions Grows exponentially with N

Best Known Solutions

The following table provides the best known solutions for small clusters of
silicon modeled by the Bolding-Andersen potential. Results are compiled from
Bolding and Andersen (1990) and Niesse and Mayne (1996).

N Energy (kcal/mol)
3 -186.1
4 -300.0
5 -382.7
6 -466.0
7 -568.2
8 -676.9
9 -789.2
10 -901.3
 Chapter 9

Bilevel Programming
Problems

9.1 Introduction
A problem where an optimization problem is constrained by another one is
classified as aBiLevel Programming Problem, BLPP, and is of the general
form:

min F(z, y)
:1:, Y
S.t.
G(z, y) < 0
H(z, y) = 0
minJ(z, y)
y
S.t.
g(z,y) < 0
h(z,y) 0
z E X ~ R n1 , y E Y ~ R n2

where J,F : Rnl x Rn2 --+ R, 9 = [gl, .. ,gJ] : R n1 x R n2 --+ Rj, G =

[GI, .. , GJ] : Rn1 x Rn2 --+ RJ', h = [h 1, .. , hI] : R n1 x R n2 --+ Ri , H =
[H1, .. , Hl'] : Rn1 xRn2 --+ Ri'. F, G and H are the outer (planner's or leader's)
problem objective function, inequality and equality constraints, and J, g, and
h are the inner (behavorial or follower's) problem objective, inequality and
equality constraints, respectively. The decision variables of the outer problem
are z and y and of the inner problem are y.
The literat ure on the application of bilevel programming problems, BLPP,
is quite extensive and diverse, such as economics, (Cassidy et al., 1971; Candler
and Norton, 1977; Luh et al. , 1982; Kolstad, 1985; Hobbs and Nelson, 1992;
Suh and Kim, 1992; Onal et al. , 1995; Mathur and Puri, 1995; Aiyoshi and
205

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
206 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Shimizu, 1981; Bard et al., 1998), civil engineering, (Fisk, 1984; LeBlanc and
Boyce, 1986; Ben-Ayed et al. , 1988; Suh and Kim, 1992; Yang and Yagar,
1994; Yang et al. , 1994; Yang and Yagar, 1995; Migdalas, 1995); chemical
engineering (Grossmann and Floudas, 1987; Clark and Westerberg, 1990a,b,
1983; Gm~ and Ciric, 1997) or warfare (Bracken and McGill, 1973). Good
reviews of bibliography on BLPP can be found in Vicente and Calamai (1994)
and Koistad (1985). Further algorithms and applications are presented in Bard
(1998).

9.1.1 Terminology and Properties

The general terminology and basic properties of the BLPP are as follows:

The relaxed BLPP feasible set (or constraint region):

o = {(x, y) : G(x, y) ::; 0, H(x, y) = 0, g(x, y) ::; 0, h(x, y) = O}
For x E X, the inner problem feasible set (follower's solution set, or
follower's feasible region):
O(x) = {y: y E Y, h(x, y) = 0, g(x, y) ::; O}
For x E X, inner rational reaction set (follower's rational reaction set):
RR(x) = {y E argminf(x, y) : y E O(x)}

The BLPP feasible set (the induced, or inducible region):

IR = {(x, y) : (x, y) E 0, Y E RR(x)}.

For the BLPP to be well-posed, it is generally assumed that 0 is nonempty

and compact, and RR(x) =I- <p.
Using the above terminology, the BLPP can be written as

min F(x, y)
s.t. (x, y) E IR.

Several studies have proved that the solution of even the linear BLPP is
an NP-hard problem (Jeroslow, 1985; Ben-Ayed and Blair, 1990; Bard, 1991).
It has been further shown that the BLPP is strongly NP-hard (Hansen et al.,
1992).

9.1.2 Solution Techniques

In general, there has been two basic techniques for the solution of the linear
BLPP: enumeration and reformulation techniques.
Enumeration techniques are based on the property that the inducible re-
gion of the linear BLPP is composed of connected faces of 0 and that a solution
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 207

lies at a vertex (Bard, 1984). The following approaches are of this type: Ver-
tex enumeration (Candler and Townsley, 1982); K-th best algorithm; (Bialas,
1984); and branch and bound (Bard and Moore, 1990).
The reformulation techniques are based on the transformation of the bilevel
problem into a single level one by replacing the inner problem with its Karush-
Kuhn-Tucker optimality conditions. For the solution of the transformed prob-
lem, the following approaches have been proposed: mixed integer programming
(Fortuny-Amat and McCarl, 1981); branch and bound (Hansen et al. , 1992;
Bard and Falk, 1982); penalty methods (Bard and Falk, 1982; Anandalingam
and White, 1990; White and Anandalingam, 1993; nal, 1993); parametric
complementary pivoting (Judice and Faustino, 1992); and global optimiza-
tion (Floudas and Zlobec, 1998; Visweswaran et al. , 1996; Tuy et al. , 1993;
AI-Khayyal et al. , 1992).
For the solution of the bilevel linear problems with quadratic objective
functions, transformation based branch and bound (Bard, 1988; Edmunds and
Bard, 1991), descent (Vicente et al., 1994), and global optimization (Visweswaran
et al., 1996) methods have been proposed. More information on solution tech-
niques can also be found in recent books by Bard (1998) and Shimizu et al.
(1997).
Procedures for generating random BLPP test problems are presented in
Calamai and Vicente (1993, 1994) and Moshirvaziri et al. (1996).

9.2 Bilevel Linear Programming Problems

When all the functions are linear, the resulting problem is abilevel linear
problem of the form:

mln
z, y
cr x + d[ Y
s.t.
Alx+Bly < bl
Glx + Dly = el

mincf x
y
+ a..J y
s.t.
A2x + B2 y :::; ~
G2x + D2Y = e2
x E X ~ R nl , y E Y ~ R n2

where we have Cl, C2 E Rnl , d l , ~ E R n2 , bl E Rj, b2 E RJ', ej E R i , e2 E

Ri',A I E Rjxnl,B I E R j xn2, GI E Rixnl,DI E Rixn2,A2 E RJ'xnl,B2 E
RJ'xxn2, and D2 E R i'xn2.
Detailed reviews of the properties of the linear BLPP and existing solution
approaches can be found in Shimizu et al. (1997), Vicente and Calamai (1994),
Ben-Ayed (1993) and Wen and Hsu (1991).
208 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

The examples presented in this chapter have all been solved by replacing
the inner problem with its equivalent Karush-Kuhn-Thcker conditions. There-
fore, abrief description follows.

9.2.1 Karush-Kuhn-Tucker Approach

Consider the bilevel linear programming problem. At fixed x, the necessary
and sufficient conditions for (y*, >'*, /L*) to be an optimal solution to the inner
level problem is that the following set of optimality condition equations are
satisfied:

A,ex + B,ey* + 8 b,e

C,eX + D,ey = e,e
C,e + >.T B,e + /LT D,e 0
>. T* 8* = 0
>. *, 8* > O.
where >. *, /L* are, respectively, the KKT multiplier vectors of the inequality and
equality constraints. It follows that the necessary and sufficient conditions for
(x* , y* , >. *, /L*) to be the optimal solution of the BLPP, (y*, >. * , /L*) must satisfy
the above conditions at fixed x = x* (Bard and Falk, 1982). From this line of
reasoning, the bilevel programming problem can be transformed into a single
level problem of the form:

min
:z: y
eJx + dJy
S.t.
A 1 x+B 1 y:::;b1
C1 x + D 1 y = el
C,e + >.TB,e +/LTD,e = 0
A,ex+B,ey+8=b,e
C,ex+D,ey=e,e
>.T 8 =0,
>., 8:2: 0,
x EX, Y E Y,

Notice that even for the linear BLPP, the resulting problem is nonlinear
and nonconvex due to the complementarity conditions. The complementarity
conditions are difficult to handle by most of the commercial nonlinear codes.

9.2.2 Test Problem 1

This test problem is taken from Clark and Westerberg (1990a).
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 209

Formulation

Outer objective function

min -x - 3Yl
z,y
+ 2Y2

Inner objective function

min -Yl
Y

Inner constraints

-Yl < 0
Yl < 4
-2x + Yl + 4Y2 < 16
8x + 3Yl - 2Y2 < 48
-2x + Yl - 3Y2 < -12

Variable bounds

Problem Statistics
No. of continuous variables 3
No. of inner variables 2
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 5
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -13

Inner objective function: -4

Continuous variables
x =5 Y = (4,2)T
210 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

9.2.3 Test Problem 2

This problem is taken from Liu and Hart (1994).

Formulation

Outer objective function

min -x - 3y
x,y

Inner objective function

miny
y

Inner Constraints

-y ~O
-x+y < 3
x+2y < 12
4x-y :s 12

Variable bounds

O~x
O~y

Variable definition

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 4
No. of linear outer equalities
No. of linear outer inequalities
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 211

Best Known Solution

Outer objective function: -16
Inner objective function: 4
Continuous variables
x=4 y=4
Behavior
The optimum is at a boundary feasible extreme point.

9.2.4 Test Problem 3

This problem is taken from Candler and Townsley (1982)
Outer objective function

min 4Yl - 40Y2 - 4Y3 - 8Xl - 4X2

x,y

Inner objective function

min Yl + Y2 + 2Y3 + Xl + 2X2

Y

Inner constraints

-y ~ 0
H1 T y+Hp7x b
where

BI =
[ -1
[
BI = ~1
-1 1
-1
1
2
11
-0.5
11 00 01
0 1 0
n
n
2 -1 -0.5 0 0 1

H2~ [~ ~l
and bT = (1,1,1).
Variable bounds

O~x
O~y
212 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Problem Statistics
No. of continuous variables 8
No. of inner variables 6
No. of outer variables 2
No. of linear inner equalities 3
No. of linear inner inequalities 6
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 29.2

Inner objective function: 3.2

Continuous variables

x = (0,0.9f
Y = (0,0.6,004,0,0, O)T

9.2.5 Test Problem 4

This test problem is taken from Clark and Westerberg (1988), and has been
used for testing purposes in Chen and Florian (1995) and Yezza (1996).

Formulation

Outer objective function

min x - 4y
x,y

Inner objective function

min y
y

Inner constraints

-2x+y < 0
2x + 5y - 108 < 0
2x - 3y+4 < 0
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 213

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 3
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -37

Inner objective function: 14

Continuous variables
x = 19 Y = 14

Behavior
The global maximum of the problem occurs at (1,2).

9.2.6 Test Problem 5

This problem is taken from Bard (1991), and has also been used in Shimizu
et al. (1997).
Outer objective function

min -x + lOYl - Y2
x,y

Inner objective function

min -Yl - Y2
Y

Inner constraints

X -Yl < 1
X+Y2 < 1
Yl +Y2 < 1
-Yl < 0
-Y2 < 0
214 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Variable bounds

O:::;x
O:::;y

Problem Statistics
No. of continuous variables 3
No. of inner variables 2
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 5
No. of outer equalities 0
No. of outer inequalities

Best Known Solution

Outer objective function: -1

Inner objective function: 0

Continuous variables
x = 1 Y = (O,Of

9.2.7 Test Problem 6

This test problem is taken from Anandalingam and White (1990), and is
slightly different that its original version given by Bialas (1982).

Formulation

Outer objective function

min -x - 3y
x,y

Inner objective function

min -x + 3y
y

Inner constraints
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 215

-x-2y < -10

x+2y < 6
2x-y < 21
x+2y < 38
-x+2y < 18
-y < 0

Variable bounds

O:S;x
O:S;y

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of out er variables 1
No. of linear inner equalities
No. of linear inner inequalities 6
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -49

Inner objective function: 17

Continuous variables
x = 16 y = 11

9.2.8 Test Problem 7

This test problem is taken from Bard and Falk (1982), and is also used in
Shimizu et al. (1997) and Bard and Moore (1990).

Formulation

Outer objective function

216 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Inner objective function

min Xl + 2X2 + Yl + Y2 + 2Y3

Y

Inner Constraints

-Yl + Y2 + Y3 < 1
2Xl - Yl + 2Y2 - 0.5Y3 < 1
2X2 + 2Yl - Y2 - 0.5Y3 < 1
-Yl < 0
-Y2 < 0
-Y3 < 0

Variable bounds

Problem Statistics
No. of continuous variables 5
No. of inner variables 3
No. of outer variables 2
No. of linear inner equalities
No. of linear inner inequalities 6
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -26.0

Inner objective function: 3.2

Continuous variables
a: = (0.0,0.9)T Y = (0.0,0.6, 0.4)T
9.2.9 Test Problem 8
This problem is taken from Bard and Falk (1982).
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 217

Formulation

Outer objective function

Inner objective function

min Xl + X2 - 4Yl + Y2
Y

Inner Constraints

-2Xl + Yl - Y2 ~ -2.5

Xl - 3X2 + Y2 < 2
Xl + X2 < 2

Variable bounds

Variable definition

Problem Statistics
No. of continuous variables 4
No. of inner continuous variables 2
No. of outer continuous variables 2
No. of outer equalities
No. of outer inequalities
No. of inner equalities
No. of inner inequalities 3

Optimum solution
Outer objective function: -1.75
Inner objective function: 0.0
Continuous variables
z = (l,O)T Y = (0.5,1)T
Two other local solutions exist, each occuring at different vertices.
218 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

9.2.10 Test Problem 9

This test problem is taken from Visweswaran et al. (1996).
Outer objective function

min x+y
x,y

Inner objective function

min -5x - y
y

Inner Constraints

-x - 0.5y < -2
-0.25x +y < 2
x + 0.5y < 8
x-2y < 2
-y < 0

Variable bounds

O~x
O~y

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 5
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 3.111
Inner ohjective function: 6.667
Continuous variables
x = 0.889 y = 2.222
9.2. BILEVEL LINEAR PROGRAMMING PROBLEMS 219

9.2.11 Test Problem 10

This test problem is taken from Thy et al. (1993).

Formulation

Outer objective function

min -2Xl
z,y
+ X2 + 0.5Yl

Inner objective function

min -4Yl + Y2 + Xl + X2
Y

Inner constraints

-2Xl +Yl-Y2 < -2.5

Xl - 3X2 +Y2 < 2
-Yl < 0
-Y2 < 0

Variable bounds

Problem Statistics
No. of continuous variables 4
No. of inner variables 2
No. of outer variables 2
No. of linear inner equalities
No. of linear inner inequalities 4
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -3.25

Inner objective function: -7.0

220 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Continuous variables
z = (2,0)T Y = (1.5,0)T
Behavior
In the paper in which this problem is originally posed, a non-optimal
solution is located at z = (1,0), Y = (0.5,1) with an outer objective
function value of -1.75 (Bard and Falk, 1982).

9.3 Bilevel Quadratic Programming Problems

9.3.1 Introduction
The linear-quadratic and quadratic-quadratic bilevel optimization problems
are of the form:

min F(z, y)
z, "
S.t.
A 1 z+B 1 y :s; b1
C1 z + D 1 y = el
mJnc1z + d[y + zTQ:y + yT Q:y
S.t.
+ B.1lY :s;
A.1l z b.1l
C!z+D!y = e!
z E X ~ R n1 , y E Y ~ R n2
where for the linear-quadratic case F(z, y) = cl z+dJ y and for the quadratic-
quadratic case F(z,y) = clz + dJy + zTQfy + yTQ4y and the inner
objective is assumed to be a convex quadratic function. Replacing the inner
problem with the necessary and sufficient KKT optimality conditions results
in the single level problem:
min F(z, y)
z, "
S.t.
A 1 z+B 1 y:S;bl
C1 z + D 1 y = el
2y T Q:+ zT + C.1l Q: + >,T B! + J.LT D! =
A.1l z + B .1lY+s = b.1l
C.1l z + D !y=e!
>,T s = 0,
>., s ~ 0,
z EX, Y E Y.
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 221

9.3.2 Test Problem 1

This problem is the simplest version of the quadratic-quadratic BLPP, where
Fis strictly convex, Gis convex, f is quadratic and 9 is affine (Shimizu et al.,
1997).

Formulation

Outer objective function

min (x - 5)2
x,y
+ (2y + 1)2

Inner objective function

min (y - 1)2 - 1.5xy

y

Inner constraints

-3x+y < -3
x - 0.5y< 4
x+y < 7

Variable bounds

0::; x
O::;y

Problem Statistics
No. of continuous variables 2
No. of continuous outer variables 1
No. of continuous inner variables 1
No. of linear outer equalities
No. of convex out er inequalities
No. of linear inner equalities 3
No. of linear inner inequalities
No. of convex inner inequalities
222 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Best Known Solution

Outer objective function: 17

Inner objective function: 1

Continuous variables
x = 1,y = 0

9.3.3 Test Problem 2

This test problem is taken from Visweswaran et al. (1996), and is originally
from Shimizu and Aiyoshi (1981).

Formulation

Outer objective function

min x 2
X,y
+ (y - 10)2

Inner objective function

min (x + 2y - 30)2
Y

Outer constraints

-x+y < 0

Inner constraints

x+y < 20
-y < 0
Y < 20

Variable bounds

o~ x ~ 15
0~y~20
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 223

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 3
No. of linear outer equalities
No. of linear outer inequalities 1

Best Known Solution

Outer objective function: 100.0

Inner objective function: 0.0

Continuous variables
x = 10.0 Y = 10.0

9.3.4 Test Problem 3

This test problem is taken from Visweswaran et al. (1996).

Formulation

Outer objective function

min 2Xl
x,y
+ 2X2 - 3Yl - 3Y2 - 60

Inner objective function

min (Yl - Xl + 20)2 + (Y2 - X2 + 20)2

Y

Outer constraints

Xl
Xl + X2
X2 + Yl
Yl -- 2Y2 -- 40
2Y2 < 0
224 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Inner constraints

-Xl + 2YI < -10

-X2 + 2Y2 < -10
-YI < 10
y1 < 20
-Y2 < 10
Y2 :::; 20

Variable bounds

0:::; Xl:::; 50
0:::; X2 :::; 50
-10 :::; YI :::; 20
-1O:::;Y2:::;20

Problem Statistics
No. of continuous variables 4
No. of inner variables 2
No. of outer variables 2
No. of linear inner equalities
No. of linear inner inequalities 6
No. of nonlinear equalities
No. of nonconvex inequalities 1

Best Known Solution

Outer objective function: 0

Inner objective function: 200

Continuous variables
z = (O,Of Y = (-10, -lOf

Behavior
The problem has a local solution at z = (25, 30)T, Y = (5, lO)T for
which the outer objective function is 5. The paper in which this problem
is presented originally locates this local optimum solution (Aiyoshi and
Shimizu, 1984). Shimizu et al. (1997) also report its local minimum as
the solution of the problem.
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 225

9.3.5 Test Problem 4

This test problem is taken from Yezza (1996).

Formulation

Outer objective function

Inner objective function

min 0.5y? + Y2

Inner constraints

Yl + Y2 X

-Yl < 0
-Y2 < 0

Variable bounds

O::;y

Problem Statistics
No. of continuous variables 3
No. of inner variables 2
No. of outer variables 1
No. of linear inner equalities 1
No. of linear inner inequalities 2
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 0.5

Inner objective function: 2.5

Continuous variables
x =3 y = (1, 2)T
226 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

9.3.6 Test Problem 5

This problem is taken from Clark and Westerberg (1990a).

Formulation

Outer objective function

min (x - 3)2 + (y - 2)2

x,y

Inner objective function

min (y - 5)2
Y

Inner constraints

-2x+y < 1
x - 2y < 2
x + 2y < 14

Variable bounds

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 3
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 5
Inner objective function: 4
Continuous variables
x=1 y=3
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 227

9.3.7 Test Problem 6

This test problem is taken from Falk and Liu (1995).

Formulation

Outer objective function

.22 2 22
mm +X2 - 2 X2 +YI +Y2
a;,y Xl - Xl

Inner objective function

Inner constraints

-YI < -0.5

-Y2 < -0.5
YI < 1.5
Y2 < 1.5

Variable bounds

O:S;y

Problem Statistics
No. of continuous variables 4
No. of inner variables 2
No. of outer variables 2
No. of linear inner equalities
No. of linear inner inequalities 4
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: -1

Inner objective function: 0

228 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Continuous variables
a: = (0.5,0.5)T Y = (0.5,0.5)T

Behavior
The same problem with the outer problem objective:

mm.23
a:,y YI - YI + Y22 - 3
Y2 + Xl2+2
x2

has its optimum solution at a: = (0.5V3, 0.5V3, 0.5V3, 0.5V3)T

9.3.8 Test Problem 7

This test problem is taken from Visweswaran et al. (1996).

Formulation

Outer objective function

min (x - 5)2
x,y
+ (2y + 1)2

Inner objective function

min (y - 1)2 - 1.5xy

y

Inner constraints

-3x+y < -3
x - 0.5y< 4
x+y < 7
-y < 0

Variable bounds

O~x
O~y
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 229

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 4
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 17

Inner objective function: 1

Continuous variables
x=1 y=O

Behavior
The problem also has a local solution at (5, 2).

9.3.9 Test Problem 8

This test problem is taken from Yezza (1996).

Formulation

Outer objective function

min -(4x - 3)y + (2x + 1)

x,y

Inner objective function

min -(1 - 4x)y - (2x

y
+ 2)

Inner constraints

-y < 0
y < 1

Variable bounds
230 CHAPTER 9. BILEVEL PROGRAMMING PROBLEMS

Problem Statistics
No. of continuous variables 2
No. of inner variables 1
No. of outer vari:J.bles 1
No. of linear inner equalities
No. of linear inner inequalities 2
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 1.5
Inner objective function: -2.5

Continuous variables
x = 0.25 Y =0
9.3.10 Test Problem 9
This test problem is taken from Bard (1991).

Formulation

Outer objective function

min X+Y2
x,y

Inner objective function

min 2Yl + XY2

Y

Inner constraints

x - Yl - Y2 ::; 0
-Yl < 0
-Y2 < 0
9.3. BILEVEL QUADRATIC PROGRAMMING PROBLEMS 231

Variable bounds

0-:::; x
O:::;y

Problem Statistics
No. of continuous variables 1
No. of inner variables 2
No. of outer variables 1
No. of linear inner equalities
No. of linear inner inequalities 3
No. of linear outer equalities
No. of linear outer inequalities

Best Known Solution

Outer objective function: 2

Inner objective function: 12

Continuous variables
x=2 y=(6,0)T.
 Chapter 10

Complementarity Problems

10.1 Introduction
A complementarity problem aims at determining x E !Rn, such that

z 2 0, F(z) 2 0, ..1

where F : !Rn -+ !Rn and the symbol ..1 denotes orthogonality, that is,

Applications of the complementarity problem are quite extensive and diverse,

covering areas of mathematical programming, game theory, economics and en-
gineering. The linear complementarity problem is applied to the development
of theory and algorithms in linear programming, quadratic programming and
bimatrix game problems, where the quadratic programming problems have
been extensively used in science and engineering for the solution of the con-
tact, porous flow, obstacle, journal bearing, elastic-plastic torsion, circuit sim-
ulation, and other free boundary problems (Luo et al., 1997; Isac, 1992; Cottle
et al. , 1992; Murty, 1988). The nonlinear complementarity problem applica-
tions include the solution of Nash equilibrium, trafik assignment or network
equilibrium, spatial price equilibrium and the general or Walrasian equilibrium
problems (Harker and Pang, 1990).
A large collection of complementarity test problems, referred as the Mixed
Complementarity Problems LIBrary, MCPLIB, is publicly available via anony-
mous ftp from

ftp.cs.wisc.edu:-/pub/mcplib/

and are presented in GAMS modeling language (Brooke et al. , 1988). The
origin and structure of the problems in MCPLIB, and additional information
is given in Dirkse and Ferris (1995). Furthermore, the standard GAMS dis-
tribution (Brooke et al., 1988) includes a library of test problems that is also
available from
233

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
234 CHAPTER 10. COMPLEMENTARITY PROBLEMS

http://www.gams.com/modlib/modlib.htm

Both collections contain problems with known types of computational dif-

ficulties or real life applications. A testing environment consisting of many
test problems from both libraries, along with GAMS and MATLAB interfaces,
and a comparison of eight large mixed complementarity problem solution al-
gorithm implementations is included in Billups et al. (1997). Selected test
problems from both libraries are presented here.

10.2 Linear Complementarity Problems

Let M be a square matrix in !Rnxn and q be a column vector in !Rn. Then,
the linear complementarity problem, LCP, is to find a vector x E !Rn such that

x ~ 0, (q + Mx) ~ 0

and is denoted by the symbol (q, M). For each q E !Rn, the LCP (q, M) has a
unique solution if and only if all the principal minors of M E !Rnxn are positive
(Le., M is a P-matrix).
For the solution of small to medium size linear complementarity problems,
pivoting methods have been proposed, while for large scale LCPs, iterative
methods have been developed. The pivoting algorithms include principal piv-
oting method (Dantzig and Cottle, 1967; Cottle and Dantzig, 1968; Graves,
1967) and Lemke and Howson's complementary pivoting algorithm (Lemke
and Howson, 1964) that is originally developed for obtaining the Nash equi-
librium points for bimatrix games. An extension of this algorithm addresses
the solution of general linear complementarity problems (Lemke, 1968). Iter-
ative methods have been applied to the analysis of elastic bodies in contact,
free boundary problem for journal bearings and network equilibrium problems
(Cottle et al. , 1992). Most of the linear complementarity problem solution
algorithms make use of the properties of the matrix M.

10.2.1 Test Problem 1

This test problem is taken from MCPLIB (Dirkse and Ferris, 1995).

Formulation

M Matrix
10.2. LINEAR COMPLEMENTARITY PROBLEMS 235

1 2 2 2

1 2 2

M= 2
2
1 2
1

q vector
qi = -1, i = 1, .. , 16.

Problem Statistics
No. of continuous variables 16
No. of linear functions 16
No. of known solutions 1

Best Known Solution

Continuous variables
:z: = (O,O,O, .. ,O,1)T .
Behavior
It was proven by Murty (1988) that the number of pivots in Lemke's
algorithm is exponential in the number of problem variables.

10.2.2 Test Problem 2

This test problem is taken from Cottle et al. (1992).

Formulation

M Matrix

M=[~ ~]
q vector
qi = -l,i = 1,2.
Problem Statistics
No. of continuous variables 2
No. of linear functions 2
No. of known solutions 3
236 CHAPTER 10. COMPLEMENTARITY PROBLEMS

Best Known Solution

Continuous variables
Solution 1: x = (1,0)T
Solution 2: x = (0,1)T
Solution 3: x = (~, ~)T

Behavior
LCPs with positive semi-definite M may have multiple solutions. This
problem has three solutions that are all given above.

10.2.3 Test Problem 3

This test problem is taken from Cottle et al. (1992).

Formulation

M Matrix

M = [ ~ -~ -~ 1
-1 1 0

q vector

q = (-3,6,-I)T

Problem Statistics
No. of continuous variables 3
No. of linear functions 3
No. of known solutions 1

Best Known Solution

Continuous variables
x = (0,1,3f

Behavior
Note that the matrix M is sufficient.

10.2.4 Bimatrix Games

The LCP formulation is applied to the solution of bimatrix games.
10.2. LINEAR COMPLEMENTARITY PROBLEMS 237

Bimatrix Formulation

-em + Ay ~ 0, x ~ 0, .1

where A and B are positive m x n matrices and em and en are vectors with
every component equal to one.

LCP Formulation

M Matrix

q vector

q= (-em,-enf
10.2.5 Test Problem 4
This test problem is taken from Cottle et al. (1992) and is abimatrix game
with:

A = B T = [10 20]
30 15

LCP Formulation

M Matrix

A = BT = [~10 20~ !~0 1520o 1

30 15 0 o
q vector

qT = (-I,-I,-I,-lf
238 CHAPTER 10. COMPLEMENTARITY PROBLEMS

Problem Statistics
No. of continuous variables 4
No. of linear functions 4
No. of known solutions 3

Best Known Solution

Continuous variables
Solution 1: :z: = (1~' 0, 1~' O)T
Solution 2: :z: = (0, A,0, A)T
Solution 3: :z: = (da, 15' 15
9~' f
Behavior
The third solution point can not be achieved by the application of the
Lemke-Howson algorithm (Cottle et al., 1992).

10.2.6 Equilibrium Transportation Model

Dantzig's equilibrium transportation model at fixed supply and demand (Dantzig,
1963) can be reformulated as a LCP.

Formulation
Zero profit conditions

Vi E I, j E J Wi - Pj + Cij ~ 0, :z: ~ 0 .1.

Supply limit at plant i

J
Vi E I - '"' x- - + a- >
L...J'J - 0, W~O .1.
j=l

Fixed demand at market j

I
Vj E J '"'x---b->O
L...J'J J - , p?,O 1..
i=l

Variable Definition

Wi - shadow price at supply node i

Pj - shadow price at demand node j

Xij - shipment quantities in cases from supply i to demand j

Parameter Definition
10.2. LINEAR COMPLEMENTARITY PROBLEMS 239

ai - capacity of plant i in cases at unit price

bj - demand at market j in cases at unit price

Cij - transportation cost in thousands of dollars per case

f - freight in dollars per case per thousand miles

10.2.7 Test Problem 5

This test problem of Dantzig's equilibrium transportation model is taken from
GAMSLIB.

Data

a = (325, 575f

b = (325,300, 275f

D = [2.5 1. 7 1.8]
7.5 1.8 1.4

f = 90

fD
C = 1000

Problem Statistics
No. of continuous variables 11
No. of linear functions 11

Best Known Solution

Continuous variables

p = (0.225,0.153, 0.126)T

w = (O.O,O.O)T

25.0 300.0 0.0 ]

[
x = 300.0 0.0 275.0
240 CHAPTER 10. COMPLEMENTARITY PROBLEMS

10.2.8 Traffic Equilibrium Problem

The simplified trafik equilibrium model is taken from Cottle et al. (1992) where
the transportation network is modeled as a digraph with nodes N and ares A.
The problem is to move D > 0 units of flow from origin node i to destination
node j via specific paths in the network.

Formulation
Feasible path demand eondition
n
2:Fk =D
k=l

Thaffie eongestion eost on eaeh are

Are-path relationship

Arc-path incidenee matrix

1 if are a is eontained in path Pk

= (dak) where dak = { 0 otherwise.

Path Costs

Parameter Definition

A - set of ares

N - set of nodes

a - index on ares

i, j - indices on nodes

ba, da - scalars

D - diag(da )

Variable Definition

Fk - traffic flow along the path Pk (k = 1, .. , n)

10.2. LINEAR COMPLEMENTARITY PROBLEMS 241

F - path flows vector

Ca - cost function

ia - flow on arc a
G - minimum cost between O-D pair (i,j)

Explicit Formulation

F? 0,
L:~=l Fk D
C(F) - Ge > 0
Fk(Ck(F) -G) 0k = 1, .. ,n.
which is a mixed complementarity problem with a single equality constraint
and a free variable G. This formulation can be turned into an LCP by trans-
forming the equality constraint into an inequality, restricting G to be nonneg-
ative, and imposing a complementarity constraint between G and the demand
constraint.

10.2.9 Test Problem 6

'!'his test problem is taken from Cottle et al. (1992). The trafik network is
given in figure 10.1. There are 3 paths that join the O-D pair and the demand
D between the O-D pair is 6, and the cost on are e32 is parametrie.
M Matrix

o
11 10
10 -1
-1 1
10 21 -1
1 1 1

q vector

Problem Statistics
No. of continuous variables 4
No. of linear functions 4
No. of known solutions parametrie
242 CHAPTER 10. COMPLEMENTARITY PROBLEMS

D 6
+ +

D 6
Figure 10.1: Trafik Network

Best Known Solution

Coritinuous variables
For A E [0,23]
F 1 -- .c'2
l:' _
-
(A+16) l:'
13 ,.c'3
_
-
2(23-A)
13 '
G_
-
(1286-9A)
13
For A ~ 23,
F1 = F2 = 3, F3 = 0, G = 83.

10.3 Nonlinear Complementarity Problems

Let F : !Rn --+ !Rn. Then, the Norilinear Complementarity Problem, NCP, is
the problem of finding z E !Rn that satisfies

z~O, F(z);:::O

For the solution of the nonlinear complementarity problems, reformulation

of the original NCP as a minimization problem (Kanzow, 1996; Mangasarian
and Solodov, 1993; More, 1996), as a nonlinear system of equations, or as a
parametrie problem have been proposed. Smoothing methods have also been
developed for both linear and norilinear complementarity problems (Chen and
10.3. NONLINEAR COMPLEMENTARITY PROBLEMS 243

Mangasarian, 1996). A review of theory, algorithms and applications of NCPs

is given in Ferris and Pang (1997). For abrief description of many of NCP
applications, see Luo et al. (1997), Harker and Pang (1990) and Nagurney
(1993).

10.3.1 Test Problem 1

This test problem is taken from MCPLIB (Dirkse and Ferris, 1995) .

Formulation

Function

F(x) = arctan(x - s)

s == 10

Problem Statistics
No. of continuous variables 1
No. of linear functions
No. of nonlinear functions 1
No. of known solutions 1

Best Known Solution

Continuous variables
x = 10
Behavior
If s ~ 1.39174521, cycling may occur in Lemke's algorithm.

10.3.2 Test Problem 2

This problem is taken from Dirkse and Ferris (1995) and is originally from
Kojima and Shindo (1986).

Formulation

Function

3x~ + 2XlX2 + 2x~ + X3 + 3X4 - 6]

F(x) = [ 2x~ + x~ + Xl + lOx3 + 2X4 - 2
3x~ + XlX2 + 2x~ + 2X3 + 9X4 - 9
x~ + 3x~ + 2X3 + 3X4 - 3
244 CHAPTER 10. COMPLEMENTARITY PROBLEMS

Problem Statistics

No. of continuous variables 4
No. of linear functions
No. of nonlinear functions 4
No. of known solutions 2

Best Known Solution

Continuous variables
Solution 1: z = (v'6/2, 0, 0, 0.5)T
Solution 2: z = (1,0,3,0)

Behavior

is used, since the LCP formed by linearizing F around x = has no

solution.

The difficulty in solving this problem arises when a Newton-type method

10.3.3 Test Problem 3

This problem is from Pang and Gabriel (1993), (originally by Mathiesen (1987))
and describes a simple Walrasian equilibrium model for testing purposes. Man-
gasarian and Solodov (1993) have also used this problem to test their solution
method.

Formulation

Functions

-Xl + X2 + X3
X4 - O.75(X2 + X3)
1
F(z) =
[
1 - X4 - O.25(X2 + X3)/X2 (1)
- X 4

Problem Statistics
No. of continuous variables 4
No. of linear functions 2
No. of nonlinear functions 2
No. of known solutions multiple

Best Known Solution

Functions: For = 0.5 F(z) = (O,O,O,of
= 2 F(z)
For = (0,0,0, 1.25)T
10.3. NONLINEAR COMPLEMENTARITY PROBLEMS 245

Continuous variables
For = 0.5, x = (0.5,4.83201,1.61067, 3.22134f
For = 2, x == (0.75, 0.78425, 0.78425, O)T
This problem has multiple nondegenerate solutions, of which two are
given in Pang and Gabriel (1993) and recited above.

10.3.4 Nash Equilibrium

The Nash equilibrium problem formulation is taken from Murphy et al. (1982).

Formulation
Find a vector x* such that Vi:

mJn -XiP(Xi + Lxi - h(Xi))

j#i

Xi ~
which, when replaced by the equivalent Karush-Kuhn-Tucker conditions can
be reformulated as an NCP problem.
N CP Formulation

Vi, V'!i(Xi)-P(Q)-XiV'p(Q)~O, Xi~O, .1

Functions
Inverse demand function for the market

P(Q) = 5000~Q-~

Total cost function for firm i

!i(Xi) = CiXi + ( ,-.

i
, /J \
)L-!i
t

Parameter Definition

Ci,L i , i - scalar parameters

'Y - elasticity of demand with respect to price
'Y

Variable Definition
246 CHAPTER 10. COMPLEMENTARITY PROBLEMS

P - market price

Xi - output of firm i

Q - total quantity produced, such that L: Xi = Q

i

10.3.5 Test Problem 4

This test problem is taken from GAMSLIB and is used for testing purposes
in Harker (1988), where Nash equilibrium of a 5-firm noncooperative game is
modeled (Murphy et al., 1982; Harker, 1984). Another Nash equilibrium prob-
lem of firms differentiated by the product they produce instead of production
costs is given in Choi et al. (1990) and included in MCPLIB (Dirkse and Ferris,
1995).

c = (10.0,8.0,6.0,4.0, 2.0)T
L = (5.0,5.0,5.0,5.0, 5.0)T
= (1.2,1.1,1.0,0.9, 0.8)T
'Y = 1.1

Problem Statistics
No. of continuous variables 5
No. of linear functions
No. of nonlinear functions 5

Best Known Solution

Continuous variables
x = (36.933,41.818,43.707,42.659, 39.179)T

Behavior
The solution reported above is taken from the GAMSLIB. The original
paper by Murphy et al. (1982) reports very slightly different values.

10.3.6 Test Problem 5

This example problem is taken from MCPLIB (Dirkse and Ferris, 1995), and is
also used for testing purposes in Harker (1988). This problem is the Cournot-
Nash equilibrium problem with 10 firms.
Data

c = (5.0,3.0,8.0,5.0,1.0,3.0,7.0,4.0,6.0, 3.0)T
L = (10.0,10.0,10.0,10.0,10.0,10.0,10.0,10.0,10.0, 1O.of
10.3. NONLINEAR COMPLEMENTARITY PROBLEMS 247

= (1.2,1.0,0.9,0.6,1.5,1.0,0.7,1.1,0.95, 0.75)T
"I = 1.2

Problem Statistics
No. of continuous variables 10
No. of linear functions
No. of nonlinear functions 10
No. of known solutions 1

Best Known Solution

Continuous variables
z = (7.442,4.098,2.591,0.935,17.949,4.098,1.305, 5.590, 3.222, 1.677f

10.3.7 Invariant Capital Stock Problem

In the invariant capital capital stock problem, an economy is assumed to grow
over an infinite number of time periods where the technology remains constant
(Dirkse and Ferris, 1995). For more details on the problem, see Hansen and
Koopmans (1972).

Formulation

s.t. (2)
AXt - Zt < 0
-BXt +ZtH < 0
CXt - w < 0
Xt ~ o.

NCP Formulation
It is assumed that the objective function is convex and continuously differ-
entiable.

-Vv(z)+(A-aBfy+CTU~O, z~O, .1
(B - A)z ~ 0, y ~ 0 .1
- Cz + w ~ 0, u ~ 0 .1

Parameter Definition

A - capital input matrix

248 CHAPTER 10. COMPLEMENTARITY PROBLEMS

A ij - units of eapital good i required for running proeess j

B - eapital output matrix

Bij - units of eapital good i produeed for running proeess j

C - resouree input matrix

Crj - units of resouree good r for running proeess j
w - a eonstant on the resourees available at the beginning of eaeh time
period

0: - discount factor 0: E (0,1).

Variable Definition

Xj - aetivity level for produetion proeess j

Yi - dual variable to eapital input/output eonstraint on eapital good i

Ur - dual variable to resouree eonstraint on resouree type r

v - utility function

10.3.8 Test Problem 6

This invariant eapital stock problem is taken from MCPLIB (Dirkse and Ferris,
1995).
Data

v(z) = (Xl + 2.5x2)O.2(2.5x3 + X4)O.2(2x5 + 3X6)o.2

2 2 2 2 2 2 2 2
A=(~ 3 2 2 1 1 1 0.5 1 0~5 )
1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 )
B =( 1.5
2.7 2.7 1.8 1.8 0.9 0.9 0.9 0.4 2 1.5

1 1 1 1 1 1 1 1
C =( 0~5 1.5 1.5 0.5 0.5 1.5 1.5 0.5 0.5 1~5 )
p = 0.20
0: = 0.7
w = (0.8,0.8)T
10.3. NONLINEAR COMPLEMENTARITY PROBLEMS 249

Problem Statistics
No. of continuous variables 14
No. of linear functions 4
No. of nonlinear functions lO

Best Known Solution

Continuous variables
z = (0,0.086,0.111,0,0,0.155, 0.lO7, 0, 0, 0.074)T
Y = (0.359,0.683)T
U = (0,0.023)T
 Chapter 11

Semidefinite Programming
Problems

11.1 Introduction
Semidefinite programming involves the minimization of a linear function sub-
ject to the constraint that an affine combination of symmetrie matrices is pos-
itive semidefinite. Several types of problems can be transformed to this form.
This constraint is in general nonlinear and nonsmooth yet convex. Semidefinite
programming can be viewed as an extension of linear programming and reduces
to the linear programming case when the symmetrie matrices are diagonal.

11.1.1 Problem formulation

The semidefinite programming problem can be expressed as:

m
subject to A o- 2: YiAi t 0
i=l

where the vector y E lRm represents the decision variables, b E lRm defines
the objective function, and Ao, . .. ,Am are symmetrie n x n matrices. The
relation A t B where A and B are real symmetrie matrices denotes A - B
as being positive semidefinite. The constraint in this problem is often referred
to as a linear matrix inequality. Multiple linear matrix inequalities can be
combined into a single block diagonal matrix inequality. The dual semidefinite
programming problem can be expressed as:
251

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
252 CHAPTER 11. SEMIDEFINITE PROGRAMMING PROBLEMS

minAoX
x
subject to Ai. X = bi for i = 1, ... ,m

Xto.

The. denotes the operation of taking the inner product of matrices,

A.B:= 'E-AijBij = trace (ATB).
i,j
Provided both primal and dual programs are feasible and there is a strictly
feasible point in either the primal or the dual, the optimal solution to the
primal equals the optimal solution of the dual. In other words, if primal and
dual optimal solutions are X* and (y*, S*) respectively, then C. X* = bT y*.
Approaches developed for the solution of SDP's include: the ellipsoid
method (Yudin and Nemirovsky, 1977) (Nesterov, 1977), general methods for
convex optimization (Hiriart-Urruty and Lernarechal, 1993; Shor, 1985; Ki-
wiel, 1985) and interior point methods which are reviewed by Vandenberghe
and Boyd (1996), Alizadeh (1995), and Pardalos and Wolkowicz (1998).

11.1.2 Semidefinite Programming Applications

The two main areas of application for semidefinite programming are in combi-
natorial optimization (Alizadeh, 1995; Ramana and Pardalos, 1996; Pardalos
and Wolkowicz, 1998) and control theory (Boyd et al., 1994). Applications in
combinatorial theory include:

relaxation of the maximum cut problem;

relaxation of the graph equipartition problem;

relaxation of the maximum clique problem.

A set of maximum cut test problems generated by Helmberg and Rendl (1998)
using G. Rinaldi's machine independent graph generator is available on the
internet at:

ftp://dollar.biz.uiowa.edu/pub/yyye/Gset

These problems involve graphs having 800 to 3000 vertices.

Toh et al. (1998) provide MATLAB files to generate random instances of
semidefinite programming applications. These files, available at:

http://www.math.cmu.edu/-reha/sdpt3.html
generate instances of the following problem types:
11.1. INTRODUCTION 253

maximum eigenvalue determination;

matrix norm minimization;

Chebychev approximation problem for a matrix;

logarithmic Chebychev approximation;

Chebychev approximation on the complex plane;

control and system problem;

relaxation of the maximum cut problem;

relaxation of the stable set problem;

the educational testing problem.

Other applications of semidefinite programming are (Vandenberghe and Boyd,
1996):

pattern separation by ellipsoids;

geometrical problems involving quadratic forms;

truss topology design problems.

A library of semidefinite programming test problems SDPLIB can be found on

the internet at
http://www.nmt.edu/-borchers/sdplib.html

A subclass of semidefinite programming problem is the second order cone

problem:

subject to IIA i y + fill ~ cT~ + di i = 1, ... ,m,

where ~ ERn, b E Rn, Ai E Rnixn, f E Rni , and Ci ERn. This type of
problem may be posed as a semidefinite programming problem:

maxbTy
y

subject to cT
[((A
iY
y++ fd;))~
i
A,j,y + fdi
ci Y + i
] >- 0
-
i = 1, ... ,m.

Applications of the second order cone problem include:

antenna array weight design (Lebret and Boyd, 1997);

254 CHAPTER
GHAPTER 11. SEMIDEFINITE PROGRAMMING PROBLEMS

robotic hand grasp force optimization (Buss et al., 1996);

FIR filter design (Wu et al., 1996);
portfolio optimization with loss risk constraints (Lobo et al., 1998);
truss design (Ben-Tel and Bends0e,
Bends!1le, 1993);
equilibrium calculation of a system with piecewise linear springs (Lobo
et al., 1998).
Second order cone programming test problems may be found on the internet
at:
http://wvw.princeton.edu/-rvdb/ampl/nlmodels/
http://vwv.princeton.edu/-rvdb/ampl/nlmodels/

11.2 Educational Testing Problem

The educational testing problem (Fletcher, 1981a) involves the estimation of a
measure of reliability of a student's total score in an examination that involves
dass of n students who each take a set of
a number of subtests. Consider a class
m tests. Let Xij represent the score obtained by student i on test j. As a
measurement the score obtained will be subject to error:

Xij = Tij + Nij.

Here Tij
1ij and Nij are respectively the "true" value and the "noise" associated
T, has a zero mean and
with X ij . Assume that the noise is uncorrelated with T,
a diagonal covariance matrix E N . The covariance matrix of X, EX, can then
be wdtten ~, the covariance matrix of T and, IJN:
written in terms of E, ~N:

IJX = IJ + IJN.
The reliability of the examination is the quantity:
eTIJe
eT~e
p
p== eTIJx
eT~Xe'e'
where e is vector with all components equal to 1. If there is a large number
of students we may estimate IJx.
~x. However IJ
~ and IJN
~N are unknown. A lower
bound on p may be obtained by using the fact that a covariance matrix is
positive semidefinite and determining IJN
~N via the problem:

~w- trace(IJN)

subject to IJ - IJN !:: 0

IJN !:: o.
11.2. EDUCATIONAL TESTING PROBLEM 255

11.2.1 General formulation

The test problems in this section are based on the tabulated data which origi-
nate from Woodhouse (1976). In the data table column j represents the scores
obtained for subtest j and row i represents the scores obtained by student i.
Test problems are derived from this data by considering cases in which
the overall examination is constituted by different subsets of tests. Let vector
v = (VI, . .. ,Vm '), where m' :::; m, represent the subtests that are included in
the overall examination. These problem definition vectors are given in the first
column of the solution table. The covariance matrix :Ex for a given set of
subtests v is defined as:

(1)

n
where Xj = L: Xij/n is the mean score for subtest j.
i=l
Objective function
The objective function is the trace of the measurement noise covariance
matrix.

where e E !Rm = (1, ... ,1f and ei E !Rm is the ith unit vector.
Constraints
The covariance matrix :E must be positive definite:
m
:Ex - L: Yieief t 0
i=l

where ei E !Rm is the ith unit vector.

Variable bounds
The measurement noise covariance matrix :E N must be positive semidefi-
nite.

y~O

Variable definition
Scalar variable Yi represents the diagonal entry i of :E N .

Data
Data for all twelve educational testing problems is contained in the following
table which defines the matrix X of the student scores for all the subtests.
256 CHAPTER 11. SEMIDEFINITE PROGRAMMING PROBLEMS

15 25 20 28
Data ror matrix X rstudent ri! by test
35 50 21 18 22 28 28 12 15 40
W!
18 23 14 16 15 10
21 27 32 32 41 42 30 35 33 32 64 16 24 38 34 13 15 17 28 18
23 35 40 22 55 48 36 40 46 18 38 18 26 37 24 24 17 20 19 26
23 29 50 36 42 52 44 32 24 19 32 24 20 46 32 23 11 12 40 38
34 37 42 19 36 46 17 26 35 28 39 54 21 47 29 42 18 18 30 20
36 60 70 45 55 54 32 30 32 29 41 28 20 47 36 28 20 20 18 24
36 35 46 27 50 40 60 34 39 46 48 63 20 48 40 19 21 24 40 22
38 70 44 50 45 42 20 28 29 16 55 40 22 49 42 25 23 26 30 28
39 46 52 24 37 60 53 30 46 43 54 54 23 46 44 22 35 22 48 30
40 74 65 60 72 41 33 36 24 52 64 36 28 50 46 26 25 23 30 30
40 48 32 23 58 52 23 40 37 24 58 38 29 54 44 28 11 27 41 34
41 12 24 so 47 48 41 42 37 28 56 57 32 51 43 25 17 24 32 39
46 52 76 48 70 58 20 so 28 42 76 58 28 58 45 34 27 35 18 56
46 73 84 63 38 57 33 56 42 18 72 77 21 52 48 32 35 32 32 19
47 42 74 28 60 57 36 42 48 52 63 46 36 53 49 53 29 30 42 40
47 82 72 70 39 64 21 25 44 26 44 44 37 51 46 33 38 35 37 42
47 40 42 50 48 61 40 40 26 29 61 44 30 56 47 52 46 27 48 40
48 70 65 48 42 57 35 58 50 46 60 32 34 58 54 35 36 31 16 18
49 65 60 55 62 56 52 50 52 28 50 48 34 58 53 41 45 40 38 52
50 30 35 28 62 54 41 46 50 21 65 33 32 58 54 38 50 44 43 50
52 42 54 33 42 64 40 40 56 44 64 38 34 60 44 34 38 30 44 38
52 72 70 65 72 68 62 38 56 44 58 46 56 58 46 36 55 20 48 47
52 44 64 72 44 62 35 44 56 46 62 39 30 61 46 38 40 42 24 80
53 25 42 28 68 52 41 45 44 26 28 43 51 62 47 35 42 48 50 40
54 48 60 58 36 51 63 41 64 29 63 49 32 58 47 39 43 58 48 49
55 64 62 30 42 57 34 47 52 34 57 37 43 63 48 38 47 20 54 65
58 30 24 62 51 51 44 36 43 25 36 54 41 65 48 43 40 35 50 42
58 16 40 45 42 58 44 42 58 36 58 52 40 64 49 36 18 45 53 28
58 44 56 51 68 68 46 48 72 38 62 34 32 68 49 42 47 47 28 70
39 58 58 50 74 52 36 58 60 28 44 56 34 72 51 33 48 58 54 58
60 32 35 48 40 56 52 32 40 37 72 57 36 61 52 51 42 46 17 51
60 78 80 62 52 54 58 47 80 32 64 39 45 66 53 42 70 40 50 18
60 38 55 66 42 52 30 54 62 42 90 38 38 63 56 46 62 48 55 44
61 48 64 68 70 53 42 40 38 45 73 56 50 64 54 46 45 42 50 22
62 86 94 50 49 62 48 56 74 33 84 36 52 67 52 47 41 70 57 53
63 35 38 55 38 58 46 59 63 48 62 58 38 68 53 47 48 75 44 25
64 79 65 76 68 57 32 33 52 46 72 62 52 54 54 48 55 35 60 54
64 50 52 35 60 56 52 64 76 36 63 44 48 56 52 49 63 38 48 46
65 37 42 70 50 58 58 62 66 34 53 64 62 72 53 50 74 45 62 57
65 82 74 63 36 62 60 58 60 38 57 63 58 70 51 51 55 55 64 58
67 44 46 54 52 58 55 68 80 40 28 65 50 71 54 52 65 70 74 68
67 48 56 80 44 64 62 35 70 62 58 72 56 72 60 41 78 38 75 68
68 62 78 56 50 53 62 54 80 64 62 48 54 74 62 76 58 45 80 36
69 39 30 42 38 62 40 32 68 56 68 71 58 73 56 43 79 47 73 70
70 52 20 76 69 61 64 56 69 54 64 66 58 78 52 72 32 47 71 71
72 54 72 38 54 51 66 65 76 72 54 49 60 74 57 42 68 62 22 46
72 42 48 70 70 57 42 40 68 53 62 74 60 78 58 52 55 48 72 75
74 64 66 70 42 60 40 78 53 48 69 67 76 77 59 68 58 55 16 60
78 68 62 63 35 56 63 80 74 43 71 78 62 76 59 68 70 75 72 75
79 37 42 28 64 52 40 38 72 72 56 52 58 82 60 61 75 66 58 58
80 62 30 65 59 51 68 57 74 48 78 71 42 54 61 62 78 69 78 58
82 85 80 52 44 57 70 69 64 71 85 76 64 56 63 67 44 70 60 26
82 40 74 52 52 64 74 64 76 46 64 46 51 80 63 68 65 55 70 67
84 42 76 70 55 61 90 80 56 41 58 82 72 72 67 73 85 60 76 78
84 42 54 60 42 58 42 60 52 70 77 68 68 74 59 71 79 65 44 75
86 85 88 80 37 63 80 72 79 42 73 42 68 82 65 73 85 62 80 82
87 53 51 62 68 56 60 42 78 42 62 74 62 84 65 75 63 75 68 83
89 41 60 40 60 54 88 88 83 57 84 64 64 80 62 65 90 78 88 52
90 73 78 77 52 42 56 50 58 59 72 84 70 84 62 63 82 85 78 87
90 81 74 64 48 38 86 52 80 63 66 68 60 62 63 74 75 81 84 94
96 85 88 90 72 44 58 62 70 74 64 74 72 64 64 84 85 78 88 54
97 56 55 35 68 70 78 76 56 72 83 69 65 86 65 76 82 89 92 90
99 75 65 88 54 42 80 90 88 58 78 88 70 88 63 82 72 71 98 80
100 65 75 70 70 60 83 85 70 62 72 90 72 84 64 78 88 80 80 72

The matrix EX can be computed from the X matrix using equation 1.

Best known solution

The solutions shown in the following table are from Chu and Wright (1995).
11.2. EDUCATIONAL TESTING PROBLEM 257

Problem Number of Optimal Solution y* Objective

Vector v Variables Value
1,2,5,6 4 173 237 104 29 543

1,3,4,5 4 156.232 240.935 128.742 107.248 633.158

1,2,3,6, 6 0.000 102.020 19.877 31.461 305.482

8, 10 82.283 69.840

1,2,4,5, 6 59.623 214.032 69.805 115.733 564.463

6, 8 47.040 58.230

1:6 6 152.706 54.476 82.631 99.643 535.362

104.655 40.953

1:8 8 14.032 38.542 95.099 158.901 641.839

120.382 28.371 106.775 79.736

1:10 10 0.000 43.892 80.717 132.887 690.781

126.862 28.030 92.610 56.620
61.336 67.826

1:12 12 18.633 61.863 63.427 127.568 747.489

99.973 30.770 96.535 45.288
41.601 45.329 64.041 52.460

1:14 14 0.000 59.499 62.912 109.924 671.275

99.949 32.719 79.073 31.738
47.421 33.789 41.953 63.596
4.252 4.451

1:16 16 0.000 63.487 52.389 108.192 663.4619

92.395 34.562 85.755 21.957
37.549 32.967 28.511 54.571
12.930 4.104 6.706 27.387

1:18 18 0.000 58.380 62.162 107.231 747.5058

80.287 25.383 70.703 24.217
52.438 41.695 24.292 39.176
15.761 6.862 3.259 14.593
68.804 52.162
258 CHAPTER 11. SEMIDEFINITE PROGRAMMING PROBLEMS

Problem Number of Optimal Solution y* Objective

Vector v Variables Value
1:20 20 0.000 47.373 76.582 101.002 820.343
63.450 13.382 41.483 4.300
56.365 33.983 33.770 29.960
17.597 0.000 4.328 13.690
45.587 51.586 57.207 128.698

11.3 Maximum Cut Problem

This section introduces the statement of the maximum cut problem, describes
how an upper bound to the problem's optimal objective value can be obtained
via a semidefinite programming relaxation and provides a set of such problems
with their best known solutions. First some graph theoretic terminology is
introduced.
A gmph G = (V, e) consists of a finite nonempty set V of nodes and a
e
finite set of edges. An unordered pair of nodes (i,j) is associated with each
edge. A weighted graph is one in which a scalar value Wij is associated with
each edge (i,j), when (i,j) is not an edge Wij = O. For a given graph G the
cut 8(S) induced by the vertex set S consists of the set of edges with exactly
one endpoint in S. The weight of a cut, w(8(S)) is the sum of the weights
associated with the set of edges in the cut, w(8(S)) = L Wij.
(i,j)Et5(S)
The maximum cut problem involves finding a cut 8(S) of maximum weight.
The maximum cut problem can be formulated as an integer quadratic program.

m:X4 L L
1
wij(l - XiXj)
iEV jEV
subject to x E {O, l}n.
In this formulation Xi = 1 if i E Sand Xi = -1 if i E V\S. The above
formulation is equivalent to the following continuous quadratically constrained
global optimization problem:

1
max -LeX
r1J,X 4
subject to Xii = 1 for all i E V.
X =xxT ,
where x E !Rn; X E !Rnxn , and the weighted Laplacian matrix L E !Rnxn is
composed of diagonal elements Lii = L Wij for i = 1, ... , n, and off-diagonal
jEV
elements Lij = -Wij for i = 1, ... , n; j = 1, ... , n; i =1= j. This nonconvex
problem can be approximated as a convex semidefinite programming problem
by relaxing the constraint X = xxT to X ~ o.
11.3. MAXIMUM CUT PROBLEM 259

11.3.1 General formulation

The dual formulation of the above maximum cut problem relaxation is pre-
sented in this section.
Objective function

maxeTy
y

eE!Rn ,ei=l, i=l, ... ,n

y E!Rn

Constraints

Ei E !Rnxn has all zero elements except (Edii = 1.

Variables
The vector y is the variable in this problem.
Data
Tables 11.1 and 11.2, from Shor (1998), define the edges and weights of
weighted graphs with 10 and 20 nodes respectively. Weights Wij where (i, j) is
not an edge are set to zero.
Best known solution
Table 11.3 provides data on the solutions to the maximum cut problems
and their semidefinite programming relaxations.

In table 11.3 w SDP (<5(S)) refers to the optimal objective function to the semidef-
inite relaxation of the maximum cut problem, as reported by Shor (1998). S*
refers to the set of nodes corresponding to the best known feasible solution to
the maximum cut problem itself, and w(<5(S*)) is the weight of this cut.
260 CHAPTER 11. SEMIDEFINITE PROGRAMMING PROBLEMS

edge weight Wij

(i,j) Pl P2 P3 P4 P5 P6 P7 Ps Pg
(1,2) 1 5 2 5 1 5 2 5 1
(1,5) 1 1 3 6 8 2 7 2 3
(1,10) 1 3 4 3 1 2 2 2 3
(2,3) 1 2 7 2 3 1 3 6 8
(2,9) 1 2 6 2 5 5 1 5 9
(3,4) 1 3 1 6 7 3 4 3 1
(3,8) 1 4 3 4 2 1 2 1 2
(4,5) 1 1 2 1 2 2 6 2 5
(4,7) 1 1 5 1 3 4 3 4 2
(5,6) 1 5 1 5 9 3 1 6 7
(6,8) 1 5 5 5 4 1 5 1 3
(6,9) 1 3 1 3 1 5 5 5 4
(7,9) 1 2 2 2 3 3 1 3 1
(7,10) 1 1 2 1 8 1 2 1 8
(8,10) 1 1 1 1 1 1 1 1 1

Table 11.1: Weighted graphs with 10 vertices and 15 edges

edge weight Wij edge weight Wij

(i, j) Tl T2 T3 T4 T5 (i, j) Tl T2 T3 T4 n
(1,2) 1 7 7 2 6 (10,12) 1 7 1 2 1
(1,18) 1 7 7 3 5 (10,20) 1 9 7 4 2
(2,3) 1 7 7 5 4 (11,12) 1 1 7 5 5
(2,16) 1 7 9 4 3 (11,13) 1 7 1 6 5
(2,18) 1 7 9 1 3 (12,13) 1 7 7 1 4
(3,4) 1 7 7 5 4 (13,14) 1 7 7 1 4
(4,5) 1 1 7 5 5 (13,15) 1 9 1 1 1
(5,6) 1 7 7 3 6 (13,19) 1 9 1 3 1
(5,9) 1 1 1 2 1 (13,20) 1 7 9 4 2
(5,19) 1 7 9 2 3 (14,15) 1 9 7 2 2
(5,20) 1 7 9 4 4 (14,16) 1 7 1 3 5
(6,7) 1 7 7 5 6 (15,16) 1 1 7 2 3
(6,8) 1 1 1 1 3 (15,20) 1 9 9 4 2
(7,8) 1 7 7 3 3 (16,17) 1 7 7 2 1
(8,9) 1 1 7 3 2 (16,18) 1 9 9 3 3
(8,19) 1 7 7 2 1 (16,20) 1 9 7 2 2
(9,10) 1 7 7 1 2 (17,18) 1 7 7 1 1
(10,11) 1 1 7 1 1 (19,20) 1 9 9 3 3

Table 11.2: Weighted graphs with 20 vertices and 36 edges

11.3. MAXIMUM CUT PROBLEM 261

Problem w~Ul' (o(S)) w(o(S*)) S* V \S*

Pl 12.71280 12.0 1,2,4,6,7,8 3,5,9,10
P2 35.70171 35.0 2,4,5,8,9,10 1,3,6,7
P3 40.91025 40.0 2,5,7,8,10 1,3,4,6,9
P4 43.81359 43.0 1,3,6,7 2,4,5,8,9,10
P5 52.76558 51.0 1,2,4,6,10 3,5,7,8,9
P6 34.97500 34.0 2,3,5,6,7,10 1,4,8,9
P7 40.13301 39.0 2,3,5,6,7,10 1,4,8,9
Ps 41.75555 41.0 1,3,5,7,8,9 2,4,6,10
P9 52.99566 52.0 1,2,4,6,7,8 3,5,9,10
Tl 29.93753 29.0 1,3,5,7,8,10 2,4,6,9,11,12
13,15,16 14,17,18,19,20
T2 192.08551 185.0 2,4,6,9,11,12, 1,3,5,7,8,10,
14,17,18,19,20 13,15,16
T3 189.71114 181.0 1,2,4,6,8,10, 3,5,7,9,11,13,
12,14,17,20 15,16,18,19
T4 88.07765 87.0 2,4,6,9,11,14, 1,3,5,7,8,10,
17,18,19,20 12,13,15,16
T5 91.79930 90.0 2,4,6,9,11,12,14, 1,3,5,7,8,10,
15,17,18,19,20 13,16

Table 11.3: Solutions to the maximum cut problems

 Chapter 12

Mixed-Integer Nonlinear
Programming Problems
(MINLPs)

12.1 Introduction
Mixed-integer problems are those that involve both eontinuous and integer vari-
ables. The introduction of integer variables allows the modeling of eomplex
deeisions through graph theoretic representations denoted as superstructures
(Floudas, 1995). This representation leads to the simultaneous determination
of the optimal structure of a network and its optimum operating parameters.
Thus MINLPs find applieations in engineering design sueh as heat exehanger
network synthesis, reaetor-separator-reeycle network synthesis or pump net-
work synthesis (Floudas, 1995; Grossmann, 1996), in metabolie pathway en-
gineering (Hatzimanikatis et al., 1996a,b; Dean and Dervakos, 1996), or in
moleeular design (Maranas, 1996; Churi and Aehenie, 1996).
The solution of MINLPs is not a trivial matter. The presenee of binary
or integer variables results in a eombinatorial explosion of the number of solu-
tions, adding to the diffieulties arising from the noneonvexity of the problem.
Several teehniques have been devised for eonvex MINLPs: braneh-and-bound
approaehes (Beale, 1977; Gupta and Ravindran, 1985; Ostrovsky et al. , 1990;
Borehers and Mitehell, 1991; Quesada and Grossmann, 1992; Fleteher and
Leyffer, 1998), the Generalized Benders Deeomposition and its variants (Ge-
offrion, 1972; Paules and Floudas, 1989; Floudas et al., 1989), the Outer Ap-
proximation algorithm and its variants (Duran and Grossmann, 1986; Koeis
and Grossmann, 1987; Viswanathan and Grossmann, 1990), and the Gen-
eralized Outer Approximation (Fleteher and Leyffer, 1994). While some of
these teehniques ean also be used to find solutions of noneonvex MINLPs,
global optimality eannot be guaranteed. In the last few years, several groups
have eoneentrated on designing global optimization algorithms for MINLPs.
The methods proposed, of varying applieability, include the interval-based ap-
263

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
264 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

proach of Vaidyanathan and EI-Halwagi (1996), the Extended Cutting Plane

algorithm of Westerlund et al. (1998), and the branch-and-bound algorithms
of Ryoo and Sahinidis (1995), Smith and Pantelides (1997) and Adjiman et al.
(1998d).
The reported global solutions in this chapter are taken from the work of
Adjiman et al. (1998d). A small set oftest problems for mixed integer quadratic
problems can be found in Fleteher and Leyffer (1998). For a description of
the theoretical and algorithmic issues of global optimization approaches for
nonconvex MINLP problems, the reader is directed to the article of Adjiman
et al. (1998d), and the forthcoming book by Floudas (2000).

12.2 Literature Problems

12.2.1 Test Problem 1
This example was first proposed by Kocis and Grossmann (1988).

Formulation

Objective function

Constraints

X~ +YI 1.25
x~5 + 1.5Y2 3
Xl +YI< 1.6
< 3
1. 333x2 + Y2
-YI - Y2 +Y3 < 0

Variable bounds

XI,X2 > 0
y E {0,1}3.

Problem Statistics
No. of continuous variables 2
No. of binary variables 3
No. of convex inequalities 3
No. of nonlinear equalities 2
No. of known solutions 2
12.2. LITERAT URE PROBLEMS 265

Global Solution
Objective function: 7.6672
Continuous variables: :v = (1.12, 1.31)T.

Binary variables: y = (0,1, l)T.

12.2.2 Test Problem 2

This example is example 6.6.5 from Floudas (1995).

Formulation

Objective function

min -0.7y + 5(Xl - 0.5)2 + 0.8

:V,y

Constraints
_e(x l -O.2) - < 0
X2

X2 + 1.1y :::; 1
Xl -1.2y < 0.2

Variable bounds

0.2:::; Xl :::; 1
-2.22554::; X2 ::;-1
Y E {0,1}

Problem Statistics
No. of continuous variables 2
No. of binary variables 1
No. of convex inequalities 2
No. of nonconvex inequalities 1
No. of known solutions 1

Global Solution
Objective function: 1.0765.

Continuous variables: :v = (0.9419, -2.1f.

Binary variables: y = 1.
266 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

12.2.3 Test Problem 3

This example was proposed by Yuan et al. (1988).

Formulation

Objective function

min (YI - 1)2 + (Y2 - 2)2 + (Y3 - 1)2 -ln(Y4 + 1)

x,y
+ (Xl - 1)2 + (X2 - 2)2 + (X3 - 3)2

Constraints

+ Y2 ++ Y3 ++ Xl
YI + :s
+ X2 ++ X3 < 5
Xl +
2 2 2 2
Y~ +
Y3 + xl + x2
x~ + x~ ++ x3
x~ < 5.5
YI +XI < 1.2
Y2 +X2 < 1.8
Y3 +X3 < 2.5
Y4 +XI < 1.2
y~
Y~ +x~ < 1.64
y~
Y~ +x~ < 4.25
y~
Y~ +x~ :s::; 4.64

Variable bounds

XI,X2,X3 > 0
~
y
Y E {0,1}4.

Problem Statistics
No. of continuous variables 3
No. of binary variables 4
No. of convex inequalities 5
No. of nonconvex inequalities 4
No. of known solutions 2

Global Solution
Objective function: 4.5796.

Continuous variables: x = (0.2,0.8, 1.908f.

Binary variables: y == (1,1,0, 1f.
12.2. LITERATURE PROBLEMS 267

12.2.4 Test Problem 4

This example was first presented by Berman and Ashrafi (1993).

Formulation

Objective function

Constraints

Xl + 0.IYIO.2Y20.15Y3 1
X2 + 0.05Y40.2Y50.15Y6 1
X3 + 0.02Y7 O.06 Y8 = 1
<
-YI - Y2 - Y3 -1
<
-Y4 - Y5 - Y6 -1
<
-Y7 - Y8 -1
3YI + Y2 + 2Y3 + 3Y4 + 2Y5 + Y6 + 3Y7 + 2Y8 < 10

Variable bounds

o ::; Xl, X2, X3

::; 1
YE{0,1}8.

Problem Statistics
No. of continuous variables 3
No. of binary variables 8
No. of convex inequalities 4
No. of nonlinear equalities 3
No. of known solutions 3

Global Solution
Objective function: -0.94347

Continuous variables: x = (0.97,0.9925, 0.98)T.

Binary variables: y = (0,1,1,1,0,1,1, O)T.

12.2.5 Test Problem 5

This example is a purely integer nonconvex problem presented in Prn et al.
(1997).
268 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Formulation

Objective function

Constraints

yt 2yp - 7Yl - < 24

9Y2
-Yl - 2Y2 < 5
-3Yl + Y2 < 1
4Yl - 3Y2 < 11

Variable bounds

Yl,Y2 E [1,5] nN

Problem Statistics
No. of integer variables 2
No. of convex inequalities 3
No. of nonconvex inequalities 1
No. of known solutions 2

Global Solution
Objective function: 31.

Integer variables: y = (3, If.

12.2.6 Test Problem 6

This example is also taken from Prn et al. (1997).

Formulation

Objective function

min 3y- 5x
x,y

Constraints

2y 2 - 2y O.5 - 2x O.5 y2 + 11y + 8x < 39

-y+x < 3
2y + 3x < 24
12.3. HEAT EXCHANGER NETWORK SYNTHESIS 269

Variable bounds

l:5x:5l0
y E [1,6] nN

Problem Statistics
No. of continuous variables 1
No. of integer variables 1
No. of convex inequalities 2
No. of nonconvex inequalities 1
No. of known solutions 1

Global Solution
Objective function: -17.

Continuous variable: x = 4.

Integer variable: y = 1.

12.3 Heat Exchanger Network Synthesis with the

ehen Approximation
12.3.1 Introduction
The goal of heat exchanger network synthesis problems is to minimize the annu-
alized cost of a heat exchanger network by determining which hot stream/cold
stream, hot stream/cold utility and cold stream/hot utility matches should
take place and what heat loads should be assigned to each heat exchanger.
Binary variables are used to represent the existence of each heat exchanger. A
variety of models based on this principle have been studied in the literat ure
(Floudas et al., 1986; Floudas and Ciric, 1988; Gundersen and Naess, 1988;
Yee and Grossmann, 1991; Quesada and Grossmann, 1993; Zamora and Gross-
mann, 1998). Several expressions have been used to compute the area of a heat
exchanger which enters into the calculation of the capital cost of the network.
In the present problem, taken from Yee and Grossmann (1991), the logarith-
mic mean temperature difference which appears in the exact area expression
is approximated with the formula derived by Chen (1987).
The design of a heat exchanger network involving two hot streams, two
cold streams, one hot and one cold utility is given. There are two temperature
intervals. The superstructure for this problem is shown in Figure 12.1. The
annualized cost of the network is expressed as the summation of the utility
costs, the fixed charges for the required heat-exchangers and an area-based
cost for each heat-exchanger. Since the Chen approximation for the logarithmic
270 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Stage 1 Stage 2
Cl

C2

H2

Figure 12.1: Superstructure for the heat exchanger network problem with the
Chen approximation.

mean of the temperature difference is used Chen (1987), the area is a highly
nonlinear function of the heat duty and the temperature differences at both
ends of the heat exchanger.

12.3.2 General Formulation

Objective function

min L CCUQCU,i + L CHUQHU,j

T,Q,.tl..T,z iEHP JECP
+ L L L CFijZijk + L CFi,CUZCU,i + L CFj,HUZHU,j
iEHPjECPkESI iEHP JECP
+ L L L CijQijk I
iEHP JECP kESI Uij[TijkTijk+l (Tij k+ Tijk+l)/2]l
+ L Ci,CUQCU,i I
iEH P UCU,i [TCU,i (Tout,i - Tin,cu ) (TCU,i +Tout,i-Tin,cU ) /2]l
+ L Cj,HUQHU,j I
JECP UHU,j [THU,j (Tin,HU-Tout,j) (THU,j+Tin,HU-Tout,j )/2]l

Constraints

L L Qijk + QCU,i
Qijk QCU,i (1in,i - Tout,i) FCPi, 'V i E HP
kESI JECP

L L Qijk + QHU,j
Qijk QHU,j (Tout,j -1in,j) Fcpj, 'V j E CP
kESliEHP
12.3. HEAT EXCHANGER NETWORK SYNTHESIS 271

(Ti,k-1i,k+l}Fcpi- L Qijk 0, VkESI, ViEHP

JECP

(Tj,k - Tj,k+l) FCPj - L Qijk = 0, V k E SI, V jE CP

iEHP
1i,1 Tin,i, Vi E HP
Tj,NS+l = Tin,j, \f jE CP
Ti,k - Ti,k+l > 0, VkESI, ViEHP
Tj,k - Tj,k+l > 0, Vk E SI, Vj E CP
1i,NS+l > Tout,i, Vi E HP
Tj,l < Tout,j, VjE CP
FCPi Ti,NS+l - QCU,i = FCPi Tout,i, Vi E HP
Fcpj Tj,l + QHU,j Fcpj Tout,j, VjE CP
max Zijk < 0,
Qijk - Qij
Vk E SI, Vi E HP, Vj E CP
QCU,i - Qf:fJUZCU,i < 0, Vi E HP
Q HU,j - Qmax
HU,jZHU,j < 0, VjE CP
Tijk - Ti,k + Tj,k + T[jax Zijk < Trn
~J
ax
'
Vk E SI, Vi E HP, Vj E CP
Trnax
~J '
Vk E SI, Vi E HP, Vj E CP
TCU,i -Ti,NS+l + T[~JtjZCU,i < Tr~Jtj - Tout,cu, Vi E HP
THU,j + Tj,l + TJI{jjzHU,j < TJI{jj + Tout,HU, V j E C P

where T[jax = max {O, Tj,in - Ti,in, Tj,in -1i,out, Tj,out - Ti,in, Tj,out - T;,out},
for all i E HP U {HU} and jE CP U {CU}; and
Qflax = min{(Tin,i - Tout,i} FCPi, (Tout,j -Tin,j) Fcpj}, for all i E HPU{HU}
and jE CP U {CU}.
Variable bounds

O~ Qijk ~ Qflax, V k E SI, Vi E HP, V jE CP

O~ QCU,i ~ Q"!:!!Ju, Vi E HP
O~ QHU,j ~ Q7iu,j, V jE CP
1i,out ~ Tik ~ Ti,in, V k E ST, Vi E HP
Tj,in ~ Tjk ~ Tj,out, V k E ST, V jE CP
Tmapp $ Tijk ~ T[;aX, V k E ST, Vi E HP, V j E C P
Tmapp ~ TCU,i ~ T[~Jtj, Vi EHP
Tmapp ~ THU,j ~ TJf{jj, V jE CP
Zijb ZCU,i, ZHU,j E{D,l}, V k E SI, Vi E HP, V j E C P
272 GHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Variable definitions
SI is the set of temperature intervals or stages, NS is the number of stages,
HP is the set of hot process streams, G P is the set of cold process streams,
and ST is the set of temperature locations.
The continuous variables are Tik, the temperature of hot stream i at the
temperature location kj Tjk, the temperature of cold stream j at the tem-
perature location kj Qijk, the heat exchanged between hot stream i and cold
stream j in stage kj QCU,i, the heat exchanged between hot stream i and the
cold utilitYj Q HU,j, the heat exchanged between cold stream j and the hot util-
itYj tlTijk, the temperature approach for the match of hot stream i and cold
stream j at temperature location kj tlTCU,i, the temperature approach for the
match of hot stream i and the cold utilitYj tlTHU,j, the temperature approach
for the match of cold stream j and the hot utility. The binary variables are
Zijk, for the existence of a match between hot stream i and cold stream j in
stage kj ZCU,i, for the existence of a match between hot stream i and the cold
utilitYj ZHU,j, for the existence of a match between cold stream j and the hot
utility.
The parameters are 1in, the inlet temperature of a streamj Tout , the out let
temperaturej Fcp, the heat capacity flowrate of a streamj Qmax, the upper
bound on heat exchangej tlT max , the upper bound on the temperature dif-
ferencej GHU, the steam utility costj Gcu, the cooling water costj GFij, the
fixed charge for the heat exchanger between hot stream i and cold stream j j
G Fi,cu, the fixed charge for the heat exchanger between hot stream i and
the cold utilitYj G Fj,HU, the fixed charge for the heat exchanger between cold
stream j and the hot utilitYj Gij, the coefficient for the area-dependent cost of
the heat exchanger between hot stream i and cold stream j j Gi,CU, the coeffi-
cient for the area-dependent cost of the heat exchanger between hot stream i
and the cold utilitYj Gj,HU, the coefficient for the area-dependent cost of the
heat exchanger between cold stream j and the hot utilitYj Uij, the overall heat
transfer coefficient for the hot stream i-cold stream j unitj UCU,i, the overall
heat transfer coefficient for the hot stream i-cold utility exchangerj U HU,j, the
overall heat transfer coefficient for the hot utility-cold stream j unitj tlTmapp ,
the minimum approach temperature.
12.3. HEAT EXCHANGER NETWORK SYNTHESIS 273

12.3.3 Test Problem 1

A specific formulation for heat exchanger network synthesis using the ehen
approximation is as follows.

Formulation

Objective function

mm 15QcU,HI + 15QcU,H2 + 80QHU,CI + 80QHU,C2 + 5500ZHl,CI,l

T,Q,AT,z
+5500ZHl,CI,2 + 5500ZHl,CI,3 + 5500ZHl,C2,1 + 5500ZHl,C2,2
+5500ZHl,C2,3 + 5500zH2,CI,1 + 5500zH2,Cl,2 + 5500zH2,CI,3
+5500zH2,C2,1 + 5500zH2,C2,2 + 5500zH2,C2,3 + 5500zCU,Hl
+5500zCU,H2 + 5500ZHU,CI + 5500ZHU,C2
300QHI Cl I
+ ' , 1

0.5 [~THl,CI,I~THl,cl,2 (~THl,CI,1 + ~THl,CI,2) /2)3

300QHl Cl2
+ ' , 1

0.5 [~THl,Cl,2~THI,CI,3 (~THl,CI,2 + ~THl,CI,3) /2]a

300QHI C21
+ ' , 1

0.5 [~THl,C2,I~THl,C2,2 (~THl,C2,1 + ~THl,C2,2) /2)3

300QHI C22
+ ' , 1

0.5 [~THl,C2,2~THl,C2,3 (~THl,C2,2 + ~THl,C2,3) /2)3

300QH2 Cil
+ ' , 1

0.5 [~TH2,CI,l~TH2,CI,2 (~TH2,CI,1 + ~TH2,Cl,2) /2)3

300QH2 C12
+ ' , 1

0.5 [~TH2,Cl,2~TH2,CI,3 (~TH2,CI,2 + ~TH2,Cl,3) /2]3

300QH2 C21
+ ' , 1

0.5 [~TH2,C2,l~TH2,C2,2 (~TH2,C2,1 + ~TH2,C2,2) /2]a

300QH2 C22
+ ' , 1

0.5 [~TH2,C2,2~TH2,C2,3 (~TH2,C2,2 + ~TH2,C2,3) /2)3

300Qcu HI
+ ' 1
0.83333 [70~Tcu,HI (~TCU,Hl + 70) /2]a
300Qcu H2
+ ' 1
0.83333 [706Tcu,H2 (~TCU,H2 + 70) /2]a
300QHUCI
+ ' 1
0.5 [306THU,CI (~THU,CI + 30) /2]a
300QHUC2
+ ' 1
0.5 [180~THU,C2 (~THU,C2 + 180) /2]a
274 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Constraints

QHl,Cl,l + QHl,C2,1 + QHl,Cl,2 + QHl,C2,2 + QCU,Hl = 2800

QH2,Cl,l + QH2,C2,1 + QH2,Cl,2 + QH2,C2,2 + QCU,H2 4400
QHl,Cl,l + QH2,Cl,l + QHl,Cl,2 + QH2,Cl,2 + QHU,Cl = 3600
QHl,C2,1 + QH2,C2,1 + QHl,C2,2 + QH2,C2,2 + QHU,C2 1950
lO(THl,l - THl,2) - QHl,Cl,l - QHl,C2,1 0
1O(THl,2 - THl,3) - QHl,Cl,2 - QHl,C2,2 0
20(TH2,1 - TH2,2) - QH2,Cl,l - QH2,C2,1 0
20(TH2 ,2 - T H2 ,3) - QH2,Cl,2 - QH2,C2,2 0
15(Tcl,1 - TCl,2) - QHl,Cl,l - QH2,Cl,l 0
15(Tcl,2 - TCl,3) - QHl,Cl,2 - QH2,Cl,2 0
13(Tc2,1 - TC2,2) - QHl,C2,1 - QH2,C2,1 0
13(Tc2,2 - TC2,3) - QHl,C2,2 - QH2,C2,2 0
THl,l 650
TH2,1 590
TCI,3 410
TC2,3 350
-THl,1 + THl,2 < 0
-THl,2 + THl,3 < 0
-TH2,1 + TH2,2 < 0
- TH2,2 + TH2,3 < 0
-TCI,1 + TCI,2 < 0
-TCI,2 + TCI,3 < 0
-TC2,1 + TC2,2 < 0
-TC2,2 + TC2,3 < 0
-THl,3 < -370
-TH2,3 < -370
TCI,1 < 650
TC2,1 < 500
10THl,3 - QCU,HI = 3700
20TH2 ,3 - QCU,H2 7400
15TcI,1+ QHU,CI 9750
13Tc2,1 + QHU,C2 6500
12.3. HEATEXCHANGERNETWORKSYNTHESm 275

QHI,CI,I - 2800ZHl,CI,I :5 0
QHI,CI,2 - 2800ZHI,GI,2 :5 0
QHl,C2,I - 1950ZHl,C2,I :5 0
QHl,C2,2 -1950ZHl,C2,2 :5 0
QH2,CI,1 - 3600zH2,CI,1 < 0
QH2,CI,2 - 3600ZH2,GI,2 :5 0
QH2,G2,1 - 1950zH2,G2,1 :5 0
QH2,C2,2 - 1950zH2,c2,2 < 0
QCU,HI - 2800zCU,HI :5 0
QCU,H2 - 4400zCU,H2 :5 0
QHU,GI - 3600ZHU,GI :5 0
QHU,C2 -1950ZHU,C2 :5 0
THl,CI,I - THl,I + TCI,I + 280ZHl,CI,l < 280
THl,CI,2 - THl,2 + TCI,2 + 280ZHl,GI,1 :5 280
THl,CI,2 - THl,2 + TCI,2 + 280ZHl,CI,2 :5 280
THI,CI,3 - THl,3 + TCI,3 + 280ZHl,CI,2 :5 280
THl,C2,l - THI,I + TC2,1 + 300ZHl,C2,l :5 300
THI,C2,2 - THI,2 + TC2,2 + 300ZHI,G2,1 :5 300
THl,C2,2 - THl,2 + TC2,2 + 300ZHl,C2,2 :5 300
THl,C2,3 - THl,3 + TC2,3 + 300ZHl,C2,2 :5 300
TH2,CI,1 - TH2,1 + TCI,I + 280zH2,GI,1 < 280
TH2,GI,2 - TH2,2 + TCI,2 + 280zH2,CI,1 < 280
TH2,CI,2 - TH2,2 + TCI,2 + 280ZH2,CI,2 < 280
.6.TH2,CI,3 - TH2,3 + TCI,3 + 280zH2,CI,2 < 280
TH2,C2,1 - TH2,1 + TC2,1 + 240zH2,C2,1 :5 240
TH2,C2,2 - TH2,2 + TC2,2 + 240zH2,C2,1 :5 240
.6.TH2,G2,2 - TH2,2 + TC2,2 + 240ZH2,G2,2 :5 240
TH2,C2,3 - TH2,3 + TC2,3 + 240zH2,C2,2 :5 240
TCU,HI - THl,3 + 350zcU,HI < -670
TCU,H2 - TH2,3 + 290zcU,H2 :5 -610
THU,CI + TCI,I + 270ZHU,CI :5 950
THU,C2 + TC2,1 + 330ZHU,C2 < 1010
276 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Variable bounds

O~ QH1,Cl,k ~ 2800, V k E {1,2}

O~ QH1,C2,k ~ 1950, Vk E {1,2}
O~ QH2,Cl,k ~ 3600, Vk E {1,2}
O~ QH2,C2,k ~ 1950, Vk E {1, 2}
O~ QCU,Hl ~ 2800
O~ QCU,H2 ~4400

O~ QHU,Cl ~ 3600
O~ QHU,C2 ~ 1950
370~ TH1,k ~ 650, Vk E {1,2,3}
370 ~ TH2,k ~ 590, Vk E {1, 2, 3}
410 ~ TC1,k ~ 650,Vk E {1,2,3}
350~ TC1,k ~ 500, V k E {1, 2, 3}
1O~ 6.TH1,Cl,k ~ 280, V k E {1, 2, 3}
10 ~ 6.TH1,C2,k ~ 300, Vk E {1, 2, 3}
1O~ 6.TH2,Cl,k ~ 280, V k E {1, 2, 3}
1O~ 6.TH2,C2,k ~ 240, Vk E {1,2,3}
10 ~ 6.TCU,Hl ~ 350
10 ~ 6.TCU,H2 =:;; 290
10 ~ 6.THU,Cl ~ 270
1O~ 6.THU,Cl ~ 330
Zijk, ZCU,i, ZHU,j E {0,1}, Vk E {1,2}, Vi E {H1,H2}, Vj E {C1,C2}

Data
Stream l1n (K) Taut (K) Fcp (kW/K)
Hot 1 650 370 10.0
Hot 2 590 370 20.0
Cold 1 410 650 15.0
Cold 2 350 500 13.0
Steam 680 680
Water 300 320

CFij = $5500/yr CFi,CU = $5500/yr CFj,HU = $5500/yr

C ij = S3OO/yr Ci,CU = S300/yr Cj,HU = $3OO/yr
U ij = 0.5 kW/m 2 K UCU,i = 0.83333 kW /m2 K UHU,j = 0.5 kW/m2 K
CHU = $80/kW-yr Ccu = $15/kW-yr 6.T mal!.P =1OK
12.3. HEAT EXCHANGER NETWORK SYNTHESIS 277

Stage 1 Stage 2
Cl

H2

Figure 12.2: Optimum configuration for the heat exchanger network problem
with the Chen approximation.

Problem Statistics
No. of continuous variables 40
No. of binary variables 12
No. of linear equalities 20
No. of convex inequalities 44
No. of known solutions 9

Global Solution
The global optimum configuration involves six heat exchangers and is shown
in Figure 12.2 (see also Adjiman et al. (1998d .

Objective function: $ 154,997/yr.

Continuous variables

k=1 k=2 k=3

THl,k 650.00 K 581.77 K 386.77 K
T H2 ,k 590.00 K 590.00 K 468.67 K
TC1,k 617.26 K 571.77 K 410.00 K
TC2,k 500.00 K 500.00 K 350.00 K

QHl,Cl,l = 682.30 kW QH2,Cl,1 = 0

QH1,Cl,2 = 0 QH2,Cl,2 = 2426.55 kW
QH1,C2,1 = 0 QH2,C2,1 = 0
QHl,C2,2 = 1950.00 kW QH2,C2,2 = 0
278 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

QCU,Hl = 167.70 kW QCU,H2 = 1973.45 kW

QHU,Cl = 491.15 kW
= QHU,C2 =0
=

~THl,Cl,l = 32.74 K ~TH2,Cl,1 = -

~TH1,cl,2 = 10.00 K ~TH2,Cl,2 = 18.23 K
~THl,cl,3 = - ~TH2,Cl,3 = 58.67 K
~THl,c2,1 = - ~TH2,C2,1 = -
~TH1,C2,2 =
= 81.77 K ~TH2,C2,2 =-
~TH1,C2,3 =
= 36.77 K ~TH2,C2,3 = -=-
~TCU,Hl = 66.77 K ~TCU,H2 = 148.67 K
~THU,Cl = 62.74 K ~THU,C2 =-
Binary variables

ZHl,Cl,l = 1 ZHl,C2,1 = 0 ZH2,Cl,1 = 0 ZH2,C2,1 = =0

ZHl,Cl,2 = 0 ZHl,C2,2 = 1 ZH2,Cl,2 = 1 ZH2,C2,2 =0
ZCU,Hl = 1 ZCU,H2 =1 ZHU,Cl =1
= ZHU,C2 = 0

12.4 Heat Exchanger Network Synthesis with Arith-

metic Mean Temperature Difference
12.4.1 Introduction
The general formulation for these test problem is taken from Zamora and
Grossmann (1998). It is similar to that of the Ohen approximation. The
greatest difference lies in the expression used for the area of the heat exchang-
ers, which is based on the arithmetic mean temperature difference rat her than
the Ohen approximation. In addition, the assumption of no stream splitting is
imposed by adding two new sets of linear constraints.

12.4.2 General Formulation

Objective function

min ~ CCUQCU,i + ~ CHUQHU,j

T,Q,T,% iEHP JECP
+ ~ ~ ~ CFijZijk + ~ CFi,CUZCU,i + ~ CFj,HUZHU,j
iEHP JECP kESI iEHP JECP
+ ~ ~ ~ Vi" CiiQijk
iEH P JECP kESI T (~T;ik+~T;ik+l)
+ " W
Ci,CUQcu,i
VCUi ( )
iEHP ~ ~TCU,i+To"t,i-Tin,CU
+ " W VHU "
Ci,HUQHU,j
.
JECP ~(~THU,i+Tin,HU-To"t,i)
12.4. HEAT EXCHANGER NETWORKS: ARITHMETIC MEAN 279

Constraints

L L Qijk + QCU,i = (Tin ,i - Tout,i) FCPi, Vi E HP

kESI JECP
L L Qijk + QHU,j = VjE CP
(Tout,j - Tin,j) Fcpj, V
kESliEHP
(Ti,k - Ti,k+1) FCPi - L Qijk = 0, V k E SI, Vi E HP
JECP

(Tj,k - Tj,k+1) Fcpj - L Qijk = 0, V k E SI, V j E CP

iEHP
Ti,!
Ti,l = Tin,i, 'Vi EE HP
1in,i, Vi
Tj,NS+1 Tin,j, V jE CP
Ti,k -1i,k+1 >
2: 0, V k E SI, Vi E HP
Tj,k - Tj,k+l >
2: 0, V k E SI, V jE CP
1i,NS+1 >2: Tout,i, Vi E HP
Tj,l < Tout,j, V jE CP
FCPi Ti,NS+1 - QCU,i FCPi Tout,i, Vi E HP
Fcpj Tj,l + QHU,j = VjE CP
Fcpj Tout,j, V
Qijk - Q
Q Qmax
ij Zijk < 0,
VkESI, ViEHP, VjECP
QCU,i - Qi:CUZCU,i <
~ 0, Vi E HP
QHU,j - Q7J,jZHU,j <
~ 0, Vj E CP
tlTijk - Ti,k f:::.Tr;ax,
+ Tj,k + tlTt;aX Zijk :S tlTt;ax,
VkESI, ViEHP, VjECP
tlTijk+1 - Ti,k+1 + T j ,k+1 + flTr;ax Zijk < tl.T!'!-ax
flT!'!-ax
tJ '
VkESI, ViEHP, VjECP
tl.TCU,i Ti,NS+1
flTcu,i - T i ,NS+1+ flT[~Jl}zcU,i <
tl.T[~Jl}zcU,i ~ tl.T[:G'l} - Tout,cu, Vi E HP
tlT[:G'lJ
flTHU,j + Tj,l + tl.TJf{jJZHU,j
tl.THU,j flTJf{jJZHU,j < flTJf{jJ + Tout,HU, V
tl.TJf{jJ VjE CP
L
L Zi,j,k ~< 1, Vi E HP, Vk E SI
JECP

L Zi,j,k < 1, V jE CP, Vk E SI

~
iEHP

where tl.T[jax = max {O, Tj,in -1i,in, Tj,in - Ti,out, Tj,out - Ti,in, Tj,out - Ti,out},
flT[jax =
with i E HP U {HU} and jE CP U {CU}; and
QiJax = min{(1in,i - Tout,i) FCPi, (Tout,j - Tin,j) Fcpj}, Vi E HPU{HU}, Vj E
280 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Figure 12.3: Superstructure for the heat exchanger network problem with the
arithmetic mean temperature difference - Test problem 1.

CPU{CU}.
Variable bounds

0:::; Qijk :::; Qr;ax,Vk E SI, Vi E HP, Vj E CP

0:::; QCU,i :::; Q7:t:f;, Vi E HP
0:::; QHU,j :::; Qliif,j, V j E CP
Ti,out :::; Tik :::; Ti,in,Vk E ST, Vi E HP
Tj,in :::; Tjk :::; Tj,out,Vk E ST, Vj E CP
6.Tmapp :::; 6.Tijk :::; 6.T[;aX, V k E ST, Vi E HP, V j E C P
6.Tmapp :::; 6.TCU,i :::; 6.T{'cu, Vi E HP
6.Tmapp :::; 6.THU,j :::; 6.TWf}j, V j E CP
Zijk, ZCU,i, ZHU,j E {O, I}, VkESI, ViEHP, VjECP

Variable definitions
The sets, variables and paramaters have the same meaning as for the ehen
approximation formulation.

12.4.3 Test Problem 1

This corresponds to example 4 of Zamora and Grossmann (1998). Two hot

streams and two cold streams are considered in a network with three temper-
ature intervals. The superstructure is shown in Figure 12.3.
12.4. HEAT EXCHANGER NETWORKS: ARlTHMETIC MEAN 281

Stage 1 Stage 2 Stage 3

HI

CI

Cl

CW

HZ

Figure 12.4: Optimum configuration for the heat exchanger network problem
with the arithmetic mean temperature difference - Test problem 1.

Data
Stream Tin (K) Taut (K) Fcp (kW/K)
H1 443 333 30
H2 423 303 15
Cl 293 408 20
C2 353 413 40
Steam 450 450
Water 293 313

CFij = $6250/yr CFi,CU = $6250/yr CFj,HU = $6250/yr

Cij = $83.26/yr Ci,CU = $83.26/yr Cj,HU = $99.91/yr
Uij = 0.8 kW/m2 K UCU,i = 0.8 kW/m2 K UHU,j = 1.2 kW/m 2 K
CHU = $80/kW-yr GCU = $20/kW-yr 6.Tmapp =lK

Problem Statistics
No. of continuous variables 52
No. of binary variables 16
No. of linear equalities 24
No. of convex inequalities 72
No. of known solutions 3

Global Solution
The global optimum configuration involves five heat exchangers and is shown
in Figure 12.4 .

Objective function: $74711/yr.

Continuous variables
282 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

k=1 k=2 k=3 k=4

THl,k 443.00 K 436.64 K 356.64 K 333.00 K
TH2,k 423.00 K 423.00 K 329.67 K 329.67 K
TCl,k 408.00 K 398.45 K 328.45 K 293.00 K
TC2,k 413.00 K 413.00 K 353.00 K 353.00 K

QHl,Cl,l = 190.93 kW QH2,Cl,1 = 0

QHl,Cl,2 = 0 QH2,Cl,2 = 1400.00 kW
QHl,Cl,3 = 709.07 kW QH2,Cl,3 =0
QHl,C2,1 = 0 QH2,C2,1 = 0
QHl,C2,2 = 2400.00 kW QH2,C2,2 =0
QHl,C2,3 = 0 QH2,C2,3 = 0

QCU,Hl = 0 QCU,H2 = 400 kW

QHU,Cl = 0 QHU,C2 = 0

~THl,Cl,l = 35.00 K ~TH2,Cl,1 = -

~THl,Cl,2 = 38.18 K ~TH2,Cl,2 = 24.55 K
~THl,Cl,3 = 28.18 K ~TH2,Cl,3 = 1.21 K
~THl,Cl,4 = 40.00 K ~TH2,Cl,4 = -
~THl,C2,1 = - ~TH2,C2,1 = -
~THl,C2,2 = 23.64 K ~TH2,C2,2 = -
~THl,C2,3 = 36.36 K l1TH2 ,c2,3 = -
~THl,C2,4 = - ~TH2,C2,4 =-

~TCU,Hl =- ~Tcu,H2 = 16.67 K

~THU,Cl =- ~THU,C2 =-

Binary variables

ZHl,Cl,l = 1 ZHl,C2,1 = 0 ZH2,Cl,1 = 0 ZH2,C2,1 = 0

ZHl,Cl,2 = 0 ZHl,C2,2 = 1 ZH2,Cl,2 = 1 ZH2,C2,2 = 0
ZHl,Cl,3 = 1 ZHl,C2,3 = 0 ZH2,Cl,3 =0 ZH2,C2,3 =0
ZCU,Hl = 0 ZCU,H2 = 1 ZHU,Cl = 0 ZHU,C2 = 0

12.4.4 Test Problem 2

This corresponds to example 5 of Zamora and Grossmann (1998). This time,
three hot streams and two cold strearns are considered over three temperature
intervals. The superstructure is shown in Figure 12.5.
 ARlTHMETIC MEAN
12.4. HEAT EXCHANGER NETWORKS: ARITHMETIC 283

Stage 1 Stage 2 Stage 3

Cl
Hl

~
iOs
~~I:P-
,

H2 : 'iii' J I~;
11"'1~ H2~:H-!n(:.w
, ~- C2 ' H2 '
I : C2 :

C2
,
,,
,

H3
H-
~i~
Figure 12.5: Superstructure for the heat exchanger network problem with the
arithmetic mean temperature difference - Test problem 2.

Data

Stream 1in (OC) Taut

Toot (OC) Fcp (kWrC)
(kWrC) h {kW
(kW /(m'2
/(m 2 C))
C
H1 159 77 2.285 0.10
H2 267 80 0.204 0.04
H3 343 90 0.538 0.50
Cl 26 127 0.933 0.01
C2 118 265 1.961 0.50
Steam 300 300 0.05
Water 20 60 0.20

The overall heat transfer coefficients are computed ba.sed

based on the film heat
transfer coefficients h, listed in the above table. Thus, Uij = (t + ~i)
~j) -1.

CFij =
= $7400/yr CFi,CU =
= $7400/yr CFj,HU = $7400/yr
Cij = $80/yr Ci,cu
Ci,CU = $80/yr Cj,HU = $80/yr
CHU = $110/kW-yr Ccu
CCU = $lO/kW-yr
$10/kW-yr ilT
6.Tmapp =
= 1C
284 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Stage 1 Stage :z Stage 3

_H_l______________~------------__------------~~.~~

_w~--------------~------------~------------~~w
CI

C2

Figure 12.6: Optimum configuration for the heat exchanger network problem
with the arithmetic mean temperature difference - Test problem 2.

Problem Statistics
No. of continuous variables 72
No. of binary variables 23
No. of linear equalities 30
No. of convex inequalities 99
No. of known solutions 7

Global Solution
The global optimum configuration involves five heat exchangers and is shown
in Figure 12.6. There exists another solution with the same objective function
but a different network structure .

Objective function: $82043/yr.

Continuous variables

k=l k=2 k=3 k=4

THl,k 159.00 K 159.64 K 159.64 K 159.00 K
TH2,k 267.00 K 267.00 K 267.00 K 267.00 K
T H3 ,k 343.00 K 265.15 K 265.15 K 90.00 K
TCl,k 127.00 K 127.00 K 127.00 K 26.00 K
TC2,k 139.36 K 118.00 K 118.00 K 118.00 K
12.5. PUMP NETWORK SYNTHESIS 285

QHl,Gl,l =0 QH2,Cl,1 =0 QH3,Cl,l =0

QHl,Cl,2 =0 QH2,Cl,2 =0 QH3,Cl,2 =0
QHl,Cl,3 = 0 QH2,Cl,3 = 0 QH3,Cl,3 = 94.23 kW
QHl,C2,1 = 0 QH2,C2,1 = 0 QH3,C2,1 = 41.88 kW
QHl,C2,2 = 0 QH2,C2,2 = 0 QH3,C2,2 =0
QHl,C2,3 = 0 QH2,C2,3 = 0 QH3,C2,3 =0

QCU,Hl = 187.37 kW QCU,H2 = 38.15 kW QCU,H3 =0

QHU,Cl = 0 QHU,C2 = 246.39 kW

THl,Cl,l = - TH2,Cl,1 =- TH3,Cl,1 = -

THl,Cl,2 = - TH2,Cl,2 =- TH3,Cl,2 = -
THl,Cl,3 = - TH2,Cl,3 =- TH3,Cl,3 = 138.15 K
THl,Cl,4 = - TH2,Cl,4 =- TH3,Cl,4 = 64 K
THl,C2,1 = - TH2,C2,1 =- TH3,C2,1 = 203.64 K
THl,C2,2 = - TH2,C2,2 =- TH3,C2,2 = 147.15 K
THl,C2,3 = - TH2,C2,3 =- TH3,C2,3 = -
THl,C2,4 = - TH2,C2,4 =- TH3,C2,4 = -

TCU,Hl = 99.00 K TCU,H2 = 207.00 K TCU,H3 =-

THU,Cl =- THU,C2 = 160.64 K
Binary variables

ZHl,Cl,l =0 ZH2,Cl,2 = 0 ZH3,Cl,1 =0

ZHl,Cl,2 = 0 ZH2,Cl,2 = 0 ZH3,Cl,2 = 0
ZHl,Cl,3 = 0 ZH2,Cl,3 =0 ZH3,Cl,3 = 1
ZHl,C2,1 = 0 ZH2,C2,1 =0 ZH3,C2,1 = 1
ZHl,C2,2 = 0 ZH2,C2,2 =0 ZH3,C2,2 = 0
ZHl,C2,3 =0 ZH2,C2,3= 0 ZH3,C2,3 = 0
ZCU,Hl =1 ZCU,H2 = 1 ZCU,H3 =0
ZHU,Cl = 0 ZHU,C2 = 1

12.5 Pump network synthesis

12.5.1 Introduction
In this example, taken from Westerlund et al. (1994), the aim is to identify
the least costly configuration of centrifugal pumps that achieves a pre-specified
pressure rise based on a given total flowrate. Figure 12.7 shows a schematic
representation of a three level pump network superstructure, where each level
corresponds to a different pump type. A major hurdle in solving this problem
stems from the nonconvex participation of integer variables which leads to the
presence of many local minima.
286 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Figure 12.7: Superstructure for the pump network synthesis problem

12.5.2 Test Problem 1

Formulation

The goal is to minimize the annualized network cost given three different pump
types.
Objective function

3
min ~)Ci + Ci'~)NpiNsiZi
z,v,w,P,t:t..p,Np,N 8,Z i=l
12.5. PUMP NETWORK SYNTHESIS 287

Constraints

PI - 19.9( ~)3 - 0.161O( ~ )2 VI + 0.000561( ~ )v~ 0

Wmax Wmax Wmax
P2 - 1.21( ~)3 - 0.0644( ~ )2V2 + 0.000564( ~ )v~ = 0
Wmax Wmax Wmax
P3 - 6.52( ~)3 - 0.1020( ~ )2V3 + 0.000232( ~ )v~ 0
Wmax Wmax Wmax
API - 629( ~)2 - 0.696( ~ )VI + 0.0116v~ 0
Wmax Wmax
AP2 - 215( ~)2 - 2.950( ~ )V2 + 0.115v~ = 0
Wmax Wmax
AP3 - 361( ~)2 - 0.530( ~ )V3 + 0.00946v~ 0
Wmax Wmax
Xl + X2 + X3 = 1
Vi N Pi - Xi Viot = 0, i = 1,2,3
APtotZi - APiN Si = 0, i = 1,2,3
Wi - WmaxZi < 0, i = 1,2,3
~ - Pmax,iZi < 0, i = 1,2,3
APi - APtotZi < 0, i = 1,2,3
Vi - ViotZi < 0, i = 1,2,3
Xi - Zi < 0, i = 1,2,3
NSi - 3Zi < 0, i = 1,2,3
NPi - 3zi < 0, i = 1,2,3

Variable bounds

o ~ Xi ~ 1, i = 1,2,3
o ~ Vi ~ Viot, i = 1,2,3
o ~ Wi ~ wmax , i = 1,2,3
o ~ Pi ~ prax , i = 1,2,3
o ~ APi ~ APtot , i = 1,2,3
NPi E {O, 1,2, 3}, i = 1,2,3
N Si E {O, 1, 2, 3}, i = 1,2,3
Zi E {O, I}, i = 1,2,3

Variable definitions
The parameters are Viot, the total volumetrie Howrate through the network,
APtot , the prescribed pressure rise, Wmax , the maximum rotation speed of a
pump, and prax , i = 1, ... ,3, the maximum power requirement for a pump
on level i.
288 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Figure 12.8: Global optimal configuration for pump network synthesis problem

The binary variables Zi, i = 1, ... ,3, denote the existence of level i. The
integer variables NPi, i = 1, ... ,3, denote the number of parallellines at level
i. The integer variables N Si, i = 1, ... ,3, denote the number of pumps in
series at level i.
The continuous variables are the fraction of total fiow going to level i, Xi,
the fiowrate on each line at level i, Vi, the rotation speed of all pumps on level
i, Wi, the power requirements at level i, I{ and the press ure rise at level i, D.Pi.

Data
vtot = 350 m 3Jh = 400 kPa W max = 2950 rpm
p:nax = 80kW = 25kW pmax = 35 kW
1 3

Pump 1 Pump 2 Pump 3

Fixed cost (FIM) 38,900 15,300 20,100
Ci (FIM) 6,329.03 2,489.31 3,270.27
Ci' (FIM/kW) 1,800 1,800 1,800

The capital costs are annualized by multiplying the fixed pump costs by a
factor of 0.1627 which corresponds to a 10% interest rate and a 10 year life.
The operating costs are based on 6000 hours per year and an electricity cost
of 0.3 FIM/kWh. Here FIM denotes Finmark.

Problem Statistics
No. of continuous variables 15
No. of binary variables 3
No. of integer variables 6
No. of linear equalities 1
No. of nonlinear equalities 12
No. of known solutions 37

Global Solution
Figure 12.8 shows the lowest cost network (see Adjiman et al. (1998d)), which
involves two levels and four pumps.
12.6. TRlM
TRIM LaSS MINIMIZATION 289

Objective function: 128894 FIM/yr.

Continuous variables

Xl =
= 0.914 X2 == 0.086 X3 = 0.000
VI = 160 m 3 /h V2 == 30 m 3 /h V3 =0
= 2855.10 rpm
WI = W2 = 2950.00 rpm W3 = 0
PI = 28.27 kW P2 == 2.63 kW
P2 P3 = 0 kW
P3
PI = 400 kPa P2 = 200 kPa P3 = 0

Binary variables

Zl =1 Z2 =1 Z3 =0
Integer variables

NPI = 2 NP2 =1 NP3 =0

NS I = 1 NS 2 = 2 NS 3 =0

12.6 Trim Loss Minimization

12.6.1 Introduction
Trim loss minimization problems arise in the paper cutting industry and have
been studied in the context of integer nonconvex optimization by Westerlund
et al. (1994) and Harjunkoski (1997). These problems involve only integer and
binary variables, which appear in linear and bilinear functions. They have a
very large number of solutions, many of which are redundant. The main task
is to cut out some paper products of different sizes from a large raw paper roll,
in order to meet a customer's order. Each product paper roll is characterized
by its width b. All product rolls are assumed to be of equal length. The
raw paper roll has a width B max . In general, it is not possible to cut out
an entire order without throwing away some of the raw paper. The optimum
cutting scheme minimizes the waste paper or trim loss. In order to identify
the best overall scheme, a maximum number of different cutting patterns P
is postulated, where a pattern is defined by the position of the knives. Each
cutting pattern may have to be repeated several times in the overall scheme
to meet the demand. There are N different product sizes and ni rolls of size
bi must be cut. The existence of each pattern is denoted by a binary variable
Zj, j = 1, ... , P. The number of repeats of pattern j is denoted by the integer

n' n
variable mj. The number of products of size i in pattern j is given by the
integer variable rij. A sampIe
sampie cut is shown in Figure 12.9 for

~ n~ ( ~( D
0 2
N 3, m and r = ( 4 1
o 0
290 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

t
I
A

b2
!
2 2
t 1
1
3
1.. j3
2 2
1

3 1 3
3

2
~ >

Pattern 1 Pattern 1 Pattern 2 Pattern 3

Trim loss

Figure 12.9: Trim loss minimization problem.

Some additional constraints are imposed on the cutting patterns. For in-
stance, each pattern must have a minimum total width of B max - . The
number of knives is limited to Nkmax .

12.6.2 General Formulation

Objective function

p
min Lmj +0.1 j Yj
m,y,r
j=l

Constraints
12.6. TRIM LaSS MINIMIZATION 291

Lmj Tij ;:::

> ni, i = 1, ... ,N
j

(Bma:z: - .6.)
(Bmax Yj - L
tl.)Yj bi Tij < O,j=l, ... ,P
Lbi

~ Bmax Yj ::;
L bi Tij - Bma:z: < 0, j = 1,
l, ... ,P

Yj - LTij < 0, j = 1, ... ,P

::;

~ Nkmax Yj ::;
L Tij - Nkma:z: < 0, j = 1, ... ,P
< 0, j = 1,
Yj -mj l, ... ,P
<
mj -MjYj :5 0, j = 1, ... ,P

Lmj
j
>
~
max
UN~maz'
E~ 1rE~nil}
~maz
Yk+l - Yk ::;
Yk+l- < 0, k = 1, ... ,P - 1
< 0, k
m k :5
mk+1- mk = 1, ... ,P - 1

Variable bounds

Yj E {O,l},j=l, ... ,P
mj
mj E [O,Mj] nN,j = 1, ... ,P

Tij E [O,Nkmax ]nN,i=l, ... ,N, j=l, ... ,P

ma:z:]nN,i=l,

The reported global solutions for the following four case studies are taken
from Adjiman et al. (1998d).

12.6.3 Test Problem 1

FormulatioD
Formulation

Objective function

min
m,y,r
ml + m2 + m3 + m4 + O.lYl + 0.2Y2 + 0.3Y3 + 0.4Y4

Constraints
292 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

ml r U + m2 r 12 + m3 r 13 + m4 r 14 > 15
m2 r 21 + m2 r 22 + m3r23 + m4r24 > 28
m3 r 31 + m2 r 32 + m3 r 33 + m4 r 34 > 21
m4 r 41 + m2 r 42 + m3r43 + m4r44 > 30
1750Yl - 290ru - 315r21 - 350r31 - 455r41 < 0
1750Y2 - 290r12 - 315r22 - 350r32 - 455r42 < 0
1750Y3 - 290r13 - 315r23 - 350r33 - 455r43 < 0
1750Y4 - 290r14 - 315r24 - 350r34 - 455r44 < 0
290ru + 315r21 + 350r31 + 455r41 - 1850Yl < 0
290r12 + 315r22 + 350r32 + 455r42 - 1850Y2 < 0
290r13 + 315r23 + 350r33 + 455r43 - 1850Y3 < 0
290r14 + 315r24 + 350r34 + 455r44 - 1850Y4 < 0
Yl - ru - r21 - r31 - r41 ~ 0
Y2 - r12 - r22 - r32 - r42 < 0
Y3 - r13 - r23 - r33 - r 43 ~ 0
Y4 - r14 - r24 - r34 - r44 < 0
rU + r21 + r31 + r41 - 5 Yl ~ 0
r12 + r22 + r32 + r 42 - 5 Y2< 0
r13 + r23 + r33 + r43 - 5 Y3 < 0
r14 + r24 + r34 + r44 - 5 Y4 < 0
Yl - ml < 0
Y2 -m2 ~ 0
Y3 -m3 < 0
Y4 -m4 < 0
12.6. TRIM LOSS MINIMIZATION 293

ml - 30Yl < 0
m2 - 3Oy2 :5 0
m3 - 30Y3 < 0
m~ - 30Y4 :5 0
ml +m2 +m3 +m4 ~ 19
YI-Y2 ~ 0
Y2 -Y3 ~ 0
Y3 -Y4 > 0
ml-m2 > 0
m2- m 3 > 0
m3- m 4 > 0

Variable bounds

(Yl,Y2,Y3,Y4) E {0,1}4
(mb m2,m3,m4) E [0,30]4 nN4
rij E {0,1,2,3,4,5}, i = 1, ... ,4,j = 1, ... ,4.

Data
N = 4 P= 4 B max = 1850 mm A = 100 mm

n = (15,28,21, 30)T

b = (290,315,350, 455)T

M = (30,30,30, 30)T

Problem Statistics
No. of binary variables 4
No. of integer variables 20
No. of convex inequalities 31
No. of nonconvex inequalities 4
No. of known solutions more than 40
294 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Global Solution
There exist several degenerate solutions to this problem. Only one is presented
here.

Objective function: 19.6.

Binary variables: y = (1,1,1, of.

Integer variables

m = (9,7,3,O)T

r = (~ ! o2 0)
1 0
0
2 1 2 0

12.6.4 Test Problem 2

Formulation

Objective function

min ml
m,y,r
+ m2 + m3 + m4 + O.lYl + O.2Y2 + O.3Y3 + 0.4Y4

Constraints

ml Tn + m2 T12 + m3T13 + m4T14 > 9

m2 T21 + m2 T22 + m3 T23 + m4 T24 > 7
m3 T31 + m2 T32 + m3 T33 + m4 T34 > 12
m4 T41 + m2 T42 + m3 T43 + m4T44 > 11
1700Yl - 330Tn - 360T21 - 385T31 - 415T41 < 0
1700Y2 - 330T12 - 360T22 - 385T32 - 415T42 < 0
1700Y3 - 330T13 - 360T23 - 385T33 - 415T43 < 0
1700Y4 - 330T14 - 360T24 - 385T34 - 415T44 < 0
330T1l + 360T21 + 385T31 + 415T41 - 1900Yl < 0
330T12 + 360T22 + 385T32 + 415T42 - 1900Y2 < 0
330r13 + 360T23 + 385r33 + 415T43 - 1900Y3 < 0
330T14 + 360T24 + 385T34 + 415T44 - 1900Y4 < 0
12.6. TRIM LOSS MINIMIZATION 295

Yl - Tn - T21 - T31 - T41 < 0

Y2 - T12 - <
T22 - T32 - T42 0
Y3 - T13 - T23 - T33 - T 43 < 0
Y4 - T14 - T24,- T34 - T44 < 0
Tn + T21 + T31 + T41 - 5 Yl < 0
T12 + T22 + T32 + T42 - 5 Y2 < 0
T13 + T23 + T33 + T43 - 5 Y3 < 0
T14 + T24 + T34 + T44 - 5 Y4 < 0
Yl-ml < 0
Y2 -m2 < 0
Y3 -m3 :::; 0
Y4- m 4 < 0
ml - 15Yl < 0
m2 -12Y2 < 0
m3 - 9Y3 < 0
m4 - 6Y4 < 0
ml +m2 +m3 +m4 > 8
Yl -Y2 > 0
Y2 - Y3 > 0
Y3 -Y4 > 0
ml-m2 > 0
m2- m 3 > 0
m3- m 4 > 0

Variable bounds

(Yb Y2, Y3, Y4) E {0,1}4

ml E {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}
m2 E {0,1,2,3,4,5,6,7,8,9,10,11,12}
m3 E {0,1,2,3,4,5,6,7,8,9}
m4 E {0,1,2,3,4,5,6}
Tij E {O, 1,2,3,4, 5}, i = 1, ... ,4, j = 1, . .. ,4.

Data

N = 4 P = 4 B max = 1900 mm = 200 mm

296 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

n = (9,7,12, l1)T

b = (330,360,385, 415)T

M = (15,12,9, 6)T

Problem Statistics
No. of binary variables 4
No. of integer variables 20
No. of convex inequalities 31
No. of nonconvex inequalities 4
No. of known solutions more than 60

Global Solution
There exist several degenerate solutions to this problem. Only one is presented
here.

Objective function: 8.6.

Binary variables: y = (1,0,0, O)T.

Integer variables

m = (11,0,0, O)T

12.6.5 Test Problem 3

Formulation

Objective function

min
m,y,r
ml + m2 + m3 + m4 + m5 + O.1Yl + 0.2Y2 + 0.3Y3 + 0.4Y4 + O.5Y5
12.6. TRIM LOSS MINIMIZATION 297

Constraints
mlrU + m2 r I2 + m3 r I3 + m4 r I4 + m5 r I5 > 12
mI r 2I + m2 r 22 + m3 r 23 + m4 r 24 + m5 r 25 > 6
mIr3I + m2r32 + m3 r 33 + m4 r 34 + m5r35 > 15
mIr4I + m2 r 42 + m3 r 43 + m4 r 44 + m5 r 45 > 6
mI r 5I + m2 r 52 + m3 r 53 + m4 r 54 + m5 r 55 > 9
1800YI - 330ru - 360r2I - 370r3I - 415r4I - 435r5I < 0
1800Y2 - 330rI2 - 360r22 - 370r32 - 415r42 - 435r52 < 0
1800Y3 - 330rI3 - 360r23 - 370r33 - 415r43 - 435r53 < 0
1800Y4 - 330rI4 - 360r24 - 370r34 - 415r44 - 435r54 < 0
1800Y5 - 330rI5 - 360r25 - 370r35 - 415r45 - 435r55 < 0
330rn + 360r2I + 370r3I + 415r4I + 435r5I - 2000YI < 0
330rI2 + 360r22 + 370r32 + 415r42 + 435r52 - 2000Y2 < 0
330rI3 + 360r23 + 370r33 + 415r43 + 435r53 - 2000Y3 < 0
330rI4 + 360r24 + 370r34 + 415r44 + 435r54 - 2000Y4 < 0
330rI5 + 360r25 + 370r35 + 415r45 + 435r55 - 2000Y5 < 0
YI - ru - r2I - r3I - r4I - r5I < 0
Y2 - rI2 - r22 - r32 - r 42 - r52 < 0
Y3 - rI3 - r23 - r33 - r43 - r53 < 0
Y4 - rI4 - r24 - r34 - r44 - r54 < 0
Y5 - rI5 - r25 - r35 - r45 - r55 < 0
rU + r2I + r3I + r4I + r5I - 5 YI < 0
rI2 + r22 + r32 + T42 + r52 - 5 Y2 ::; 0
rI3 + r23 + r33 + r 43 + r53 - 5 Y3 < 0
rI4 + r24 + r34 + r44 + r54 - 5 Y4 < 0
rI5 + r25 + r35 + r45 + r55 - 5 Y5 < 0
YI -mI < 0
Y2 -m2 < 0
Y3 -m3 < 0
Y4 -m4 < 0
Y5 -m5 < 0
ml - 15YI < 0
m2 -12Y2 < 0
m3 - 9Y3 < 0
m4 - 6Y4 < 0
m5 - 6Y5 < 0
298 GHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

ml +m2 +m3 +m4 +m5 > 10

Yl - Y2 > 0
Y2 - Y3 > 0
Y3 -Y4 > 0
Y4 - Y5 > 0
ml-m2 > 0
m2- m 3 > 0
m3- m 4 > 0
m4- m 5 > 0

Variable bounds

(Yl,Y2,Y3,Y4,Y5) E {0,1}5
ml {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}
E
m2 E {0,1,2,3,4,5,6,7,8,9,10,11,12}
m3 E {0,1,2,3,4,5,6,7,8,9}

m4 E {0,1,2,3,4,5,6}
m5 E {0,1,2,3,4,5,6}
rij E {0,1,2,3,4,5}, i = 1, ... ,5,j = 1, ... ,5.

Data
N =5 P =5 B max = 2000 mm = 200 mm

n = (12,6,15,6, 9f

b= (330,360, 370,415,435)T

M = (15,12,9,6, 6)T

Problem Statistics
No. of binary variables 5
No. of integer variables 30
No. of convex inequalities 39
No. of nonconvex inequalities 5
No. of known solutions more than 40
12.6. TRlM LaSS MINIMIZATION 299

Global Solution
There exist several degenerate solutions to this problem. Only one is presented
here.

Objective function: 10.3.

Binary variables: y = (1,0,0,0, O)T.

Integer variables

m = (15,0,0,0,0)T

12.6.6 Test Problem 4

Formulation

Objective function

mm ml +m2 +m3 +m4 +m5 +m6

m,y,r
+ O.IYI + 0.2Y2 + 0.3Y3 + 0.4Y4 + 0.5Y5 + 0.6Y6

Constraints

mI T l1 + m2 T12 + m3 T13 + m4 T14 + m5T15 + m6T16 > 8

mI T 2I + m2 T22 + m3 T23 + m4T24 + m5T25 + m6 T26 > 16
mI T 3I + m2 T 32 + m3 T 33 + m4 T34 + m5T35 + m6T36 > 12
mI T4I + m2 T42 + m3 T43 + m4 T44 + m5T45 + m6T46 > 7
mI T5I + m2 T52 + m3 T53 + m4 T 54 + m5T55 + m6T56 > 14
mI T 6I + m2 T62 + m3 T 63 + m4 T64 + m5T65 + m6T66 > 16
<

2100YI - 330Tl1 - 360T2I - 380T3I - 430T4I - 490T5I - 530T61

<

2100Y2 - 330T12 - 360T22 - 380T32 - 430T42 - 490T52 - 530T62

<

2100Y3 - 330T13 - 360T23 - 380T33 - 430T43 - 490T53 - 530T63
300 CHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

2100Y4 - 330r14 - 360r24 - 380r34 - 430r44 - 490r54 - 530r64 < 0

2100Y5 - 330r15 - 360r25 - 380r35 - 430r45 - 490r55 - 530r65 < 0
2100Y6 - 330r16 - 360r26 - 380r36 - 430r46 - 490r56 - 530r66 < 0
330ru + 360r21 + 380r31 + 430r41 + 490r51 + 530r61 - 2200Yl < 0
330r12 + 360r22 + 380r32 + 430r42 + 490r52 + 530r62 - 2200Y2 < 0
330r13 + 360r23 + 380r33 + 430r43 + 490r53 + 530r63 - 2200Y3 < 0
330r14 + 360r24 + 380r34 + 430r44 + 490r54 + 530r64 - 2200Y4 < 0
330r15 + 360r25 + 380r35 + 430r45 + 490r55 + 530r65 - 2200Y5 < 0
330r16 + 360r26 + 380r36 + 430r46 + 490r56 + 530r66 - 2200Y6 < 0
Yl - ru - r21 - r31 - r41 - r51 - r61 < 0
Y2 - r12 - r22 - r32 - r42 - r52 - r62 < 0
Y3 - r13 - r23 - r33 - r43 - r53 - r63 < 0
Y4 - r14 - r24 - r34 - r44 - rM - r64 < 0
Y5 - r15 - r25 - r35 - r45 - r55 - r65 < 0
Y6 - r16 - r26 - r36 - r46 - r56 - r66 < 0
ru + r21 + r31 + r41 + r51 + r61 - 5 Yl < 0
r12 + r22 + r32 + r 42 + r52 + r62 - 5 Y2 < 0
r13 + r23 + r33 + r 43 + r53 + r63 - 5 Y3 < 0
r14 + r24 + r34 + r44 + rM + r64 - 5 Y4 < 0
r15 + r25 + r35 + r 45 + r55 + r65 - 5 Y5 < 0
r16 + r26 + ra6 + r46 + r56 + r66 - 5 Y6 < 0
Yl -mI < 0
Y2 -m2 < 0
Ya -ma < 0
Y4 -m4 < 0
Y5 -m5 < 0
Y6 -m6 < 0
ml - 15Yl < 0
m2 -12Y2 < 0
m3 - 8Y3 < 0
m4 -7Y4 < 0
m5 - 4Y5 < 0
m6 - 2Y6 < 0
ml + m2 + m3 + m4 + m5 + m6 > 16
12.6. TRlM LaSS MINIMIZATION 301

Yl -Y2 ~ 0
Y2 -Y3 ~ 0
Y3 - Y4 > 0
Y4 -Y5 > 0
Y5 - Y6 ~ 0
ml-m2 > 0
m2- m 3 > 0
m3- m 4 ~ 0
m4- m 5 > 0
m5- m 6 ~ 0

Variable bounds

(Yl,Y2,Y3,Y4,Y5,Y6) E {0,1}6
ml E {0,1,2,3,4,5,6,7,8,9, 10, 11, 12, 13, 14, 15}
m2 E {0,1,2,3,4,5,6,7,8,9,10,11,12}
m3 E {0,1,2,3,4,5,6,7,8}
m4 E {0,1,2,3,4,5,6,7}
m5 E {0,1,2,3,4}
m6 E {O,1,2}
rij E {O, 1, 2, 3, 4, 5}, i = 1, ... ,6, j = 1, ... ,6.

Data
N=6 P=6 B max =2200mm A=100mm

n = (8,16,15,7,14, 16)T

b= (330, 360, 380, 430, 490, 530)T

M = (15, 12,8,7,4, 2)T

Problem Statistics
No. of binary variables 6
No. of integer variables 42
No. of convex inequalities 47
No. of nonconvex inequalities 6
No. of known solutions more than 25
302 GHAPTER 12. MIXED-INTEGER NONLINEAR PROBLEMS

Global Solution
There exist several degenerate solutions to this problem. Only one is presented
here.

Objective function: 15.3.

Binary variables: y = (1,1,0,0,0, O)T.

Integer variables

m = (8,7,0,0,0,0)T

1
2

2
r=
1
2
1

Chapter 13

Combinatorial Optimization
Problems

13.1 Modeling with Integer Programming

Combinatorial optimization problems possess a discrete special structure, such
that it is very diflicult to develop general purpose test problems, as weIl as
general purpose software for solving them. For the exact solution of these
problems, usually an equivalent integer programming formulation is provided
to an IP solver, that uses branch and bound to solve it. For a sub optimal so-
lution, many heuristic procedures have been refined over the years, and there
exist procedures designed to provide sub optimal solutions to general combina-
torial optimization problems, given that the problem has been put into some
prespecified format.
The general discrete (or combinatorial) optimization problem has the form:

min f(x)
XEFCD

where F is a set of constraints, typically in the form of inequalities, and D is

the set of n-dimensional integral vectors zn. Discrete optimization problems
are classified according to the form that f, Fand D have. For example the
o- 1 linear programming problem can be stated as
min cT x
s.t. Ax :S b
x E B n := {O, l}n

The set of n-dimensional binary vectors Rn is used frequently to model prob-

lems where the decision variables x are constrained to take true or false values.
An immediate observation for the above 0 - 1 linear program, is that the feasi-
ble set is finite, and one may conclude that the solution to the problem can be
303

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
304 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

found trivially by examining all solutions. From the algorithmie point of view
however, this is an impossibility, sinee the number of feasible solutions usually
grows exponentially with respeet to the problem data (A, c and b).
Using the simple faet that z E {O, I} {::} z + w = 1, z ;::: 0, W ;::: 0, zw =
o it is clear that integer eonstraints are equivalent to eontinuous noneonvex
(eomplementarity) eonstraints Pardalos (1996).
Many problems arising from diverse areas ean be formulated as integer
programming problems. One ean argue that combinatorial optimization and
integer programming are synonymous terms (Du and Pardalos (1998)). This
is beeause the majority (if not all) of the eombinatorial optimization problems
are integer programming problems, usually involving binary variables.

13.1.1 Test problems in the Internet

In reeent years many eollections of test problems are available at the internet
(Beasley (1996)). For example, the OR-Library, is a eolleetion of several eombi-
natorial optimization test problems (including integer programming, sehedul-
ing problems, set eovering problems, and graph planarization). Information
on the OR-Library ean be obtained by email at

o.library~ic.ac.uk.

or at the web address:

http://mscmga.ms.ic.ac.uk/.

13.2 Quadratic Integer Programming

Many eombinatorial optimization problems (including maximum clique and
graph eoloring) ean be formulated as quadratic integer programming problems.
The general 0 - 1 quadratic programming is defined below:

1
min c T X + -xTQx
x 2
s.t. Ax ::; b
xEB n

where Q in an rational n x n symmetrie matrix. When only the eonstraint

x E B n is present, we have a pure (or uneonstrained) 0 - 1 integer program.
It is clear that any bivalent problem is equivalent to the 0 - 1 problem by a
simple linear transformation.

13.2.1 Quadratic 0-1 Test problems

A test problem generator for eonstrained 0 - 1 quadratic problems has been
proposed by Pardalos (1991) (see also Pardalos (1987)). For any selected fea-
13.2. QUADRATIC INTEGER PROGRAMMING 305

sible 0 - 1 point of a given domain, the method generates quadratic functions

whose global minimum over the given domain occurs at the selected point.
QIPI Test Problem - nonlinear constraints:

1
min cTx+ -xTQx
x 2
s.t. -1 ~ + X3X4 1
XIX2 :s
-3 ~ Xl + x2 + x3 + x4 ~ 2
xE {-1,1}4

where c = (6,8,4, -2) and

Q~ [1
2 0
1 -1 2
2
0
-1
2 Jl 1
The global solution is the point x* = (-1, -1, -1, 1) with f(x*) = -20.
In Pardalos (1991) a standardized random test problem generator is also
given for unconstrained 0 - 1 programming.
QIP2 Test Problem - unconstrained:

min xTQx
x
s.t. xE {O, l}IO

and

-1 -2 2 8 -5 1 -4 0 0 8
-2 2 0 -5 4 -4 -4 -5 0 -5
2 0 2 -3 7 0 -3 7 5 0
8 -5 -3 -1 -3 -1 7 1 7 2
-5 4 7 -3 1 0 -4 2 4 -2
Q= 1 -4 0 -1 0 1 9 5 2 0
-4 -4 -3 7 -4 9 3 1 2 0
0 -5 7 1 2 5 1 0 -3 -2
0 0 5 7 4 2 2 -3 2 3
8 -5 0 2 -2 0 0 -2 3 3
The global solution is the point x* = (1,1,0,0,1,0,1,1,0,0) with f(x*) =
-29.
At the present time there exist efficient algorithms for solving pure 0 - 1
quadratic integer problems. QOlSUBS solves unconstrained quadratic 0 -
1 problems both for dense and sparse matrices, including concave quadratic
minimization problems with box constraints. Q01SUBS is available at the
NEOS Server which can be accessed via the web:
306 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

http://www.mcs.anl.gov/home/otc/Server.
The code QOlSUBS can be used to provide large size test instances of
quadratic 0 - 1 problems of certain difficulty by controlling the diagonal dom-
inance of the input data matrix.

13.3 Satisfiability Problems

Many propositional calculus problems can be formulated as integer programs
(Cavalier et al. (1990); Du et al. (1997); Williams (1987. Consider a boolean
variable li which can take the values of either True of False, and a binary
variable Xi where

1 if li is True,
{
Xi = 0 if li is False.

Then the logical expression hl\l2 is equivalent to the algebraic expression X1X2,
and h V l2 equivalent to Xl + X2 - X1X2. Furthermore, Z; the negation of the
boolean variable li can be represented as 1 - Xi. Applying the just mentioned
rekttionships recursively one can transform any propositional calculus problem
into an equivalent 0 - 1 polynomial programming problem. Furthermore, any
product of binary variables can be replaced by a single binary variable, and
two additional constraints. Specifically, the product niEl Xi can be replaced
by the binary variable y E {O, 1} and the two linear 0 - 1 equations:

- LXi +Y + 1 ~ 1
1 1

iEl

and

LXi - yl/I ~ 0,
iEl

where III stands fr the cardinality of the set I. The SatisjiabiLity problem
(SAT) is probably the most significant problem in propositionallogic. It has
many applications in the areas of artificial intelligence, mathematical logic,
computer vision, and it was the first problem shown to be NP-complete Cook
(1971). Assume that we have boolean variables Li, i = 1, ... ,m, and define a
dause Cj as

where Ij, IT are index sets of the unnegated and negated boolean variables in
clause j respectively. In SAT we are given a set n of clauses, and the problem
13.3. SATISFIABILITY PROBLEMS 307

is to find whether or not there is a truth assignment of the variables such that
all clauses are satisfied (Le. have a truth value). SAT can be transformed
into an equivalent integer programming feasibility problem as folIows. A truth
assignment of the boolean variables Li satisfies
n
1\ Cj
j=l

if and only if the binary variables Xi as defined in (13.3) satisfy the following
system of inequalities
LXi-Lxi~l-IIj-l, i=l, ... ,n.
iE1t iElj-
Resolving feasibility for the above system of integral inequalities, will solve the
SAT problem. A related problem which is equally important, is the weighted
Maximum SatisfiabiLity problem (MAX-SAT), where we are given a set of n
clauses with weights Wj, and we have to find a truth assignment of the boolean
variables such that the sum of the weights of the satisfied clauses is maximized.
It is clear that SAT is a special case of MAX-SAT. Setting Wj = 1, Vj if the
solution of the MAX-SAT is n then the set of clauses is satisfiable with the
truth assignment indicated by the solution. MAX-SAT also admits a integer
programming formulation. Define m binary variables Xi as in (13.3), and
also define a continuous variable Yj = 1 if clause Cj is satisfied and Yj = 0
otherwise. Then the MAX-SAT is equivalent to the following mixed linear-
integer programming formulation:
m
max LWiYi
i=l

subject to: LXi +L (1- Xi) ~ Yj, j = 1, ... ,n,

iE1t iElj-
Xi E {O, I}, i = 1, ... ,m,
o ::; Yi ::; 1, i = 1, ... ,n,
where It and IT denotes the set of boolean variables appearing unnegated
(resp. negated) in clause Cj'

13.3.1 SAT Test Problems

During the second DIMACS Implementation Challenge a set of instances were
chosen to provide the satisfiability benchmarks. These problems can be found
in the web address:
http://dimacs.rutgers.du/Challenges/index.html
The most complete information on test problems and algorithms for all
versions of the satisfiability problem can be found in the book by Du et al.
(1997).
308 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

13.4 The Traveling Salesman Problem

Consider a complete graph G:= (V,E), where V = {V1,V2, ... ,vn } is the set
of vertices and E := {(Vi, Vj) lVi, Vj E V, i, j = 1,2 ... , n} is the set of edges.
A simple cycle of the graph G is a sequence of vertices (ViI' Vi2' ... , Vik' ViI)'
such that the vertices VipVi2' ... ,Vik are distinct, while a Hamiltonian cycle,
or a tour of G is a simple cycle wich contains every vertex of V. Assume that
each node in the graph represents a city, while R nxn 3 D := (d ij ) contains the
Euclidean distances between any two nodes i and j. The traveling salesman
problem (TSP) is to find a tour that visits every city once with the minimum
total distance. Given that 8 n is the set of all permutations of the integers
{1, 2, ... , n} the TSP formulated as a combinatorial optimization problem is
n-1
min Ld p(i)p(i+1) + dp(n)p(l)
i=l
s.t. pE IIN.

The TSP is a weIl known NP-Complete combinatorial optimization problem

which has been studied extensively (Lawler et al. (1995)). Despite its com-
pactness, little can be done in terms of solving the TSP by using the above
formulation however. An integer programming formulation of the TSP is the
following
n
min L dijXij

n
s.t. LXij = 1, j = 1,2, ... ,n,

n
LXij=l, i=1,2, ... ,n,
j=l

L Xij :5 181 - 1, V 8 c V,8 :I V,

i,jES
Xij E {0,1}, i,j = 1,2, ... ,n,

where 8 represents a subtour or a disjoint cycle. The decision variables Xij in

the above formulation take the following values

x.. = {1 if node j immediately follows i in the tour,

ZJ 0 otherwise.

The first set of constraints are called the assignment constraints and ensure
that in the tour each node is entered and left only once. The assignment
constraints are used frequently in formulating particularly location problems.
The third set of constraints eliminates any subtours in the solution. The above
13.5. ASSIGNMENT PROBLEMS 309

formulation of the TSP is considered very large in terms of the number of

constraints, mainly because the number of the subtour elimination constraints
is nearly 2n . In a more compactformulation one replaces the last constraint
with

Ui - Uj + nXij :$ n - 1, 1:$ i, j :$ n and i i- j,

where the Ui are real variables unrestricted in sign.

13.4.1 TSP test Problems

There is a large collection of TSP test problem instances available in the in-
ternet. TSPLIB is such a collection of TSP instances. The web address of
TSPLIB is:

http://www.iwr.uni-heidelberg.de/iwr/comopt/soft/TSPLIB95/TSPLIB.html.

13.5 Assignment Problems

There is a large number of assignment problems arising mainly from location
theory, that can be formulated as integer programs. Probably the simplest of
all assignment problems is the Linear Assignment Problem (LAP). Consider
that we are given n facilities and locations, and a matrix R nxn 3 C := (Cij)
where Cij represents the cost of placing facility i to location j. Our objective
is to find the optimal assignment of facilities to locations such that the total
placement cost is minimized. Using the decision variables Xij E {O, 1} as

X .. = { 1 if facility i is assigned to location j,

'3 0 otherwise,

we can formulate the LAP as an integer 0 - 1 program

n

min L CijXij
i,j=l
n
s.t. LXij=l, j=1,2, ... ,n,
i=l
n
LXij = 1, i = 1,2, ... ,n,
j=l
Xij E {O,l}, i,j = 1,2, ... ,no

In contrast with the TSP the LAP can be solved efficiently in O(n 3 ) time.
Consider now that we have the same set of facilities to place, and we are given
a flow matrix F := (fij) where fij is the flow of material or information between
facilities i and j, and a distance matrix D := (dij ) where dij is the distance
310 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

between facilities i and j. The Quadratic Assignment Problem (QAP) is to find

the assignment of facilities to locations, such that the total cost associated with
the distance and flow between any two locations, plus the cost of placement
of facilities to locations is minimized. Formulated as an integer 0 - 1 program
the QAP can be stated as follows

n n n n n
min L L L L fijdklXikXjl + L CijXij
i=1 j=1 k=1 1=1 i,j=1
n
s.t. L Xij = 1, j = 1,2, ... , n,
i=1
n
LXij = 1, i = 1,2, ... ,n,
j=1
Xij E {0,1}, i,j = 1,2, ... ,no

Although extensive research has been done for more than three decades, the
QAP, in contrast with its linear counterpart the linear assignment problem
(LAP), remains one of the hardest optimization problems and no exact algo-
rithm can solve problems of size n > 20, in reasonable computational time
(Pardalos and Wolkowicz (1994)). Note that the TSP is a special case of the
QAP, where the distance matrix D is the corresponding distance matrix of the
TSP, the flow matrix F is the adjacency matrix of a complete cycle of length
n, while there is no fixed cost matrix C.

13.5.1 QAP Test Problems

Test Problem Generators

A QAP test problem generator with a known optimal solution has been pre-
sented in Li and Pardalos (1992). The Fortran code of this generator can be
obtained by sending an e-mail message to "[email protected]", and in the
body of the message put "send 92006". Using this generator we present next
two small instances of QAP.
QAPI Test Problem - Symmetrie matriees:

Flow matrix:
13.5. ASSIGNMENT PROBLEMS 311

0 5 3 7 9 3 9 2 9 0
5 0 7 8 3 2 3 3 5 7
3 7 0 9 3 5 3 3 9 3
7 8 9 0 8 4 1 8 0 4
9 2 2 8 0 8 8 7 5 9
3 2 4 4 9 0 4 8 0 3
8 4 1 1 8 4 0 7 9 5
3 2 2 8 6 8 6 0 5 5
9 6 9 0 7 0 9 5 0 5
0 7 2 4 9 1 4 7 4 0

Distance matrix:

0 7 4 6 8 8 8 6 6 5
7 0 8 2 6 5 6 8 3 6
4 8 0 10 4 4 7 2 6 7
6 2 10 0 6 6 9 3 2 6
8 6 4 6 0 6 4 8 8 6
8 5 4 6 6 0 3 8 3 2
8 6 7 9 4 3 0 6 7 8
6 8 2 3 8 8 6 0 8 8
6 3 6 2 8 3 7 8 0 9
5 6 7 6 6 2 8 8 9 0

An optimal permutation is (9 1 8 3 6 7 2 5 4 10). The optimal objective

function value is 2227.
QAP2 Test Problem - Asymmetrie matriees:

Flow matrix:

0 9 4 2 2 9 7 4 4 3
8 0 6 9 0 9 7 2 5 8
1 7 0 1 9 9 7 9 3 5
0 0 9 0 4 8 8 8 8 2
5 4 6 2 0 8 8 7 2 7
7 8 0 4 4 0 9 3 4 3
8 5 5 9 8 3 0 9 5 2
7 1 5 0 3 5 4 0 9 9
7 5 3 2 6 3 3 4 0 0
3 7 1 3 3 3 5 8 9 0

Distance matrix:
312 GHAPTER 13. GOMBINATORIAL OPTIMIZATION PROBLEMS

0 8 7 6 8 8 6 10 5 9
7 0 8 2 10 2 8 8 4 7
1 7 0 10 7 7 8 1 5 10
7 1 9 0 6 6 10 3 3 2
8 2 1 5 0 5 1 7 10 10
7 8 1 7 6 0 4 9 3 1
9 4 6 8 6 2 0 4 6 9
2 8 3 4 9 6 8 0 8 8
7 2 7 2 7 4 9 8 0 9
2 4 3 10 1 3 7 9 10 0

An optimal permutation is (9 4 5 10 7 2 6 3 1 8). The optimal objective

function value is 2025.

QAPLIB Test Problems

There is a large collection QAPLIB of electronically available data instances
for QAP (Burkard et al. (1997)). The data (and the updated best known
solution) are available in the following web page:

http://vvv.diku.dk/ $\sim$ karisch/qaplib

The QAPLIB contains a large collection of instances including the first

benchmark set of problems called Nugent. Next we provide the one such
instance of size 15:
QAP3 Test Problem - nug15:

Flow matrix:

0 1 2 3 4 1 2 3 4 5 2 3 4 5 6
1 0 1 2 3 2 1 2 3 4 3 2 3 4 5
2 1 0 1 2 3 2 1 2 3 4 3 2 3 4
3 2 1 0 1 4 3 2 1 2 5 4 3 2 3
4 3 2 1 0 5 4 3 2 1 6 5 4 3 2
1 2 3 4 5 0 1 2 3 4 1 2 3 4 5
2 1 2 3 4 1 0 1 2 3 2 1 2 3 4
3 2 1 2 3 2 1 0 1 2 3 2 1 2 3
4 3 2 1 2 3 2 1 0 1 4 3 2 1 2
5 4 3 2 1 4 3 2 1 0 5 4 3 2 1
2 3 4 5 6 1 2 3 4 5 0 1 2 3 4
3 2 3 4 5 2 1 2 3 4 1 0 1 2 3
4 3 2 3 4 3 2 1 2 3 2 1 0 1 2
5 4 3 2 3 4 3 2 1 2 3 2 1 0 1
6 5 4 3 2 5 4 3 2 1 4 3 2 1 0
13.6. GRAPH COLORING 313

Distance matrix:

0 10 0 5 1 0 1 2 2 2 2 0 4 0 0
10 0 1 3 2 2 2 3 2 0 2 0 10 5 0
0 1 0 10 2 0 2 5 4 5 2 2 5 5 5
5 3 10 0 1 1 5 0 0 2 1 0 2 5 0
1 2 2 1 0 3 5 5 5 1 0 3 0 5 5
0 2 0 1 3 0 2 2 1 5 0 0 2 5 10
1 2 2 5 5 2 0 6 0 1 5 5 5 1 0
2 3 5 0 5 2 6 0 5 2 10 0 5 0 0
2 2 4 0 5 1 0 5 0 0 10 5 10 0 2
2 0 5 2 1 5 1 2 0 0 0 4 0 0 5
2 2 2 1 0 0 5 10 10 0 0 5 0 5 0
0 0 2 0 3 0 5 0 5 4 5 0 3 3 0
4 10 5 2 0 2 5 5 10 0 0 3 0 10 2
0 5 5 5 5 5 1 0 0 0 5 3 10 0 4
0 0 5 0 5 10 0 0 2 5 0 0 2 4 0

An optimal solution is (1 2 13 8 9 4 3 14 7 11 10 15 6 5 12 ). The optimal

objective function value is 1150.

13.6 Graph Coloring

Let G = (V, E), with vertex set V = {I, 2, ... ,n} and edge set E, be an
undirected graph with no loops or multiple edges. Vertices i and j are said to
be adjacent ifthe edge, denoted by (i,j), is an element of E. A proper coloring
of G is defined to be an assignment of colors to the vertices so that no pair
of adjacent vertices share the same color. Equivalently, a proper coloring is a
mapping f of V into {I, 2, ... ,n} such that for any (i, j) E E, f(i) i- f(j). A
coloring naturally induces a partition of V such that the members of each set in
the partition are pairwise non-adjacentj these sets are precisely the subsets of
vertices assigned the same color. If there exists a coloring of G that utilizes no
more than k colors, we say that G admits a k-coloring, or that G is k~colorable.
The chromatic number, X(G), of Gis the smallest k for which G admits a k-
coloring. Given a graph and an integer k, the graph coloring problem refers to
the problem of determining whether X( G) ~ k.
There exist several mathematical programming formulations for finding the
chromatic number of a graph. The graph coloring problem associated with G
can be formulated as the following linear mixed 0-1 integer program P:

n
min LYk
k=l
314 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

s.t. E~=l Xik = 1 Vi E V

Xik +Xjk ~ 1 V (i,j) E E
Yk 2: Xik Vi E V, k = 1, ... , n
Xik E {O, 1} Vi E V, k = 1, ... , n.

In the above model a binary variable is associated with each vertex of a graph
G = (V, E), where IVI = n. The first set of constraints ensure that exactly
one color is assigned to each vertex. The second set of constraints ensure that
adjacent vertices are assigned different colors. The optimal objective function
value to the above program is the chromatic number of the graph. In the
solution, Xik equals 1 if vertex i is colored k, and is zero, otherwise. Moreover,
the sets

Bk = {i : Xik = 1},
for all k with Yk > 0 comprise an optimal partition of the vertices.
Another integer programming model is:

min 'Y

s.t. Xi ~ 'Y
Xi - Xj - 1 2: -ndij V (i,j) E E
Xj - Xi - 1 2: -n(1 - dij) V(i,j)EE
ij E {O, I}
Xi E Z+ Vi E V.

The second and third set of constraints ensure that adjacent vertices are not
assigned the same color. This can be seen by noting that if Xi = Xj then no
feasible assignment of dij leads to simultaneous satisfaction of the constraints.
The optimal objective function yields the chromatic number of the graph under
consideration.

13.6.1 Test Problems

During the second DIMACS Implementation Challenge a set of 32 graph in-
stances were chosen to provide the coloring benchmarks. These problems can
be found in the web address:

http://dimacs.rutgers.du/Challenges/index.html.

13.7 Maximum Clique Problem

Let G = (V,E), with vertex set V = {1,2, ... ,n} and edge set E ~ V x V,
be an arbitrary undirected and weighted graph. For each vertex i E V, a
13.7. MAXIMUM CLIQUE PROBLEM 315

positive weight Wi is associated with i. AG = (aij)nxn is the adjacency matrix

of G, where aij = 1 if (i,j) E Eis an edge of G, and aij = if (i,j) tJ. E.
The complement gmph 0/ G = (V, E) is the graph G = (V, E), where E =
{(i,j) I i,j E V, i =f. j and (i,j) tJ. E}. For a subset 8 ~ V, we define the
weight of 8 to be W(8) = :EiES Wi. We call G(8) = (8, En8 x 8) the subgmph
induced by 8. Graph G is complete if all its vertices are pairwise adjacent, i.e.
Vi, jE V, (i,j) E E. A clique Cis a subset of V such that G(C) is complete.
The maximum clique problem asks for a clique of maximum cardinality. The
maximum weight clique problem asks for a clique of maximum weight. These
problems are N P-complete on arbitrary graphs.
The maximum clique problem has many equivalent formulations as an inte-
ger programming problem, or as a continuous nonconvex optimization problem
Pardalos (1996). The simplest one is the following edge /ormulation:

s.t. Xi + Xj ~ 1, V (i,j) E E,

Xi E {O, I}, i = 1, . " ,n.

This simple formulation of the Maximum Clique Problem and a linear 0 - 1

problem, has not been useful in practice.
Recently, new efficient algorithms and heuristics have been developed us-
ing nontrival nonlinear integer formulations of the Maximum Clique Problem.
Next we discuss quadratic zero-one formulations of the Maximum Clique Prob-
lem. To facilitate our discussion, define a transformation T from {O, l}n to 2v ,

T(x) = {i I Xi = 1, i E V}, "Ix E {O,l}n.

Denote the inverse of T by T- 1 If X = T- 1 (8) for some 8 E 2v then

Xi = 1 if i E 8 and Xi = 0 if i rt 8, i = 1, ... ,n. We can rewrite the
maximum problem (13.7) as a minimization problem (when Wi = 1) with
f(x) = - :Ef=l Xi Using the fact that XiXj = 0 for all (i,j) E E since for
Xi, Xj E {O, I} Xi + Xj ~ 1 if and only if XiXj = 0, the constraints in (13.7) can
be removed by adding two times the quadratic terms to the objective' function,
which is now
n
f(x) =- LXi +2 L XiXj = xT(Aa - I)x.
i=l (i,j)EE, i>j

The quadratic terms represent penalties for violations of XiXj = 0. This leads
to the following global quadmtic zero-one formulation:

global minf(x) = x T Ax,

x
316 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEM8

s.t. :z: E {O,l}n, where A = AG - I.

If:z:* solves this problem, then the set C defined by C = T(x*) is a maximum
clique of G with ICI = -z = - f(x*).
An independent set (stable set, vertex packing) is a subset of V whose
elements are pairwise nonadjacent. The maximum independent set problem
asks for an independent set of maximum cardinality. The size of a maximum
independent set is the stability number of G (denoted by a(G)). The maxi-
mum weight independent set problem asks for an independent set of maximum
weight. Following the equivalence of the maximum clique problem with the
maximum independent set problem, we have that the maximum independent
set problem is equivalent to the following global quadratic zero-one problem:

global minf(:z:) = :z:T A:z:

s.t. :z: E {O, l}n, where A = AG - I.

If:z:* solves this problem, then the set 8 defined by 8 = T(:z:*) is a maximum
independent set of G with 181 = -z = - f(:z:*).
The above formulations for the maximum clique problem and the maximum
independent set problem can be regarded as a special case by taking Wi = 1, i =
1,2, ... ,n. The maximum ~eight independent set problem is equivalent to the
following global quadratic zero-one problem:

global minf(:z:) = :z:T A:z:,

s.t.:Z: E {O, l}n,

where aii = -Wi, i = 1, ... , n, aij = ~(Wi + Wj), 'V (i,j) E E, and aij =
0, 'V (i,j) E E. Let :z:* solve this problem, then the set 8 defined by 8 = T(:z:*)
is a maximum independent set of G with weight W(8) = -z = - f(:z:*).
The Motzkin-Strauss formulation of the Maximum Clique Problem is the
quadratic program:

s.t. :z:T e = 1,

:z:~O.

It is weIl known (Pardalos (1996)) that the global optimum value of this QP
is (1 - l/w(G)), where w(G) is the clique number of G.
13.7. MAXIMUM CLIQUE PROBLEM 317

13.7.1 Maximum Clique: Coding Theory Test Problems

In Coding Theory, one wishes to find a binary code as large as possible that can
correct a certain number of errors for a given size of the binary words (vectors),
see Hasselberg et al. (1993). In order to correct errors, the code must consist
of binary words among which any two differ in a certain number of positions
so that a misspelled word can be detected and corrected. A misspelled word
is corrected by replacing it with the word from the code that differs the least
from the misspelled one.
The Hamming distance between the binary vectors U = (UI, U2, ,un )
and v = (VI, V2, . .. ,Vn ) is the number of indices i such that 1 ~ i ~ n and
Ui =I- Vi We denote the Hamming distance by dist(u, V).
It is weIl known that a binary code consisting of a set of binary vectors any
two of which have Hamming distance greater or equal to d can correct ld 2I J
errors (see Hasselberg et al. (1993)). Thus, what a coding theorist would like
to find is the maximum number of binary vectors of size n with Hamming
distance d. We denote this number by A(n, d).
Another problem arising from co ding theory, closely related to the one
mentioned above, is to find a weighted binary code, that is, to find the max-
imum number of binary vectors of size n that have precisely w 1's and the
Hamming distance of any two of these vectors is d. This number is denoted by
A(n, w, d). A binary code consisting of vectors of size n, weight wand distance
d, can correct w - ~ errors.
We define the Hamming graph H (n, d), of size n and distance d, as the
graph with vertex set the binary vectors of size n in which two vertices are
adjacent if their Hamming distance is at least d. Then, A(n, d) is the size of a
maximum clique in H(n, d). The graph H(n, d) has 2n vertices, 2n - 1 I:~=d (7)
edges and the degree of each vertex is I:~=d m.
We define the Johnson graph, J(n, w, d), with parameters n, wand d, as
the graph with vertex set the binary vectors of size n and weight w, where
two vertices are adjacent if their Hamming distance is at least d. Then, sim-
ilar to Hamming graph, the size of the weighted code, A(n, w, d), equals the
size ofthe maximum clique in J(n, w, d). The graph J(n, w, d) has (:) vertices,
~ (:) I:~=r~l (~) (nkw) edges and the degree of each vertex is I:~=r~l (~) (nkw).

13.7.2 Maximum Clique: Keller Graphs

A family of hypercubes with disjoint interiors whose union is the Euclidean
space Rn is a tiling. A lattice tiling is a tiling for which the centers of the
cubes form a lattice.
In the beginning of the century, Minkowski conjectured that in a lattice
tiling of Rn by translates of a unit hypercube, there exist two cubes that share
a (n - l)-dimensional face. About fifty years later, Hajs (Hasselberg et al.
(1993)) proved Minkowski's conj ect ure.
At 1930, Keller suggested that Minkowski's conjecture holds even in the
318 CHAPTER 13. COMBINATORlAL OPTIMIZATION PROBLEMS

absence of the lattice assumption. Ten years later Perron (Hasselberg et al.
(1993)) proved the correctness of Keller's conjecture for n ::; 6. Since then,
many papers have been devoted to prove or disprove this conjecture and re-
cently, Lagarias and Shor (Hasselberg et al. (1993)) proved that Keller's con-
jecture fails for n ~ 10. Thus, it is left to prove whether the conjecture holds
for n = 7,8,9.
We define the graph r n as a graph with vertex set Vn = {(dl, d 2 , ... , dn ) :
di E {O,1,2,3}, i = 1,2, ... ,n} where two vertices u;:::: (d 1 ,d2,'" ,dn ) and
v = (d~,d2"" ,d~) in Vn are adjacent ifand only if

3 i, 1::; i ::; n : di - d~ = 2 mod 4

and

Corrooi and Szab6 presented a graph theoretic equivalent of Keller's con-

jecture. It is shown that, there is a counterexample to Keller's conjecture if
and only if there exist a n E N+ such that r n has a clique of size 2n .
r n has 4n vertices, !4n (4 n - 3n - n) edges and the degree of each node is
4 - 3n - n. r n is very dense and has at least 8n n! different maximum cliques.
n
It can be shown (see Hasselberg et al. (1993)) that the maximum clique size
of r n is less than or equal to 2n .

13.8 Steiner Problems in Networks (SPN)

The Steiner problem in graphs (SPG) consists of connecting a subset of given
nodes on a graph with the minimum cost tree.
Instance Formulation ofthe Steiner Problem in Undirected Graphs:
Let G ;:::: (N, A, C) be an undirected graph, where
N = {I, ... , n} is a set of nodes, and
Ais a set of undirected arcs (i,j) with each arc incident to two nodes,
and
C is a set of nonnegative costs Cij associated with undirected arcs (i, j).
Then, the SPG is defined as follows.

Instance: A graph G = (N, A, C), anode subset REN.

Question: Find the minimum cost tree, on G, that would connect all the
vertices in R.
Instance Formulation of the Steiner Problem in Directed Graphs:
The SPDG is the directed graph version of the Steiner problem. Consider
Gd = (N,Ad,C d) where
N = {l, ... , n} is a set of nodes, and
A d is a set of directed arcs (i,j) incident to nodes in N.
13.8. STEINER PROBLEMS IN NETWORKS (SPN) 319

Cd = {cfj} is a set of eosts associated with A d.

Furthermore, let n (a set of regular nodes) be a subset of N, and let r (a root
node) be an arbitrary node in n. Also, define a direeted Steiner tree T rooted
at node r, with respeet to n on G, to be a directed tree where all nodes in
n \{r} ean be reaehed from r via direeted paths in T, and all leaf nodes are in
n. Next, we define the SPDG.
Instance: A graph Gd = (N, A d, Cd), anode subset n and a root node r
in n.
Question: Find the minimum eost directed Steiner tree, on G with respect
to n, that has node r as root node.
In the light of the SPG and SPDG definitions given above, it is not hard
for one to see that the SPDG is a generalization of the SPG. In fact, the SPG
ean be transformed to a special ease of the SPDG where the are eost strueture
is symmetrie. This transformation ean be done as follows. Given a graph
G = (N,A,C), ereate the eorresponding direeted graph Gd = (N,Ad,C d),
where every undirected are (i,j) E A eorresponds to two oppositely direeted
ares (i,j) and (j,i) E A d associated with the same eost (c~ = cji = Cij).
There are several equivalent formulations for the Steiner problems in graphs.
The different mathematical formulations ean be divided into three eategories:
mixed integer formulations, integer formulations and eontinuous formulations.
Therefore different algorithmie approaehes ean be tested by using the same
test problems.

13.8.1 Test Problems

An automatie test problem generator for problems with a known optimal so-
lution has been proposed by Khoury et al. (1993).
Using this generator, next we provide a set of test problems: The network
is represented by a table where eolumn 1 entries are are numbers. Column 2
and eolumn 3 give the two end nodes of an are. Are eosts are represented in
eolumn 4.
320 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

Test problem spgtpl

# of nodes: 20
# of ares: 40
~= {2,3,7,8,18}
Minimum Steiner tree set of ares
Minimum Steiner tree eost: 986
={8, 15, 22, 29, 31, 38}

Are # Endl # End2 # Are Cost

1 11 10 289
2 2 6 275
3 1 15 347
4 11 13 109
5 3 5 110
6 15 12 161
7 11 9 275
8 8 11 252
9 12 19 354
10 13 17 241
11 19 20 180
12 2 14 228
13 19 16 202
14 10 4 116
15 8 1 188
16 1 17 314
17 12 18 190
18 9 10 159
19 5 4 227
20 18 19 190
21 10 8 319
22 11 7 123
23 16 12 234
24 17 6 222
25 15 3 168
26 13 6 211
27 9 17 134
28 10 5 160
29 1 2 101
30 14 10 179
31 1 3 215
32 18 16 238
33 13 2 313
34 15 8 336
35 6 19 124
36 9 3 175
37 7 16 176
38 3 18 107
39 5 6 238
40 9 18 260
13.8. STEINER PROBLEMS IN NETWORKS (SPN) 321

Test problem spgtp2

# of nodes: 20
# of arcs: 40
~= {2,5,6,7,16}
Minimum Steiner tree set of ares
Minimum Steiner tree east: 3562
={8, 15, 19,24,26,32,36, 38}
I Are # I Endl # I End2 # I Are Cast
1 6 14 576
2 6 18 897
3 2 10 720
4 2 8 495
5 16 15 947
6 7 13 348
7 12 19 547
8 16 11 442
9 2 1 612
10 17 20 553
11 11 4 1187
12 17 3 307
13 1 6 1211
14 2 12 1459
15 11 9 153
16 8 5 1244
17 16 5 790
18 8 17 1032
19 11 17 830
20 9 10 1238
21 5 15 509
22 8 4 760
23 1 18 213
24 11 12 103
25 13 4 336
26 9 6 223
27 11 14 512
28 9 3 763
29 15 17 644
30 14 9 534
31 8 6 1263
32 17 7 524
33 6 11 496
34 17 13 1438
35 20 4 740
36 12 5 442
37 10 8 752
38 6 2 845
39 1 17 770
40 9 4 1301
322 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

Test problem spgtp3

# of nodes: 20
# of ares: 40
~= {3,6,10, 13, 14, 16, 18}
Minimum Steiner tree set of ares
Minimum Steiner tree eost: 2213
={5, 11, 19,22,27,31,33,34,37, 40}

Are # Endl # End2 # Are Cost

1 6 15 417
2 6 1 427
3 13 7 181
4 10 4 476
5 16 11 252
6 15 8 248
7 11 5 580
8 8 2 170
9 2 19 293
10 16 20 326
11 11 9 424
12 5 2 325
13 9 7 522
14 4 16 560
15 6 8 488
16 16 8 482
17 9 4 248
18 15 9 376
19 11 17 101
20 3 1 448
21 18 11 620
22 11 12 154
23 5 13 167
24 14 8 279
25 3 15 439
26 16 14 637
27 11 6 215
28 9 8 394
29 4 6 549
30 9 18 660
31 9 14 107
32 13 20 316
33 17 3 105
34 12 13 450
35 4 12 534
36 9 1 123
37 6 10 223
38 17 9 589
39 4 7 435
40 14 18 182
13.8. STEINER PROBLEMS IN NETWORKS (SPN) 323

Test problem spgtp4

# of nodes: 20
# ofarcs: 40
~= {2,8, 10,13, 14, 15, 16, 18}
Minimum Steiner tree set of arcs
Minimum Steiner tree cost: 2864
={1, 4, 6, 8, 9,13,15,20,21,23,29,35, 37}
I Are # I End1 # I End2 # I Are Cast
1 16 11 252
2 9 4 260
3 16 4 88
4 11 12 215
5 1 16 469
6 11 5 107
7 15 17 319
8 17 14 450
9 12 18 223
10 13 4 400
11 6 13 372
12 8 4 363
13 2 15 152
14 1 18 472
15 6 10 182
16 17 19 305
17 14 1 339
18 17 20 306
19 11 7 267
20 11 9 101
21 11 17 154
22 17 3 274
23 5 8 355
24 2 10 663
25 16 6 420
26 8 17 404
27 9 10 658
28 5 16 290
29 11 6 289
30 8 3 373
31 11 15 277
32 14 9 865
33 8 6 448
34 6 9 400
35 9 2 105
36 17 10 651
37 5 13 279
38 20 4 84
39 10 11 661
40 9 5 289
324 CHAPTER 13. COMBINATORIAL OPTIMIZATION PROBLEMS

Test problem spgtp5

# of nodes: 20
# of ares: 40
~= {1,3,4,5,6, 10, 16, 17, 18, 19}
Minimum Steiner tree set of ares
Minimum Steiner tree eost: 2776
={3, 8, 13, 22, 26, 27, 28, 32, 33, 35, 36, 37}

Are # Endl # End2 # Are Cost

1 19 12 274
2 6 11 167
3 4 2 252
4 16 15 430
5 1 7 348
6 11 9 335
7 2 20 252
8 4 13 123
9 17 14 130
10 7 17 239
11 2 9 433
12 7 11 354
13 4 8 101
14 16 3 569
15 15 8 468
16 18 2 260
17 14 13 411
18 14 16 342
19 15 4 492
20 16 8 385
21 12 14 403
22 2 19 154
23 14 19 392
24 17 11 476
25 13 19 266
26 2 5 289
27 2 16 252
28 13 6 431
29 15 19 524
30 15 12 286
31 15 10 113
32 13 18 135
33 8 10 351
34 7 8 416
35 19 17 223
36 10 3 147
37 17 1 318
38 9 7 151
39 16 7 463
40 2 14 380

Remark: Steiner tree problems can be used also to generate hard instances
of MAX-SAT problems. The exact translation of Steiner problems into MAX-
SAT is given in H.Kautz et al. (1997).
 Chapter 14

Nonlinear Systems of
Equations

The solution of nonlinear systems of equations is an important problem in

engineering. Applications include identifying the multiple steady states of a
reactor network (Folger, 1986; Schlosser and Feinberg, 1994), predicting the
azeotropes formed in a nonideal mixture (Fidkowski et al. , 1993; Maranas et
al., 1996), finding the steady-states of a flowsheet or part of a flowsheet (Zhang,
1987; Bullard and Biegler, 1991; Wilhelm and Swaney, 1994; Bekiaris et al. ,
1993), identifying equiIibrium points in multiphase systems (Heidemann and
Mandhane, 1973; Seader et al., 1990).
Many techniques are available to find one of the solutions of a given system
such as Newton and quasi-Newton methods (PoweIl, 1970; ehen and Stadtherr,
1984; Duff et al. , 1987; More, 1987; Paloschi and Perkins, 1988), homotopy
continuation techniques (Lahaye, 1934; Leray and Schauder, 1934; Davidenko,
1953; Klopfenstein, 1961; Garcia and Zangwill, 1981), interval-Newton meth-
ods (Kearfott and Novoa, 1990; Kearfott et al. , 1994; Seader and Kumar,
1994) and optimization-based techniques (Horst and Tuy, 1993; Maranas and
Floudas, 1995). The approaches based on interval arithmetic or global opti-
mization are the only ones that can guarantee the enclosure of all solutions.
For a presentation of global optimization approaches for enclosing all solutions
of constrained systems of nonlinear equations, the reader is directed to the
article of Maranas and Floudas (1995), and the book by Floudas (2000).
The general formulation of these problems is given below.

h(x) = 0
g(x) < 0
x E [xL,x U).

325

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
326 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

This ean be transformed into an optimization problem of the form:

min S
8, z
S.t. h(z) - S < 0
-h(z) - S < 0
g(z) < 0
z E [ZL, zU]

In this ehapter, we foeus on loeating all solutions of nonlinearly eonstrained

systems of equations and present (i) test problems that are taken from the
literature and (ii) test problems that eorrespond to the enclosure of all homo-
geneous azeotropes.

14.1 Literature problems

In this seetion we present nine test problems taken from the literat ure. Test
problems 6, 7, 8, and 9 are regarded as ehallenging problems.

14.1.1 Test Problem 1: Himmelblau function

The goal of this example is to find all stationary points of the Himmelblau
function (Reklaitis and Ragsdell, 1983).

Formulation

Equalities

4x~ + 4XIX2 + 2x~ - 42xl 14

4x~ + 2x~ + 4XIX2 - 26x2 = 22

Variable bounds

Problem Statistics
No. of eontinuous variables 2
No. of linear equalities
No. of nonlinear equalities 2
No. of known solutions 9
14.1. LITERATURE PROBLEMS 327

Known Solutions
Xl -3.7793 -3.0730 -2.8051 -0.2709 -0.1280
X2 -3.2832 -0.0814 3.1313 -0.9230 -1.9537

Xl 0.0867 3.0 3.3852 3.5844

X2 2.8843 2.0 0.0739 -1.8481

14.1.2 Test Problem 2: Equilibrium Combustion

This problem aims to identify the concentrations of the products of a hydro-
carbon combustion process at equilibrium (Meintjes and Morgan, 1990).

Formulation

Equalities

2X1X2 + Xl + 3RlOX~
+ R7X2X3 + R9X2x4 + R8X2 - RX5
+X2X~ 0
2X2X~ + R7X2X3 + 2R5X~ + R6X3 - 8X5 0
R9X2X4 + 2x~ - 4Rx5 0
X1X2 + Xl + RlOX~ + X2X~ + R7X2X3

+ R9X2x4 + R8X2 + R5X~ + ~X3 + x~ 1

Variable bounds
0.0001 ~ Xi ~ 100, i = 1, ... ,5

Data
R = 10 R5 = 0.193 ~ = 4.10622 10- 4
R7 = 5.45177 10-4 R8 = 4.4975 10- 7 Rg = 3.40735 10-5
R lO = 9.615 10- 7

Problem Statistics
No. of continuous variables 5
No. of linear equalities
No. of nonlinear equalities 5
No. of known solutions 1

Known Solution

z = (0.003431,31.325636,0.068352,0.859530, 0.036963)T
328 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

14.1.3 Test Problem 3

A two equation system taken from Bullard and Biegler (1991) is solved in this
example.

Formulation

Equalities

104XIX2 -1 0
e- X1 + e- X2 - 1.001 0

Variable bounds

5.49 X 10-6 ~ Xl ~ 4.553

2.196 x 10- 3 ~ X2 ~ 18.21

Problem Statistics
No. of continuous variables 2
No. of linear equalities
No. of nonlinear equalities 2
No. of known solutions 1

Known Solution

x = (1.450 X 10- 5 , 6.8933353)T

14.1.4 Test Problem 4

A system of two nonlinear equations taken from Ferraris and Tronconi (1986)
is solved here.

Formulation

Equalities

0.5 sin(xlx2) - 0.25x2/7r - 0.5XI 0

(1 - 0.25/7r)(exp(2xI) - e) + eX2/7r - 2exl 0

Variable bounds

0.25 ~ Xl ~ 1
1.5 ~ X2 ~ 27r
14.1. LITERATURE PROBLEMS 329

Problem Statistics
No. of continuous variables 2
No. of linear equalities
No. of nonlinear equalities 2
No. of known solutions 2

Known Solutions
Solution 1: :J: = (0.29945, 2.83693)T.

Solution 2: :J: = (0.5, 3.14159)T.

14.1.5 Test Problem 5

This example is Brown's almost linear system (Kearfott and Novoa, 1990).

Formulation

Equalities

2XI + X2 + X3 + X4 + X5 = 6
Xl + 2X2 + X3 + X4 + X5 6
Xl + X2 + 2X3 + X4 + X5 6
Xl + X2 + X3 + 2X4 + X5 6
XIX2X3X4X5 1

Variable bounds

-2 ::; Xi ::; 2, i = 1, ... ,5

Problem Statistics
No. of continuous variables 5
No. of linear equalities 4
No. of nonlinear equalities 1
No. of known solutions 2

Known Solutions
Solution 1: :J: = (1,1,1,1, 1)T
Solution 2: :J: = (0.916,0.916,0.916,0.916, 1.418)Y.

14.1.6 Test Problem 6

This is a robot kinematics problem taken from Kearfott and Novoa (1990).
330 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Formulation

Equalities

4.7311O- 3XIX3 - 0.3578x2X3 - 0.1238x I

+X7 - 1.637 1O- 3x2 - 0.9338x4 - 0.3571 o
0.2238xIX3 + 0.7623x2X3 + 0.2638x I
-X7 - 0.07745x2 - 0.6734x4 - 0.6022 o
X6XS + 0.3578xI + 4.731 1O- 3x+2 o
-0.7623xI + 0.2238x2 + 0.3461 o
x~ + x~-1 o
x~ + x~-1 o
x~ + x~-1 o
x~ + x~-1 o

Variable bounds

-1 :$ Xi :$ 1, i = 1, ... ,8

Problem Statistics
No. of continuous variables 8
No. of linear equalities 1
No. of nonlinear equalities 7
No. of known solutions 16

Known Solutions
Xl X2 X3 X4 X5 X6 X7 Xs
0.1644 -0.9864 -0.9471 -0.3210 -0.9982 -0.0594 0.4110 0.9116
0.1644 -0.9864 -0.9471 -0.3210 -0.9982 0.0594 0.4110 -0.9116
0.1644 -0.9864 -0.9471 -0.3210 0.9982 -0.0594 0.4110 0.9116
0.1644 -0.9864 -0.9471 -0.3210 0.9982 0.0594 0.4110 -0.9116
0.1644 -0.9864 0.7185 -0.6956 -0.9980 -0.0638 -0.5278 0.8494
0.1644 -0.9864 0.7185 -0.6956 -0.9980 0.0638 -0.5278 -0.8494
0.1644 -0.9864 0.7185 -0.6956 0.9980 -0.0638 -0.5278 0.8494
0.1644 -0.9864 0.7185 -0.6956 0.9980 0.0638 -0.5278 -0.8494
0.6716 0.7410 -0.6516 -0.7586 -0.9625 -0.2711 -0.4376 0.8992
0.6716 0.7410 -0.6516 -0.7586 -0.9625 0.2711 -0.4376 -0.8992
0.6716 0.7410 -0.6516 -0.7586 0.9625 -0.2711 -0.4376 0.8992
14.1. LITERATURE PROBLEMS 331

Xl X2 X3 X4 X5 X6 X7 X8
0.6716 0.7410 -0.6516 -0.7586 0.9625 0.2711 -0.4376 -0.8992
0.6716 0.7410 0.9519 -0.3064 -0.9638 -0.2666 0.4046 0.9145
0.6716 0.7410 0.9519 -0.3064 -0.9638 0.2666 0.4046 -0.9145
0.6716 0.7410 0.9519 -0.3064 0.9638 0.2666 0.4046 -0.9145
0.6716 0.7410 0.9519 -0.3064 0.9638 -0.2666 0.4046 0.9145

14.1.7 Test Problem 7

This example represents a circuit design problem (Ratschek and Rokne, 1993).

Formulation

Equalities

(1 - XIX2)X3 {exp [X5 (glk - g3kX7 10- 3 10- 3 ) 1- 1 }

- g5kx8
- g5k + g4k x 2 = 0,k = 1, ... ,4
(1 - XIX2)X4 {exp [X6 (glk - g2k - g3kx7 10- 3 - g4kX9 10- 3 ) 1-1}
- g5kxl + g4k = 0, k = 1, ... ,4
XIX3 - X2X4 = 0

Variable bounds

o :::; Xi :::; 10, i = 1, ... ,9

Data

9 =
( 0.~50 0.7520
0.3690
5.2095
1.2540
0.8690
0.7030
10.0677 22.9274
09820
1.4550
20.2153
1
23.3037 101.7790 111.4610 191.2670
28.5132 111.8467 134.3884 211.4823

Problem Statistics
No. of continuous variables 9
No. of linear equalities
No. of nonlinear equalities 9
No. of known solutions 1

Known Solution

x = (0.89999,0.44999,1.00001,2.00007,7.99997,
7.99969,5.00003,0.99999,2.00005)T
332 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

14.1.8 Test Problem 8

The steady state of aseries of CSTRs with recycle must be identified in this
example (Kubicek et al., 1980).

Formulation

Equalities

(l-R) CO(l~d -<Pl)exP(l+l~t~d,) -<PI 0

<PI - (1 + fh) <P2
+ (1 - R) (D/lO - l<Pl - (1 + fh)<P2) exp C+l~t;2/,) 0

Variable bounds

Data

, = 1000 D = 22 l =2 fh =2

The parameter R is varied between 0 and 1.

Problem Statistics
No. of continuous variables 2
No. of linear equalities
No. of nonlinear equalities 2

The number of known solutions varies between 1 and 7 for 13 different

values of R.
14.1. LITERATURE PROBLEMS 333

Known Solutions
R rPl rP2 R rPl rP2
0.935 0.724987 0.245241 0.965 0.298523 0.168212
0.940 0.724234 0.245134 0.034546 0.049423
0.945 0.079754 0.664390 0.034546 0.302260
0.172234 0.591344 0.034546 0.689784
0.723330 0.244990 0.716707 0.243634
0.950 0.062799 0.103831 0.970 0.027994 0.039357
0.062799 0.195656 0.027994 0.338384
0.062799 0.675510 0.027994 0.690874
0.206004 0.557881 0.333780 0.165215
0.722227 0.244797 0.713383 0.242837
0.955 0.051212 0.078028 0.975 0.022196 0.030790
0.051212 0.235148 0.022196 0.379801
0.051212 0.682347 0.022196 0.689588
0.236330 0.520375 0.374632 0.166725
0.720847 0.244533 0.708353 0.241556
0.960 0.042125 0.061755 0.980 0.016978 0.023300
0.042125 0.268726 0.016978 0.431009
0.042125 0.686930 0.016978 0.683991
0.266590 0.178424 0.425213 0.172848
0.266590 0.327276 0.699758 0.239253
0.266590 0.461132 0.985 0.012223 0.016625
0.719074 0.244164 0.012223 0.504131
0.012223 0.666204
0.496354 0.186263
0.680842 0.233986
0.990 0.007848 0.010593
0.995 0.003789 0.005081

14.1.9 Test Problem 9

The goal of this example taken from Smith (1985) is to find all steady state
temperatures of a nonisothermal CSTR.

Formulation

Equalities

b
-Texp { -C }
- b( 1 + aTo) exp { -C }
+ -T -1 = 0
To T aTo T To

Variable bounds
100 ::; T ::; 1000

Data
-1000
a = 3tlH b = 1.344 X 109 C = -7548.1193 T o = 298
334 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Three different values of the reaction enthalpy 6.H are used.

Problem Statistics
No. of continuous variables 1
No. of linear equalities
No. of nonlinear equalities 1

The number of known solutions varies between 1 and 3 for three different
values of 6.H.

Known Solutions
6.H T
-50,000 300.42
347.41
445.49
-35,958 299.63
380.46
380.51
-35,510.3 299.61

14.2 Enclosing All Homogeneous Azeotropes

14.2.1 Introduction
The ability to predict the presence and composition of an azeotropes in a
mixture is essential for the design of nonideal separation processes. In addition,
when some experimental data about the azeotropes in a system is available, a
technique for finding an azeotropes can be very useful for determining which
thermodynamic equation models the system most accurately.
The thermodynamic conditions for azeotropy can be written as a nonlin-
ear system of equations. The number of solutions to the system is not known
apriori, and there may be no solution at all. Recently, several methods for
endosing an homogeneous azeotropes have been presented in the literat ure.
Harding et al. (1997) developed a deterministic global optimization approach
for finding all azeotropes and illustrated the method through several examples
using the Wilson, NRTL, UNIQUAC and UNIFAC activity coefficient equa-
tions. Fidkowski et al. (1993) developed a homotopy continuation approach
for solving the system of nonlinear equations. They applied this method for
the Wilson and modified Wilson equations. Chapman and Goodwin (1993)
presented a local search method for finding homogeneous azeotropes. Their
method employs a Levenberg-Marquardt algorithm.
The test problems for this section are taken from Harding et al. (1997).
They cover a wide range of activity coefficient equations and contain up to five
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 335

components. Global optimization-based approaches for enclosing all homoge-

neous azeotropes are described in detail in Harding et al. (1997) and the book
by Floudas (2000).

14.2.2 General Formulation - Activity Coefficient Equations

The necessary and sufficient conditions for homogeneous azeotropy are:

the vapor phase and liquid phase must be in equilibrium,

the composition of the vapor phase must equal the composition of the
liquid phase, and

the material balance must be satisfied.

These conditions form a system of nonlinear equations:

Inyl -lnx'l _lnpsat

l + InP -ln""Tl o 'V i E N
Yi - Xi o 'V i E N
LXi 1
iEN

LYi 1
iEN

In order to locate all homogeneous azeotropes, all solutions to this system

of nonlinear equations must be located. Maranas and Floudas (1995) proposed
a method for finding all solutions to systems of nonlinear equations. In their
approach, the nonlinear equations are allowed to be violated. This is done
by writing each nonlinear equality as two inequalities and introducing a non-
negative slack variable for the amount of violation. An optimization problem
can be formulated where the objective is to minimize the violation of the non-
linear equalities. There is a one-to-one correspondence between global minima
of the optimization problem, s' = 0, and solutions to the original system of
nonlinear equations.
The goal is to enclose all homogeneous azeotropes in an N component
mixture. This corresponds to finding all global minima where s* = 0 for the
following nonconvex NLP.
Objective function
The objective is to minimize the maximum violation of the nonlinear equal-
ity constraints.

mms
z,T,s
336 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Constraints
Equilibrium Expressions
This formulation is derived from the equivalence of chemical potentials of
each component in the liquid and vapor phases. An activity coefficient equation
is used to model the nonideal behavior in the liquid phase, and the vapor phase
is considered to be ideal. Each nonlinear equality is split into two inequalities
and a slack variable is subtracted from the left-hand side of each. When the
slack variable is equal to zero then the original equality is satisfied.

, - In 'V." - S <- - In P
- In psat Vi E N
ln~sat + In'Yi - s :s InP Vi E N

Mole Balance

LXi 1
iEN

Variable bounds

Variable definition

s is a slack variable corresponding to the violation of the nonlinear equal-

ity constraints.

prt denotes the saturated vapor pressure, and is a function of temper-

ature. For the examples below, the Antoine Equation is used,

b
lnpsat = a_ --'-
, 'T+Ci

In 'Yi is the liquid phase activity coefficient of component i.

Xi is the liquid mole fraction of component i.

T is temperature.
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 337

14.2.3 Wilson Equation

Wilson (1964) developed an activity coeflicient equation that is often used to

model solutions containing both polar and nonpolar components. The equation
contains a single nonsymmetric binary interaction parameter.

In 'Yi = 1 - In ( L
JEN
xjA ij ) -?= I:
JEN kEN
XjAji
XkAjk

14.2.4 Test Problem 1

The ternary system (Acetone(l) - Methyl Acetate(2) - Methanol(3)) is com-

monly used in the literat ure of phase equilibria. Experimental data show that
this system contains three binary and one ternary azeotrope. This is a diflicult
example due to the number of solutions that must be located.

Explicit Formulation

mm s

subject to -al + T~C! + In(x l Al1 + X2 A 22)

+ XIAUXl +X2A12
Au + X1A21X2 +X2
A 21
A 22
_ S < 1-
-
In P

al - T~C! -ln(x l Al1 + X2 A 22)

- xlAu
xlAu
Xl Au +X2A12
Xl
__ x2A21
x2A21
X1A21
X1A21 +X2
-
A 22 -
+X2A22
S < -1
S :::;
-
+ In
-1 + InPP

-a2 + T~C2 + In(x l A2l + X2 A22)

+ XIXIAU
xlA12
AU +x2A12
+ X1A21X2+X2
A nn
+X2A22
< 1 - In P
-s<l-lnP
A 22 - S _
-

a2 - T~C2 - In(x l A21 + X2 A22)

-
Xl
xlA12
xlA12
Xl Au +x2A12
+x2A12
__ X2An
X2 A n
xlA21 +X2
_ Ss
A 22 -
+X2A22
< -1
:::;
-
+ In
-1 + InPP
Xl + X2 == 1
Xl, x2 2: 0
338 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Data

P = 760 mmHg
R = 1.98721 (cal)mol-lK- l
a (16.388, 16.268, 18.607f
b (2787.50, 2665.54, 3643.31f
c (229.66,219.73,239.73f

A = C 1.55
0.48
o 1.0 0.768
0.566 0.65 1.0
0.544
)

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of nonconvex inequalities 6
No. of known solutions 4

Known Solutions

Objective function: s = 0 corresponding to a valid solution of the original

system of nonlinear equations.

Xl X2 X3 T
0.532 0.468 0.000 55.675
0.747 0.000 0.253 54.505
0.000 0.677 0.323 54.356
0.272 0.465 0.253 54.254

14.2.5 Test Problem 2

This example is for the quaternary system (Methanol(1) - Benzene(2) - i-

Propanol(3) - n-Propanol(4)). Physical experiments have located three binary
azeotropes.
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 339

Data

P 760 mmHg
R = 1.98721 (cal)mol-1K- 1
a = (18.607,15.841,20.443, 19.293)T
b (3643.31,2755.64,4628.96, 4117.07f
c = (239.73,219.16,252.64, 227.44f
1.0 0.192 2.169 1.611)
A ( 0.316 1.0 0.477 0.524
0.377 0.360 1.0 0.296
0.524 0.282 2.065 1.0

Problem Statistics
No. of continuous variables 5
No. of linear equalities 1
No. of nonconvex inequalities 8
No. of known solutions 3

Known Solutions

Objective function: s = 0 corresponding to a valid solution of the original

system of nonlinear equations.

0.624 0.376 0.000 0.000 58.129

0.000 0.586 0.414 0.000 71.951
0.000 0.780 0.000 0.220 76.946

14.2.6 Test Problem 3

This example is a five-component system (Acetone(1) - Chloroform(2) - Methanol(3)

- Ethanol(4) - Benzene(5)). From experiments, the system has been found to
contain six binary azeotropes, three ternary azeotrope and one quaternary
azeotrope. This is an especially challenging example due to its size, and to the
number of soIutions that exist.
340 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Data

P = 760 mmHg
R = 1.98721 (cal)mol-IK- I
a = (16.388,16.014,18.607,18.679,15.841f
(2787.50,2696.25,3643.31,3667.70,
b = (2787.50, 2696.25,3643.31,3667.70, 2755.64f
c = (229.66,226.23,239.73,226.18,219.16)T
1.269 0.696
( 1.0 1.269 0.696 0.636
0.636 0.590)
0.590 )
( 1.552
1.0 1.0 0.697 1.005 1.273
A 0.767 0.176 1.0 1.676 0.188
0.826 0.188 0.531 1.0 0.198
0.990 0.928 0.308 0.461 1.0

Problem Statistics
No. of continuous variables 6
No. of linear equalities 1
No. of nonconvex inequalities 10
No. of known solutions 9

Known Solutions
= 0 corresponding to a valid solution of the original
Objective function: s =
system of nonlinear equations.

Xl X2 X3 X4 Xs
X5 T
0.371 0.629 0.000 0.000 0.000 64.656
0.802 0.000 0.198 0.000 0.000 55.457
0.000 0.631 0.369 0.000 0.000 53.070
0.000 0.842 0.000 0.158 0.000 59.250
0.000 0.000 0.624 0.000 0.376 58.015
0.000 0.000 0.000 0.454 0.546 67.700
0.375 0.189 0.436 0.000 0.000 57.218
0.374 0.438 0.000 0.188 0.000 63.180
0.295 0.148 0.463 0.000 0.094 57.154

14.2.7 NRTL Equation

The non-random two-liquid (NRTL) equation was developed by Renon and
Prausnitz (1968). This equation contains two binary interaction parameters.

L L
+L
TjiGjiXj TkjGkjXk
1n"/i = JEN GijXj (kEN
Tij -
)
L GjiXj . L GkjXk L GkjXk
JEN JEN kEN kEN
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 341

14.2.8 Test Problem 4

The ternary system (Ethanol(l) - Methyl Ethyl Ketone(2) - Water(3)) contains
three binary azeotropes and a ternary azeotrope.

Data

P 760 mmHg
R 1.98721 (cal)mol-IK- I
a = (18.679,16.264,18.585)T
b (3667.70, 2904.34, 3984.92)T
C (226.18,221.97,233.43)T
0.094 -0.254 )
( 0.673
00
T 0.0 0.979
2.092 2.628 0.0
0.972 1.078 )
1.0
( 0.817
G 1.0 0.707
0.539 0.395 1.0

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of nonconvex inequalities 6
No. of known solutions 4

Known Solutions
Objective function: s = 0 corresponding to a valid solution of the original
system of nonlinear equations.

Xl X2 X3 T
0.486 0.514 0.000 74.076
0.952 0.000 0.048 78.275
0.000 0.657 0.343 73.388
0.187 0.560 0.253 72.957

14.2.9 Test Problem 5

This example is for the quaternary system (Methanol(l) - Benzene(2) - i-
Propanol(3) - n-Propanol(4)). Physical experiments have located three binary
azeotropes. The NRTL equation predicts a composition for the methanol-
benzene azeotrope that is quite different from the Wilson equation (test prob-
lem 2).
342 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Data

P 760 mmHg
R 1.98721 (cal)mol-IK- I
a (18.607,15.841,20.443,19.293f
b = (3643.31,2755.64,4628.96,4117.07f
c = (239.73,219.16,252.64,227.44)T
0.040 -0.191
( 0.0 0035 )
-0.671 0.0 1.010 1.212
T
0.093 0.508 0.0 -0.934
0.013 0.396 1.375 0.0
0.988 1.060
( 1.223
1.0 0990 )
1.0 0.745 0.704
G =
0.972 0.863 1.0 1.321
0.996 0.892 0.664 1.0

Problem Statistics
No. of continuous variables 5
No. of linear equalities 1
No. of nonconvex inequalities 8
No. of known solutions 3

Known Solutions
Xl X2 X3 X4 T
0.063 0.937 0.000 0.000 80.166
0.000 0.588 0.412 0.000 71.832
0.000 0.776 0.000 0.224 77.131

Objeetive function: s = 0 corresponding to a valid solution of the original

system of nonlinear equations.

14.2.10 UNIQUAC Equation

The UNIQUAC equation, Abrams and Prausnitz (1975), is based on two ef-
feets, a combinatorial effect due to differences in the shapes of the components
and a residual effect from the energetic interactions between the components.
This equation contains one binary interaction parameter, three pure compo-
nent parameters, and one additional parameter.

where,
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 343

C <l>i Z Oi <l>i ' "

In-Yi =ln----:-+-2qiln"'" +li-----:- L....,.ljxj
x, '1" x, JEN
and,

, '" TijOj )
ln-ytl = qi(1-ln(LTji Oj) - L....,. " 0' TkJ"
, " N U k
JEN JE kEN

where,

0"= ~ ViEN
, j~N qjXj

0'=
,
~
2.., qiXj
ViEN
JEN J
<l>i = ti~iX
J J
Vi E N
JEN

14.2.11 Test Problem 6

The ternary system (Ethanol(l) - Benzene(2) - Water(3)) has been widely
studied due to the importance of the azeotropic distillation process used to
separate ethanol and water using benzene as an entrainer. The system con-
tains three binary homogeneous azeotropes. A fourth solution to this system
of equations exists, however it does not correspond to a stable homogeneous
azeotrope.

Data

p
P = 760 mmHg
R = 1.98721 (cal)mol-1K- 1
z 10
a = (18.912, 15.901, 18.304f
b (3803.98,2788.51, 3816.44f
c (231.47,220.79,227.02)T

q =
= C
( 1.472 1.214)
1.0 1.472
o 1.214 )
0.050 1.0 0.002
0.319 0.708 1.0
(1.970,2.400,1.400)T
(1.970,2.400,1.400f
q' = (0.920,2.400,1.000f
r (2.110, 3.190, 0.920f
0.920)T
(-0.410,1.760, -2.320)T
344 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of nonconvex inequalities 6
No. of known solutions 4

Known Solutions
Objective function: s = 0 corresponding to a valid solution of the original
system of nonlinear equations.

Xl X2 X3 T
0.428 0.572 0.000 67.331
0.000 0.608 0.392 61.317
0.886 0.000 0.114 78.153
0.013 0.604 0.383 61.583

14.2.12 Test Problem 7

There is experimental data for five binary azeotropes in the four-component
system (Ethanol(l) - Methylcyclopentane(2) - Benzene(3) - Hexane(4)). The
UNIQUAC equation predicts six binary, four ternary and one quaternary
azeotrope. This is a difficult example because eleven solutions to the system
of equations must be located.

Data

p 760 mmHg
R 1.98721 (cal)mol-lK- l
z 10
a (18.912,15.810, 15.901, 15.912)T
b = (3803.98,2735.59, 2788.51, 2739.25f
c (231.47, 226.28, 220.79, 226.28)T
1.804 1.472
( 0.025
1.0 1.536 )
1.0 0.618 1.946
T =
0.050 1.217 1.0 1.082
0.042 0.441 0.687 1.0
q (1.970,3.01O,2.400,3.860f
q' (0.920,3.010, 2.400, 3.860)T
r = (2.110, 3.970, 3.190,4.500)T
l (-0.410,1.830,1.760, -0.300f
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 345

Problem Statistics
No. of continuous variables 5
No. of linear equalities 1
No. of nonconvex inequalities 8
No. of known solutions 11

Known Solutions
Objective function: s = 0 corresponding to a valid solution of the original
system of nonlinear equations.

Xl X2 X3 X4 T
0.316 0.684 0.000 0.000 63.489
0.441 0.000 0.559 0.000 67.292
0.334 0.000 0.000 0.666 61.265
0.000 0.738 0.262 0.000 70.367
0.000 0.654 0.000 0.346 72.892
0.000 0.000 0.307 0.693 66.919
0.301 0.545 0.154 0.000 63.029
0.344 0.418 0.000 0.238 64.401
0.300 0.000 0.172 0.528 60.669
0.000 0.458 0.333 0.209 71.235
0.322 0.322 0.222 0.133 63.558

14.2.13 UNIFAC Equation

The UNIFAC equation is a group-contribution method developed by Fre-
denslund et al. (1975) for systems where insufficient data is available to obtain
accurate binary interaction parameters. In this method, moleeules are con-
sidered to be collections of groups of elements (e.g., CH3 ) with a total of
G groups. Interactions between groups are obtained from a standard table,
Gmehling et al. (1990). The UNIFAC equation contains two group parame-
ters, one group binary interaction parameter, and one additional parameter.

where,

C
In"' CPi
" = ln- + -q'ln-
Z
2~
()i CPi
,/.. +i- - - Z
L lx
JJ
Xi 'f'i Xi JEN

and,
346 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

E
- L:
Wgm QmVmjXj
In 'Yf = L QgVgi
QgVgi {
mEG lEG
EW;~NE~-
kEN
gEG

-In (L: L:
\f!mg Qmvm jXj )
mEG JEN

+ In (L: qjXj)
JEN

+ In (L: \f!mgQmvmi)
mEG
-lnqi

+ mEG
' " WOmQmVmi}
L..J ~ W,mQ,V/i
lEG

where,

(}i = ~ q;Xj ViE N

JEN

4Ji =
rx
"liEN
~ JEN

14.2.14 Test Problem 8

The ternary system (Acetone(l) - Methyl Ethyl Ketone(2) - Cyclohexane(3))

contains two binary azeotropes.
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 347

Data

p = 760 mmHg
R 1.98721 (cal)mol-1K- 1
z = 10
a = (16.388,16.264, 15.753)T
b (2787.50, 2904.34,2766.63)T
C = (229.66,221.97, 222.65)T
0
v =
UD 1
6
1.0 0.0105 )
( 1.0
'I! = 1.0 1.0 0.0105
0.0002 0.0002 1.0
q = (2.336,2.876,3.240)T
r = (2.5735,3.2479,4.0464)T
l (-0.386, -0.3884, 0.9856)T
Q = (0.848,0.540,0.640)T
R = (0.9011,0.6744, O. 7713)T

Problem Statistics
No. of continuous variables 4
No. of linear equalities 1
No. of nonconvex inequalities 6
No. of known solutions 2

Known Solutions
Objective function: s = 0 corresponding to a valid solution of the original
system of nonlinear equations.

0.878 0.000 0.122 55.726

0.000 0.424 0.576 71.822

14.2.15 Test Problem 9

This is the same four-component system as in test problem 5, (Ethanol(l)
- Methylcyclopentane(2) - Benzene(3) - Hexane(4)). However, in this case
the UNIFAC equation predicts only the five binary azeotropes which have
348 CHAPTER 14. NONLINEAR SYSTEMS OF EQUATIONS

been reported experimentally. The presence of five solutions to the system of

nonlinear equations and the complexity of the UNIFAC equation makes this
example especially challenging.

Data

p 760 mmHg
R 1.98721 (cal)mol-1K- 1
z = 10
a = (18.912,15.810,15.901, 15.912f
b = (3803.98,2735.59,2788.51, 2739.25f
=

n
c (231.47,226.28,220.79,226.28)T
0 0 1

U
4 1 0
v
0 0 0
4 0 0

( LO
1.0
1.0
1.0
1.0
1.0
0.0611
0.0611 0.9082
0,0082)
-q, 1.0 1.0 1.0 0.0611 0.9082
0.6009 0.6009 0.6009 1.0 0.7783
0.9552 0.9552 0.9552 0.1202 1.0
q - (1.970,3.010, 2.400, 3.860f
r (2.110, 3.970, 3.190, 4.500f
I = (-0.410,1.830,1.760, -0.300f
Q (0.848,0.540,0.228,1.124,0.400)T
R (0.9011,0.6744,0.4469,1.2044,0.5313f

Problem Statistics
No. of continuous variables 5
No. of linear equalities 1
No. of nonconvex inequalities 8
No. of known solutions 5

Known Solutions
Objective function: s = 0 corresponding to a valid solution of the original
system of nonlinear equations.
14.2. ENCLOSING ALL HOMOGENEOUS AZEOTROPES 349

Xl X2 X3 X4 T
0.371 0.629 0.000 0.000 60.632
0.450 0.000 0.550 0.000 67.841
0.357 0.000 0.000 0.643 59.054
0.000 0.865 0.135 0.000 71.651
0.000 0.000 0.065 0.935 68.900
 Chapter 15

Dynamic Optimization
Problems

15.1 Introduction
A dass of problems for which there has been a scarcity of work in the area of
global optimization are the dynamic optimization problems. These problems
involve dynamic variables whose values change in time. Such problems exist in
the areas of optimal control, parameter estimation for dynamic models, reactor
network synthesis where the dynamic models arise from the differential mod-
eling of the tubular reactors (plug flow reactors), and for dynamic simulation
and optimization.

15.1.1 General Formulation

The problems that are considered here conform to the following general for-
mulation:
mm J(z(td, Z(ti), U(ti), x)
s.t. j(z(t), z(t), u(t), x, t) = 0
c(z(to),x) = 0
h'(Z(ti), Z(ti), U(ti), x) = 0
g' (z (ti), z( td, u( ti), x) < 0
h" (x) = 0
g"(x) ~ 0
x E x~np
ti E [to, tNl
where J is the objective function, j is the set of dynamic state equations, c
are the initial conditions, g' and h' are the point constraints, and g" and h"
are the general algebraic constraints. The variables in the problem are x, the
vector of time invariant continuous variables, z, the vector of dynamic state
variables, and u, the vector of control variables. The independent variable is
t.
351

C. A. Floudas et al., Handbook of Test Problems in Local and Global Optimization

Springer Science+Business Media Dordrecht 1999
352 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

15.1.2 Solution Techniques (Local)

There are several ways to solve the optimal control problem. These methods
are (a) the solution of the necessary conditions, (b) dynamic programming, (c)
complete discretization, and (d) control parameterization. The solution of the
necessary conditions and the application ofPontryagin's maximum principle as
described in Bryson and Ho (1988) leads to the solution of two-point bound-
ary value problems. Dynamic programming approaches employing grids for
both the state and control variables have been proposed and applications are
reported in Luus (1990b,a)j Luus and Rosen (1991)j Luus (1993). Complete
discretization methods discretize aH the dynamic equations with respect to all
dynamic variables. The DAEs are thus transformed into algebraic equations,
and the problem can be solved using nonlinear programming techniques. Finite
element collocation techniques were proposed by Cuthrell and Biegler (1987)
for the fuH discretization of DAE problems. The collocation methods were ex-
tended in Logsdon and Biegler (1989) and Vasantharajan and Biegler (1990) to
include appropriate error constraints. The control parameterization techniques
apply discretization only to the control variables. These methods are described
in Sargent and Sullivan (1977) and in Vassiliadis et al. (1994a,b). The prob-
lem is formulated as an NLP where the control parameters are determined
through the optimization procedure and the DAE system is solved through an
integration technique. The use of the control parameterization techniques for
solving optimization problems in process synthesis are described by Schweiger
and Floudas (1997, 1998b).
Since the methods used for solving the dynamic optimization problems of-
ten involve the use of approximations, the best solution that can be obtained
is closely related to the solution algorithm. In other words, the same prob-
lem may have different solutions and a different global solution when different
solution techniques are used.
The examples presented in this chapter have all been solved using the
control parameterization approach. Because of this, abrief description of the
control parameterization is provided.

15.1.3 Control Parameterization

The basic idea behind the control parameterization approach is to express
the control variables u(t) as functions of time invariant parameters. This
parameterization can be done in terms of the independent variable t (open
loop):

u(t) = 4J(w, t)

or in terms of the state variables z(t) (closed loop):

u(t) = 'ljJ(w, z(t))

15.1. INTRODUCTION 353

For the following example problems, open loop parameterization will be con-
sidered.
A convenient choice for the control parameterization function ep(w, t) is
a polynomial expression such as Lagrange polynomials. The time horizon
is divided into intervals with the controls defined as polynomials over each
interval. The Lagrange polynomial expression of order M in interval i has the
form
n
epi(W, t) = L Wij for N =1
j=1
n M __
epi(W, t) = L Wik TI . t~~tk for N ~2
J=1 k=l,k#J
where t is the normalized time over the interval i
- t - ti-l
t = ----=--=-
ti - ti-l
This parameterization allows for polynomial expressions of various order to be
used. For example, a piecewise linear expression with continuity between the
intervals is expressed as
t - ti+! t - ti
epi = Wi + Wi+! =-----=-
ti - ti+! ti+l - ti
Since the control parameters change from one interval to the next discontinu-
ities arise in the DAE system.
The set of time invariant parameters x is now expanded to include the
control parameters:
x = {x,w}
The set of DA Es (f) is expanded to include parameterization functions
f() = {J(.), ep(.)}
and the control variables are converted to dynamic state variables:
z = {z,u}
Through the application of the control parameterization, the control vari-
ables are effectively removed from the problem and the following NLP /DAE
results:
mm J(Z(ti), Z(ti), x)
S.t. J(z(t), z(t), x, t) 0
c(z(to), x) 0
h'(Z(ti), Z(ti), x) = 0
g' (z (ti), z( ti), x) < 0
h"(x) 0
g" (x) < 0
x E x~np
ti E [ta, tN]
354 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

This problem is an optimization in the space of the x variables where J,

g', and h', are implicit functions of x variables through the solution of the
DAE system.

15.1.4 Solution of NLP /DAE

Standard gradient based methods for solving NLPs locally such as reduced
gradient methods, conjugate gradient methods, and sequential quadratic pro-
gramming methods, require function evaluations along with the gradient eval-
uations. For this problem, the function evaluations and gradients with respect
to x are required for J, h', g', h", and g". For h", and g", analytical ex-
pressions can be obtained. However, for J, h', g' the function evaluations and
gradients are determined as implicit functions of x through the solution of the
DAE system. Integrating the DAE system along with the variational equations
provides the values of the state variables at the time instances, z(td, as well
as the gradients ~~. With this information known, the functions J, g', and h'
are evaluated directly, and the gradients are determined by applying the chain
rule:

dJ
dz = (g;) (g;) +(g; )
dh'
dz (~)(~)+(~)
!YL
dz (~)(~)+(~)
In problems where the time horizon needs to be optimized, the independent
variable is scaled and the integration is performed over this scaled variable. The
scaling factor becomes a variable which can be optimized.

15.2 Chemical Reactor Network Problems

15.2.1 Introduction
Reactor network synthesis has been a widely studied problem in the area of
process synthesis. It deals with determining the types, sizes, and interconnec-
tions of reactors which optimize a desired performance objective. The reac-
tor network synthesis problem has been discussed previously in Section 8.4 of
Chapter 8.
Problem Statement
In reactor network synthesis, the goal is to determine the reactor network
that transforms the given raw materials into the desired products. The follow-
ing information is assumed to be given in the problem definition:
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 355

the reaction mechanism and stoichiometry

the chemical kinetic data

the rate laws

the energetic data

the inlet stream(s) conditions

the performance objective (output).

The synthesis problem is to determine:

the type, size, and interconnection of reactor units

the stream owrates, compositions, and temperatures

the composition and temperatures within the reactors

the heating and cooling requirements

the optimal performance objective.

The two reactor types that are often encountered in reactor network syn-
thesis are the Continuous Stirred Tank Reactor (CSTR) and the Plug Flow
Reactor (PFR). The CSTR is assumed to be perfectly mixed such that there are
no spatial variations in concentration, temperature, and reaction rate within
the reactor. The PFR is a tubular reactor where there are no radial variations
in concentration, temperature, and reaction rate. Other types of reactors used
in the synthesis problem are the Maximum Mixed Reactor (MMR), Segre-
gated Flow Reactor (SFR), and the Cross Flow Reactor (CFR). The CFR is
a generalization of the PFR, MMR, and SFR.
Since the mathematical modeling of the tubular reactors results in differen-
tial equations, the resulting problem is formulated as a dynamic optimization
problem.
The reactor network consists of mixers, splitters, CSTRs and CFRs. Split-
ters are used to split the feed and the outlet of each reactor. Mixers are
situated at the inlet to each reactor to mix the streams which come from the
feed splitter and the splitters after each reactor. A final mixer is used to mix
the streams which go from the splitter after each reactor unit to the product
stream. A owsheet which has two CSTRs is shown in Figure 15.1

15.2.2 General Formulation

In the general formulation, two assumptions will be made: constant density
and, for nonisothermal systems, constant heat capacity which is the same for
all components.
356 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

lr
Feb
Fee
F ab F b c~ r b F l c~ T S

/ IIIIIIIII
F ac ..... F c c~ rcJ CFR 1F c~
t Tl l/ Fell

/ I Fr c~ TI V'
J
,/

IIIIIIIII FI cl rl rI' F'h

l
FO cf TO
Flc ~
Fa c~ r a } - Fgc F" c 11 r"
Fgb

lc. "
F'd
Fld

Fad ,.J CSTR IF9 c9 r9 --! F"h

l VIII
J

Figure 15.1: Example of a reactor network superstructure consisting of one

CSTR and one CFR

The general formulation has been derived for any number of CSTRs and
CFRs, any number of reaction components, any number of reaction paths, any
number of feed streams, and any number of product streams.
In the general formulation, the following sets will be used:
Set description
I Components
J Reactions
K CFR reactor units
L CSTR reactor units
R Feeds
P Products

Constraints
Feed splitter

L-t F r,k + F r,k + ~

=~ Vr E R
Fa ab ac
r L-t Fad
r,1
kEK IEL
CFR sidestream mixer total balance
F~ = L F:,% + L Fk'~k + Ff,~k + L Fj~~ Vk E K
rER k'EK IEL
CFR sidestream mixer component balance
b Fb ~ a F ab
ck,i k = L-t cr,i r,k
+ ~ e F eb
L-t ck',i k',k
+ ck',i
f Ff b + ~ ~q Fg b
k',k L-t Cf,i l,k \.J'
vZ E
I \.Jk
v E
K
rER k'EK IEL
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 357

CFR main mixer total balance

Fkc = '~ ac
" Fr,k+ '~
" Fec " p9l,kC VkEK
k',k+ '~
k',k+ Ffc
rER k'EK IEL

CFR main mixer component balance

r,k + '" + ck',i c + '" JI p9C

c FC
ck,i k = '~
" cr,i
a Fac ~ e Fec
ck',i k',k f Ffk',k ~ q,i I,k W
vZ E I wk
v E K
rER k'EK IEL

CSTR feed mixer total balance

Fjd = L F~1 + L F:j + Ft,t + L F q1 Vi E L l

rER kEK l'EL

CSTR feed mixer component balance

d pd
cl,i I -_ '~
" Cr,i
a Fad
r,1 + '"
~ Ck,i k,l + Ck,i
e Fed f Ffk,1d + '" ~Q p9
~ q',i
b
1',1 w
vZ E I wi
v EL
rER kEK I'EL

CSTR total balance

F/9 = Fl Vi E L

CSTR component balances

F 19c'f.i = Flc~i + Vi m L Vi,jrrJ Vi E I Vi E L

JE}

CSTR reaction rates

rrJ = fJ(cl'J, Tt) Vj E J Vi E L

CFR total balance

dF) = _ dF: _ dFk
dV dV dV Vk E K
CFR component balance

Leaving sidestream

dcL t dFk
F k d-i = (Ck,i - Ck,i) V Vi E I Vk E K

Initial conditions (V = 0)
358 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Ck8 ,I. = Ckb ,I. 'Vi E I 'Vk E K

F~ = Ff 'Vk E K

Ckt ,'l. = Ckc ,1. Vi E I 'Vk E K

Point constraints (V = 1)

ViE I Vk E K

'Vi E I Vk E K

CFR reaction rates

rL = fJ(cL, T~) Vj E J 'Vk E K

CFR main splitters

F: = L F:~I +F:J.,+ LF:j + LF:; 'Vk E K

k'EK lEL pEP

CFR side exit splitters

Ft = L F!.~, + F!.~I + L F!.t + L F!.Z 'Vk E K

k'EK lEL pEP

CSTR product splitters

Fl = L Fj~Z + Fj~Z + L Fl~~ + L Fl~; Vl E L

kEK l'EL pEP

Product mixer total balance

F ph = "~ F.k,p
eh + Ff h + " F g h Vp E P
k,p ~ l,p
kEK IEL

Product mixer component balance

Fph cp,i
h _ " Feh e
- ~ k,pck,i
h f
+ Ffk,pck,i + "FghJ1
~ l,p q,i Vi E I Vp E P
kEK IEL
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 359

Control Parameterization
The control parameterization requires the selection of the control variables.
For these problems, the sidestream fiowrates for the CFR and the temperature
in the CFR are selected as the control variables. For the control parameteri-
zation, 10 equaHy spaced intervals are used and piecewise linear polynomials
are used for the fiowrates and a piecewise quadratic polynomial is used for the
temperature. Since the derivatives are needed explicitly, they are calculated
as weH. Continuity in the first derivatives of the temperature is also enforced.
Control parameterization equations for the CFR side feed stream

s _ sB tt -- ti,2
ti,2 sB t - ti,o
t - ti,o wk E K Wo. E CI
DS
F k =
L'k - "'k i
, ti,O - ti,2
+ "'k''ti,2
i+l
i+l=---=-
t i , 2 - ti,O
v EK
Vk v. E CI
Vz

dFB
dF
_k k= ",Bk .
"'k,i,t + ",S
"'k,i+1
k,i+1 Vk E K Vi E CI
dv ti,O - ti,2 ti,2 - ti,O

Control parameterization equations for the CFR side exit stream

F o ti 2
t - ti,2 t - ti,O k .
'++ "'k''ti,2
DO 0 0
L'kk = "'k i i+1 Vk E K
V Vi E CI
Vz
, ti,O - ti,2 ti,2 - ti,O

o
dFk =
0
"'k,i + _"'k,i+! Vk E K Vi E CI
dv ti,O - ti,2 ti,2 - ti,o

Control parameterization equations for the CFR temperature

Tt - ",t (t-fi,I)(t-fi,2)
(t-fi,J)(t-fi,2)
k - k,i,O (ti,o-ti,tl(ti,o-ti,2)

+",tt (t-fi,o)(t-fi,2)
+"'k,i,l (ti,l -ti,O)(ti,!
k,i,l (ti,! -ti,O)(ti,l -ti,2)
ti,2)

t (t-fi,O)(t-fi,I)
+"'k,i+1,O (ti,2-ti,O)(ti,2 ti,I)

Vk E K Vi E CI

~ _ ",tt (t-f
(t-fi,J)+(t-fi,2)
i I)+(t-fi 2)
dv = k,i,O (ti,O-t;,tl(ti,O-ti,2)
- "'k,i,O (ti,O-ti,tl(ti,O-ti,2)

+",tt (t-fi.!0)+(t-fi,2)
+"'k,i,l (ti,l-ti,O)(ti,1-ti,2)
k,i,l (ti,! ti,O)(ti,! ti,2)

t (t-fi,O)+(t-li,I)
+"'k,i+1,O (ti,2 -ti,O)(ti,2 ti,I)

Vk E K Vi E CI
360 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Continuity in the first derivatives of the temperature

t fi 1,2 -fi - 1,O

+Kk,i-l,l (ti-1,1- ti-1,O)(ti-1,1 ti-1,2)

t ( 2fi 1,2-f. 1,1-fi - 1,O)

+Kk,i,O (ti-1,2- ti-1,O)(ti-1,2- ti-1,1)

_
- K, t
(2f." 0-[-" 2-[-
(2f',o-fi,2- fi,!)
.,
1)
- Kk,i,O (ti,0-ti,Il(ti,0-ti,2)
k,i,O (ti,O-ti,!)(t.,O-ti,2)

t (fi,O-fi ,2)
+Kk "i 1 (t " 1 ti ,0)(ti,1 ti,2)

t (f',o-fi ,!)
+K k ,i+1,O (ti,2- t.,O)(ti,2 ti,d

Vk E K Vi E CI,i > 0

The array of collocation points is

0.00 0.05 0.10

0.10 0.15 0.20
0.20 0.25 0.30
0.30 0.35 0.40
- =
I 0.40 0.45 0.50
tCi,co
tCi,co = 0.50
0.50 0.55
0.55 0.60
0.60
0.60 0.65 0.70
0.70 0.75 0.80
0.80 0.85 0.90
0.90 0.95 1.00

Note that there are other possible ways to handle the control parameter-
ization. For example, the control parameterization could have been done for
the derivatives and the integrator could have been used to calculate the flow
rates and the temperature. This would require that the initial conditions for
these variables be specified. Since they would depend on the parameters, they
would need to be optimized.
Objective Function
The objective function varies from problem to problem. In many cases it
is the maximization of the yield of the desired product. It may be based upon
economic criteria reflecting the value of the products, cost of reactants, cost of
the utilities, and the cost of the reactors.
Variables
Time Invariant Variables
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 361

Flowrates
ab
Fr,k "IrE R VkEK Volumetrie flowrates from feed
splitters to CFR side mixers
Fae "IrE R VkEK Volumetrie flowrates from feed
r,k
splitters to CFR main mixers
Fad "IrE R Vi E L Volumetrie flowrates from feed
r,1
splitters to CSTR mixers
Fkb VkEK Volumetrie flowrates into CFR side inlets
Fek VkEK Volumetrie flowrates into CFR main inlets
F,d Vi E L Volumetrie flowrates into CSTRs
I
Fek VkEK Volumetrie flowrates out of CFR main outlet
Ffk VkEK Volumetrie flowrates into CFR side outlet
p9 Vi E L Volumetrie flowrates out of CSTRs
I
Feb VkEK Vk' E K Volumetrie flowrates from CFR main
k ,k'
splitter to CFR side inlet mixers
Fee VkEK Vk' EK Volumetrie flowrates from CFR main
k,k'
splitter to CFR main mixers
Fed VkEK Vi E L Volumetrie flowrates from CFR main
k,1
splitters to CSTR mixers
fb ,
Fkk VkEK Vk' E K Volumetrie flowrates from CFR side
,
splitters to CFR side inlet mixers
Ff e VkEK Vk' E K Volumetrie flowrates from CFR side
k,k'
splitters to CFR main inlet mixers
Ff d VkEK Vi E L Volumetrie flowrates from CFR side
k,1
splitters to CSTR inlet mixers
p9b Vi E L VkEK Volumetrie flowrates from CSTR outlet
I,k
splitters to CFR side inlets mixers
p9C Vl E L VkEK Volumetrie flowrates from CSTR outlet
I,k
splitters to CFR main inlet mixers
p9 d Vi E L Vi' E L Volumetrie flowrates from CSTR out let
1,1'
splitters to CSTR inlet mixers
Feh VkEK VpEP Volumetrie flowrates from CFR main outlet
k,p
splitters to product mixer
Ffk,ph VkEK VpEP Volumetrie flowrates from CFR side outlet
splitters to product mixer
F9I,ph Vi E L VpEP Volumetrie flowrates from CSTR outlet
splitter to product mixer
Fh VpEP Volumetrie flowrates of the product streams
2
362 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Concentrations
b VkEK Vi E I Concentrations of species
ck,i
in the CFR side streams
c~,i VkEK Vi E I Concentrations of species
in the CFR main inlet streams
d Vl E L Vi E I Concentrations of species
cl,i
in the CSTR inlet streams
ck,i VkEK Vi E I Concentrations of species
in the CFR main outlet streams
f VkEK Vi E I Concentrations of species
ck,i
in the CFR side outlet streams
q. ,t
Vl E L Vi E I Concentrations of species
in the CSTR outlet streams
h VpEP Vi E I Concentrations of species
cp,i
in the product streams
Temperatures
r,m
I Vl E L Temperatures
in the CSTRs
Reactor Volumes
Vom
I VI E L Volume of the CSTRs
v;tk VkEK Volume of the CFRs
Reaction Rates
rF
,1
VI E L Vj E J Rate of reaction j in CSTR reactor l
Control Parameters
Vk E K Vi E CI Control parameters for the side feed
Vk E K Vi E CI Control parameters for the side exit
Vk E K Vi E CI Control parameters for the temperature

Dynamie Variables

Flowrates
Fkt VkEK Volumetrie flowrates within CFRs
FkS VkEK Volumetrie flowrates of CFR side inlets
FOk VkEK Volumetrie flowrates of CFR side outlets
Concentrations
t VkEK Vi E I Concentrations of species
ck,i
within CFRs
c k,t VkEK Vi E I Concentrations of species
within CFR side outlet
Temperatures
Tt VkEK Temperature within the CFRs
I>
Reaction Rates
t VkEK Vj E J Rate of reaction j in CSTR reactor l
rk,i
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 363

Variable Bounds
The variable bounds vary from one problem to the next. The flowrates,
concentrations, temperatures, volumes and reaction rates have a lower bound
of zero due to physical restrictions. The upper bounds on the flow rates are set
arbitrarily (generally 10 times the feed flowrate) to help the solution algorithm.
The upper bounds on the concentrations are set to the maximum possible
concentration that any species could achieve. The temperature ranges for the
nonisothermal problems are usually specified for each problem. The upper
bounds on the reactor volumes and reaction rates are set arbitrarily to help
the solution algorithm.
Since the recycle around a CSTR unit is not necessary, the upper and lower
bounds for these flowrates are set to zero.

15.2.3 Specific Information

Specific information needs to be provided for each problem. This information
includes the reaction mechanism, the reaction paramters, the rate expressions,
and any additional information required for the problem.
Reaction Mechanism
For each problem, areaction mechanism indicating the species and reaction
steps involved is provided.
Parameters
The parameters for the problems are the stoichiometric coefficient matrix,
the kinetic constants, and the feed conditions. The stoichiometric coefficient
matrix, Vi,j indicates the relative number of moles of products and reactants
that participate in a given reaction.
Reaction constants
Vi E I vj E J Stoichiometric coefficient matrix
Vj E J Kinetic rate constant
Vj E J Activation Energy
Vj E J Heat of reaction
Feed conditions
Fra Vr E R Flowrate of feed streams
c~,i Vr E R Vi E I Concentrations of species in feed streams

Rate Expressions
The rate expressions UJ{Ci, T)) are specified for each problem. These ex-
pressions often involve the products of the concentrations of the species. The
rate expressions have the form

fJ{Ci, T) = kj{T) rr
iEI
f
c~i,j
364 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

where v!,j correspond to the positive coeflicients in the stoichiometric matrix

representing the forward reactions. (For equilibrium reactions, both the for-
ward and reverse reactions must appear separately and the rate constants for
both must be provided.)
For isothermal reactions, the values of kj are fixed. For most nonisothermal
problems, Arrhenius temperature dependence will be assumed:

, (-E-)
kj (T) = kj exp R,J.

Additional Information
For most problems, this is all of the information that will be required.
Some problems will have additional restrictions on the sizes of the reactors,
the amount of conversion, etc.

15.2.4 Problem Characteristics

This problem has bilinearities which result from the mass balances of the mix-
ers. The rate expressions usually involve the products of the concentrations
and thus lead to squared or bilinear terms for first order kinetics, trilinear
terms for second order kinetics, and so on. For nonisothermal problems, the
temperature dependence is an exponential which is multiplied by the product
of the concentrations.
For each problem, the number of CSTRs and CFRs that can potentially ex-
ist in the network must be given. Existence of the CSTRs in the network could
be handled with binary variables, however, this would lead to an MINLP. Ex-
istence of the reactors is handled through the volume of the reactors. Reactors
with zero volume are mixers and do not exist.

15.2.5 Test Problems

Unless otherwise stated, all of the example problems use one CSTR and one
CFR in the reactor network superstructure. The same set of test problems
that were used in Section 8.4 of Chapter 8 are presented here. These problems
have been formulated in the MINOPT modeling language and the solutions
provided have been obtained using MINOPT (Schweiger and Floudas (1998c,
1997)). These problems have multiple local optima which have been found by
using MINOPT to solve the problem with random starting points. The best
known solution and simplest configuration are given for each problem.

15.2.6 Test Problem 1 : Nonisothermal Van de Vusse Reaction

Case I
Problem Information

Reaction Mechanism
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 365

k1 k2
A----- B ----- C
k3
2A--- D

Objective:
Maximize the yield of B

h
maxcl,B

Parameters

Rate Constants
reaction k E
1 5.4x109 h 15.84 kcaljmol -84 K Ljmol
2 1.6x10 12 h- 1 23.76 kcaljmol -108 K Ljmol
3 3.6x105 Lj{mol h) 7.92 kcaljmol -60 K Ljmol
Feed Conditions
F: 100 Ljs
c~,i 1.0 moljL A, 0 moljL B, 0 moljL C

Rate Expressions

Additional Information
The temperatures in the reactors are bounded between 300 K and 810 K.
366 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Explicit Formulation

Fa
1
-- F
-
ab
F1,1
ab + FF1,1
ac + Fad
ac Fad
]1
F b
k = Ff~ + Ff~ + FF1 ~ + Ff~
b Fb __ a'
a' Fab ab' 'eb + J Flb
+' Cee1,A F'eb Ib + ~Q Fggbb
_Q
C1,A 1 -~A~I+~A~I+~A~I+SA~1
- C1,A 1,1 1,1 Cl A 1,1 GJ.,A 1,1
b Fb
C1,B 1
-_ a'
a
- C1,B
F~b
Fab
C1,B 1,1
+ e F~b
e'
C1,B
C1,B 1,1
Feb + I
C1,B
Flb
Fi
C1,B 1 1
b + -Q Fg
-Q
S,B 1,1
Fg bb
b
C1 ,C
Fb
1
__ aa FabFab + C1e ,C Feb
- C1 ,C 1,1
Feb
1,1
+ ClII C Flb
Flb
1,1
+ Cl,c
JI-Q
S,C 1,1
Fg bb
Fg
b
C1,D
Fb
1
__ aa Fab Fab
Cl ,D 1,1
- C1,D 1,1
+ C1,De FebFeb
Cl'Df 1,1 1'1
+ C1,D
I Flb
Flb
Cl ,D 1,11,1
+ Cl,D
-Q Fg1 bb1
Fg
(;;1 ,D 1,1,
ac + Fec +
1,1 + Ff,l1,1 + Fl,~ + Fg
lc c
F 1C == FF~I F
1,1 + Ff~
11
c
C1,A
FC
1
__ aa Fac
- C1,A
Fac + C1,A
C1,A 1,1
e FecFec + ClJ A Flc
C1,A 1,1
Flc
1,1
+ S,A
~Q Fg
-Q Fg cc
GJ.,A 1,1
ci,BFf __ aa Fac
- C1,B
Fac + C1,B
C1,B 1,1
e FecFec + C1,B
C 1,B 1,1
I Flc
1 Flc + Cl,B
C 1,B 1 1
-Q
S,B 1,1
Fg cc
Fg

Ci,cFf __ aa FacFac + C1e ,C Fec

- C1 ,C 1,1
Fec
1,1
+ ClII C Flc
Flc
1,1
+ Cl,c
-Q
S,C 1,1
Fg cc
Fg

Ci,DFf __ aa Fac Fac + Cle D Fec

- Cl ,D 1 ,1
Fec11
+ Cl1I D Flc
Flc
1 '1
+ Cl-Q Fg cc
Fg
(;;1 ,D 1 ,1
, lf' gdd
F1,1,t + FFf1 1
_ ~d 'ed ' i ' , ,
Fd
1 =- FF~11,1 + F 1,1 + F
Ff,1
d
C1,A
Fd
1
__ aa Fad
- C1,A
Fad + C1,A
C 1,A 1,1
e FedFed + ClJA Fld
C1,A 1,1
Fld
1,1
+ Cl,A
~Q Fg
-Q
S,A 1,1
Fg bb
d
C1,B
Fd
1
_ a
_ a Fad
- C1,B
Fad
C 1,B 1,1
+ e
C1,B
C1,B 1,1
Fed
Fed + I
C1,B
Fld
Fld
C 1,B 1 1
+ JI
S,B
Fg bb
Fg
GJ.,B 1,1
d
C1 ,C
Fd
1
__ aa FadFad + C1e ,C Fed
- C1 ,C 1,1
Fed
1,1
+ ClII C Fld
Fld
1,1
+ S,C
-Q Fg bb
Fg
GJ.,C 1,1
d
Cl D
Fd
1
__ aa Fad
- C1,D
Fad + C1,D
C1,D 1,1
e Fed
C1,D 1,1
Fed + C1,D Fld + Cl,D
Ij Fld
C 1,D 1,1
-Q
S,D 1,1
Fg bb
Fg
, Fg
1 =Ff
=Ff
FfdfA = Ffct,A + V 1m (VA,l r rl + VA,2 r r2 + VA,3 r r3)
= Ffct,B + Vr(VB,l r rl + VB,2 r r2 + VB,3 r r3)
9 -q'
F 1 ClB =
Ffdf'c i'\ + VC,2r~2 + VC,3r~3)
= Ffct,c + V 1m (VC,l r rl
= Ffct,D + V 1 (VD,l rl + VD,2 r r2 + VD,3 r r3)
9 -Q'
m
,
F 1 ClD r
rrl = 5.4 X 10g e -7971.82/T[" df A
10ge-7971.82/T[" cf
cf B
rr2 = 1.6
= lOI2 e -1l957.7/T["
X 1012
X B

rr3 = 3.6
= X 105 e -3985.91/T["
105e-3985.91/T[" c?I,A' c?I,A

dFS dFo
~ = - =.;..
dV
- =.;..
dV
1V
t dC1,A )~
F ldV = - ( Cs1,A - Ct
1,A dV + TTt(
VI VA,lrl,1 + VA,2rl,2 + VA, 3rl,3
t t t )
tdCLB t) ~
C1,B dV + vI VB,lrl,1 + VB,2rl,2 + VB, 3rl,3
T Tt ( t t t )
F ldV = - ( Cs1,B -
tdCLc t) ~
C1,C dV + vI VC,lrl,1 + VC,2rl,2 + VC, 3rl,3
t )
F ldV = - ( Cs1,C - TTt( t t
tdCLD t) ~
C1,D dV + vI VD,lrl,1 + VD,2rl,2 + VD, 3rl,3
TTt( t t t )
F1 dY = - ( Cs1,D -
F Odc1,A _t 0 dFo
1 dV - (C1,A - C1,Ar=if
o dC'!,B t dFo
F 1 dJ' = (C 1,B - cr,BrV-
o dc1,C _t 0 dFo
F ldV - (C 1,c - C1,d:V-
o dCl,D _t 0 dFo
F 1 dV - (C1,D - C1,D):V-
15.2. GHEMIGAL REAGTOR NETWORK PROBLEMS 367

FS1 =Ff F 1t =Ff

ctA -- cb1A Ci ,A = ctA
b' t
cf,B = Cl B c1,B = ctB
b' t
ctG = c 1,G c1,G = ctG
b
cf,D -- c 1,D Ci D = ctD
Fe1 =Fi F'I1 =Ff
I
Cr ,A = ci ,A Cl A = cf ,A
ctB = Ci ,B
J'
Cl B = cf,B
ctG -- ct1,G l
Cl G = ctG
ctD -- Ct1,D J'
C1,D -- CO1,D

Ti ,1 = 5.4 xX 10g e-7971.82/Tf ci A

t
T1,2 = 1.6 xX 1012 e -11957.7/T1c
= t '
i
B
t
T1,3
= 3.6 xX 105 e -3985.91/T1 c 1,A c 1,A
t t ' t

Ff - F eb + Fec + F ed + Feh
- jt j~ jJ jh
FI = F 11,1,1 + F1,1 + F 1,1 + F 11,1,1
Fe gb Fgc Fg d Fg h
= F11+
= 11+ 11+ 11
Ft' e'h Fih Fgh
= F 1,1 + 1,1 + 1 1
'
F1h ch1,A eh Flh J Fg
= F1,1 C1 ,A + 1 1Cl A + 1,1 S,A
e h JI

F1h ch1,B eh e + Flh 1 + Fg h ~Q

= F 1,1 ,B
C1,B
C1 1 1CC1,B
1,B 1,1 S,B
Ff'c~,G =
= F eh
1,1 C1 e
C1,G
,G + Fi h f + Fg hh -- Q
1 1Cl G 1,1 S,G
h h eh C1e ,D + F}h 1,D + Fg1,1hS,D
F 1 c1,D = F1,1
= C1,D C1,D
1,1 C1
-q

Ti = I\;L,o(t - 7i,t}(t Ti,2)/(Ti,O - 7i,1)/(7i,0

Ti,t}(t - 7i,2)/(7i,0 Ti,l)/(Ti,O - 7i,2)
Ti,2)
+I\;L,l (t - 7i,0)(t
Ti,O)(t - 7i,2)/(7i,1
Ti,2)/(Ti,1 - 7i,0)/h,1
Ti,O)/h,l - 7i,2)
Ti,2)
Ti,O)/(t - 7i,1)/(7i,2
+1\;i,i+l,O(t - 7i,0)/(t Ti,1)/(Ti,2 - 7i,0)/(7i,2
Ti,O)/(Ti,2 - 7i,t}
Ti,t}

W
dT t
= l\;i,i,O((t - 7i,t} + (t - 7i,2))/(7i,0 - 7i,1)/h,0 - 7i,2)
+I\;i i 1((t - 7i,0) + (t - 7i,2))/(7i,1 - 7i,0)/h,1 - 7i,2)
+1\;(i~l,O((t - 7i,0) + (t - 7i,1))/(7i,2 - 7i,0)/(7i,2 - 7i,l)

Ft = I\;f,i(t - 7i,2)/(7i,0
Ti,2)/(Ti,O - 7i,2)
Ti,2) + 1\;f,i+1(t - 7i,0)/(7i,2
Ti,O)/(Ti,2 - 7i,0)
Ti,O)
W- I\;td(Ti,O - 7i,2)
= I\;td(7i,0 Ti,2) + l\;f,i+l/h,2 - 7i,0)
Ti,O)
Fk = I\;ti(t - 7i,2)/(7i,0 Ti,2) + I\;L+l (t - 7i,0)/(7i,2
Ti,2)/(Ti,O - 7i,2) Ti,O)/(Ti,2 - 7i,0)
Ti,O)
~ = 1\;f,;/(7i,0
l\;f,d(Ti,O - 7i,2)
Ti,2) + I\;ti+1/h,2 - 7i,0)
Ti,O)
368 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Problem Statistics

No. of continuous variables 100

No. of dynamic variables 18
No. of linear equalities 20
No. of linear inequalities 20
No. of nonlinear equalities 23
No. of point equalities 9
No. of dynamic equations 18

Best Known Solution

Objective function: 0.8267309 mol/L

Variables

F:,% = 0.00 F:'k = 100.000 F:1 = 0.00

Ft = 0.00 F k = 100.000 Fi = 100.000
F: = 100.000 Ft = 0.00 F/9 = 100.000
F:~, = 0.00
jb
F: k, = 0.00 F:1 = 100.00
jd = 0.00
F kk, = 0.00 Fl,~, = 0.00 Fkl
'b
Fj9k = 0.00 Fj~~ = 0.00 Fj9~ = 0.00
F:ph h
,p = 0.00
= 100.00
fh
Fk,p = 0.00 F/~; = 100.000
p
4 i = (1.000000,0.000000,0.0000000, O.OOOOOOO)T
Ck:i = (1.000000,0.000000,0.0000000, O.OOOOOOO)T
C1i = (0.110190,0.748042,0.0639591, 0.0389047)T
ci i = (0.110190,0.748042,0.0639591, 0.0389047)T

C{i = (1.000000,0.000000,0.0000000, O.OOOOOOO)Y

df i = (0.00819274,0.826731,0.0763064, 0.0443850)T
c;,i= (0.00819274,0.826731,0.0763064, 0.0443850)T
Vjm = 10000.0 V~ = 2.45039 11m = 323.968
rIJ
= (0.000910363,0.000123473, 0.000109607)Y
K k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 369

K,% ,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

582.4623 300.0000
325.3572 451.5398
471.5539 447.4510
441.2828 441.4477
436.3444 430.4997
K,t
k,m,t
. = 428.4403 427.5221
425.1011 422.8199
422.3212 421.5077
418.2821 416.9589
421.8527 395.3906
300.0000 400.0000

15.2.7 Test Problem 2 Isothermal Van de Vusse Reaction

Case I
Problem Information

Reaction Mechanism

A ----- B
k1

k3
--=---- C
k2

2A----- D

Objective:
Maximize the yield of B:

maxc~,B

Parameters

lIi,j = [-1 0-2]

1 -1
0 1
0
0
001

Rate Constants
k1 10 s-1 (first order)
k2 1 s-l (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
Far 100 L/s
c~,i 0.58 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D
370 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Rate Expressions

Ir = k1cA
12 =k 2CB
Ir = k3~
Problem Statistics
No. of continuous variables 77
No. of dynamic variables 16
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 21
No. of point equalities 9
No. of dynamic equations 16

Best Known Solution

Objective function: 0.4370782 mol/L

Variables

F:i = 0.00 FaCk

T,
= 100.000 F:1 = 0.00
F kb = 0.00 F~ = 100.000 EId = 0.00
F: = 100.000 Ft = 0.00 F t9 = 0.00
Fk~' = 0.00 F kk, = 0.00 F:1 = 0.00
jb
Fk,k' = 0.00 Fh,, =0.00 jd = 0.00
F kl
Fr! = 0.00 F'rZ = 0.00 Fr~ = 0.00
Fk;h = 100.00 Jh
Fk,p = 0.00
'h
Fl~p = 0.00
,p
Fph = 100.00

ci i = (0.580000,0.000000,0.000000, O.OOOOOO)T
C~'i = (0.580000,0.000000,0.000000, O.OOOOOO)T
d'
cl i = (0.580000,0.000000,0.000000,0.000000)
T
ci i = (0.0437085,0.437078,0.0831118, 0.00805073f
C(i = (0.580000,0.000000,0.000000, O.OOOOOO)T
df i, =(0.580000,0.000000,0.000000,0.000000)
T

C~,i = (0.0437085,0.437078,0.0831118, 0.00805073f

Vjm = 0.00 V1 = 25.3347
rIJ = (0.00,0.00, O.OO)T
!\;k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, 0.00) T
!\;k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, o.oof
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 371

15.2.8 Test Problem 3 : Isothermal Van de Vusse Reaction

CaseII
Problem Information

Reaction Mechanism
k1 k2
A--"'~ B ---'---.... C
k3
2A--... D
Objective:
Maximize the yield of B:
h
maxc1,B

Parameters

Rate Constants
k1 10 s 1 (first order)
k2 1 s-l (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
Far 100 L/s
c~!i 0.58 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

Ir = k 1cA
12 = k2 CB
13 = k3C~
Problem Statistics
No. of continuous variables 77
No. of dynamic variables 16
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 21
No. of point equalities 9
No. of dynamic equations 16
372 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Best Known Solution

Objective function: 0.442966 mol/L

Variables

FT,abk = 0.00 = 100.000 F:r = 0.00

F:
F~~
= 0.00 FI = 100.000 Eid = 0.00
F: = 100.000 Ft = 0.00 Fr = 0.00
F: bk , = 0.00 F: k, = 0.00 F:1 = 0.00
fb
F k k' = 0.00 Fh,, =0.00
fd
F kl = 0.00
Ei~~ = 0.00
'b
Fl9,k = 0.00 Fl9Z = 0.00
jh 'h
F:~ = 100.000 Fk,p = 0.00 Fl~p = 0.00
Fi: = 100.00

<icri4
i = (0.580000,0.000000,0.000000, O.oooooof
= (0.580000,0.000000,0.000000, O.OOOOOO)T
= (0.580000,0.000000,0.000000, O.OOOOOO)T

ceci =
i = (0.0442974,0.442966,0.0845351, 0.00820189f
(0.580000,0.000000,0.000000, O.OOOOOO)T
df~ = (0.580000,0.000000,0.000000, O.OOOOOO)T
C;,i = (0.0442974,0.442966,0.0845351, 0.00820189)T
~m = 0.00 v: = 25.4573
riJ = (0.00,0.00, O.Oo)T
K~,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
K%,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.9 Test Problem 4 : Isothermal Van de Vusse Reaction

Case III
Problem Information

Reaction Mechanism
k1 k2
A----- B ----- C
k3
2A----- D

Objective:
Maximize the yield of B:

h
maxcl,B
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 373

Parameters

Rate Constants
k1 10 s 1 (first order)
k2 1 s-l (first order)
k3 0.5 L/(mol s) (second order)
Feed Conditions
FaT 100 L/s
C~,i 5.8 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D

Rate Expressions

fr = k 1cA
f2 = k 2cB
f 3 = k3~
Problem Statistics
No. of continuous variables 77
No. of dynamic variables 16
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 21
No. of point equalities 9
No. of dynamic equations 16

Best Known Solution

Objective function: 3.68185 mol/L
Variables

F T,abk = 0.00 Fac

T, k
= 0.00 F:1 = 100.000
Ft = 0.00 F k = 100.000 F/J = 100.000
Ff = 100.000 Ft = 0.00 Fr = 100.000
Ffbk, = 0.00 Ff~1 = 0.00 Ff1 = 0.00
Ib = 0.00
F kkl Fl.~1 = 0.00 Id
F k I = 0.00
b = 0.00
F/g,k F'lZ = 100.00 Fj~~ = 0.00
Ff~ = 100.000 !h
Fk,p = 0.00 F,~: = 0.00
F; = 100.00
374 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

cL = (5.80000,0.00000,0.000000, O.OOOOOO)T
c~ i = (2.40827,2.45481,0.278627, 0.329144)T
, T
c/,i = (5.80000,0.00000,0.000000,0.000000)
ck i = (0.368196,3.68185,0.847795, 0.451080)T
C{i = (5.80000,0.00000,0.000000, O.OOOOOO)T
d/ i, = (2.40827,2.45481,0.278627,0.329144) T
.~
cp,i = (0.368196,3.68185,0.847795,0.451080) T
Vi m = 11.3502 vi
= 16.9843
rlJ
= (24.0827,2.45481, 2.89989)T
!\;~,m =
(0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
!\;k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.10 Test Problem 5 : Isothermal Van de Vusse Reaction

Case IV
Problem Information

Reaction Mechanism
k1 k2
A----- B ----- C
k3
2A----- D

Objective:
Maximize the yield of B:
h
maxcl,B

Parameters

vi,i =
[-1~ 0~ -2]~
1 -1 0

Rate Constants
kl 1 s 1 (first order)
k2 2 s-l (first order)
k3 10 L/ (mol s) (second order)
Feed Conditions
Far 100 L/s
c~,i 1 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 375

Rate Expressions

Ir = klcA
12 = k 2CB
Ir = k3C~
Problem Statistics
No. of continuous variables 77
No. of dynamic variables 16
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 21
No. of point equalities 9
No. of dynamic equations 16

Best Known Solution

Objective function: 0.07026654 mol/L

Variables

Fr,abk = 0.00 F:,k = 0.00 Frar = 100.000

F~ = 0.00 Fk = 100.000 Fl' = 100.000
F: = 100.000 F{ = 0.00 F/9 = 100.000
F:bk , = 0.00 F:k, = 0.00 F:1 = 0.00
FI~, = 0.00 Fh, =0.00 fd
F kl = 0.00
Fr~ = 0.00 F/g~ = 100.00 Fl~~ = 0.00
jh 'h
Fi~ = 100.000 Fk,p = 0.00 Fl~p = 0,00
F; = 100.00
4 i = (1.00000,0.000000,0.000000, O.OOOOOO)T
4:ci ==
Ck'i

i
(0.316222,0.0586978,0.0346578, 0.295211)T
= (1.00000,0.000000,0.000000, O.oooooof
(0.140537,0.0702665,0.0557588, 0.366719f
C{i = (1.00000,0.000000,0.000000, O.OOOOOO)T
df: = (0.316222,0.0586979,0.0346578,0.295211 f
C;,i = (0.140537,0.0702665,0.0557588, 0.366719f
Vi m = 29.5222 V1
= 15.7581
rIJ = (0.316222,0.117396, 0.999964)T
~k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, o.oof
~k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
376 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

15.2.11 Test Problem 6 : Isothermal Trambouze Reaction

Problem Information

Reaction Mechanism

The first reaction is zero order, the second is first order, and the third is second
order.
Objective:
Maximize the selectivity of C to A

Parameters

Rate Constants
kl 0.025 gmolj(L min) (first order)
k2 0.2 min- 1 (first order)
k3 0.4 Lj(mol min) (second order)
Feed Conditions
F: 100 Ljmin pure A
c~,i 1 moljL A, 0 moljL B, 0 moljL C, 0 moljL D

Rate Expressions

Ir = k 1
12 =k2 CA
II = k3 c3t
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 377

Problem Statistics
No. of continuous variables 85
No. of dynamic variables 19
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 27
No. of point equalities 11
No. of dynamic equations 19

Best Known Solution

Objective function: 0.5

Variables

Fra~ = 0.00 , = 0.00

Fra~ Frar = 100.000
Ff = 0.00 F k = 0.00 Fl- = 100.000
F: = 0.00 F[= 0.00 Fr = 100.000
F:~, = 0.00 F:ck , = 0.00 F:1 = 0.00
FIt,b = 0.00 Fh, , =0.00
fd
F k / = 0.00
F/:k = 0.00 Frzfh = 0.00 F/gt, = 0.00
Fkeh = 0.00 Fk,p = 0.00 Fj~: = 100.000
F; = 100.000
,p

4,i = (1.000000,0.000000,0.000000, O.oooooof

cl; i = (1.000000,0.000000,0.000000, O.OOOOOO)T
c1~ = (1.000000,0.000000,0.000000, O.OOOOOO)T

cecid/ ==
i

,
(1.000000,0.000000,0.000000, O.OOOOOo)T
(1.000000,0.000000,0.000000, O.OOOOOO)T
i = (0.250000,0.187500,0.375000,0.187500)
T

C~,i = (0.250000,0.187500,0.375000, 0.187500)T

Vr = 750.00 V1
= 0.00 rrJ = (0.0250000,0.0450000, 0.0250000)T
I\:k m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, o.oof
I\:k:m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.12 Test Problem 7 : Isothermal Denbigh Reaction Case I

Problem Information

Reaction Mechanism
378 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

D E
Objective:
Maximize the selectivity of B to D:

max-'
dtB
dtD,
Parameters
0 -1 0
[ 0.5
-1 -1 0 -1
1 0 0
";J ~ ~ 0 1 0
0 0 1

Rate Constants
k1 1.0 L/(mol 8} (second order)
k2 0.6 S-l (first order)
k3 0.6 8- 1 (first order)
k4 0.1 L/(mol s} (second order)
Feed Conditions
F: 100 L/s
c~!i 6.0 mol/L A, 0 mol/L B, 0 mol/L C, 0.6 mol/L D, 0 mol/L E

Rate Expressions
f1 = k1St
/'2 = k2cB
/3 = k3 CA
/r = k4C~
Problem Statistics
No. of continuous variables 85
No. of dynamic variables 19
No. of linear equalities 13
No. of linear inequalities 20
No. of nonlinear equalities 27
No. of point equalities 11
No. of dynamic equations 19
15.2. GHEMIGAL REAGTOR NETWORK PROBLEMS 379

Best Known Solution

Objective function: 1.321759

Variables

F r,abk = 0.00 Fr~~ = 100.000 Fra? = 0.00

Ff = 0.00 F k = 100.000 Fjd = 0.00
Ff = 100.000 Ft = 0.00 Fj9 = 0.00
Ff~' = 0.00 Ff~' = 0.00 F:1 = 0.00
fb
F kk , = 0.00 Fh,, =0.00
fd
F k / = 0.00
Fl 'b
,k = 0.00 FlZ = 0.00 F/9t = 0.00
Ff~ = 100.00
jh
Fk,p = 0.00 Fi: = 0.00
Fph = 100.00
4i = (6.00000,0.00000,0.000000,0.60000, O.OOOOOOO)T
C~:i = (6.00000,0.00000,0.000000,0.60000, O.OOOOOOO)T
Cl,i = (6.00000,0.00000,0.000000,0.60000, O.OOOOOOO)T
ck,i = (2.44063,1.40768,0.117826,1.06500, 0.0216797)T
c! i = (6.00000,0.00000,0.000000,0.60000, O.OOOOOOO)T
d/: = (6.00000,0.00000,0.000000,0.60000, O.OOOOOOO)T
Si,i = (2.44063,1.40768,0.117826,1.06500, 0.0216797)T
V;m = 0.00 V: = 20.7497 rlj = (0.00, 0.00, o.oof
II:k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
II:~ ,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.13 Test Problem 8 : Isothermal Denbigh Reaction Case

11
Problem Information

Reaction Mechanism

Objective:
Maximize the production of C subject to 95% conversion of A
h
maxcl,C
380 GHAPTER 15. DYNAMIG OPTIMIZATION PROBLEMS

Parameters

Rate Constants
k1 1.0 L/(mol s) (second order)
k2 0.6 S-l (first order)
k3 0.6 S-l (first order)
k4 0.1 L/(mol s) (second order)
Feed Conditions
F: 100 L/s pure A
c~,i 6.0 mol/L A, 0 mol/L B, 0 mol/L C, 0 mol/L D, 0 mol/L E

Rate Expressions

f[ = k 1 St
I~ = k2CB
Ir = k3CA
Ir =k4c1
Additional Information
For this problem, two CSTRs and one CFR are considered. There is an
additional constraint due to the requirement of 95% conversion of A:

c~,1 = 0.3mol/L

Problem Statistics
No. of continuous variables 111
No. of dynamic variables 19
No. of linear equalities 19
No. of linear inequalities 20
No. of nonlinear equalities 39
No. of point equalities 11
No. of dynamic equations 19

Best Known Solution

Objective function: 3.539890 mol/L

Variables
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 381

Fab
r, k
= 0.00 F~k = 13.4112 F:r = 86.5888,0.00
Ff = 0.00 F~ = 95.0043 F/- = 86.5888,95.0043
F k = 95.0043 Ft = 0.00 Fr = 86.5888,95.0043
Fkbkl = 0.00 Fkkl = 0.00 F k1= 0.00,95.0043
jb jd
F kk, = 0.00 Fl,~1 = 0.00 Fk I = 0.00,0.00
F l9,b
'd
k = 0.00 Fr~ = 81.5931,0.00 = 0.00,0.00
F l911
jh 'h
F k,ph = 0.00 Fk,p = 0.00 Fl~p = 4.99574,95.0043
F!: = 100.00
4 i = (6.00000,0.000000,0.000000,0.000000, O.OOOOOO)T
Ck'i, = (6.00000,0.000000,0.000000,0.000000, O.OOOOOO)T

d ( 6.00000 0.000000 0.000000 0.000000 0.0000 )

cl,i = 0.7994710 2.527592 1.199291 0.9305851 0.5430603

cL = (0.799471,2.52759,1.19929,0.930585, 0.543060)T
ced( = (6.00000,0.00000,0.00000,0.000000, O.OOOOOO)T
,
= (6.00000,0.00000,0.00000,0.000000,0.000000)
i
T

dfi = (0.000269045,0.000850987,3.72603,1.72943, 0.543419)T

C~,i = (0.300000,0.000808474,3.53989,1.643031, 0.516271)T
Vi m = 0.00, 470143 v~ = 73.5632
m ( 36.000 0.0000 3.6000 0.0000 )
rl,j = 7.2385 x 10- 8 0.00051059 0.00016143 7.2418- 8

K km = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

Kic:m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.14 Test Problem 9 : Isothermal Levenspiel Reaction
Problem Information

Reaction Mechanism
k
A + B ----. 2B
Objective:
Maximize the production of B
h
maxct,B

Parameters
382 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Rate Constants
1.0 L/(mol s) (second order)
Rate Constants
F.ar 100 L/s
c~,i 0.45 mol/L A, 0.55 mol/L B

Rate Expression

Additional Information
The reactor volume is bounded below 100 L.

Problem Statistics
No. of continuous variables 61
No. of dynamic variables 10
No. of linear equalities 12
No. of linear inequalities 20
No. of nonlinear equalities 11
No. of point equalities 5
No. of dynamic equations 10

Best Known Solution

Objective function: 0.7686438 mol/L
Variables

Fabk = 0.00 F:'~ = 100.000 F:1 = 0.00

F: = 0.00
r,
Ff = 100.000 Fzd = 0.00
Ff = 100.000 F! = 0.00 Fr = 0.00
Ff~' = 0.00 Ffk' = 0.00 Ff1 = 0.00
/b
F k k' = 0.00 Fl.~, = 0.00 /d = 0.00
F kl
'b
Ftk = 0.00 Fjg~ = 0.00 Fl~~ = 0.00
F.k~h,p = 100.000 ih = 0.00
Fk,p Fj~; = 0.00
F; = 100.000
cL = (0.450000,0.550000)T
cti = (0.450000,0.550000f
c1i = (0.450000,0.550000)T
cii = (0.231356, O. 768644)T
C{i = (0.450000,0.550000f
df,i = (0.450000,0.550000)T
S!,i = (0.231356, O. 768644f
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 383

v;m = 0.00 Vi = 100.000 riJ= (O.229493)T

I\:~,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
I\:k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.15 Test Problem 10 : Q-Pinene Reaction

Problem Information

Reaction Mechanism

Objective:
Maximize the selectivity of C to D

max-'
die
di,D
Parameters

-~ I
-1 0 0 -2 0
[ -1 0 1 0 0 -1
ViJ ~ ~ 0 0 1
1 -1 -2
1
0
0
1
0 0 0 0 0

Rate Constants
k1 0.33384 s 1 (first order)
k2 0.26687 S-l (first order)
k3 0.14940 S-l (second order)
k4 0.18957 Lj(mol s) (second order)
k5 0.009598 Lj(mol s} (second order)
k6 0.29425 S-1 (second order)
k7 0.011932 s-1 (second order)
Feed Conditions
F: 100 Ljs pure A
C:,i 1.00 molfL A, 0 moljL B, 0 moljL C, 0 moljL D, 0 moljL E
384 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Rate Expressions

f[ = k1cA
12 = k 2CA
Ir = k 3CD
Ir = k 4Cb
1[' = k5C~
I~ = k 6cB
f!i = k 7 Cc
Additional Information
The sum of the reactor volumes is bounded below 6000 L.

Problem Statistics
No. of continuous variables 88
No. of dynamic variables 22
No. of linear equalities 16
No. of linear inequalities 21
No. of nonlinear equalities 27
No. of point equalities 11
No. of dynamic equations 22

Best Known Solution

Objective function: 1.557033
Variables

F:% = 0.00 F~k = 10.2219 F:1 = 89.7781

Ft = 0.00 F k = 10.2219 Fjd = 89.7781
F k = 10.2219 F! = 0.00 Fr = 89.7781
F kbkl = 0.00 Fk~1 = 0.00 Fk1= 0.00
lb = 0.00
F kkl
'b
Fh,, =0.00 Fkld1= 0.00
Fj9k = 0.00 Fj9Z = 0.00 Fr;~ = 0.00
Fi~ = 10.2219 fh
Fk,p = 0.00 Fj~: = 89.7781
Fph = 100.000
4i = (1.000000,0.000000,0.000000,0.000000, O.OOOOOO)T
Ck'i = (1.000000,0.000000,0.000000,0.000000, o.oooooof
c1.~ = (1.000000,0.000000,0.000000,0.000000, O.OOOOOO)T
C~ i = (0.000000,0.0497595,0.152595,0.0980035, 0.547048f
C{i = (1.000000,0.000000,0.000000,0.000000, O.OOOOOO)T
4; = (1.000000,0.000000,0.000000,0.000000, o.oooooof
C;,i = (0.897781,0.00508637,0.0155981, 0.100178, 0.0559187f
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 385

Vjm = 0.00 V: = 6000.00

rij = (0.33384,0.26687,0.0000,0.0000,0.0095980, 0.0000, O.OOOO)T
~~ m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, o.oof
~k:m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

15.2.16 Test Problem 11 : Nonisothermal Van de Vusse Reac-

tion Case 11
Problem Information

Reaction Mechanism
k1 k2
A - -... B - -... C
k3
2 A - -... D
Objective:
Maximize the yield of B
maxdt.,B

Parameters

/.Ii,j = [-1 0-2]

1 -1
0 1
0
0
001
Rate Constants
reaction k E
1 5.4 x 109 h 15.84 kcal/mol 84 K L/mol
2 3.6 x 105 h- 1 7.92 kcal/mol 108 K L/mol
3 1.6 x10 12 L/(mol h) 23.76 kcal/mol 60 K L/mol
Feed Conditions
F: 100 L/s
c~!i 1.0 mol/L A, 0 mol/L B, 0 mol/L C

Rate Expressions

Additional Information
The temperatures in the reactors are bounded between 450 K and 810 K.
386 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Problem Statistics

No. of continuous variables 100

No. of dynamic variables 18
No. of linear equalities 20
No. of linear inequalities 20
No. of nonlinear equalities 23
No. of point equalities 9
No. of dynamic equations 18

Best Known Solution

Objective function: 0.8229079 mol/L

Variables

F:~ = 0.00 Fr~~ = 0.00 Frat = 100.000

F kb = 0.00 Ft = 100.000 Fjd = 100.000
Ft = 100.000 Ft = 0.00 Fj9 = 100.000
F:fb = 0.00
bk , F: k, = 0.00 fd F:1 = 0.00
F k k' = 0.00 Fl.~, = 0.00 F k 1= 0.00
Fj9~ = 100.00 Fj9~ = 0.00
'b
FI~k = 0.00
jh 'h
Ft~ = 100.000 Fk,p = 0.00 Fj~p = 0.00
F; = 100.000
4,i = (1.000000,0.000000,0.000000, O.OOOOOO)T
c~ i = (0.931253,0.0640282,0.00227142, 0.00122377)T
ct~ = (1.000000,0.000000,0.000000, 0.000000)Y
ci
ce = i = (0.00831126,0.822908,0.0794098, 0.0446855)T
(1.000000,0.000000,0.000000, O.OOOOOO)T
q~ = (0.931253,0.0640282,0.00227142, 0.00122377)Y
C~,i = (0.00831126,0.822908,0.0794098, 0.0446855)T
l-r = 0.0736545 V1 = 0.0807310
11m = 446.889
rl,j =
(90.0143,3.08389, 3.32300)T
I\:~ ,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 387

I\;k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

479.235 481.342
483.339 485.439
487.856 490.382
492.812 495.611
499.249 502.868
I\;t
k,m,i -
- 505.612 509.531
516.674 523.016
524.534 530.571
550.471 565.379
556.442 609.401
810.000 800.000

15.2.17 Test Problem 12 : Nonisothermal Naphthalene Reac-

tion
Problem Information

Reaction Mechanism

Objective:
Maximize the yield of B

h
maxcl,B

Parameters

-1 -1 0

-! 1
v- . _ [ 1 o -1
',3 - 0 1 1
o o 0
388 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Rate Constants
reaction k E I5.H
pCp
1 2.0 x 10 13 h 1 38.00 kcal/mol 364 K L/mol
2 2.0 x 10 13 h- 1 38.00 kcal/mol 129 K L/mol
3 8.15x10 17 h- 1 50.00 kcal/mol 108 K L/mol
4 2.1 x 105 h- 1 20.00 kcal/mol 222 K L/mol
Feed Conditions
F: 100 L/h
c~,i 1.0 mol/L A, 0 mol/L B, 0 mol/L C

Rate Expressions

Additional Information
The temperatures in the reactors are bounded between 900 K and 1500 K.

Problem Statistics
No. of continuous variables 101
No. of dynamic variables 19
No. of linear equalities 20
No. of linear inequalities 20
No. of nonlinear equalities 24
No. of point equalities 9
No. of dynamic equations 19

Best Known Solution

Objective function: 0.9999585 mol/L

Variables

F:,% = 0.00 F:,'ic = 100.000 F:1 = 0.00

F: = 0.00 F k = 100.000 Fjd = 0.00
F k = 100.000 Ft = 0.00 Flg = 0.00
Fkbk , = 0.00 F k1' = 0.00 1
F k = 0.00
fb fd
Fk = 0.00
k' Fl~,, = 0.00 Fk I= 0.00
FlZ, = 0.00 Flg~ = 0.00 Fjg~ = 0.00
jh Fgl,ph = 0.00
Fk~ = 100.000 Fk,p = 0.00
Fph = 100.00
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 389

4,i = (1.000000,0.000000,0.000000, O.OOOOOO)T

cA: i = (1.000000,0.000000,0.000000, O.OOOOOO)T
d' T
Ct,i = (1.000000,0.000000,0.000000,0.000000)
cki = (3.22531 X 10- ,4.96095 X 10- ,0.999959,3.82322 x 1O- 5 f
6 8

cL ,
= (1.000000,0.000000,0.000000, O.OOOOOO)T
T
df,i = (1.000000,0.000000,0.000000,0.000000)
C~,i = (3.22531 X 10- 6 ,4.96095 X 10- 8 ,0.999959,3.82322 x 10- 5 ) T
Vjm = 0.00 Vf = 0.000189577 T"m = 900.000
rrJ
= (0.00,0.00, O.OO)T
K: km = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

K:'k'm = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

,

900.000 900.000
1242.86 1500.00
1242.86 900.000
900.000 1008.96
992.983 930.514
K: tk,m,.t. = 900.000 900.000
929.074 952.697
936.343 909.086
900.000 900.000
900.000 900.000
900.000 900.000

15.2.18 Test Problem 13 : Nonisothermal Parallel Reactions

Problem Information

Reaction Mechanism

Objective:
Maximize the yield of C while minimizing the volume of the reactor

max 100c~,c - L vF
tEL
390 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Parameters

[
-1
-1 -1
-1 1
Vi,j = ~ ~

Rate Constants
reaction k E
1 5.4 x 107 19.138 kcal/mol 10 K L/mol
2 3.6 x 105 9.569 kcal/mol 20 K L/mol
Feed Conditions
Ff 50 L/s pure A
F~ 50 L/s pure B
cti 1.0 moljL A, 0 mol/L B
~!i 0 mol/L A, 1.0 mol/L B

Rate Expressions

Additional Information
The temperatures in the reactors are bounded between 450 K and 800 K.

Problem Statistics
No. of continuous variables 102
No. of dynamic variables 17
No. of linear equalities 21
No. of linear inequalities 20
No. of nonlinear equalities 22
No. of point equalities 9
No. of dynamic equations 17

Best Known Solution

Objective function: 44.78686

Variables
15.2. CHEMICAL REACTOR NETWORK PROBLEMS 391

F:~ = 0.0,49.0076 F:'k = 50.0,0.992433 F:1 = 0.0,0.0

F k6 = 49.0076 F k = 50.9924 Eid = 0.00
F: = 100.000 Ff = 0.00 Fl = 0.00
F:bk , = 0.00 F: ck , = 0.00 F:1 = 0.00
fb
F k k' = 0.00 Fh,, =0.00
fd
F k I = 0.00
Fl'bk = 0.00 Fjg~ = 0.00 Fl~~ = 0.00
Fi ,p = 1000.00
h fh
Fk,p = 0.00 Ei~: = 0.00
F; = 1000.00

cti = (0.000000,1.000000,0.000000, O.OOOOOO)T

cki = (0.980538,0.0194624,0.000000, O.OOOOOO)T
c1~ = (0.500000,0.500000,0.000000, o.oOOOOO)T
ci
ce i

c't,~ =
=
=
(0.0119436,0.0119435,0.481617, 0.00643913)T
(0.500000,0.500000,0.000000, O.OOOOOO)T
(0.500000,0.500000,0.000000, O.OOOOOO)T
C;,i = (0.0119436,0.0119435,0.481617, 0.00643913)T
Vjm = 0.00 v1
= 3.37487
Tim = 800.0
rLJ = (0.000332841,0.135224f
K,k,m = (49.0076,30.8666,20.3423, 13.4839,8.72020,5.25364
2.73586,0.268785, 0.00, 0.00, O.OO)T
K,~ ,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

800.000 800.000
800.000 800.000
800.000 800.000
800.000 800.000
800.000 800.000
K,t .= 800.000 800.000
k,m,'I.
800.000 800.000
800.000 800.000
800.000 800.000
800.000 800.000
800.000 800.000

15.2.19 Test Problem 14 : Sulfur Dioxide Oxidation

Problem Information

Reaction Mechanism
392 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

8pecies 1 is 802, species 2 is 02, species 3 is 803, and species 4 is N2.

Objective:
Maximize the yield of 80 3
h
maxc1,3

Parameters

.. _
[
'~1 1
-0.5 0.5
1
I/',J - 1-1
o 0

Rate Constants
reaction k E
1 6.284 x 10 mol/(hr kgcat) 15.500 kcal/mol
11 96.5 K kg/mol
2 2.732 X 10 16 mol/(hr kgcat) 26.7995 kcal/mol 96.5 K kg/mol
Feed Conditions
F: 7731 kg/h
c~,i 2.5 mol/kg S02, 3.46 mol/kg O 2,
o mol/kg S03, 26.05 mol/kg N 2
Rate Expressions
The rates are given in terms of kg mol 80 3 per kg catalyst.

Additional Information
The temperatures in the reactors are bounded between 300 K and 1200 K
and the size of the reactor bounded below 50000 kg of catalyst.

Problem Statistics
No. of continuous variables 99
No. of dynamic variables 17
No. of linear equalities 21
No. of linear inequalities 20
No. of nonlinear equalities 22
No. of point equalities 9
No. of dynamic equations 17
15.3. PARAMETER ESTIMATION PROBLEMS 393

Best Known Solution

Objective function: 2.497518 mol/kg
Variables

Fra~ = 0.000 F:~

, = 7731.00 F:1 = 0.00
F kb = 0.00 Ff. = 100.000 F,.d = 0.00
F: = 7731.00 Ft = 0.0000 Fr = 0.00
F:bk, = 0.00 F:)., = 0.00 F:1 = 0.00
fb
F k k' = 0.00 Fh,, = 0.00 fd
F kl = 0.00
'b
, = 0.00
Frk F?Z = 0.00 F,.g~ = 0.00
F:~.= 7731.00
fh =
Fk,p 0.00 Fl~: = 0.00
F; = 7731.00
c~ i = (2.50000,3.46000,0.00000, 26.0500)T
cti, = (2.50000,3.46000,0.00000,26.0500) T
Cl,i = (2.50000,3.46000,0.00000, 26.0500f
c% i = (0.00248248,2.21124,2.49752, 26.0500)T
cI'
k i = (2.50000,3.46000,0.00000, 26.0500)T
, T
df,i = (2.50000,3.46000,0.00000,26.0500)
C~,i =(0.00248248,2.21124,2.49752, 26.0500f
Vi m = 0.00 V1 = 50000.0 ~m = 1200.0
rrJ
= (O.OO,O.OO)T
I\:k,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, o.oof

1\:% ,m = (0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00, 0.00, O.OO)T

791.584 713.502
673.290 654.609
641.118 631.775
625.537 620.599
615.160 609.744
I\:t -
k,m,i - 604.874 601.064
598.823 596.981
594.365 591.607
589.340 586.898
583.615 581.643
583.131 1200.00

15.3 Parameter Estimation Problems

15.3.1 Introduction
The problem of estimating parameters in dynamic models is as important and
even more difficult than with algebraic models. The extra difficulty arises from
394 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

the inclusion of differential-algebraic equations in the optimization problem.

This type of problem arises most often in the estimation of kinetic constants
from time series data. The statistical approach used is similar to that pre-
sented for the algebraic problems in section 8.5, with the exception that all the
variables receive an equal weighting in the objective function.
A general formulation of the problem will be presented where the only
difference between examples is in the definition of the model equations. Next
a set of example problems will be given. The first two problems represent
rat her simple models of irreversible and reversible first order series reactions.
The next two problems add the difficulty of nonlinear reaction kinetics to the
problem. The next problem combines the difficulties imposed by nonlinear
kinetics with relatively large residuals at the solution. The final problem is
the most difficult since the dynamic system exhibits a limit cycle and a large
number of local minima have been identified.

15.3.2 General Formulation

Objective Function
r m
!llin L L
;1)/,,8 JL==l i==l
(XJL,i - X JL ,i)2

Constraints
Dynamic Model
dz
f( dt ,z,8) o z(to) Zo

Point constraints

@ t = tJL Zi - XJL,i =0 i=l, ... ,m JL=l, ... ,r

Variable Definitions
f is a system of l differential-algebraic functions which represent the non-
linear model, z is a vector of i dynamic variables, 8 (J is a vector of p parameters,
and xJL
xJl is a vector of i fitted data variables at the JLth data point. xJL
xJl is a vector
of i experimentally observed values at the JL th data point, and er is a vector of
i values which represent the experimental standard deviation of these observa-
tions. tt E (to, tt f) and tJL
tJl is the time associated with the JL th observation.

15.3.3 Test Problem 1

This model represents the first order irreversible chain reaction:

A~B~C
15.3. PARAMETER ESTIMATION PROBLEMS 395

Only the concentrations of components A and B were measured, therefore

component C does not appear in the model used for estimation. This model
appears in Tjoa and Biegler (1991a).

Formulation

Objective Function

Constraints
Dynamic Model
dZl
= -(hzl
dt
dZ2
= (hzl - fh z2
dt
Point constraints

i = 1,2 J.L = 1, ... ,10

Initial conditions

%0 = (1,0)

Variable Bounds

~ :l:p ~ 1
0~(J~1O

Variable Definitions
Zl and Z2 are the mole fractions of the components A and B respectively.
fh and fh are the rate constants of the first and second reactions respectively.

Data
J.L 1 2 3 4 5 6 7 8 9 10
t# 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Xl 0.606 0.368 0.223 0.135 0.082 0.050 0.030 0.018 0.011 0.007
X2 0.373 0.564 0.647 0.669 0.656 0.624 0.583 0.539 0.494 0.451
396 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Problem Statistics
No. of continuous variables 2
No. of dynamic variables 2
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 2
No. of point constraints 20

Best Known Solution

Objective Function: 1.18584 x 10- 6

Parameters
= (5.0035,1.0000)T
9=

Fitted Data Variables

Matches data to the given precision

15.3.4 Test Problem 2

This model represents the first order reversible chain reaction:

A~B::C
k2 k4

This model appears in Tjoa and Biegler (1991a).

Formulation

Model

dZ 1
dt
= -(hZl + ()2 Z2
dZ2
dt
()lZl - (()2 + ()3)Z2 + ()4 Z3
dZ3
dt
-()4 Z3 + ()3 Z2

Initial conditions

Zo = (1,0,0)
15.3. PARAMETER ESTIMATION PROBLEMS 397

Variable Bounds
o :::; xJ.L :::; 1
(0,0,10,10) :::; () :::; (10,10,50,50)

Variable Definitions
Zl, Z2, and Z3 are the mole fractions of the components A, B, and C
respectively. (h and (}3 are the rate constants of the first and second forward
reactions, and (}2 and (}4 are the rate constants of the reverse reactions.

Data
Il 1 2 3 4 5 6 7
tJ.L 0.05 0.1 0.15 0.2 0.25 0.3 0.35
XJ.L,l 0.8241 0.6852 0.5747 0.4867 0.4166 0.3608 0.3164
XJ.L,2 0.0937 0.1345 0.1654 0.1899 0.2094 0.2249 0.2373
XJ.L,3 0.0821 0.1802 0.2598 0.3233 0.3738 0.4141 0.4461
Il 8 9 10 11 12 13 14
tJ.L 0.4 0.45 0.5 0.55 0.6 0.65 0.7
XJ.L,l 0.2810 0.2529 0.2304 0.2126 0.1984 0.1870 0.1870
XJ.L,2 0.2472 0.2550 0.2613 0.2662 0.2702 0.2733 0.2759
XJ.L,3 0.4717 0.4920 0.5082 0.5210 0.5313 0.5395 0.5460
Il 15 16 17 18 19 20
tJ.L 0.75 0.8 0.85 0.9 0.95 1.0
XJ.L,l 0.1709 0.1651 0.1606 0.1570 0.1541 0.1518
XJ.L,2 0.2779 0.2794 0.2807 0.2817 0.2825 0.2832
XJ.L,3 0.5511 0.5553 0.5585 0.5612 0.5632 0.5649

Problem Statistics
No. of continuous variables 4
No. of dynamic variables 3
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 3
No. of point constraints 60

Best Known Solution

Objective Function: 1.8897 x 10- 7
Parameters
() = (4.000,2.000,40.013, 20.007)T
398 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Fitted Data Variables

Matches data to the given precision

15.3.5 Test Problem 3 : Catalytic Cracking of Gas Oil

This model represents the catalytic cracking of gas oil (A) to gasoline (Q) and
other side products (S).

Only the concentrations of A and Q were measured, therefore the concentration

of S does not appear in the model for estimation. This model appears in Tjoa
and Biegler (1991a).

Formulation

Initial conditions
Zo (1,0)

Variable Bounds
o :::; xp. :::; 1
0:::;0:::;20

Variable Definitions
Zl and Z2 are the mole fractions of the components A and Q respectively.
(h, (h, and (h are the rate constants of the respective reactions.

Data
J.L 1 2 3 4 5 6 7
tp. 0.025 0.05 0.075 0.10 0.125 0.150 0.175
Xp.,l 0.7307 0.5982 0.4678 0.4267 0.3436 0.3126 0.2808
Xp.,2 0.1954 0.2808 0.3175 0.3047 0.2991 0.2619 0.2391
15.3. PARAMETER ESTIMATION PROBLEMS 399

J.1. 8 9 10 11 12 13 14
tJ-l 0.20 0.225 0.250 0.30 0.35 0.40 0.45
XJ-I,l 0.2692 0.2210 0.2122 0.1903 0.1735 0.1615 0.1240
XJ-I,2 0.2210 0.1898 0.1801 0.1503 0.1030 0.0964 0.0581
J.1. 15 16 17 18 19 20
tJ-l 0.50 0.55 0.65 0.75 0.85 0.95
XJ-I,l 0.1190 0.1109 0.0890 0.0820 0.0745 0.0639
XJ-I,2 0.0471 0.0413 0.0367 0.0219 0.0124 0.0089

Problem Statistics
No. of continuous variables 3
No. of dynamic variables 2
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 2
No. of point constraints 40

Best Known Solution

Objective Function: 2.65567 x 10- 3

Parameters
(J = (12.214,7.9798, 2.2216)T

Fitted Data Variables

J.1. 1 2 3 4 5 6 7
XJ-I,l 0.7348 0.5808 0.4802 0.4093 0.3566 0.3159 0.2836
XJ-I,2 0.2014 0.2822 0.3079 0.3064 0.2912 0.2696 0.2456

J.1. 8 9 10 11 12 13 14
XJ-I,l 0.2573 0.2354 0.2170 0.1876 0.1652 0.1476 0.1334
XJ-I,2 0.2213 0.1980 0.1763 0.1386 0.1085 0.0850 0.0669

J.1. 15 16 17 18 19 20
XJ-I,l 0.1217 0.1119 0.0963 0.0846 0.0754 0.0680
XJ-I,2 0.0530 0.0424 0.0280 0.0193 0.0140 0.0106

15.3.6 Test Problem 4 : Bellman's Problem

This problem describes a reversible homogeneous gas phase reaction of the

form:
400 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

The model equation is the result of many normalizations. This model originally
appeared in Bellman et al. (1967) as well as in Tjoa and Biegler (1991a) and
Varah (1982).

Formulation

~ 2 2
dt = (h(126.2 - z)(91.9 - z) - (hz

Initial conditions

zo 0

Variable Bounds

XJ.I - 5 ~ xJ.l ~ xJ.I +5

o ~ () ~ 0.1

Data
f.L 1 2 3 4 5 6 7
tJ.l 1.0 2.0 3.0 4.0 5.0 6.0 7.0
xJ.I 1.4 6.3 10.4 14.2 17.6 21.4 23.0
f.L 8 9 10 11 12 13 14
tJ.l 9.0 11.0 14.0 19.0 24.0 29.0 39.0
xJ.I 27.0 30.4 34.4 38.8 41.6 43.5 45.3

Problem Statistics
No. of continuous variables 2
No. of dynamic variables 1
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 1
No. of point constraints 14

Best Known Solution

Objective Function: 22.03094

Parameters
() = (4.5704 x 10-6 ,2.7845 x 1O- 4 )T
15.3.
15.3. PARAMETER ESTIMATION PROBLEMS 401

Fitted Data Variables

f.L 1 2 3 4 5 6 7
x/l 4.54 8.51 12.01 15.12 17.89 20.38 22.62

f.L 8 9 10 11 12 13 14
x/l 26.49 29.69 33.55 38.18 41.32 43.49 46.07

15.3.7 Test Problem 5 : Methanol-to-Hydrocarbons Process

This model represents the conversion of methanol to various hydrocarbons.
The simplified kinetic model used here is:

A ~ B
A+B ~ C
C+B ~ P
A ~ C
A ~ P
A+B ~ P

In this model, A represents the oxygenates, B is H2 , C is the olefins, and P

denotes the paraffins, aromatics, and other products. The differential equation
model is formulated under the assumption of simple kinetics and a quasi-steady
state for the intermediate B. The model appears in Maria (1989).

Formulation

Model

dZ 1
- (2(}1 - (() (}(}lZ)2 + (}3 + (}4) Zl
dt 2 + 5 Zl + Z2
dZ2 (}lZl ((}2Z1 - Z2)
+ ()3 Z1
--=-..::..0.....:=-=----'::':"-
dt ((}2 + (}5)Zl + Z2
dZ3 (}lZl (Z2 + (}5 Z1)
--=--=-..:....--,----'-- + 4 Zl
()
dt ((}2 + (}5)Zl + Z2

Initial conditions

Zo = (1,0,0)

Variable Bounds

o :::; x/l :::; 1

o :::; () :::; 100
402 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Variable Definitions
Zl, Z2, and Z3 are the mole fractions of the components A, C, and P
respectively. The parameter vector (J is defined as:

Data
f..L 1 2 3 4 5 6 7 8
tJ.L 0.050 0.065 0.080 0.123 0.233 0.273 0.354 0.397
XJ.L,l 0.461 0.426 0.383 0.305 0.195 0.170 0.139 0.112
XJ.L,2 0.114 0.135 0.157 0.194 0.231 0.234 0.228 0.228
XJ.L,3 0.018 0.035 0.045 0.047 0.084 0.095 0.111 0.134
f..L 9 10 11 12 13 14 15 16
tJ.L 0.418 0.502 0.553 0.681 0.750 0.916 0.937 1.122
XJ.L,l 0.112 0.090 0.082 0.066 0.053 0.043 0.041 0.029
XJ.L,2 0.226 0.220 0.214 0.178 0.188 0.183 0.184 0.166
XJ.L,3 0.168 0.148 0.157 0.206 0.206 0.214 0.213 0.230

Problem Statistics
No. of continuous variables 5
No. of dynamic variables 3
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 3
No. of point constraints 48

Best Known Solution

Objective Function: 0.10693

Parameters
(J = (5.2407,1.2176,0,0, O)T

Fitted Data Variables

f..L 1 2 3 4 5 6 7 8
XJ.L,l 0.607 0.528 0.460 0.316 0.133 0.100 0.058 0.044
XJ.L,2 0.169 0.195 0.214 0.243 0.246 0.240 0.226 0.220
XJ.L,3 0.018 0.028 0.037 0.066 0.125 0.140 0.163 0.172
15.3. PARAMETER ESTIMATION PROBLEMS 403

J.L 9 10 11 12 13 14 15 16
XJL,l 0.039 0.023 0.017 0.008 0.006 0.002 0.002 0.000
XJL,2 0.217 0.207 0.203 0.196 0.194 0.191 0.190 0.189
XJL,3 0.176 0.187 0.192 0.200 0.202 0.205 0.206 0.207

15.3.8 Test Problem 6 : Lotlca-Volterra Problem

This problem involves a two parameter predator-prey model similar to those
used in ecology. The model determines the population over time of the two
different species (the predators, and the prey) given parameters relating to the
interaction between them. The state profiles at the solution are cyclic and out
of phase with each other. The parameter estimation problem in this model
was studied by Luus (1998).

Formulation

dZ 1
dt
= (h Zl(1 - Z2)

dZ2
dt
= (hZ2(Zl - 1)

Initial conditions
%0 (1.2,1.1)

Variable Bounds
0.5 :::; xJL :::; 1.5
O:::;(J:::;lO

Variable Definitions
Zl represents the population of the prey, and Z2 the population of the
predator. The parameters (J are related to the birth, death, and interaction
rates of the two species.

Data
J.L 1 2 3 4 5
tJL 1 2 3 4 5
XJL,l 0.7990 0.8731 1.2487 1.0362 0.7483
XJL,2 1.0758 0.8711 0.9393 1.1468 1.0027
J.L 6 7 8 9 10
tJL 6 7 8 9 10
XJL,l 1.0024 1.2816 0.8944 0.7852 1.1527
XJL,2 0.8577 1.0274 1.1369 0.9325 0.9074
404 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Problem Statistics
No. of continuous variables 2
No. of dynamic variables 2
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 2
No. of point constraints 20
No. of known solutions 23

Best Known Solution

Objective Function: 1.24924 x 10- 3

Parameters
() = (3.2434,0.9209f
Fitted Data Variables
/J 1 2 3 4 5
:1;J.I,l 0.7980 0.8643 1.2538 1.0385 0.7608
:1;J.I,2 1.0794 0.8876 0.9402 1.1443 1.0032

/J 6 7 8 9 10
:1;J.I,l 0.9976 1.2785 0.8841 0.7907 1.1515
:1;J.I,2 0.8670 1.0282 1.1288 0.9315 0.8883

15.4 Optimal Control Problems

15.4.1 Introduction
Optimal Control problems arise often in many different fields in the engineering
and sciences. This dass ofproblems has been well studied from both theoretical
and computational perspectives. The books by Hager and Pardalos (1998),
Polak (1997, 1971), Dontchev (1983), Alekseev et al. (1987), and Bryson and
Ho (1988) offer the reader a sampIe of the available material. Examples taken
from the chemical engineering field, indude such problems as the temperature
control of a batch or plug flow reactor system, or the optimal catalyst mix in a
tubular reactor. The models used to describe these types of systems are almost
always nonlinear in nature. This often results in the existence of multiple local
minima in the area of interest.
All of the example problems which follow exhibit at least two local minima.
The first is a simple single differential equation model with constant control
and the minimization of a concave objective function. The second problem
adds the difficulty of a time varying control, while the third, which deals with
the control of a CSTR, adds a very nonlinear differential algebraic system. The
15.4. OPTIMAL CONTROL PROBLEMS 405

fourth problem is not only an optimal temperature control problem in a PFR,

but also a design problem since the size of the reactor is also an optimization
variable. The final example, the bifunctional catalyst problem, is the most
challenging since it has been shown to contain well over 100 local minima.
Additional difficulty arises from the .small separation, on the order of 10- 5 ,
between many of the local minima.

15.4.2 Test Problem 1

This example is of a simple control problem with one state and one control
variable.

Formulation

Objective function

Constraints

dz 2
- = -z +u
dt

Initial Conditions

Control
u is a constant over the whole time horizon.
Variable bounds

-5 ::; u ::; 5
-12 ::; z ::; 9
tE [0,1]

Problem Statistics
No. of continuous variables 1
No. of dynamic variables 1
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 1
No. of known solutions 2
406 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Best Known Solution

Objective function: -8.23623

Control variable
u= -5

15.4.3 Test Problem 2 : Singular Control Problem

This example represents a nonlinear singular control problem and appears in
Luus (1990b) as weH as in Rosen and Luus (1992).

Formulation

Objective function

Constraints

-Z3U + 16t - 8

= U

Zl2 + Z22 + 0.0005 ( Z2 + 16t - 8 - 0.1z3 u 2)2

Initial Conditions

z(t o ) = (0, -1, -v'5, 0)

Control
u is piecewise constant over 10 intervals:
U = Wi for ti:::; t < ti+1
ti = (0,0.1,0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, I)

Variable bounds

-4 :::; w :::; 10
tE[O,I]
15.4. OPTIMAL CONTROL PROBLEMS 407

Problem Statistics
No. of continuous variables 10
No. of dynamic variables 4
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 4
No. of known solutions 2

Best Known Solution

Objective function: 0.120114

Control variables
w = (10,7.728,5.266, -0.764,0.132,0.133, -0.835,6.764,6.221,5.414)

15.4.4 Test Problem 3 : CSTR Problem

This example represents a continuous stirred tank reactor model and appears
in Luus (1990b).

Formulation

Objective function

Constraints

dZ 1
dt
= -(Zl + 0.25) + (Z2 + 0.5) exp [ Zl25z1
+ 2 ] - (1 + U)(Zl + 0.25)
dZ2
0.5 - Z2 - (Z2 + 0.5) exp [ Zl25z1
+ 2]
dt
dZ3
dt
Z? + z~ + 0.lu2

Initial Conditions

Z(t o ) = (0.09,0.09,0)

Control
U is piecewise linear over 10 intervals:

U = Wi + ~:!~ =~i (t - ti) for ti ~ t < ti+1

408 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

ti = (0, 0.078, 0.156, 0.234, 0.312, 0.39, 0.468, 0.546, 0.624, 0.702, 0.78)
Variable bounds

-0.5 :::; w :::; 5

tE [0,0.78]

Variable Definitions
Zl represents the deviation from the dimensionless steady-state tempera-
ture, Z2 the deviation from the dimensionless steady-state eoneentration, and
u is the sealed eoolant Howrate.

Problem Statistics
No. of eontinuous variables 11
No. of dynamie variables 3
No. of linear equalities
No. of eonvex inequalities
No. of nonlinear equalities
No. of noneonvex inequalities
No. of dynamie eonstraints 3
No. of known solutions 2

Best Known Solution

Objective funetion: 0.13317

Control variables
w = (4.27,2.22,1.38,0.887,0.584,0.379,0.237,0.137, 0.068, 0.023, -0.002)

15.4.5 Test Problem 4 : on Shale Pyrolysis

This model deseribes the determination of the optimal temperature profile and
residenee time for the following reaetion system:

Al ~ A 2
A2 ~ A3
Al +A2 ~ A 2 +A2
Al +A2 ~ A 3 +A2
Al +A 2 ~ A 4 +A2

Only eomponents Al and A 2 are included in the model. This example was
studied by Luus (1990b); Rosen and Luus (1992); Carraseo and Banga (1997).
15.4. OPTIMAL CONTROL PROBLEMS 409

Formulation

Objective function

min -Z2(t/)
w,p

Constraints

dZl
-p (k1z1 + k3Z1Z2 + k 4z 1Z2 + k 5z 1Z2 )
dt
dZ2
dt
P (k1z 1 - k 2z 2 + k 3 z 1Z2)
ki aiexp
,
-bi]
a"exp [-
Ru'
Ru [-b"] Vi = 1. .. 5

Initial Conditions

z(to ) == (1,0)
Z(t

Control
u is piecewise constant over 10 intervals:
u = Wi for ti::; ti:::; t < ti+1
ti+l
ti =
= (0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1)
Variable bounds

698.15 ::;w ::;

:::; :::; 748.15
7 ::; :::; 11
p ::;
:::;
tE[O,I]
t E [0,1]

Variable Definitions
Zl and Z2 are the mole fractions of components Al and A 2 respectively.
The control u is the temperature of the reactor, and p is the residence time.
ai and bi
bi are the Arrhenius constants for reaction i.

Data

lnai = (8.86, 24.25, 23.67, 18.75, 20.70)

bdR == (10215.4, 18820.5, 17008.9, 14190.8, 15599.8)
b;j
410 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Problem Statistics
No. of continuous variables 12
No. of dynamic variables 7
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 7
No. of known solutions 5

Best Known Solution

Objective function: -0.353606
Residence time: p = 8.3501
Control variables
w = (698.15, 698.15, 698.15, 748.15, 748.15, 699.20, 698.15,
698.15, 698.15, 698.15)

15.4.6 Test Problem 5 : Bifunctional Catalyst Blend Problem

This example concerns the optimization of a bifunctional catalyst in converting
methylcyclopentane to benzene. The catalyst contains a hydrogenation com-
ponent and a isomerization component. The object is to determine the mixt ure
of the two along the length of the reactor which maximizes the concentration
of the desired product, A 7 , in the following reaction scheme:
A6 A7

A 3~2 k
k~X 8 10
k

1: '
7
3 9

k, :,1:,
Al ~
This problem has been studied by Luus et al. (1992), Bojkov and Luus (1993),
and Luus and Bojkov (1994). Even though this example is relatively small (7
states and 1 control), it has been shown to exhibit a very large number of loeal
minima.

Formulation

Objective function
15.4. OPTIMAL CONTROL PROBLEMS 411

Constraints
dZ 1
= -k1z1
dt
dZ2
= k1z1 - (k2 + k3)Z2 + k4 Z5
dt
dZ3
= k2 Z2
dt
dZ4
dt
= -k6Z4 + k5 Z5
dZ5
dt
= k3Z2 + k6Z4 - (k4 + k5 + ks + k 9 )Z5 + k7Z6 + klO Z7
dZ6
= ksz5 - k7Z6
dt
dZ7
= k 9 z 5 - k lO z 7
dt
ki = (Ci,l + Ci,2U + Ci,3U2 + Ci,4 U3 ) V r Vi = 1 ... 10

Initial Conditions

Z(t o ) = (1,0,0,0,0,0,0)

Control
U is piecewise constant over 10 intervals:
U = Wi for ti ~ t < ti+1
ti = (0,0.1,0.2,0.3,0.4,0.5, 0.6,0.7,0.8,0.9, 1)
Variable bounds

0.6 ~ w ~ 0.9
tE [0,1]

Variable Definitions
Zi is the mole fraction of the component Ai, and the control, u, is the
mass fraction of hydrogenation component present in the catalyst. V r is the
characteristic volume of the reactor determined by the total mass of catalyst
in the reactor divided by the molar flowrate of methylcyclopentane into the
reactor. Cij are experimentally determined constants which relate the catalyst
blend to the reaction rate constants using a cubic equation.

Data

gh
Vr = 2000-1
mo
412 CHAPTER 15. DYNAMIC OPTIMIZATION PROBLEMS

Ci,j =
0.2918487xlO- 2 -0.8045787xlO- 2 0.6749947xlO- 2 -0. 1416647xlO- 2
0.9509977 X 101 -0.3500994 X 102 0.4283329 X 102 -0.1733333 X 10 2
0.2682093 X 10 2 -0.9556079 X 102 0.1130398 X 10 3 -0.4429997 X 10 2
0.2087241 X 103 -0.7198052 X 103 0.8277466 X 103 -0.3166655 X 10 3
0.1350005 X 10 1 -0.6850027 X 10 1 0.1216671 X 10 2 -0.6666689 X 10 1
0.1921995xlO- 1 -0.794532OxlO- 1 0.1105666 -0.5033333xlO- 1
0.1323596 -0.4696255 0.5539323 -0.2166664
0.7339981 x 10 1 -0.2527328 x 10 2 0.2993329 x 102 -0.1199999 x 10 2
-0.3950534 0.1679353 x 10 1 -0.1777829 X 10 1 0.4974987
-0.2504665xlO- 4 0.1005854x1O- 1 -0.1986696x10- 1 0.9833470xlO- 2

Problem Statistics
No. of continuous variables 10
No. of dynamic variables 17
No. of linear equalities
No. of convex inequalities
No. of nonlinear equalities
No. of nonconvex inequalities
No. of dynamic constraints 17
No. of known solutions > 300

Best Known Solution

Objective function: -10.09582 x 10- 3

Control variables
w = (0.66595, 0.67352, 0.67500, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9)
Bibliography

Abate M., Barmish B., Murillo-Sanchez C., and Tempo R., 1994, Application
of some new tools to robust stability analysis of spark ignition engines : A
case study. IEEE Trans. Contr. Byst. Teeh. 2, 22.

Abrams D. and Prausnitz J., 1975, Statistical thermodynamics of liquid mix-

tures: A new expression for the excess gibbs energy of partly or completely
miscible systems. AICHE J. 21, 116.

Achenie L.K.E. and Biegler L.T., 1986, Algorithmic synthesis of chemical re-
actor networks using mathematical programming. Ind. Eng. Chem. Fund.
25, 621.

Achenie L.K.E. and Biegler L.T., 1988, Developing targets for the performance
index of a chemical reactor network. Ind. Eng. Chem. Res. 27, 1811.

Ackermann J., Kaesbauer D., and Muench R., 1991, Robust gamma-stability
analysis in a plant parameter space. Automatiea 27, 75.

Adjiman C.S., Androulakis I.P., and Floudas C.A., 1998b, Aglobai optimiza-
tion method, aBB, for general twice-differentiable NLPs - 11. Implementa-
tion and computational results. Comput. Chem. Eng. 22, 1159-1179.

Adjiman C.S., Androulakis I.P., and Floudas C.A., 1998d, Globaloptimization

of mixed-integer nonlinear problems. submitted for publication.

Adjiman C.S., Dallwig S., Floudas C.A., and Neumaier A., 1998a, Aglobai
optimization method, aBB, for general twice-differentiable NLPs - I. The-
oretical advances. Comput. Chem. Eng. 22, 1137-1158.

Adjiman C.S. and Floudas C.A., 1996, Rigorous convex underestimators for
general twice-differentiable problems. Journal 0/ Global Optimization 9, 23-
40.

Aggarwal A. and Floudas C.A., 1990, Synthesis of general distillation

sequences-nonsharp separations. Comput. Chem. Eng. 14, 631-653.

Aggarwal A. and Floudas C.A., 1992, Synthesis of heat integrated nonsharp

distillation sequences. Comput. Chem. Eng. 16, 89.
413
414 BIBLIOGRAPHY

Aiyoshi E. and Shimizu K., 1981, Hierarchical decentralized systems and its
solution by a barrier method. IEEE Trans. on Sys., Man., Cyb. AC-26.
Aiyoshi E. and Shimizu K., 1984, A solution method for the static constrained
stackelberg problem via penalty method. IEEE Trans. on A utomatic Contral
AC-29, 1111-1114.
AI-Khayyal F.A. and Falk J.E., 1983, Jointly constrained biconvex program-
ming. Math. of Oper. Res. 8, 273-286.

AI-Khayyal F.A., Horst R., and Pardalos P.M., 1992, Global optimization
of concave functions subject to quadratic constraints: An application in
nonlinear bilevel programming. Annals of Operations Research 34, 125.
Alekseev V.M., Tikhomirov V.M., and Fomin S.V., 1987, Optimal Contral.
Contemporary Soviet Mathematics, Consultants Bureau, Translated by V.
M. Volosov.
Ali M. and Smith R., 1993, The structure of small clusters ejected by ion
bombardment of solids. Vacuum 44, 377-379.
All M.M., Storey C., and Torn A., 1997, Application of stochastic global opti-
mization algorithms to practical problems. JOTA 95, 545-563.
Alizadeh F., 1995, Interior point methods in semidefinite programming with
applications to combinatorialoptimization. SIAM J. Optimization 5, 13-51.
Altmann M., 1968, Bilinear programming. Bulletin de l'Academie Polonaise
des Sciences 16, 741-745.
Anandalingam G. and White D.J., 1990, A solution method for the linear static
stackelberg problem using penalty functions. IEEE Trans. Auto. Contr. 35,
1170-1173.
Anderson T.F. and Prausnitz J.M., 1978a, Application of the uniquac equation
to calculation of mulitcomponent phase equilibria. 1: Vapor-liquid equilibria.
Ind. Eng. Chem. Prac. Des. Dev. 17, 552.
Anderson T.F. and Prausnitz J.M., 1978b, Application ofthe uniquac equation
to calculation of mulitcomponent phase equilibria. 1: Liquid-liquid equilib-
ria. Ind. Eng. Chem. Prac. Des. Dev. 17, 561.
Androulakis I.P., Maranas C.D., and Floudas C.A., 1995, aBB : AglobaI
optimization method for general constrained nonconvex problems. J. Glob.
Opt. 7, 337-363.
Androulakis I.P., Visweswaran V., and Floudas C.A., 1996, Distributed
decomposition-based approaches in global optimization. In Floudas C.A. and
Pardalos P.M. (eds.), State of the Art in Global Optimization: Computational
Methods and Applications, (pp. 285-301), Kluwer Academic Publishers.
BIBLIOGRAPHY 415

Avriel M. and Williams A.C., 1971, An extension of geometrie programming

with applications in engineering optimization. Journal of Engineering Math-
ematics 5, 187-194.
Baker T.E. and Lasdon L.S., 1985, Successive linear programming at Exxon.
Management Science 31, 264-274.

Balakrishna S. and Biegler L.T., 1992a, A constructive targeting approach for

for the synthesis of isothermal reactor networks. Ind. Eng. Chem. Res. 31,
300.
Balakrishna S. and Biegler L.T., 1992b, Targeting strategies for synthesis and
energy integration of nonisothermal reactor networks. Ind. Eng. Chem. Res.
31,2152.

Bard J.F., 1984, An investigation of the linear three-Ievel programming prob-

lem. IEEE Transactions on Systems, man and Cybernetics 14, 711-717.

Bard J.F., 1988, Convex two-Ievel optimization. Mathematical Programming

40,15-27.

Bard J.F., 1991, Some properties ofthe bilevel programming problem. Journal
of Optimization Theory and Applications 68, 371-378.

Bard J.F., 1998, Practical bilevel optimization algorithms and applications.

Nonconvex optimization and its applications, Kluwer Academic Publishers.
Bard J.F. and Falk J., 1982, An explicit solution to the multi-level program-
ming problem. Comp. Op. Res. 9, 77-100.

Bard J.F. and Moore J.T., 1990, A branch and bound algorithm for the bilevel
programming problem. SIAM J. on scientific and statistical computing 11,
281-282.

Bard J.F., Plummer J., and C. S.J., 1998, Determining tax credits for con-
verting nonfood crops to biofuels: an application of bilevel programming. In
Migdalas A., Pardalos P.M., and Varbrand P. (eds.), Multilevel Optimiza-
tion: Algorithms and Applications, vol. 20 of Nonconvex Optimization and
Its Applications, (p. 23), Kluwer Academic Publishers.

Bard Y., 1974, Nonlinear Parameter Estimation. Academic Press, New York.

Barmish B.R., 1994, New Tools for Robustness of Linear Systems. MacMillan
Publishing Company, New York, NY.

Barreln C., Gomez S., and Romero D., 1997, Lower energy icosahedral atomic
clusters with incomplete core. App. Math. Lett. 10, 25-28.

Beale E.M.L., 1977, Integer programming. In Jacobs D. (ed.), The State of the
Art in Numerical Analysis, (pp. 409-448), Academic Press.
416 BIBLIOGRAPHY

Beasley J., 1996, Obtaining test problems via the internet. Journal of Global
Optimization 8, 391-403.

Bekiaris N., Meski G.A., Radu C.M., and Morari M., 1993, Multiple steady-
states in homogeneous azeotropic distillation. Ind. Eng. Chem. Res. 23,
2023-2038.

Bellman R., Jacquez J., Kalaba R., and Schwimmer S., 1967, Quasilineariza-
tion and the estimation of chemical rate constants from raw kinetic data.
Mathematical Biosciences 1, 71.

Ben-Ayed 0., 1993, Bilevellinear programming. Comp. Op. Res. 20, 485.

Ben-Ayed O. and BIair C.E., 1990, Computational difficulties of bilevellinear

programming. Operations Research 38, 556-559.

Ben-Ayed 0., Boyce D.E., and BIair In C.E., 1988, A general bilevel linear
programming problem formulation of the network design problem. Trans.
Res. 22B, 311.

Ben-Saad S., 1989, An algorithm for a dass of nonlinear nonconvex optimiza-

tion problems. Ph.D. thesis, University of California, Los Angeles.

Ben-Tal A., Eiger G., and Gershovitz V., 1994, Global minimization by reduc-
ing the duality gap. Mathematical Programming 63, 193-212.

Ben-Tel A. and Bends\?le M.P., 1993, A new method for optimal truss topology
design. SIAM J. Optimization 3, 322-358.

Berman O. and Ashrafi N., 1993, Optimization models for reliability ofmodular
software systems. IEEE Transactions on Software Engineering 19, 1119-
1123.

Bialas W. F. Karwan M.H., 1982, On two-Ievel optimization. IEEE Transac-

tions on automatie control AC-27, 211-214.

Bialas W. F. Karwan M.H., 1984, Two-Ievellinear programming. Management

Science 30, 1004.

Billups S.C., Dirkse S.P., and Ferris M.C., 1997, A comparison of large scale
mized complementarity problem solvers. Computational Optimization and
Applications 7, 3-25.

Birewar D.B. and Grossmann LE., 1989, Incorporating scheduling in the opti-
mal design of multiproduct batch plants. Comput. Chem. Eng. 13, 141-161.

Bojkov B. and Luus R., 1993, Evaluation of the prameters used in iterative
dynamic programming. Can. J. Chem. Eng. (p. 451).
BIBLIOGRAPHY 417

Bolding B.C. and Andersen H.C., 1990, Interatomic potential for silicon clus-
ters, crystals and surfaces. Phys. Rev. B 41, 10568-10585.

Bongartz 1., Conn A.R., Could N., and Toint P., 1995, Cute - constrained
and unconstrained testing environment. ACM Transactions 0/ Mathematical
Software 21, 123-160.

Borehers B. and Mitchell J.E., 1991, An improved branch and bound algorithm
for mixed integer nonlinear programs. Tech. Rep. RPI Math Report No. 200,
Renssellaer Polytechnic Institute.

Boyd S., EI Ghaoui L., Feron E., and Balakrishnan V., 1994, Linear Matrix In-
equalities in System and Control Theory, vol. 15 of Studies in Applied Math-
ematics. Society for Industrial and Applied Mathematics, Philadelphia, PA.

Bracken J. and McGill J.T., 1973, A convex programming model for optimiza-
tion problems in the constraints. Operations Research 21, 30.

Brenner D.W., 1990, Empirical potential for hydrocarbons for use in simulating
the chemical vapor deposition of diamond films. Phys. Rev. B 42, 9458-9471.

Brooke A., Kendrick D., and Meeraus A., 1988, GAMS: A User's Guide. The
Scientific Press, South San Fransisco, CA.

Bryson A.E. and Ho Y.C., 1988, Applied Optimal Contral. Hemisphere Pub-
lishing Corporation, Washington.

Bullard L.G. and Biegler L.T., 1991, Iterative linear-programming strategies

for constrained simulation. Comput. Chem. Eng. 15, 239-254.

Burkard R., Karisch S., and Rendl F., 1997, Qaplib - a quadratic assignment
problem library. Journal 0/ Global Optimization 10, 391-403.

Buss M., Hashimoto H., and Moore J., 1996, Dextrous hand grasping force
optimization. IEEE Transactions on Robotics and Automation 12,406-418.

Calamai P.H. and Vicente L.N., 1993, Generating linear and linear-quadratic
bilevel programming problems. SIAM J. Sei. Comput. 14, 770-782.

Calamai P.H. and Vicente L.N., 1994, Generating quadratic bilevel program-
ming problems. ACM T. Math. Software 20, 103-119.

Calamai P.H., Vicente L.N., and Judice J.J., 1993, A new technique for gen-
erating quadratic programming test problems. Mathematical Pragramming
61, 215-231.

Candler W. and Norton N., 1977, Multi-level programming and development

policy. Tech. Rep. Working Paper No.258, World Bank.
418 BIBLIOGRAPHY

Candler W. and Townsley R., 1982, A linear two-Ievel programming problem.

Comp. Op. Res. 9, 59-76.

Carrasco E. and Banga J., 1997, Dynamic optimization ofbatch reactors using
adaptive stochastic algorithms. Ind. Eng. Chem. Res 36, 2252.

Cassidy R.G., Kriby M.J.L., and Raike W.M., 1971, Efficient distribution of
resources through three levels of government. Management Science 17, B-
462.

Castillo J. and Grossmann 1., 1981, Computation of phase and chemical equi-
libria. Comput. Chem. Eng. 5, 99.

Cavalier T.M., Pardalos P.M., and Soyster A.L., 1990, Modeling and integer
programming techniques applied to propositional calculus. Computers and
Operations Research 17, 561-570.

Chapman R.G. and Goodwin S.P., 1993, A general algorithm for the calcula-
tion of azeotropes in fluid mixtures. Fluid Phase Equilibrium 85, 55-69.

Chen C. and Mangasarian O.L., 1996, A dass of smoothing functions for non-
linear and mixed complementarity problems. Comput. Optim. Appl. 5, 97-
138.

Chen H. and Stadtherr M., 1984, On solving large sparse nonlinear equation
systems. Comput. Chem. Eng. 8, 1-7.

Chen J.J.J., 1987, Comments on improvements on areplacement for the loga-

rithmic mean. Chem. Eng. Sei. 42, 2488-2489.

Chen Y. and Florian M., 1995, The nonlinear bilevel programming problem:
formulations, regularity, and optimality conditions. Optimization 32, 193-
209.

Choi S., W.S. D., and Harker P., 1990, Product positioning under price com-
petition. Management Science 36, 265-284.

Chu M.T. and Wright J.W., 1995, The educational testing problem revisited.
IMA Journal 0/ Numerical Analysis 15, 141-160.

Churi N. and Achenie L.E.K., 1996, Novel mathematical programming model

for computer aided molecular design. Ind. Eng. Chem. Res. 35, 3788-3794.

Clark P.A. and Westerberg A.W., 1983, Optimization for design problems
having more than one objective. Comput. Chem. Eng. 7, 259-278.

Clark P.A. and Westerberg A.W., 1988, A note on the optimality conditions
for the bilevel programming problem. Naval research logistics 35, 413-418.
BIBLIOGRAPHY 419

Clark P.A. and Westerberg A.W., 1990a, Bilevel programming for steady-state
chemical process design-i. fundament als and algorithms. Comput. Chem.
Eng. 14, 87.
Clark P.A. and Westerberg A.W., 1990b, Bilevel programming for steady-state
chemical process design-ii. performance study for nondegenerate designs.
Comput. Chem. Eng. 14, 99.

Coleman T., Shalloway D., and Wu Z.J., 1994, A parallel buildup algorithm
for global energy minimization of molecular clusters using effective energy
simulated annealing. J. Glob. Opt. 4, 171-185.
Colville A.R., 1970, A comparative study of nonlinear programming codes.
In Kuhn H.W. (ed.), Princeton Symposium on Mathematical Programming,
Princeton Univ. Press.

Cook S., 1971, The complexity of theorem-proving procedures. Proceedings 01

the Third annual ACM Symposium on Theory 01 Computing (pp. 151-1580).

Cottle R. and Dantzig G.B., 1968, Complementary pivot theory of mathemat-

ical programming. In Dantzig G. and Veinott JR. A. (eds.), Mathematics
01 the Decision Sciences, Part 1, (pp. 115-136), Americal Mathematical
Society, Providence, Rhode Island.
Cottle R.W., Pang J., and Stone R.E., 1992, The linear complementarity prob-
lem. Academic Press, San Diego, CA.
Csendes T. and Ratz D., 1995, Sub division direction selection in interval meth-
ods for global optimization. Journal 01 Global Optimization 7, 183.

Cuthrell J.E. and Biegler L.T., 1987, On the optimization of differential-

algebraic process systems. AIChE J. 33, 1257-1270.
Cvijovic D. and Klinowski J., 1995, Taboo search - an approach to the multiple
minima problem. Science 267, 664-666.
Dantzig G.B., 1963, Linear Programming and Extensions. Princeton University
Press, Princeton, NJ.

Dantzig G.B. and Cottle R., 1967, Positive (semi- )definite programming. In
Abadie J. (ed.), Nonlinear Programming, North-Holland, Amsterdam.

Davidenko D., 1953, On a new method of numerically integrating a system of

nonlinear equations. Dokl. Akad. Nauk USSR 88, 601.

de Gaston R. and Safonov M., 1988, Exact calculation of the multi-Ioop sta-
bility margin. IEEE Trans. Automat. Contr. 33, 68.

Dean J.P. and Dervakos G.A., 1996, Design of process-compatible biological

agents. Comput. Chem. End. 20, S67-S72.
420 BIBLIOGRAPHY

Deaven D.M., Tit N., Morris J.R, and Ho K.M., 1996, Structuraloptimization
of lennard-jones cluster by a genetie algorithm. Chem. Phys. Lett. 256, 195-
200.

Dembo RS., 1976, A set of geometrie programming test problems and their
solutions. Math. Prog. 10, 192-213.

Dirkse S.P. and Ferris M.C., 1995, Mcplib: A collection of nonlinear mixed
complementarity problems. Optimization Methods and Software 5, 319-345.

Dixon L. and Szeg G.P., 1975, Towards global optimization. In Proeeedings

0/ a Workshop at the University 0/ Cagliari, Italy, North Holland.
Dixon L.C.W., 1990, On finding the global minimum of a function of one
variable. Presented at the SIAM National Meeting, Chieago.

Dontchev A.L., 1983, Perturbations, Approximation and Sensitivity Analysis

0/ Optimal Control Systems. No. 52 in Lecture Notes in Control and Infor-
mation Sciences, Springer-Verlag.
Doye J.P.K. and Wales D.J., 1995, Magic numbers and growth sequences of
small face-centered cubic and decahedral clusters. Chem. Phys. Lett. 247,
339-347.
Doye J.P.K. and Wales D.J., 1997, Structural consequences of the range of
the interatomie potential: A menagerie of clusters. J. Chem. Soe., Faraday
Trans. 93,4233-4244.
Doye J.P.K., Wales D.J., and Berry RS., 1995, The effect of the range of the
potential on the structures of clusters. J. Chem. Phys. 103, 4234-4249.
Du D., Gu J., and Pardalos P.M. (eds.), 1997, Satisfiability Problem: Theory
and Applieations. DIMACS Series Vol. 35, Ameriean Mathematical Society.
Du D. and Pardalos P.M. (eds.), 1998, Handbook in Combinatorial Optimiza-
tion. Kluwer Academie Publishers.
Duff 1., Nocedal J., and Reid J., 1987, The use of linear-programming for the
solution of sparse sets of nonlinear equations. SIAM J. Sei. Stat. Comp. 8,
99-108.
Duran M.A. and Grossmann I.E., 1986, An outer-approximation algorithm for
a class of mixed-integer nonlinear programs. Math. Prog. 36, 307-339.

Edmunds T.A. and Bard J.F., 1991, Algorithms for nonlinear bilevel mathe-
matical programs. IEEE Trans. on Systems, Man and Cybernetics 21, 83-89.

Esposito W.R and Floudas C.A., 1998a, Global optimization in parameter

estimation of nonlinear algebraie models via the error-in-variables approach.
Ind. Eng. Chem. Res. 35, 1841-1858.
BIBLIOGRAPHY 421

Esposito W.R. and Floudas C.A., 1998b, Parameter estimation of nonlinear

algebraic models via global optimization. Comput. Chem. Eng. 22, S213-
S220.

Eubank P., Elhassen A., Barrufet M., and Whiting W., 1992, Area method for
predietion of fluid-phase equilibria. Ind. Eng. Chem. Res. 31, 942.

Falk J.E., 1973, A linear max-min problem. Math. Prog. 5, 169-188.

Falk J.E. and Liu J., 1995, On bilevel programming, part i: general nonlinear
eases. Mathematieal Programming 70,47.

Falk J.E. and Soland R.M., 1969, An algorithm for separable noneonvex pro-
gramming problems. Management Seienee 15, 550-569.

Farges J., de Feraudy M.F., Raoult B., and Torehet G., 1985, Cluster models
made of double ieosahedron units. Surf. Sei. 156, 370-378.

Feinberg M. and Hildebrandt D., 1997, Optimal reaetor design from a geomet-
rie viewpoint-I. universal properties of the attainable region. Chem. Eng.
Sei. 52, 1637-1665.

Ferraris G.B. and Troneoni E., 1986, BUNLSI - a FORTRAN program for
solution of systems of nonlinear algebraie equations. Comput. Chem. Eng.
10, 129-141.

Ferris M.C. and Pang J.S., 1997, Engineering and eeonomie applieations of
eomplementarity problems. SIAM Review 39, 669-713.

Fidkowski Z., Malone M., and Doherty M., 1993, Computing azeotropes in
multicomponent mixtures. Comput. Chem. Eng. 17, 1141-1155.

Fisk C.S., 1984, Game theory and transportation systems modeling. Trans.
Res. B. 18B, 301.

Fleteher R., 1981a, A nonlinear programming problem in statisties (edueational

testing). SIAM J. Sei. Statist. Comput. 2, 257-267.

Fleteher R. and LeyfIer S., 1994, Solving mixed integer nonlinear programs by
outer approximation. Math. Prog. 66, 327.

Fleteher R. and LeyfIer S., 1998, Numerieal experienee with lower bounds for
miqp braneh and bound. SIAM J. Optim. 8, 604-616.

Floudas C.A., 1987, Separation synthesis of multieomponent feed streams into

multicomponent produet streams. AIChE J. 33, 540-550.

Floudas C.A., 1995, Nonlinear and Mixed Integer Optimization: Fundamentals

and Applieations. Oxford University Press, New York.
422 BIBLIOGRAPHY

Floudas C.A., 1997, Deterministie global optimization in design, control, and

computational chemistry. In Biegler L.T., Coleman T.F., Conn A.R., and
Santosa F.N. (eds.), IMA Volumes in Mathematies and its Applieations:
Large Scale Optimization with Applieations, Part II, vol. 93, (pp. 129-184),
New York, Springer-Verlag.
Floudas C.A., 2000, Deterministie Global Optimization: Theory, Computa-
tional Methods and Applieations. Nonconvex Optimization and its Applica-
tions, Kluwer Academie Publishers.
Floudas C.A. and Aggarwal A., 1990, A decomposition strategy for global
optimum search in the pooling problem. OSRA Journal on Computing 2,
225-234.
Floudas C.A., Aggarwal A., and Ciric A.R., 1989, Global optimal search for
nonconvex NLP and MINLP problems. Comput. Chem. Eng. 13, 1117.
Floudas C.A. and Anastasiadis S.H., 1988, Synthesis of distillation sequences
with several multieomponent feed and product streams. Chem. Eng. Sei. 43,
2407-2419.
Floudas C.A. and Cirie A.R., 1988, Global optimum issues on heat exchanger
networks synthesis. In Proeeedings of the Third International Symposium on
Proeess Systems Engineering, Sydney, Australia, (p. 104).
Floudas C.A., Ciric A.R., and Grossmann I.E., 1986, Automatie synthesis of
optimum heat exchanger network configurations. AIChE J. 32, 276.
Floudas C.A. and Grossmann I.E., 1995, Algorithmie approaches to process
synthesis: logie and global optimization. In FOCAPD'94 , vol. 91 of AICHE
Symposium Series, (pp. 198-221).
Floudas C.A. and Pardalos P.M. (eds.), 1990, A Colleetion of Test Problems
for Constrained Global Optimization Algorithms. Springer-Verlag, Lecture
Notes In Computer Science 455.
Floudas C.A. and Pardalos P.M. (eds.), 1996, State of the Art in Global Opti-
mization. Kluwer Academie Publishers.
Floudas C.A. and Visweswaran V., 1990, Aglobai optimization algorithm
(GOP) for certain classes of nonconvex NLPs : I. theory. Comput. Chem.
Eng. 14, 1397.
Floudas C.A. and Visweswaran V., 1993, A primal-relaxed dual global opti-
mization approach. Journal of Optimization Theory and Applieations 78,
187.
Floudas C.A. and Visweswaran V., 1995, Quadratic optimization. In Horst R.
and Pardalos P.M. (eds.), Handbook of Global Optimization, (pp. 217-270),
Kluwer Academie Publishers.
BIBLIOGRAPHY 423

Floudas C.A. and Zlobee S., 1998, Optimality and duality in parametrie eonvex
lexicographic programming. In Migdalas A., Pardalos P.M., and Varbrand
P. (eds.), Multilevel Optimization:Algorithms and Applications, Kluwer Aea-
demic Publishers.
Fogler H., 1986, Elements 0/ Chemical Reaction Engineering. Pentiee-Hall,
Ine.

Fortuny-Amat J. and MeCarl B., 1981, A representation and eeonomic inter-

pretation of a two-Ievel programming problem. Journal 0/ Operations Re-
search Society 32, 783.
Fredenslund A., Jones R., and Prausnitz J., 1975, Group eontribution estima-
tion of activity coeffieients in non-ideal mixtures. AIChE J. 21, 1086.

Freeman D.L. and Doll J.D., 1984, Quantum monte earlo study of the thermo-
dynamie properties of argon clusters: The homogeneous nucleation of argon
in argon vapor and "magie number" distributions in argon vapor. J. Chem.
Phys. 82, 462-471.
Gallo G. and Ulcul A., 1977, Bilinear programming: an exaet algorithm.
Mathematical Programming 12, 173-194.
Garcia C. and Zangwill W., 1981, Pathways to Solutions, Fixed Points and
Equilibria. Prentiee-Hall.
Geoffrion A.M., 1972, Generalized benders deeomposition. Journal 0/ Opti-
mization Theory and Applications 10, 237.
Glasser D., Crowe C., and Hildebrandt D., 1987, The attainable region and
optimization in eoneentration spaees. Ind. Eng. Chem. Res 26, 1803-1810.

Glinos K., Marquez F., and Douglas J., 1985, Simple, analytical eriteria for
the sequencing of distillation eolumns. AIChE Journal 31, 683.

Glover F. and de Werra D. (eds.), 1993, Tabu Search, vol. 41 of Annals 0/

operations research. J.C. Baltzer AG.
Gmehling J., Tiegs D., and Knipp U., 1990, A eomparison of the predietive
eapability of different group eontribution methods. Fluid Phase Equilibria
54, 147-165.

Goldberg D.E., 1989, Genetic Algorithms in Search, Optimization, and Ma-

chine Learning. Addison-Wesley, Reading, MA.

Goldstein A. and Price J., 1971, On deseent from loeal minima. Mathematics
0/ Computation 25, 569-574.
Graves R.L., 1967, A prineipal pivoting simplex algorithm for linear and
quadratic programming. Operations Research 15, 482-494.
424 BIBLIOGRAPHY

Green K., Zhou S., and Luks K., 1993, The fractal response of robust solution
techniques to the stationary point problem. Fluid Phase Equilibria 84, 49-
78.

Grossmann I.E. (ed.), 1996, Global Optimization zn Engineering Design.

Kluwer Academic Publishers, Dordrecht.

Grossmann I.E. and Floudas C.A., 1987, Active constraint strategy for flexi-
bility analysis in chemical processes. Comput. Chem. Eng. 11, 675-693.

Grossmann I.E. and Sargent R., 1979, Optimal design of multipurpose chemical
plants. Ind. Eng. Chem. Proe. Des. Dev. 18, 343.

Gm~ Z.H. and Ciric A.R., 1997, Reactive distillation column design with
vaporjliquidjliquid equilibria. Comput. Chem. Eng. 21, S983.
Gundersen T. and Naess L., 1988, The synthesis of cost optimal heat exchanger
networks, an industrial review of the state of the art. Comput. Chem. Eng.
12,503.

Gupta O.K. and Ravindran R., 1985, Branch and bound experiments in convex
nonlinear integer programing. Management Sei. 31, 1533-1546.

Hager W.W. and Pardalos P.M., 1998, Optimal Control: Theory, Algorithms
and Applieations. Kluwer Academic Publishers.

Hansen E., 1992, Global Optimization using Interval Analysis. Pure and Ap-
plied Mathematics, Marcel Dekker, Inc.
Hansen P., Jaumard B., and Savard G., 1992, New branch and bound rules
for linear bilevel programming. SIAM Journal on seientifie and statistieal
eomputing 13, 1194-1217.
Hansen T. and Koopmans T., 1972, On the definition and computation of a
capital stock invariant under optimization. Journal 01 Eeonomie Theory 5,
487-523.

Harding S. and Floudas C., 1997, Global optimization in multiproduct and

multipurpose batch design under uncertainty. If3EC Res. 36, 1644-1664.

Harding S.T., Maranas C.D., McDonald C.M., and Floudas C.A., 1997, Lo-
cating all azeotropes in homogeneous azeotropic systems. If3EC Res. 36,
160-178.

Harjunkoski 1., 1997, Applieation 01 MINLP methods on a seheduling problem

in the paper eonverting industry. Ph.D. thesis, Abo Akademi University,
Abo, Finland.

Harker P., 1984, A generalized spatial prize equilibrium model. Papers of the
Regional Seienee Assoeiation 54, 25-42.
BIBLIOGRAPHY 425

Harker P., 1988, Aeeelerating the eonvergenee of the diagonalization and pro-
jeetion algorithms for finite-dimentional variational inequalities. Mathemat-
ical Programming 41, 29-59.

Harker P. and Pang J.S., 1990, Finite-dimensional variational inequality and

nonlinear eomplementarity problems: a survey of theory, algorithms and
applications. Mathematical Programming 48, 161-220.

Hasselberg J., Pardalos P., and Vairaktarakis G., 1993, Test ease genera-
tors and eomputational results for the maximum clique problem. Journal
01 Global Optimization 3, 463-482.

Hatzimanikatis V., Floudas C.A., and Bailey J.E., 1996a, Analysis and design
of metabolie reaction networks via mixed-integer linear optimization. AIChE
Journal 42, 1277-1292.

Hatzimanikatis V., Floudas C.A., and Bailey J.E., 1996b, Optimization of

regulatory arehiteetures in metabolie reaetion networks. Biotechnology and
Bioengineering 52, 485-500.

Haverly C., 1978, Studies of the behavior of reeursion for the pooling problem.
ACM SIGMAP Bulletin 25, 19-28.

Heidemann R. and Mandhane J., 1973, Some properties of the NRTL equation
in eorrecting liquid-liquid equilibrium data. Chem. Eng. Sei. 28, 1213.

Helmberg C. and Rendl F., 1998, Solving quadratic (0,1)-problems by semidef-

inite programs and eutting planes. Mathematical Programming 82,291-315.

Hesse R., 1973, A heuristic seareh proeedure for estimating a global solution
of noneonvex programming problems. Op. Res. 21, 1267.

Hildebrandt D. and Glasser D., 1990, The attainable region and optimal reaetor
struetures. Chem. Eng. Sei. 45, 2161-2168.

Hildebrandt D., Glasser D., and Crowe C.M., 1990, Geometry of the attainable
region generated by by reaetion and mixing: with and without eonstraints.
Ind. End. Chem. Res. 29,49-58.

Hiriart-Urruty J.B. and Lemareehal C., 1993, Convex Analysis and Mini-
mization Algorithms II: Advanced Theory and Bundle Methods, vol. 306
of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, New
York, Berlin.

H.Kautz, Selman B., and Jiang Y., 1997, A general stochastie approach to
solving problems with hard and soft eonstraints. In Satisjiability Problem:
Theory and Applications, (pp. 573-586), DIMACS Series Vol. 35, American
Mathematieal Soeiety.
426 BIBLIOGRAPHY

Hoare M.R. and McInnes J., 1983, Morphology and statistieal staties of simple
microclusters. Adv. Phys. 32, 791-821.
Hoare M.R. and Pal P., 1971, Physical cluster meehanies: Staties and energy
surfaees for monoatomie systems. Adv. Phys. 20, 161-196.

Hoare M.R. and Pal P., 1972, Geometry and stability of "spherical" fee miero-
erystallites. Nature Phys. Sci. 236, 35-37.

Hobbs B.F. and Nelson S.K., 1992, A nonlinear bilevel model for analysis of
eleetric utility demand-side planning issues. Annals of Operations Research
34,255.

Hobday S. and Smith R., 1997, Optimisation of earbon cluster geometry using
a genetic algorithm. J. Chem. Soc., Faraday Trans. 93, 3919-3926.

Hock W. and Schittkowski K., 1981, Test Examples for Nonlinear Programming
Codes, vol. 187 of Lecture Notes in Economics and Mathematical Systems.
Springer Verlag.

Horn F., 1964, Attainable regions in ehemical reaction teehnique. In The Third
European Symposium on Chemical Reaction Engineering, Pergamon.

Horn F.J.M. and Tsai M.J., 1967, The use of adjoint variables in the develop-
ment of improvement eriteria for ehemieal reactors. J. Opt. Theory Applic.
1, 131-145.

Horst R. and Pardalos P.M., 1995, Handbook of Global Optimization. Kluwer

Aeademic Publishers, Dordreeht.
Horst R. and Tuy H., 1993, Global optimization: deterministic approaches.
Springer-Verlag, Berlin, 2nd. rev. edition.
Hua J., Brennecke J., and Stadtherr M., 1998, Enhanced interval analysis for
phase stability: Cubic equation of state models. If3EC Res. 37, 1519-1527.
Ierapetritou M. and Pistikopoulos E., 1995, Design of multiproduct bat eh
plants with uncertain demands. Comput. Chem. Eng. 19, suppl., S627-
S632.

Ierapetritou M. and Pistikopoulos E., 1996, Batch plant design and operations
under uncertainty. Ind. Eng. Chem. Res. 35, 772-787.

Isac G., 1992, Complementarity Problems. Springer-Verlag, Berlin, Germany.

Jackson R., 1968, Optimization of chemical reactors with respect to flow con-
figuration. J. Opt. Theory Applic. 2, 240-259.

Jeroslow R.G., 1985, The polynomial hierarchy and a simple model for com-
petitive analysis. Math. Programming 32, 146.
BIBLIOGRAPHY 427

Johns W., Marketos G., and Rippin D., 1978, The optimal design of chem-
ical plant to meet time-varying demands in the presence of technical and
commercial uncertainty. Trans. Inst. Chern. Eng. 56, 249-257.
Judice J.J. and Faustino A.M., 1992, A sequentiallcp method for bilevellinear
programming. Annals of Operations Research 34, 89-106.

Kanzow C., 1996, Nonlinear complementarity as unconstrained optimization.

Journal of Optirnization Theory and Applications 88, 139-155.

Kearfott R, Dawande M., Du K, and Ru C., 1994, Intlib, a portable fortran

77 interval standard function library. subrnitted .

Kearfott Rand Novoa M., 1990, INTBIS, a portable interval newton bisection
package. ACM Trans. Math. Soft. 16, 152-157.

Khoury B.N., Pardalos P.M., and Du D.Z., 1993, A test problem generator for
the steiner problem in graphs. ACM Trans. Math. Soft. 19, 509-522.

Kim 1., Liebman M.J., and Edgar T.F., 1990, Robust error-in-variables esti-
mation using nonlinear programming techniques. AIChE J. 36, 985.

Kirkpatrick S., Gellat D., and Vecci M., 1983, Optimization by simulated
annealing. Science 220, 671-680.

Kiwiel KC., 1985, Methods of Descent for Nondifferentiable Optirnization, vol.

1133 of Lecture Notes in Mathernatics. Springer-Verlag, New York, Berlin.

Klopfenstein R., 1961, Zeroes of nonlinear functions. J. Assoe. Cornput. Mach.

8,366.

Kocis G.R and Grossmann I.E., 1987, Relaxation strategy for the structural
optimization of process fiow sheets. Ind. Eng. Chern. Res. 26, 1869.
Kocis G.R. and Grossmann I.E., 1988, Global optimization of nonconvex
mixed-integer nonlinear programming (MINLP) problems in process syn-
thesis. Ind. Eng. Chern. Res. 27, 1407-1421.
Kojima K and Tochigi K, 1979, Prediction of Vapor-Liquid Equilibria by the
ASOG Method. Elsevier, Englewood Cliffs, New Jersey.

Kojima M. and Shindo S., 1986, Extensions of newton and quasi-newton meth-
ods to systems of pe 1 equations. Journal of Operations Research Society of
Japan 29, 352-374.

Kokossis A.C. and Floudas C.A., 1990, Optimization of complex reactor

networks-1. isothermal operation. Chern. Eng. Sei. 45, 595-614.

Kokossis A.C. and Floudas C.A., 1994, Optimization of complex reactor

networks-II. nonisothermal operation. Chern. Eng. Sei. 49, 1037-1051.
428 BIBLIOGRAPHY

Kolstad C.D., 1985, A review of the literat ure on bilevel mathematical pro-
gramming. Tech. Rep. LA-10284-MS, US-32, Los Alarnos National Lab.

Konno H., 1976, A cutting plane algorithm for solving bilinear programs. Math-
ematieal Programming 11, 14-27.
Kubicek M., Hofmann H., Hlavacek V., and Sinkule J., 1980, Multiplicity
and stability in a sequence of two nonadiabatic nonisothermal cstrs. Ghem.
Engng Sei. 35, 987.
Lahaye E., 1934. G.R. Aead. Sei. Paris 198, 1840.

Lakshmanan A. and Biegler L.T., 1996, Synthesis of optimal chemical reactor

networks. Ind. Eng. Ghem. Res. 35, 1344.

Lasdon 1., Waren A., Sarkar S., and Palacios F., 1979, Solving the pooling
problem using generalized reduced gradient and successive linear program-
ming algorithms. ACM SIGMAP Bulletin 26, 9-15.

Lawler L., Lenstra J., Kan H.R., and Shmoys D. (eds.), 1995, The travelling
salesman problem: a guided tour 01 eombinatorial optimization. Wiley, New
York.

Leary R.H., 1997, Global optima of lennard-jones clusters. J. Glob. Opt. 11,
35-53.

LeBlanc L.J. and Boyce D.E., 1986, A bilevel programming algorithm for exact
solution of the network design problem with user optimal flows. Trans. Res.
B. 20B, 259.

Lebret H. and Boyd S., 1997, Antenna array pattern synthesis via convex
optimization. IEEE Trans. on Signal Proeessing 45, 526-532.
Lemke C.E., 1968, On complementary pivot theory. In Dantzig G. and Veinott
JR A.F. (eds.), Mathematies 01 the Deeision Seiences, Part 1, (pp. 95-1l4),
American Mathematical Society, Providence, Rhode Island.

Lemke C.E. and Howson J.J., 1964, Equilibrium points of bi matrix games.
SIAM Journal on Applied Mathematies 12, 413-423.

Leray J. and Schauder J., 1934, Topologie et equations fonctionelles. Ann. Sei.
Eeole Norm Sup. 51, 45.

Li Y. and Pardalos P.M., 1992, Generating quadratic assignment test prob-

lems with known optimal permutations. Computational Optimization and
Applieations 1, 163-184.

Liu Y.H. and Hart S.M., 1994, Characterizing an optimal solution to the linear
bilevel programming problem. European Journal 01 Operational Research 79,
164-166.
BIBLIOGRAPHY 429

Lobo M.S., Vandenberghe L., Boyd S., and Lebret H.,

1998, Applications of second-order cone programming.
http://www.stanford.edu/-boyd/socp.html.
Logsdon J.S. and Biegler L.T., 1989, Accurate solution of differential-algebraic
optimization problems. Ind. Eng. Ghem. Res. 28, 1628-1639.

Luh P.B., Ho Y., and Muralidaran R., 1982, Load adaptive pricing: an emerg-
ing tool for electric utilities. IEEE Trans. Auto. Gontr. AC-27, 320.

Luo Z.Q., Pang J.S., and Ralph D., 1997, Mathematical Programs with Equi-
librium Gonstraints. Cambridge University Press.

Luus R., 1990a, Application of dynamic programming to high-dimensional

non-linear optimal control problems. Int. J. Gontrol 52, 239-250.

Luus R., 1990b, Optimal control by dynamic programmin using systematic

reduction in grid size. Int. J. Gontrol 5, 995-1013.

Luus R., 1993, Piecewise linear continuous optimal control by iterative dynamic
programming. Ind. Eng. Ghem. Res. 32, 859-865.

Luus R., 1998, Parameter estimation of lotka-volterra problem by direct search

optimization. Hung. J. Ind. Ghem. 26.

Luus R. and Bojkov B., 1994, Global optimization of the bifunctional catalyst
problem. Gan. J. Ghem. Eng. 72, 160.

Luus R., Dittrich J., and Keil F., 1992, Multiplicity of solutions in the opti-
mization of a bifunctional catalyst blend in a tubular reactor. Gan. J. Ghem.
Eng. 70, 780.
Luus R. and Rosen 0., 1991, Application of dynamic programming to final
state constrained optimal control problems. Ind. Eng. Ghem. Res. 30, 1525-
1530.

Magnussen T., Rasmussen P., and Fredenslund A., 1981, UNIFAC parameter
table for prediction of liquid-liquid equilibria. IBEG Proc. Des. Dev. 26,
159.

Mangasarian 0., 1964, Equilibrium points of bimatrix games. SIAM Journal

12, 778-780.

Mangasarian O. and Stone H., 1974, Two person nonzero-sum games and
quadratic programming. Journal 0/ Mathematical Analysis and Applications
9,345-355.

Mangasarian O.L. and Solodov M.V., 1993, Nonlinear complementarity as un-

constrained and constrained minimization. Mathematical Programming 62,
277-297.
430 BIBLIOGRAPHY

Manousiouthakis V. and Sourlas D., 1992, Aglobai optimization approach to

rationally constrained rational programming. Chem. Eng. Comm. 115, 127.

Maranas C., McDonald C., Harding S., and Floudas C., 1996, Locating all
azeotropes in homogeneous azeotropic systems. Comput. Chem. Eng. 20,
S413-S418.

Maranas C.D., 1996, Optimal eomputer-aided moleeular design: a polymer

design case study. Industrial f3 Engineering Chemistry Research 35, 3403-
3414.

Maranas C.D. and Floudas C.A., 1994, Global minimum potential energy con-
formations of small molecules. Journal 0/ Global Optimization 4, 135-170.

Maranas C.D. and Floudas C.A., 1995, Finding all solutions of nonlinearly
eonstrained systems of equations. Journal 0/ Global Optimization 7, 143-
182.

Maranas C.D. and Floudas C.A., 1997, Global optimization in generalized

geometrie programming. Comput. Chem. Eng. 21, 351-370.

Maria G., 1989, An adaptive strategy for solving kinetie model concomitant
estimation-reduction problems. Can. J. Chem. Eng. 67, 825.

Marketos G., 1975, The Optimal Design 0/ Chemical Plant Considering Un-
certainty and Changing Circumstances. Ph.D. thesis, ETH Zurieh.

Mathiesen L., 1987, An algorithm based on a sequenee oflinear complementar-

ity problems applied to a walrasian equilibrium model: an example. Mathe-
matical Programming 37, 1-18.

Mathur K. and Puri M.C., 1995, A bilevel bottleneck programming problem.

European Journal 0/ Operation al Research 86, 337.

McCormiek G.P., 1976, Computability of global solutions to factorable non-

convex programs : Part 1- convex underestimating problems. Math. Prog.
10, 147-175.

McDonald C.M., 1995, Global Optimization /or the Phase and Chemical Equi-
librium Problem. Ph.D. thesis, Princeton University.

MeDonald C.M. and Floudas C.A., 1995a, Global optimization for the phase
stability problem. AIChE J. 41, 1798.

MeDonald C.M. and Floudas C.A., 1995b, Global optimization and analysis
for the gibbs free energy funetion using the UNIFAC, Wilson and ASOG
equations. If3EC Res. 34, 1674.
BIBLIOGRAPHY 431

McDonald C.M. and Floudas C.A., 1995c, Global optimization for the phase
and chemical equilibrium problem: Application to the NRTL equation. Com-
put. Chem. Eng. 19, l.
McDonald C.M. and Floudas C.A., 1997, GLOPEQ: A new computational tool
for the phase and chemical equilibrium problem. Comput. Chem. Eng. 21,
1-23.

Meintjes K and Morgan A.P., 1990, Chemical-equilibrium systems as numer-

ical test problems. ACM Trans. Math. Soft. 16, 143-15l.
Michelsen M.L., 1982, The isothermal flash problem - part i. stability. Fluid
Phase Equilibria 9, l.
Migdalas A., 1995, Bilevel programming in trafik planning: models, methods
and challenge. Journal 0/ Global Optimization 7, 38l.
Moore R., Hansen E., and Ledere A., 1992, Rigorous methods for global op-
timization. In Floudas C.A. and Pardalos P.M. (eds.), Recent Advances in
Global Optimization, Princeton Series in Computer Science, (p. 321), Prince-
ton University Press.
Moore R.E., 1979, Methods and Applications o/Interval Analysis. Prentice-
Hall, Englewood Cliffs, NJ.
More J., 1987, The levenberg-marquardt algorithm: implement at ion and the-
ory. In Numerical Analysis-Proceedings 0/ the Biennial Con/erence, no. 630
in Lecture Notes in Mathematics, (p. 105), Berlin, Germany, Springer-
Verlag.
More J., 1996, Global methods for nonlinear complementarity problems. Math-
ematics 0/ Operations Research 21, 589-614.
Morse P.M., 1929, Diatomic molecules according to the wave mechanics ii.
vibrationallevels. Phys. Rev. 34, 57-67.
Moshirvaziri K, Amouzegar M.A., and Jacobsen S.E., 1996, Test problem
construction for linear bilevel programming problems. Journal 0/ Global Op-
timization 8, 235-243.
Murphy F.H., Sherali H.D., and L. S.A., 1982, A mathematical programming
approach for determining oligapolistic market equilibrium. Mathematical
Programming 24, 1-18.

Murtagh B.A. and Saunders M.A., 1993, MINOS 5.4 User's Guide. Systems
Optimization Laboratory, Department of Operations Research, Stanford
University, Technical Report SOL 83-20R.

Murty KG., 1988, Linear complementarity, Linear and Nonlinear Program-

ming. Helderman Verlag, Berlin, Germany.
432 BIBLIOGRAPHY

Nagurney A., 1993, Network Economics. A variational inequality approach ..

Kluwer Academic Publishers, Dordrecht, The Netherlands.

Nesterov N.Z., 1977, Cut-off method with space extension in convex program-
ming problems. Cybernetics 13, 94-96.

Neumaier A., 1990, Interval Methods for Systems of Equations. Encyclopedia

of Mathematics and its Applications, Cambridge University Press.

Niesse J.A. and Mayne H.R., 1996, Minimization of small silicon clusters using
the space-fixed modified genetic algorithm method. Chem. Phys. Lett. 261,
576-582.

Northby J.A., 1987, Structure and binding of lennard-jones cluster: 13 ::; n ::;
147. J. Chem. Phys. 87, 6166-6177.

Onal H., Darmawan D.H., and Johnson III S.H., 1995, A multilevel analysis of
agricultural credit distribution in east java, indonesia. Comp. Op. Res. 22,
227.

nal H.A., 1993, A modified simplex approach for solving bilevel linear pro-
gramming problems. European Journal of Operation al Research 67, 126-135.

Ostrovsky G.M., Ostrovsky M.G., and Mikhailow G.W., 1990, Discrete opti-
mization of chemical processes. Comput. Chem. Eng. 14, 111.

Paloschi J. and Perkins J., 1988, An implement at ion of quasi-newton methods

for solving sets of nonlinear equations. Comput. Chem. Eng. 12, 767-776.

Pang J.S. and Gabriel S.A., 1993, Nejsqp: A robust algorithm for the nonlinear
complementarity problem. Mathematical Programming 60, 295-337.

Pardalos P. and Wolkowicz H. (eds.), 1998, Topics in Semidefinite and Interior-

Point Methods, vol. 18 of Fields Institute Communications Series. American
Mathematical Society.

Pardalos P.M., 1987, Generation of large-scale quadratic programs for use as

global optimization test problems. A CM Trans. Math. Soft. 13, 133-137.

Pardalos P.M., 1991, Construction of test problems in quadratic bivalent pro-

gramming. ACM Trans. Math. Soft. 17,74-87.

Pardalos P.M., 1996, Continuous approaches to discrete optimization prob-

lems. In Nonlinear Optimization and Applications, (pp. 313-328), Plenum
Publishing.

Pardalos P.M. and Phillips A., 1990, Global optimization approach to the
maximum clique problem. Int. J. Computer Math. 33, 209.
BIBLIOGRAPHY 433

Pardalos P.M. and Rosen J.B., 1987, Constrained global optimization: Algo-
rithms and applications, vol. 268 of Lecture Notes in Computer Science.
Springer Verlag, Berlin, Germany.
Pardalos P.M. and Wolkowicz H. (eds.), 1994, Quadratic Assignment and Re-
lated Problems. DIMACS Series Vol. 16, American Mathematical Society.

Paules IV G. and Floudas C., 1989, A new optimization approach for phase
and chemical equilibrium problems. Paper presented at the Annual AIChE
Meeting, San Franciso, CA, November.

Paules, IV G.E. and Floudas C.A., 1989, APROS: Algorithmic development

methodology for discrete-continuous optimization problems. Oper. Res. 37,
902-915.

Pillardy J. and Piela L., 1995, Molecular dynamics on deformed potential

energy hypersurfaces. J. Phys. Chem. 99, 11805-11812.

Polak E., 1971, Computational Methods in Optimization: A Unified Approach.

Academic Press.

Polak E., 1997, Optimization: Algorithms and Consistent Approximations. No.

124 in Applied Mathematical Sciences, Springer.

Prn R, Harjunkoski 1., and Westerlund T., 1997, Convexification of different

classes of non-convex minlp problems. Submitted to Comput. Chem. Eng.

Powell M., 1970, Numerical Methods for Nonlinear Algebraic Equations. Gor-
don and Breach, London, UK.

Quesada 1. and Grossmann 1.E., 1992, An LP /NLP based branch and bound
algorithm for convex MINLP optimization problems. Comput. Chem. Eng.
16, 937-947.

Quesada 1. and Grossmann 1.E., 1993, Global optimization algorithm for heat
exchanger networks. Ie3EC Res. 32, 487.

Ramana M.V. and Pardalos P.M., 1996, Semidefinite programming. In Terlaky

T. (ed.), Interior Point Methods in Mathematical Programming, (pp. 369-
398), Kluwer Academic Publishers.

Ratschek H. and Rokne J., 1993, A circuit design problem. J. Global Opt. 3,
501.

Reinhart H. and Rippin D., 1986, The design of flexible batch chemical plants.
Presented at the 1986 AIChE Annual Meeting.

Reinhart H. and Rippin D., 1987, Design of flexible multi-product plants: A

new procedure for optimal equipment sizing under uncertainty. Presented at
the 1987 AIChE Annual Meeting.
434 BIBLIOGRAPHY

Reklaitis G. and Ragsdell K., 1983, Engineering Optimization. John Wiley &
Sons, New York, NY.

Renon H. and Prausnitz J.M., 1968, Local compositions in thermodynamie

excess functions for liquid mixtures. AIChE J. 14, 135.

Rijckaert M. and Martens X., 1978b, Comparison of generalized geometrie

programming algorithms. J. Opt. Th. Appl. 26, 205.

Rod V. and Hancil V., 1980, Iterative estimation of model parameters when
measurements of all variables are subject to error. Comp. Chem. Engng. 4,
33.

Rosen O. and Luus R., 1992, Global optimization approach to nonlinear opti-
mal control. Journal of Optimization Theory and Applieations 73, 547.

Ryoo H.S. and Sahinidis N.V., 1995, Global optimization of nonconvex NLPs
and MINLPs with applications in process design. Comput. Chem. Eng. 19,
551-566.

Sargent R.W.H. and SuIlivan G.R., 1977, The development of an efficient op-
timal control package. Proe. IFIP Con. Optim. Teeh. 8th, part 2.

Schlosser P. and Feinberg M., 1994, A theory of multiple steady-states in

isothermal homogeneous CFSTRs with many reactions. Chem. Eng. Sei.
49, 1749-1767.

Schweiger C.A. and Floudas C.A., 1997, MINOPT: A Software Paekage for
Mixed-Integer Nonlinear Optimization. Princeton University, Princeton, NJ
08544-5263, Version 2.0.

Schweiger C.A. and Floudas C.A., 1998b, Synthesis of optimal reactor network
systems. Ind. Eng. Chem. Res. submitted for publication.

Schweiger C.A. and Floudas C.A., 1998c, MINOPT A Modeling Language

and Algorithmic F'ramework for Linear, Mixed-Integer, Nonlinear, Dynamie,
and Mixed-Integer Nonlinear Optimization. Kluwer Academic Publishers, in
preparation.

Seader J. and Kumar R., 1994, Multiple solutions in process simulation. In

ASPENWORLD 94, no. November 6-9, Boston, MA.
Seader J., Kuno M., Lin W.J., Johnson S., Unsworth K., and Wiskin J., 1990,
Mapped continuation methods for computing all solutions to general systems
of nonlinear equations. Comput. Chem. Eng. 14, 71-85.

Sherali H. and Shetty C., 1980, A finitely convergent algorithm for bilinear
programming problems using polar and disjunctive face cuts. Mathematieal
Programming 19, 14-31.
BIBLIOGRAPHY 435

Shimizu K and Aiyoshi E., 1981, A new computational method for stackelberg
and min-max problems by use of a penalty method. IEEE Trans. on Aut.
Control AC-26.

Shimizu K, Ishizuka Y., and Bard J.F., 1997, Nondifferentiable and Two-Level
Mathematical Programming. Kluver Academic Publishers.

Shor N.Z., 1985, Minimization Methods for Non-differentiable Functions.

Springer Series in Computational Mathematics, Springer-Verlag, New York,
Berlin.

Shor N.Z., 1998, Nondifferentiable Optimization and Polynomial Problems.

Nonconvex Optimization and Its Applications, Kluwer Academic Publishers,
Dordrecht, The Netherlands.

Sideris A. and Sanchez, 1988, Fast computation of the multivariable stability

margin for real interrelated uncertain parameters. in Proc. ACC, Atlanta,
GA.

Smith E. and Pantelides C., 1996, Global optimisation of general process mod-
els. In Grossmann 1. (ed.), Global Optimization in Engineering Design, (pp.
355-386), Kluwer Academic Publishers.

Smith E. and Pantelides C., 1997, Global optimisation of nonconvex MINLPs.

Comput. Chem. Eng. 21, S791-S796.

Smith J., 1985, Chemical Engineering Kinetics. Butterworth, Stoneham, MA.

Soave G., 1972, Equilibrium constants from a modified redlich-kwong equation

of state. Chem. Eng. Science 27, 1197-1203.

Soland R., 1971, An algorithm for separable nonconvex programming problems

ii : Nonconvex constraints. Mgmt. Sei. 17, 759.

Stephanopoulos G. and Westerberg A.W., 1975, The use ofhestenes' method of

multipliers to resolve dual gaps in engineering system optimization. Journal
of Optimization Theory and Applications 15, 285.

Subrahmanyam S., Pekny J.F., and Reklaitis G.V., 1994, Design of batch
chemical plants under market uncertainty. Ind. Eng. Chem. Res. 33, 2688-
2701.

Suh S. and Kim T.J., 1992, Solving nonlinear bilevel programming problems
of the equilibrium network design problem: A comparative review. Annals
of Operations Research 34, 203.

Sun A. and Seider W., 1995, Homotopy-continuation method for stability anal-
ysis in the global minimization of the gibbs free energy. Fluid Phase Equi-
libria 103, 213.
436 BIBLIOGRAPHY

Tersoff J., 1988, Empirical interatomic potential for silicon with improved elas-
tic properties. Phys. Rev. B 38, 9902-9905.

Thieu T., 1980, Relationship between bilinear programming and concave pro-
gramming. Acta Mathematica Vietnamica 2, 106-113.

Tjoa I.B. and Biegler L.T., 1991a, Simultaneous strategies for data reconcili-
at ion and gross error detection of nonlinear systems. Comp. Chem. Engng.
15,679.

Tjoa T.B. and Biegler L.T., 1992, Reduced successive quadratic programming
strategy for errors-in-variables estimation. Comp. Chem. Engng. 16, 523.

Toh KC., Todd M.J., and Ttnc RH., 1998, SDPT3

a matlab software package for semidefinite programming.
http://www.math.cmu.edu/-reha/sdpt3.html.

Tsuboka T. and Katayama T., 1975, Modified wilson equation for vapor-liquid
and liquid-liquid equilibria. J. Chem. Eng. Japan 8, 187.

Tuy H., Migdalas A., and Varbrand P., 1993, AglobaI optimization approach
for the linear two-Ievel program. Journal 01 global optimization 3, 1.

Vaidyanathan Rand EI-Halwagi M., 1994, Global optimization of nonconvex

nonlinear programs via interval analysis. Comput. Chem. Eng. 18, 889-897.

Vaidyanathan Rand EI-Halwagi M., 1996, Global optimization of nonconvex

MINLP's by interval analysis. In Grossmann I.E. (ed.), Global Optimization
in Engineering Design, (pp. 175-193), Kluwer Academic Publishers.
Vaish H. and Shetty C., 1976, The bi linear programming problem. Naval Re-
search Logistics Quarterly 23, 303-309.
Vandenberghe L. and Boyd S., 1996, Semidefinite programming. SIAM Review
38,49-95.

Varah J., 1982, A spline least squares method for numerical parameter estima-
tion in differential equations. SIAM J. Sei. Stat. Comput. 3, 28.

Vasantharajan S. and Biegler L.T., 1990, Simultaneous strategies for opti-

mization of differential-algebraic systems with enforcement of error criteria.
Comput. Chem. Eng. 14, 1083-1100.
Vassiliadis V.S., Sargent RW.H., and Pantelides C.C., 1994a, Solution of a
dass of multistage dynamic optimization problems. 1. problems without path
constraints. Ind. Eng. Chem. Res. 33, 2111-2122.

Vassiliadis V.S., Sargent RW.H., and Pantelides C.C., 1994b, Solution of a

dass of multistage dynamic optimization problems. 2. problems with path
constraints. Ind. Eng. Chem. Res. 33, 2123-2133.
BIBLIOGRAPHY 437

Vicente L. and Calamai P., 1994, Bilevel and multilevel programming: A bib-
liography review. Journal oi Global Optimization 5, 291.

Vicente L., Savard G., and Judice J., 1994, Descent approaches for quadratic
bilevel programming. Journal oi Optimization Theory and Applications 81,
379.

Vicente L.N., Calamai P.H., and Judice J.J., 1992, Generation of disjointly
constrained bilinear programming test problems. Computational Optimiza-
tion and Applications 1, 299-306.

Vicino A., Tesi A., and Milanese M., 1990, Computation of nonconservative
stability perturbation bounds for systems with nonlinearly correlated uncer-
tainties. IEEE Trans: Autom. Contr. 35, 835.

Viswanathan J. and Grossmann LE., 1990, A combined penalty function and

outer approximation method for MINLP optimization. Comput. Chem. Eng.
14, 769-782.

Visweswaran V. and Floudas C.A., 1990, Aglobai optimization algorithm GOP

for certain classes of nonconvex NLPs : Ii. application of theory and test
problems. Comput. Chem. Eng. 14, 1419.

Visweswaran V. and Floudas C.A., 1992, Unconstrained and constrained global

optimization of polynomial functions in one variable. J. Global Optimization
2,73.

Visweswaran V. and Floudas C.A., 1996b, Computational results for an effi-

cient implementation of the GOP algorithm and its variants. In Grossmann
LE. (ed.), Global Optimization in Engineering Design, Kluwer Book Series
in Nonconvex Optimization and its Applications, Chapter 4.

Visweswaran V., Floudas C.A., Ierapetritou M.G., and Pistikopoulos E.N.,

1996, A decomposition-based global optimization approach for solving bi-
level linear and quadratic programs. In Floudas C.A. and Pardalos P.M.
(eds.), State oi the Art in Global Optimization, Nonconvex Optimizatin
and its Applications, (p. 139), Kluwer Academic Publishers.

Wales D.J. and Doye J.P.K., 1997, Global optimization by basin-hopping and
the lowest energy structures of lennard-jnes clusters containing up t 100
atoms. J. Phys. Chem. A 101, 5111-5116.

Walraven F. and van Rompay P., 1988, An improved phase-splitting algorithm.

Comput. Chem. Eng. 12, 777.

Wehe R.R. and Westerberg A.W., 1987, An algorithmic procedure fr the

synthesis of distillation sequences with bypass. Comput. Chem. Eng. 11,
619.
438 BIBLIOGRAPHY

Wen U.P. and Hsu S.T., 1991, Linear bi-level programming problem - a review.
Journal 0/ the Operational Research Society 42, 125.

Westerberg A.W., 1985, The synthesis of distillation based separation systems.

Comput. Chem. Eng. 9, 421.

Westerlund T., Pettersson F., and Grossmann 1.E., 1994, Optimization of

pump configuration problems as a MINLP probem. Comput. Chem. Eng.
18, 845-858.

Westerlund T., Skrifvars H., Harjunkoski 1., and Prn R., 1998, An extended
cutting plane method for a class of non-convex minlp problems. Comput.
Chem. Eng. 22, 357-365.

White D.J. and Anandalingam G., 1993, A penalty function approach for solv-
ing bilevellinear programs. Journal 0/ Global Optimization 3, 397-419.

White W., Johnson S., and Dantzig G., 1958, Chemical equilibrium in complex
mixtures. J. Chem. Phys. 28, 751.

Wilhelm C. and Swaney R., 1994, Robust solution of algebraic process model-
ing equations. Comput. Chem. Eng. 18, 511-531.

Wilkinson J.H., 1963, Rounding Errors in Algebraic Processes. Prentice-Hall,

Englewood Cliffs, New Jersey.

Wille L.T., 1987, Minimum energy configurations of atomic clusters: New

results obtained by simulated annealing. Chem. Phys. Lett. 133, 405-410.
Williams H., 1987, Linear and integer programming applied to the proposi-
tional calculus. Syst. Res. Inf. Sei. 2, 81-100.

Wilson G.M., 1964, Vapor liquid equilibrium xi: A new expression for the
excess gibbs energy of mixing. J. Am. Chem. Soc. 86, 127.

Wingo D., 1985, Globally minimizing polynomials without evaluating deriva-

tives. Intl. J. Comput. Math 17, 287.

Woodhouse B., 1976, Lower Bounds tor the Reliablity 0/ a Test. Master's thesis,
Dept. of Statistics, University of Wales, t.

Wu S.P., Boyd S., and Vandenberghe L., 1996, FIR filter design via semidefinite
programming and spectral factorization. In Proc. IEEE Conf. on Decision
and Control, (pp. 271-276).

Xue G.L., 1994, Molecular conformation on the cm-5 by parallel 2-level simu-
lated annealing. J. Glob. Opt. 4, 187-205.

Yang H. and Yagar S., 1994, Trafik assignment and trafIic control in general
freeway arterial corridor systems. Transport Res. B-Meth. 28, 463.
BIBLIOGRAPHY 439

Yang H. and Yagar S., 1995, Trafik assignment and signal control in saturated
road networks. Trans. Res. A. 29A, 125.

Yang H., Yagar S., Iida Y., and Asakura Y., 1994, An algorithm for the inflow
control problem on urban freeway networks with user-optimal flows. Trans.
Res. B. 28B, 123.

Yee T.F. and Grossmann I.E., 1991, Simultaneous optimization model for heat
exchanger network synthesis. In Grossmann I.E. (ed.), Ghemical Engineering
Optimization Models with GAMS, vol. 6 of Process Design Gase Studies
Series, CACHE.

Yezza A., 1996, First-order necessary optimality conditions for general bilevel
programming problems. Journal 0/ Optimization Theory and Applications
89, 189-219.

Yuan X., Zhang S., Pibouleau L., and S. D., 1988, A mixed-integer
nonlinear-programming method for process design. RAIRO - Recherche
Operationnelle-Operations Research 22, 331-346.

Yudin D.B. and Nemirovsky A.S., 1977, Informational complexity and efficient
methods for solving complex extremal problems. Matekon 13, 25-45.

Zamora J. and Grossmann 1., 1998, Aglobai minlp optimization algorithm

for the synthesis of heat exchanger networks with no stream splits. Gomput.
Ghem. Eng. 22, 367-384.
Index

a-pinene reaction, x, xiv, 155, 383 CSTR sequence design, ix, 89

(D.C.) function, 59
Denbigh reaction, x, xiv, 149, 151,
alkylation process design, ix, 86 377,379
all homogeneous azeotropes, xiii, 334 differential-algebraic equations, 394
ASOG equation, 79 dynamic optimization, 351
assignment problems, xiii, 309
educational testing problem, 254
equations of state, 178
batch plant design under uncertainty,
equilibrium transport at ion model, xii,
ix, 114
238
Bellman's problem, 399
biconvex, viii, 1, 59, 64 generalized geometrie programming,
bifunctional catalyst blend problem, 1, 85
410 Gibbs energy minimization, 61
bilevel linear programming, xi, 207 Goldstein and Price function, ix, 110
bilevel optimization, 1 graph coloring, xiii, 304, 313
bilevel programming, 205
bilevel quadratic programming, xi, Haverly pooling, viii, 34
220 heat exchanger design, ix, 90
bilinear, 1, 33 heat exchanger network, 51, 52, 54,
blending problems, 34 269
Bolding-Andersen potential, xi, 202
Keller graphs, xiii, 317
Brenner potential, xi, 199
Kowalik problem, x, 172
catalytic cracking of gas oil, 398 Lennard-Jones potential, xi, 188
clusters of atoms and molecules, xi, Levenspiel reaction, x, xiv, 153, 381
186 linear complementarity, xi, 234
co ding theory, 317 Lotka-Volterra problem, 403
Colville's test problem, ix, 92
combinatorialoptimization, 1, 303 maximum clique, 5, 17, 304, 315
complementarity, xi, 1, 233 maximum cut problem, 258, 259
constrained nonlinear optimization, maximum likelihood, 166
1 maximum mixed reactor (MMR), 355
continuous stirred tank reactor (CSTR), methanol-to-hydrocarbons process,
129, 355 401
control parameterization, 352 mixed-integer nonlinear optimization,
cross fiow reactor (CFR), 355 1
440
INDEX 441

mixed-integer nonlinear programming, six-hump camelback function, ix, 111

263 SRK equation, x, 182
Morse potential, xi, 193 stability margin, 97
multiproduct batch plants, 114, 115 steady-state CSTR, x, 174
multipurpose batch plants, 114 Steiner problems, xiii, 318
sulfur dioxide oxidation, x, xiv, 164,
naphthalene reaction, x, xiv, 159, 391
387
nonlinear complementarity problem, T-K-Wilson equation, 83
242 tangent plane distance minimization,
nonlinear systems of equations, 325 x, 63,179
nonsharp separation, viii, 44, 46 Tersoff potential, xi, 197
NRTL equation, xiv, 64, 66, 340 trafiic equilibrium problem, xii, 240
Trambouze reaction, x, xiv, 147, 376
oil shale pyrolysis, 408 traveling salesman problem, xiii, 308
optimal control, 1, 404 trim loss minimization, 289
optimal reactor design, ix, 91 twice continuously differentiable NLPs,
107
parallel reactions, x, xiv, 162, 389
parameter estimation, x, xiv, 166, UNIFAC equation, xiv, 75-77, 345
393 UNIQUAC equation, xiv, 70, 342
Peng-Robinson equation, x, 184 univariate polynomial, 27
pharmacokinetic model, x, 173
phase and chemical equilibrium prob- Van de Vusse reaction, x, xiv, 134,
lem, 59, 178 139,141,143,145,157,364,
plug flow reactor (PFR), 129, 355 369, 371, 372, 374, 385
pooling problems, 34, 36 van der Waals equation, x, 180
posynomials, 85 vapor-liquid equilibrium, x, 176
pseudoethane, ix, 109
pump network synthesis, xii, 285 Wilson equation, xiii, 337

quadratic 0-1 test problems, xiii, 304

quadratic integer programming, 304
quadratic programming, 1, 5
quadratically constrained, 21

reactor network synthesis, 128, 354

respiratory mechanical, x, 171
robust stability analysis, 97

satisfiability problems, xiii, 306

segregated flow reactor (SFR), 355
semidefinite programming, 1, 251,
252
Shekel function, ix, 111
singular control problem, 406