Waterflow

CORE Metadata, citation and similar papers at core.ac.
uk
Provided by ScholarBank@NUS
A WATER FLOW ALGORITHM FOR
OPTIMIZATION PROBLEMS
TRAN TRUNG HIEU
NATIONAL UNIVERSITY OF SINGAPORE
2011
A WATER FLOW ALGORITHM FOR
TRAN TRUNG HIEU

(B.Eng. (Hons.), HCMUT)
A THESIS SUBMITTED
FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
DEPARTMENT OF INDUSTRIAL AND SYSTEMS ENGINEERING
NATIONAL UNIVERSITY OF SINGAPORE
2011
ACKNOWLEDGEMENTS
First of all, I would like to express my sincerest gratitude to my supervisor, Associate
Professor Ng Kien Ming, at the Department of Industrial and Systems Engineering,
National University of Singapore, for his encouragement and guidance throughout my
PhD studying process. His invaluable advices have helped me to successfully complete
my research work as well as thesis.
Next, I would like to thank all the lecturers of the Department of Industrial and
Systems Engineering, National University of Singapore, for their lessons which helped me
to achieve necessary and useful knowledge for my research work. I would also like to
extend my acknowledgement to the officers of the department for their assistance in
handling my administrative matters.
Finally, I would like to take this chance to express my special gratitude to my beloved
girlfriend, Ms. Ry, for her constant love and continuous support throughout my PhD
studying process.
Tran Trung Hieu
November 2011
i
TABLE OF CONTENTS
ACKNOWLEDGEMENTS ................................................................................................ i
TABLE OF CONTENTS .................................................................................................... ii
ABSTRACT ..................................................................................................................... viii
LIST OF TABLES ............................................................................................................. x
LIST OF FIGURES ........................................................................................................ xiii
GLOSSARY ................................................................................................................... xvi
CHAPTER 1 INTRODUCTION ................................................................................ 1
1.1 Combinatorial Optimization Problems ................................................................... 2
1.2 Nature Inspired Algotithms..................................................................................... 4
1.3 Motivation and Research Objectives ...................................................................... 7
1.4 Main Contributions of the Thesis ........................................................................... 8
1.5 Outline of the Thesis ............................................................................................. 10
CHAPTER 2 LITERATURE REVIEW .................................................................. 13
2.1 Biologically Inspired Algorithms ......................................................................... 14
2.1.1 Evolutionary Algorithms ............................................................................ 14
2.1.2 Stigmergic Optimization Algorithms .......................................................... 19
2.1.3 Swarm-Based Optimization Algorithms ..................................................... 22
2.2 Botanically Inspired Algorithms ........................................................................... 27
ii
2.2.1 An Invasive Weed Optimization Algorithm ............................................... 27
2.2.2 A Botany-Grafting Inspired Algorithm ...................................................... 30
2.3 Water Flow Inspired Techniques .......................................................................... 30
2.3.1 Image Processing Methods Based on Water Flow Model .......................... 30
2.3.2 Intelligent Water Drops Algorithm ............................................................. 35
2.3.3 Water Flow-Like Algorithm ....................................................................... 37
2.4 Conclusions and Possible Nature-Inspired Algorithm .......................................... 39
CHAPTER 3 A GENERAL WATER FLOW ALGORITHM .............................. 43
3.1 Hydrological Cycle in Meteorology ..................................................................... 44
3.2 Erosion Process of Water Flow in Nature............................................................. 47
3.3 General Water Flow Algorithm ............................................................................ 51
3.3.1 Encoding Scheme........................................................................................ 54
3.3.2 Memory Lists .............................................................................................. 56
3.3.3 Exploration Phase ....................................................................................... 57
3.3.4 Exploitation Phase ...................................................................................... 58
3.3.4.1 Erosion Condition and Capability .................................................. 58
3.3.4.2 Erosion Process .............................................................................. 61
CHAPTER 4 WFA FOR PERMUTATION FLOW SHOP SCHEDULING ....... 65
4.1 Introduction .......................................................................................................... 66
4.2 Formulation of the PFSP ....................................................................................... 68
4.3 WFA for the PFSP ................................................................................................ 69
iii
4.3.1 Encoding Scheme........................................................................................ 69
4.3.2 Memory Lists .............................................................................................. 70
4.3.3 Exploration Phase ....................................................................................... 70
4.3.4 Exploitation Phase ...................................................................................... 71
4.3.5 A Numerical Example for Erosion Machenism .......................................... 72
4.4 Computational Experiments and Comparisons ..................................................... 78
4.4.1 Benchmark Problem Sets ............................................................................ 78
4.4.2 Platform and Parameters ............................................................................. 79
4.4.3 Performance Measure ................................................................................. 79
4.4.4 Computational Results ................................................................................ 81
4.5 Conclusions ........................................................................................................... 84
CHAPTER 5 WFA FOR FLEXIBLE FLOW SHOP SCHEDULING ................. 85
5.1 Introduction .......................................................................................................... 86
5.2 FFSP with Intermediate Buffers ........................................................................... 89
5.3 WFA for the FFSP with Intermediate Buffers ...................................................... 93
5.3.1 Encoding Scheme........................................................................................ 94
5.3.2 Memory Lists .............................................................................................. 99
5.3.3 Exploration Phase ..................................................................................... 100
5.3.4 Exploitation Phase .................................................................................... 101
5.3.4.1 Erosion Condition and Capability ................................................ 101
5.3.4.2 Erosion Process ............................................................................ 104
5.4 An Example of the FFSP in Maltose Syrup Production .................................... 104
iv
5.5 Computational Experiments and Comparisons ................................................... 107
5.5.1 Benchmark Instances and Randomly Generated Instances ...................... 107
5.5.2 Platform and Parameters ........................................................................... 109
5.5.3 Performance Measures .............................................................................. 111
5.5.4 Computational Results .............................................................................. 112
5.6 Conclusions ......................................................................................................... 121
CHAPTER 6 MOWFA FOR MULTI-OBJECTIVE SCHEDULING ............... 122
6.1 Introduction ........................................................................................................ 123
6.2 MOFFSP with Intermediate Buffers ................................................................... 125
6.3 MOWFA for the MOFFSP with Intermediate Buffers ....................................... 128
6.3.1 Encoding Scheme...................................................................................... 129
6.3.2 Memory Lists ............................................................................................ 130
6.3.3 Exploration Phase ..................................................................................... 131
6.3.3.1 Distinct Regions ........................................................................... 131
6.3.3.2 Landscape Analysis ..................................................................... 132
6.3.3.3 Seed Job Permutations ................................................................. 134
6.3.3.4 Hill-Sliding Algorithm ................................................................. 134
6.3.4 Neighborhood Structures .......................................................................... 135
6.3.5 Exploitation Phase .................................................................................... 136
6.3.5.1 Erosion Condition and Capability ................................................ 136
6.3.5.2 Erosion Process ............................................................................ 138
6.3.6 Evaporation and Precipitation ................................................................... 140
v
6.3.7 Improvement Phase ................................................................................... 140
6.4 Computational Experiments and Comparisons ................................................... 141
6.4.1 Generation of Test Problems and Benchmark Problem Set...................... 141
6.4.2 Platform and Parameters ........................................................................... 144
6.4.3 Performance Metrics ................................................................................. 145
6.4.4 Computational Results .............................................................................. 148
6.5 Conclusions ......................................................................................................... 154
CHAPTER 7 WFA FOR OTHER COMBINATORIAL OPTIMIZATION
PROBLEMS ................................................................................................................. 155
7.1 Quadratic Assignment Problem .......................................................................... 156
7.1.1 Introduction .............................................................................................. 156
7.1.2 WFA for the QAP ..................................................................................... 158
7.1.2.1 Encoding Scheme and Memory Lists .......................................... 159
7.1.2.2 Exploration Phase ........................................................................ 161
7.1.2.3 Exploitation Phase ....................................................................... 163
7.1.2.3.1 Erosion Condition and Capability ............................... 164
7.1.2.3.2 Erosion Process ........................................................... 164
7.1.2.4 Improvement Phase ...................................................................... 166
7.1.3 Computational Experiments and Comparisons ......................................... 166
7.1.3.1 Benchmark Problem Sets ............................................................. 168
7.1.3.2 Platform and Parameters .............................................................. 168
7.1.3.3 Performance Measures ................................................................. 171
vi
7.1.3.4 Computational Results ................................................................. 172
7.2 Vehicle Routing Problem .................................................................................... 180
7.2.1 Capacitated Vehicle Routing Problem ...................................................... 180
7.2.2 Two-Level WFA for the CVRP ................................................................ 182
7.2.2.1 First Level .................................................................................... 182
7.2.2.2 Second Level ................................................................................ 187
7.2.3 Preliminary Experiments .......................................................................... 189
7.3 Conclusions ......................................................................................................... 191
CHAPTER 8 CONCLUSIONS AND FUTURE RESEARCH WORK .............. 193
8.1 Conclusions ......................................................................................................... 194
8.2 Future Research Work ........................................................................................ 198
REFERENCES .............................................................................................................. 200
vii
ABSTRACT
A novel natured-inspired algorithm, called the water flow algorithm (WFA), for solving
optimization problems has been proposed in this research work. The proposed algorithm is
designed by simulating the hydrological cycle in meteorology and the erosion
phenomenon in nature. Basic operators of this algorithm are based on the raindrops
distribution simulation, the property of water flow always moving to lower positions, and
the erosion process to overcome obstacles. Depending on the structure of a specific
problem, the WFA can be appropriately customized to solve the problem efficiently.
In this thesis, we focus on solving well-known combinatorial optimization problems,
such as the permutation flow shop scheduling problem (PFSP) in production planning,
quadratic assignment problem (QAP) in facility layout design, and vehicle routing
problem (VRP) in logistics and supply chain management. The general WFA has been
customized and implemented successfully for solving these problems. For the PFSP, the
proposed algorithm obtained not only optimal solutions for several PFSP benchmark
instances taken from OR Library, but also a new best known solution for the benchmark
instance of Heller. For the QAP, the algorithm solved most QAP benchmark instances
drawn from the QAPLIB. The comparison results also show that the WFA outperforms
other algorithms in terms of solution quality for both the PFSP and the QAP. For the VRP,
the WFA is combined with solving the relaxed mathematical programming model of the
VRP to search for optimal solutions to this problem. Preliminary results show that this
viii
algorithm is able to obtain optimal solutions for some of the VRP benchmark instances
taken from the literature.
Also, the WFA has been developed to solve flexible flow shop scheduling problem
(FFSP) with intermediate buffers, which is a general case of the PFSP. The FFSP is more
complex than the PFSP since there are a number of parallel identical machines at each
stage and intermediate buffers between consecutive stages. To solve the problem, an
efficient procedure for constructing a complete schedule is required. Hence, we proposed
a procedure for constructing a complete schedule as well as determining corresponding
objective values for the FFSP. The procedure is included in the WFA to increase the
efficiency of this algorithm. The experimental results and comparisons show that the
proposed algorithm outperforms other algorithms in terms of solution quality as well as
computation time. Moreover, the WFA has obtained new upper bound solutions for
several Wittrock benchmark instances.
The WFA can also be modified to solve multi-objective optimization problems. In this
thesis, we have designed the WFA for solving multi-objective scheduling problems, called
the MOWFA. Landscape analysis as well as evaporation and precipitation processes are
integrated into the MOWFA to solve the multi-objective FFSP with limited buffers
efficiently. Experimental results show that the MOWFA outperforms an improved hybrid
multi-objective parallel genetic algorithm for the multi-objective scheduling problem.
In conclusion, the WFA is able to obtain optimal or good quality solutions to several
well-known combinatorial optimization problems within reasonable computation time. It
is thus a promising algorithm to solve other types of optimization problems as well.
ix
LIST OF TABLES
Table 2.1 Summary of Applications of the Nature-Inspired Algorithms ..................... 41
Table 4.1 UE-list of Local Optimal Job Permutations .................................................. 74
Table 4.2 Possible Erosion Directions at the Local Optimal Job Permutation ............. 75
Table 4.3 Steps of Finding Improvement Job Permutation .......................................... 76
Table 4.4 Updating the UE-list ..................................................................................... 77
Table 4.5 Parameter Sets for Benchmark Problem Sets ............................................... 80
Table 4.6 Comparison Results between the WFA and the NEH-ALA ........................ 82
Table 4.7 Average Relative Percentage Increase over the Best Known Solution for
Taillard’s Benchmarks Obtained by Meta-heuristic Algorithms .................. 83
Table 5.1 An Example of Converting FFSP with Finite Buffers to FFSP with No
Available Buffer............................................................................................ 93
Table 5.2 Problem Data for Maltose Syrup Production.............................................. 105
Table 5.3 Problem Data for the Instances in Wittrock (1988) .................................... 108
Table 5.4 Parameter Sets for Benchmark Instances ................................................... 110
Table 5.5 Comparison Results of WFA, TS-H1/Z3, and MA with CPU Time Limit for
the TS Instances .......................................................................................... 113
Table 5.6 Comparison Results of WFA, TS-H1, and MA with CPU Time Limit for
the MA Instances ........................................................................................ 114
Table 5.7 Comparison Results between WFA and TS-H1/Z3 on the Randomly
Generated TS Instances............................................................................... 117
x
Table 5.8 Comparison Results between WFA and MA on the Randomly Generated
MA Instances .............................................................................................. 118
Table 5.9 Comparison Results between WFA and TS-H1 for the FFSP with No
Available Buffer Space ............................................................................... 119
Table 5.10 Comparison Results between WFA and TS-H1 for the FFSP with Finite
Buffer Capacities ........................................................................................ 119
Table 5.11 Comparison Results between WFA and TS-Z3 for the FFSP with Unlimited
Buffers......................................................................................................... 120
Table 5.12 Computational Results of WFA for Maltose Syrup Production Problem .. 121
Table 6.1 Parameter Sets of the MOWFA for Instances of the MOFFSP .................. 145
Table 6.2 Comparison of MOWFA and IHMOPGA for the Wittrock Benchmarks
with No Buffer ............................................................................................ 149
Table 6.3 Comparison of MOWFA and IHMOPGA for the Wittrock Benchmarks
with Finite Buffers ...................................................................................... 149
Table 6.4 Comparison of MOWFA and IHMOPGA for the Randomly Generated
Instances with No Buffer ............................................................................ 153
Table 6.5 Comparison of MOWFA and IHMOPGA for the Randomly Generated
Instances with Finite Buffers ...................................................................... 153
Table 7.1 Parameter Sets of WFA for the QAP Benchmark Instances ...................... 170
Table 7.2a Comparison Results of the WFA with Other Algorithms for Burkard’s and
Christofides’ Instances ................................................................................ 173
Table 7.2b Comparison Results of the WFA with Other Algorithms for Elshafei’s,
Eschermann’s, Hadley’s, and Krarup’s Instances....................................... 174
xi
Table 7.2c Comparison Results of the WFA with Other Algorithms for Li & Pardalos’
and Skorin-Kapov’s Instances .................................................................... 175
Table 7.2d Comparison Results of the WFA with Other Algorithms for Nugent’s,
Roucairol’s, Scriabin’s, Steinberg’s, Thonemann’s, Wilhelm’s Instances . 176
Table 7.2e Comparison Results of the WFA with Other Algorithms for Taillard’s
Instances ...................................................................................................... 177
Table 7.3 Improved Results of the WFA Variants for the QAP Instances Not
Optimally Solved by 2-Opt WFA ............................................................... 178
Table 7.4 Experimental Results for the CVRP ........................................................... 191
xii
LIST OF FIGURES
Figure 1.1 A Classification of Algorithms for Combinatorial Optimization Problems ... 6
Figure 1.2 Organization of the Thesis ............................................................................ 12
Figure 2.1 Flow Chart of Genetic Algorithm ................................................................. 16
Figure 2.2 Flow Chart of Memetic Algorithm ............................................................... 17
Figure 2.3 Flow Chart of Shuffled Frog-Leaping Algorithm ......................................... 18
Figure 2.4 Flow Chart of Ant Colony Optimization Algorithm..................................... 21
Figure 2.5 Flow Chart of Bee Colony Optimization Algorithm .................................... 23
Figure 2.6 Flow Chart of Particle Swarm Optimization Algorithm ............................... 25
Figure 2.7 Flow Chart of Firefly Algorithm................................................................... 26
Figure 2.8 Flow Chart of Bat Algorithm ........................................................................ 28
Figure 2.9 Flow Chart of Invasive Weed Optimization ................................................. 29
Figure 2.10 Flow Chart of Botany-Grafting Inspired Algorithm ..................................... 31
Figure 2.11 Flow Chart of the Method Proposed by Kim et al. (2002) ........................... 32
Figure 2.12 Flow Chart of the Improved Method Proposed by Oh et al. (2005) ............. 33
Figure 2.13 Procedure of Text Image Identification (Brodic and Milivojevic, 2010)..... 34
Figure 2.14 Flow Chart of the IWD Algorithm (Shah-Hosseini, 2007)........................... 36
Figure 2.15 Flow Chart of the Water Flow-Like Algorithm (Yang and Wang, 2007) .... 38
Figure 2.16 Time Line of Nature-Inspired Algorithms ................................................... 42
Figure 3.1 Basic Components of the Hydrological Cycle .............................................. 45
Figure 3.2 Erosion from Water Flow ............................................................................ 47
Figure 3.3 Factors Affecting Erosion Capability .......................................................... 49
xiii
Figure 3.4 Illustration for Smoothed Terrain by Erosion Process ................................. 51
Figure 3.5 Flow Chart of the General WFA ................................................................... 54
Figure 3.6 Pseudocode of the General WFA .................................................................. 55
Figure 3.7 Illustration for Effect of Altitude on Erosion Capability .............................. 60
Figure 3.8 Pseudocode for General Erosion Process of the WFA ................................. 63
Figure 3.9 Illustration for “Big Valley” and “Small Valley” ......................................... 64
Figure 4.1 An Example of Solution Representation in the WFA for the PFSP ............. 69
Figure 4.2 Flow Chart of the WFA for the PFSP ........................................................... 73
Figure 4.3 The Case of Fully Blocked Position ............................................................. 78
Figure 4.4 Means Plot for Comparing the WFA and Meta-heuristic Algorithms .......... 83
Figure 5.1 The Schematic of FFSP with Limited Intermediate Buffers ........................ 89
Figure 5.2 A Gantt Chart Illustration of the FFSP with Intermediate Buffers ............... 92
Figure 5.3 Flow Chart of the WFA for the FFSP with Intermediate Buffers................. 94
Figure 5.4 An Example of Solution Representation in the WFA for the FFSP ............. 95
Figure 5.5 A Comparison Between the FAM Rule and the Modified FAM Rule ......... 97
Figure 5.6 A Comparison Between the Procedure H1 and the H1-Variant Procedure .. 98
Figure 5.7 Maltose Syrup Production Process ............................................................. 106
Figure 5.8 Illustration of the FFSP with Controlled and Limited Buffers ................... 106
Figure 5.9 Standard Deviation of the Objective Values Obtained by the WFA, TS-H1
and MA ....................................................................................................... 114
Figure 5.10 Trajectory of Solution Improvement of the WFA, TS-H1, and MA .......... 115
Figure 6.1 FFSP with Operation Stages Including Intermediate Buffers..................... 126
Figure 6.2 An Example of Solution Representation in MOWFA for the MOFFSP .... 129
xiv
Figure 6.3 An Illustration for Finding the Pareto Set Based on the Orientation Angle in
MOWFA ..................................................................................................... 133
Figure 6.4 An Illustration of the Search Direction of DOWs on Two Types of 2-opt
Neighborhood ............................................................................................. 136
Figure 6.5 An Illustration for Scheme I of the Erosion Process with the Local 2-Opt
Neighborhood ............................................................................................. 139
Figure 6.6 An Illustration for Scheme II of the Erosion Process with the Global 2-Opt
Neighborhood ............................................................................................. 139
Figure 6.7 Flow Chart of the MOWFA for the MOFFSP ............................................ 142
Figure 6.8 Plot of Pareto Fronts Obtained by MOWFA and IHMOPGA on Wittrock
Instances with No Buffer ............................................................................ 150
Figure 6.9 Plot of Pareto Fronts Obtained by MOWFA and IHMOPGA on Wittrock
Instances with Finite Buffers ...................................................................... 151
Figure 7.1 An Example of Solution Representation in the WFA for the QAP ............ 159
Figure 7.2 Flow Chart of the WFA for the QAP .......................................................... 167
Figure 7.3 Comparison of the WFA with Other Algorithms on (a) Average Percentage
Difference; and (b) Number of Best Known Solutions Obtained ............... 179
Figure 7.4 The Cross-Exchange Procedure (Taillard et al., 1997) ............................... 184
Figure 7.5 Flow Chart of the First Level of the 2LWFA for the CVRP ...................... 186
Figure 7.6 An Example of Solution Representation in the 2LWFA for the CVRP ..... 188
Figure 7.7 Flow Chart of the Modified WFA for the CVRP ....................................... 189
xv
GLOSSARY
WFA Water flow algorithm
MOWFA Multi-objective water flow algorithm
2LWFA Two-level water flow algorithm
PFSP Permutation flow shop scheduling problem
FFSP Flexible flow shop scheduling problem
MOFFSP Multi-objective flexible flow shop scheduling problem
QAP Quadratic assignment problem
VRP Vehicle routing problem
CVRP Capacitated vehicle routing problem
P0-list The best position list
UE-list Un-eroded position list
E-list Eroded position list
DOW Drop of water
MaxCloud The maximum number of clouds generated
MaxPop The maximum number of DOWs generated in each cloud
MinEro The minimum amount of precipitation allowed to star

erosion process
MaxUIE The maximum number of iterations for erosion process
xvi
Chapter 1 Introduction
CHAPTER 1
INTRODUCTION
Combinatorial optimization problems are considered as a category of general optimization
problems in which the values of decision variables are subjected to integrality constraints.
Such problems are commonly encountered in the real world, especially in the field of
industrial systems (Yu, 1998). Because of the integrality constraints, their set of feasible
solutions is finite and an enumeration method may be applied to find an optimal solution
for these problems. However, the solution search space of combinatorial optimization
problems can grow exponentially according to the size of the problems, and many of them
are classified as NP-hard problems (Papadimitriou and Steiglitz, 1982). Thus when
solving combinatorial optimization problems with large size, exact optimization methods,
such as the branch and bound method, may not obtain optimal solutions to these problems
within acceptable computation time; even heuristic algorithms may face difficulties
finding good quality solutions. Recently, meta-heuristics, especially nature-inspired
algorithms, have achieved some success with solving certain combinatorial optimization
problems in reasonable computation time (Shah-Hosseini, 2007; and Yang, 2008). The
results show the potential of such approaches in solving other types of combinatorial
optimization problems.
1
In this thesis, we propose a novel nature-inspired meta-heuristic algorithm that is able
to solve different types of combinatorial optimization problems. A brief description of
such problems that can be solved by the proposed algorithm is provided in the next section.
We also present an overview of the practical applications of combinatorial optimization
problems in this section. As meta-heuristic algorithms are able to obtain solutions for
certain combinatorial optimization problems effectively, we describe the features of meta-
heuristic algorithms, especially nature-inspired ones, as well as give a classification of the
algorithms in Section 1.2. Next, we present the motivation and research objectives of the
thesis in Section 1.3. Then, the main contributions of the thesis are summarized in Section
1.4. Finally, the organization of this thesis is provided in Section 1.5.
1.1 Combinatorial Optimization Problems
Many real-world optimization problems can be formulated as mathematical programming
models with integrality constraints. The modeling and solving of such real-world
problems are related to “Combinatorial Optimization” (Christofides et al., 1979). In a
combinatorial optimization problem, the set of feasible solutions is discrete, or can be
reduced to a discrete set. Some examples of classical combinatorial optimization problems
consist of the knapsack problem that arises in resource allocation with financial
constraints or electronic transfer of funds (Martello and Toth, 1990), traveling salesman
problem that is applied to product distribution or the production of printed circuit boards
(Applegate et al., 1994), and the multi-commodity flow problem that has applications in
production planning or warehousing (Ahuja et al., 1993). In addition, other well-known
examples of combinatorial optimization problems in production and logistics include the
2
permutation flow shop scheduling problem, flexible flow shop scheduling problem,
quadratic assignment problem and vehicle routing problem, which are described below:
1. Permutation Flow Shop Scheduling Problem: This problem involves
determining a sequence of n jobs to be processed through m machines with the
same order of jobs on the machines. Because of the same order of jobs through all
machines, the sequence of jobs processed on the first machine is considered as a
feasible solution of the problem. The most popular objective of this problem is to
minimize the completion time of jobs, also known as makespan (Cmax). An integer
programming model for the scheduling problem was presented in Pinedo (2005).
2. Flexible Flow Shop Scheduling Problem: This problem involves a set of jobs
processed through several consecutive operation stages with parallel identical
machines in each stage. A job can be processed on any idle machine at the stage
into which the job is going. In some cases, there are limited intermediate buffers
between consecutive stages. The primary objective of this problem is to find a
production schedule to minimize the completion time of jobs. There are also other
important objectives of this problem, such as to minimize the total weighted flow
time of jobs and to minimize the total weighted tardiness time of jobs. This
problem can be formulated as a mixed-integer programming model (Sawik, 2000).
3. Quadratic Assignment Problem: This problem is first stated by Koopmans and
Beckmann (1957). Given a set of facilities and a set of locations with the same size
n, the aim is to assign the facilities to the locations such that the total cost is
minimized. The total cost w is calculated using the distance between locations and
3
the flow between facilities. This problem can then be expressed as the problem of
finding a permutation π of n facilities as follows:
n n
min w (π ) = ∑ ∑ f ij d π [ i ]π [ j ] , (1.1)
π ∈Π n
i =1 j =1
where Π n denotes the set of possible permutations of N = {1, 2, …, n}, while
π [i] and π [ j ] denote the location of facilities i and j in the permutation π
respectively. Furthermore, fij is the flow between facilities i and j, and dπ [ i ]π [ j ] is
the distance between locations π [i] and π [ j ] .
4. Vehicle Routing Problem: The problem on how to service a number of
customers’ demand with a fleet of given vehicles under some specific constraints
is known as a vehicle routing problem. Such a problem was first proposed by
Dantzig and Ramser (1959), and it is currently well-known with many important
applications in the field of logistics and supply chain management. An objective of
the vehicle routing problem is to assign the vehicles to the customers and find the
efficient routes of the vehicles so that the total travel distance or time is minimized.
This problem may be formulated as an integer programming model.
1.2 Nature Inspired Algorithms
In this section, we show a classification of optimization methods for solving combinatorial
optimization problems in Figure 1.1. The classification is based on the operational
mechanism of optimization methods. Our focus is on nature-inspired algorithms, which
4
can be considered as a type of meta-heuristic algorithms. Here the term “meta-” means
“beyond” or “higher level”, while the term “heuristic” means “search” or “discover by
trial or error” (Yang, 2008). Thus, meta-heuristic algorithms are known as optimization
methods that search for optimal solutions of a problem by iteratively improving a
candidate solution with regards to a given objective function.
Meta-heuristic algorithms usually consist of two major processes, i.e., solution
exploration and solution exploitation. These two processes are iteratively performed to
search for optimal or near-optimal solutions in reasonable computation time. The
exploration process not only increases the diversity of solutions found, but also helps to
overcome local optimal solutions to obtain better or optimal ones due to its randomization.
The exploitation process aims to improve the quality of solutions obtained from the
exploration process in order to ensure that the solutions will converge to optimality. In
some meta-heuristic algorithms, this exploitation process also helps to overcome local
optimal solutions to search for better or optimal ones. The performance of meta-heuristic
algorithms depends on the appropriate combination between these two processes. Because
of the features of meta-heuristic algorithms, they can search for solutions of combinatorial
optimization problems with good quality in realistic computation time. Some well-known
applications can be found in Liao et al. (2007) and Yang (2008).
We have classified meta-heuristic algorithms into two major types, i.e., nature-inspired
algorithms and non-nature inspired algorithms. Nature has been evolving for millions of
years and hence learning from the success of nature to design meta-heuristic algorithms is
a creative idea (Yang, 2008). Nature-inspired algorithms can be further divided into
biologically inspired algorithms, botanically inspired algorithms and water flow inspired
5
Optimization
Continuous Discrete
Optimization Optimization
Exact Approximation
Algorithms Algorithms
Constructive Local search Meta-heuristic

Heuristics Methods Algorithms
Non-nature
Nature-inspired
Inspired
Algorithms
Algorithms
Biologically Botanically Water Flow

Inspired Inspired Inspired
Algorithms Algorithms Techniques
Figure 1.1 A Classification of Algorithms for Combinatorial Optimization Problems
6
techniques. A detailed description of these types of nature-inspired algorithms is given in
Chapter 2. We also present the important applications of these algorithms in the same
chapter.
1.3 Motivation and Research Objectives
Real-world problems that are formulated as combinatorial optimization problems often
face the following difficulties: (a) their problem size or dimension is large, (b) they are
computationally complex to solve, and (c) even after decomposing into simpler sub-
problems, they are still NP-hard problems. Although the search space of such problems is
determined by a finite set of feasible solutions, it grows exponentially with the size of the
problems. Hence, it is sometimes sufficient to obtain a near-optimal solution to such
problems in practice. As such, there is a need for constructing good meta-heuristic
algorithms that can search for solutions with good quality in realistic computation time.
Over the last few decades, meta-heuristic algorithms inspired by natural phenomena
have been extensively developed to become search and optimization tools in various
optimization problems. Based on the inherent features of meta-heuristic algorithms,
nature-inspired algorithms have been successful in solving many optimization problems.
However, the algorithms are only efficient on specific problems, and there is a need to
change their operational mechanism to solve other problems. There is thus a lack of such
algorithms that are able to solve diverse combinatorial optimization problems.
The success of nature-inspired algorithms for solving optimization problems has
motivated us to learn about their potential capability in constructing an algorithm inspired
7
by other natural phenomena. Thus the main research objective of this thesis is to design a
novel nature-inspired algorithm for solving combinatorial optimization problems. Also,
the algorithm has to balance between solution exploration and exploitation capabilities to
achieve optimal solutions in realistic computation time to real-world problems. Moreover,
this thesis aims to develop a nature-inspired algorithm that can solve a variety of
optimization problems with similar problem structure. Another objective of the thesis is to
design the algorithm in such a way that it is able to solve both single-objective and multi-
objective optimization problems efficiently. Finally, the algorithm should not have
difficulty hybridizing with other well-known algorithms to increase the algorithm’s
solution efficiency.
One of the well-known natural phenomena is the hydrological cycle in meteorology
and the erosion process of water flow in nature. They possess features that are suitable for
developing an efficient optimization algorithm. As such, this thesis is focused on utilizing
such natural phenomena in developing a novel nature-inspired algorithm, and then testing
the performance of the resulting algorithm with different types of NP-hard combinatorial
optimization problems.
1.4 Main Contributions of the Thesis
In this section, we summarize the main contributions of this thesis as follows:
Firstly, a novel nature-inspired algorithm, known as the water flow algorithm (WFA),
for solving NP-hard optimization problems is proposed. The proposed algorithm has
mostly imitated the characteristics of water flow in nature and the components of the
8
hydrological cycle in meteorology. This algorithm consists of two major phases, i.e., the
solution exploration and exploitation phases, inspired by the hydrological cycle and the
erosion phenomenon, respectively.
Secondly, the WFA has been developed to successfully solve several NP-hard
optimization problems, such as permutation flow shop scheduling problem (PFSP),
flexible flow shop scheduling problem (FFSP) with intermediate buffers, quadratic
assignment problem (QAP), and capacitated vehicle routing problem (CVRP). Almost all
the best known solutions of the benchmark instances used can be found by the proposed
algorithm. Moreover, this algorithm can obtain many new best known solutions for PFSP
and FFSP.
Thirdly, we have constructed a WFA which is able to solve multi-objective
optimization problems by modifying and integrating some specialized components. It was
applied to solve the multi-objective FFSP with intermediate buffers. In this algorithm,
landscape analysis based on the ellipsoid approximation was integrated to help determine
the weights of objective functions, which guide the WFA to exploit potential regions and
move towards the optimal Pareto solution set. The results obtained demonstrate the
effectiveness and efficiency of the WFA, and show that this is a promising approach to
solve other multi-objective optimization problems.
In addition, when WFA is developed to solve the FFSP with intermediate buffers, we
also propose an efficient procedure of constructing a complete schedule from a job
permutation at the first stage. The constructive procedure outperforms other procedures in
the literature. This procedure may increase the performance of any algorithm when it is
9
integrated to solve the FFSP. Also for solving the FFSP, the WFA is applied to the
maltose syrup production problem. The results obtained show the capability of the WFA
for solving problems in complex real-world applications.
Lastly, a constructive algorithm for generating initial solutions of the WFA when
solving the CVRP is proposed in this thesis. The constructive algorithm is based on
solving a relaxed CVRP. The results obtained show that this algorithm may generate good
initial solutions for the CVRP. Furthermore, this approach of finding initial solutions may
be applied to other optimization problems.
1.5 Outline of the Thesis
This thesis aims to develop a novel nature-inspired algorithm for combinatorial
optimization problems and its content is organized into eight chapters. Figure 1.2 shows
the organization and relationship among the chapters.
A detailed review of nature-inspired algorithms that have emerged in recent years is
provided in Chapter 2. There are three main groups in the literature review corresponding
to the classification of nature-inspired algorithms shown in Section 1.2. The basic ideas of
designing nature-inspired algorithms, the development of the algorithms, as well as their
successful applications are also described in this chapter.
Chapter 3 describes the phenomena of nature used to construct the WFA, i.e., the
hydrological cycle in meteorology and the erosion process of water flow in nature. Also,
the operational mechanism and the main operators of the proposed algorithm are
explained in detail.
10
Chapters 4 and 5 describe the implementation of WFA for two single-objective
combinatorial optimization problems, namely the permutation flow shop scheduling
problem and the flexible flow shop scheduling problem, respectively. Computational
results and comparisons carried out on benchmark problems are also presented and
discussed. In addition, a practical example of maltose syrup production solved by the
WFA is shown in Chapter 5. Chapter 6 presents on how the WFA can be developed to
solve a multi-objective flexible flow shop scheduling problem, and computational results
are also shown in this chapter.
Some further applications of the WFA for other single-objective combinatorial
optimization problems, such as the quadratic assignment problem and vehicle routing
problem, are described in Chapter 7. The proposed algorithms are also tested with the
benchmark instances from the literature, with the computational results and comparisons
being shown in this chapter. Finally, some conclusions are provided in Chapter 8. The
contributions of this research work are also discussed, together with some possible future
research works.
11
Introduction
(Chapter 1)
Literature Review
(Chapter 2)
Water Flow Algorithm

(Chapter 3)
Single-Objective Multi-Objective
Optimization Problems Optimization Problem
Flow Shop Scheduling (Chapter 4)
Flexible Flow Shop

(Chapter 5) (Chapter 6) Conclusions &
Scheduling
Future Research
Quadratic Assignment (Chapter 8)
Problem
(Chapter 7)
Vehicle Routing
Problem
Figure 1.2 Organization of the Thesis
12
Chapter 2 Literature Review
CHAPTER 2
LITERATURE REVIEW
Nature-inspired algorithms have become useful optimization methods for solving a variety
of real-world problems. Based on the appropriate balance between solution exploration
and exploitation capabilities, the algorithms can obtain solutions with high quality in
realistic computation time. In this chapter, we present several well-known nature-inspired
algorithms in recent years. The basic ideas of constructing and developing the algorithms,
as well as their most important applications are described in detail. Here, the nature-
inspired algorithms are classified into three main groups: biologically inspired algorithms,
botanically inspired algorithms and water flow inspired techniques. A survey of these
three groups of algorithms is presented in Section 2.1, Section 2.2 and Section 2.3
respectively. Also, some findings from the literature review are presented in this section.
Finally, a timeline of all the nature-inspired algorithms and the additional features of the
proposed nature-inspired algorithm which help to overcome the drawbacks of existing
algorithms are shown.
13
2.1 Biologically Inspired Algorithms
In this section, we present some popular meta-heuristic algorithms inspired from biology.
The biologically inspired algorithms are classified into three major groups: evolutionary
algorithms, stigmergic optimization algorithms, and swarm-based optimization algorithms.
The group of evolutionary algorithms consists of genetic algorithm, memetic algorithm,
and shuffled frog-leaping algorithm. The group of stigmergic optimization algorithms
includes termite algorithm, ant colony optimization and bee colony optimization. Finally,
the group of swarm-based optimization algorithms includes particle swarm optimization,
firefly algorithm, and bat algorithm.
2.1.1 Evolutionary Algorithms
Evolutionary algorithms are stochastic optimization methods based on the principles of
natural evolution. Natural evolution is a complex process which operates on chromosomes,
instead of organisms (Michalewicz, 1992). The chromosomes contain genetic information,
called a gene, which is passed from one generation to next generation through
reproduction. In reproduction, the most important operators are recombination and
mutation. The recombination plays a role in the exchange of genetic information among
parent individuals to produce an offspring, while the mutation aims to create
diversification of genetic information in offspring. Organisms with good chromosomes
have a higher chance to exist and develop in nature. According to Darwin’s natural
selection theory (Darwin, 1859), natural selection prioritizes the proliferation of
environment-adapted organisms, but causes the extinction of non-environment-adapted
organisms.
14
A well-known evolutionary algorithm is genetic algorithm (GA) first introduced by
Holland (1975). The idea of establishing GA originated from the evolutionism of Darwin
(1859): “Survival of the genetically fittest”. Since the introduction of this algorithm, it has
become a popular and useful optimization method for solving optimization problems.
Although many variants of GA have been developed, the general framework of this
algorithm does not have any significant change. The basic principles of GA are described
in detail by Goldberg (1989). A flow chart of the general GA is shown in Figure 2.1.
The solution exploration capability of GA increases because of the initial population.
However, a certain degree of exploiting the regions with high quality solutions is missing.
Hence, the GA is combined with a local search to overcome the drawback, which
constitutes a memetic algorithm (MA). In particular, the MA allows all chromosomes as
well as individuals to gain some experience through a local search process before they are
evolved. The MA was first introduced by Moscato (1989). In this algorithm, the genetic
information to form a chromosome is called memes and not genes. This is inspired by
Dawkins’ notion of a meme (Dawkins, 1976). A detailed description of the algorithm can
be found in Moscato and Cotta (2003). Here, a flow chart of the MA is shown in Figure
2.2.
Eusuff and Lansey (2003) proposed a shuffled frog-leaping algorithm which combines
the advantages of MA and the social behavior of frogs. In this algorithm, a set of frogs
similar to a population of individuals in GA is partitioned into subsets, called memeplexes.
The memeplexes representing the different cultures of frogs are improved through a
process of memetic evolution. Based on the social behavior of frogs, the good evolved
knowledge obtained is passed among memeplexes to help the memeplexes evolve together.
15
The process continues to be performed until stopping criteria are satisfied. A flow chart of
the shuffled frog-leaping algorithm is shown in Figure 2.3.
Start
Parameter
Initialization
Initial population
Fitness evaluation
Reproduction
No Recombination
Terminate?
Yes
Output the best

individual
End
Figure 2.1 Flow Chart of Genetic Algorithm
16
Start
Parameter
Initialization
Initial population
Fitness evaluation
Local search
Reproduction
Recombination
No
Local search
Terminate?
Yes
Output the best

individual
End
Figure 2.2 Flow Chart of Memetic Algorithm
17
Start
Parameter
Initialization
Initial population of
frogs
Fitness evaluation
Based on descending order of

the fitness to divide the
population into memeplexes
Based on the memetic evolution

process to improve the position of
the worst frog in each memeplex
No
Update the evolved memeplexes
into the population
Terminate?
Yes
Output the best frog
End
Figure 2.3 Flow Chart of Shuffled Frog-Leaping Algorithm
18
2.1.2 Stigmergic Optimization Algorithms
According to (Abraham et al., 2006) on stigmergic optimization, Grasse´ quoted: “Self-
Organization in social insects often requires interactions among insects: such interactions
can be direct or indirect. Direct interactions are the “obvious” interactions: antennation,
trophallaxis (food or liquid exchange), mandibular contact, visual contact, chemical
contact (the odor of nearby nestmates), etc. Indirect interactions are more subtle: two
individuals interact indirectly when one of them modifies the environment and the other
responds to the new environment at a later time. Such an interaction is an example of
stigmergy”.
We can observe such an indirect interaction from social insects, such as termites. In
the process of nest reconstruction, termites interact through local pheromone
concentrations on the nest structure. The state of nest structure coordinates tasks for
termites. They work together until the nest construction is completed. Another example of
stigmergy is pheromone communication in ant colony. Ants are able to leave a chemical
trail on their path to guide other ants to the food source found. Deneubourg et al. (1987)
carried out experimental studies to test the ability of ants in searching for the shortest path
to the food source. Similar to ants, honey bees can communicate by pheromones. They can
deliver a chemical message to encourage attack response to other bees. In addition, honey
bees can communicate by “waggle dances”. The so-called waggle dances play a role as a
signal system which is used to guide other bees to the path to a good food source. Seeley
et al. (1991) carried out experimental studies about the ability of bees in allocating and
collecting their flower patches.
19
Many meta-heuristic algorithms can be constructed by learning the behavior of these
social insects. We present well-known nature-inspired algorithms in this stigmergic
optimization group, such as termite algorithm, ant colony optimization and bee colony
optimization.
The first algorithm considered in this group is the termite algorithm proposed by Roth
and Wicker (2006) for mobile wireless ad-hoc networks. The proposed algorithm is
inspired by the hill building behavior of termites with four principles. Resnick (1997)
described the principles as well as a detailed example of the hill building procedure. In the
example, it is assumed that termites and pebbles are distributed on a flat surface. Since the
hill of termites is built from the pebbles, the objective of termites is to collect all the
pebbles into the same place. Termites move based on pheromone trails which are excreted
by the others. Following the trails, termites complete their hill building together. To
achieve success, a termite must conform to the rules described in Roth and Wicker (2006).
A detailed description of matching the hill building procedure of termites and the
principles of swarm intelligence to design the termite algorithm is also provided in Roth
and Wicker (2006).
The second algorithm considered in this group is the ant colony optimization (ACO)
inspired by the foraging procedure of real ants in nature. Although an ant is very tiny and
wanders aimlessly, a colony of ants expresses an extraordinarily intelligent behavior
through their nest building and foraging. ACO was first introduced by Dorigo and his
colleagues around 1991-1992. Since then, it is known as a useful nature-inspired
algorithm for combinatorial optimization problems (Dorigo and Gambardella, 1997; and
Dorigo, 1999). In the literature, there is a variety of ACO algorithms. However, all of
20
them have the same framework that is described in Grosan and Abraham (2006) in detail.
Here, a flow chart of the basic ACO algorithm is shown in Figure 2.4.
Start
Initialize pheromone
trails and parameters
Generate population
of ants
Calculate fitness
value for each ant
Determine the best

position for each ant
No
Determine the best
global ant
Update pheromone
trails
Terminate?
Yes
Output the best

global ant
End
Figure 2.4 Flow Chart of Ant Colony Optimization Algorithm
21
The third algorithm considered in this stigmergic optimization group is the bee
algorithm. The bee algorithm is inspired by the foraging behavior of honey bees. The
honey bee algorithm proposed by Nakrani and Tovey (2004) for allocating computers
among different clients and web hosting server is one of the earliest bee algorithms. The
idea of constructing this algorithm has originated from how forager bees can optimally
collect an amount of nectar if they are allocated to different flower patches. On the other
hand, Haddad et al. (2005) developed a honey-bee mating optimization algorithm for
solving a reservoir operation problem. This algorithm originated from the behavior of
queen bee in mating with other bees to form the bee colony. Farooq (2006) presented a
bee algorithm for routing in telecommunication network. The algorithm is inspired by the
way bees communicate.
Although various bee algorithms have been developed based on the different behavior
of honey bees in foraging and mating, they still keep to the same framework. A detailed
survey of bee colony algorithms is presented in Bitam et al. (2010). A flow chart of the
basic bee algorithm is shown in Figure 2.5.
2.1.3 Swarm-Based Optimization Algorithms
In this section, we present a group of swarm-based optimization algorithms. The
algorithms are inspired by the social behavior of swarm-based animals or insects,
especially those in which the property of historical information exchange among
individuals is magnified. The well-known algorithms in this group include particle swarm
optimization, firefly algorithm and bat algorithm.
22
Start
Initialize parameters
Encode the objective

function into virtual
nectar level
Define dance routine
Generate new
solutions for bees
Evaluate the new No

solutions of bees
Consider all
No dimensions?
Yes
Communicate and update

the optimal solution set
Terminate?
Yes
Decode and output

the best solution
End
Figure 2.5 Flow Chart of Bee Colony Optimization Algorithm
23
The first algorithm considered in this group is particle swarm optimization (PSO). This
is a biologically inspired algorithm that simulates the social behavior of animals, such as a
school of fish or a flock of birds. Here, the basic idea of this algorithm is to use many
autonomous agents (particles) that act together in simple ways to produce seemingly
complex behavior (Banks et al., 2007). The PSO was first formally introduced by
Kennedy and Eberhart (1995). In the paper, the authors replaced the simple goal of the
model proposed by Reynold (1987) with a more realistic goal to find food. With this new
goal, they established a general model for PSO which is used by most researchers in PSO.
Since PSO was developed from an original flocking system, various components were
incorporated into the algorithm, such as inertia and constriction. Also, some unnecessary
components were removed, such as velocity matching and collision avoidance. However,
the standard structure of PSO has generally been preserved. A detailed description of the
standard PSO structure is provided in Kennedy and Eberhart (1995). A flow chart of the
standard PSO algorithm is shown in Figure 2.6.
The second algorithm considered in this group is the firefly algorithm introduced by
Yang (2008). This algorithm is a swarm-based optimization method for solving
continuous optimization problem. The algorithm was motivated by simulating the social
behavior of fireflies. The fireflies produce short and rhythmic flashes to attract mating
partners and potential prey. Also, such flashes of the fireflies may be used to warn
potential predators about their bitter taste. A signal system including the rhythmic flash,
the rate of flashing and the amount of flashing time attracts both genders together. If
female fireflies are attracted by a male firefly, they will respond to the male firefly’s
pattern of flashing. A flow chart of the algorithm is shown in Figure 2.7.
24
Figure 2.6 Flow Chart of Particle Swarm Optimization Algorithm
25
Start
Generate the initial

population of fireflies
Determine the light

intensity of fireflies
Based on the light intensity of

firefly, update the light intensity at
all dimensions
Recalculate the
attractiveness
Evaluate new solutions

and update the light
No intensity for fireflies
Rank the fireflies and determine

the best one so far
MaxGeneration?
Yes
Output the best fireflies
End
Figure 2.7 Flow Chart of Firefly Algorithm
26
The third algorithm considered in the group of swarm-based optimization algorithm is
the bat algorithm proposed by Yang (2010). The algorithm is an optimization method
based on the echolocation behavior of bats. Among all the species of bats, the behavior of
microbats motivated the bat algorithm, since they use echolocation extensively, unlike
other bats (Richardson, 2008). Microbats’ echolocation capability helps them to detect
preys, distinguish different kinds of insects, avoid obstacles, and locate their roosting
crevices in the dark. A detailed description of the bat algorithm is provided in Yang
(2010). Here, a flow chart of the algorithm is shown in Figure 2.8.
2.2 Botanically Inspired Algorithms
In this section, we describe two meta-heuristic algorithms inspired from botany. They
consist of an invasive weed optimization algorithm and a grafting-inspired algorithm.
2.2.1 An Invasive Weed Optimization Algorithm
Invasive weed optimization (IWO) is a numerical stochastic optimization algorithm
inspired by the ecological process of weed colonization and distribution. In biology,
weeds are plants whose invasive habits of growth are vigorous. They threaten the growth
of cultivated plants, which in turn poses a threat to agriculture. Weeds are also found to be
very robust and adaptive to changes of environment. Hence, Mehrabian and Lucas (2006)
used their prominent properties, such as robustness, adaptation and randomness, to design
a simple but efficient optimization algorithm for real parameter optimization. A detailed
description of the algorithm is provided in Mehrabian and Lucas (2006). A flow chart of
the IWO is shown in Figure 2.9.
27
Start
Generate the position and

velocity for the initial bats
Determine the pulse

frequency for each bat
Generate new positions for bats by

adjusting frequency, and updating
velocities and positions
Select the bat with the best position,

and generate a local position around
the selected best position of the bat
Randomly generate a new position

for a bat chosen
No
Accept the position if it satisfies the
conditions, and then increase pulse
rate and reduce loudness
Rank the bats and determine

the best one so far
Maximum
Iterations?
Yes
Output the best bats
End
Figure 2.8 Flow Chart of Bat Algorithm
28
Start
(i = 1)
Generate a population of initial plants

(initializing a population)
Determine the fitness value of each

plant in the population
Based on the fitness value, the plants

No produce the corresponding number of seeds
(reproduction)
Randomly disperse the produced seeds over

the search area and grow new plants i=i+1
(spatial dispersal)
Is the maximum number

of plants reached?
Yes
Only the plants with lower fitness

are used to produce seeds
No
(competitive exclusion)
Is the maximum number

of iterations reached?
Yes
Output the plant with the best

fitness
End
Figure 2.9 Flow Chart of Invasive Weed Optimization
29
2.2.2 A Botany-Grafting Inspired Algorithm
Li et al. (2010) proposed a hybrid algorithm that combines an improved genetic algorithm
(IGA) and an improved particle swarm optimization (IPSO), called HIGAPSO. The
combination is inspired by the idea of grafting in botany. Based on the principle that
grafting in botany can integrate the strengths of two original branches, the hybrid
algorithm combines the advantages of IGA and IPSO together. A flow chart of the
proposed hybrid algorithm is shown in Figure 2.10.
2.3 Water Flow Inspired Techniques
In this section, we introduce some recent techniques inspired by certain factors or
processes of nature that are related to water flow. The techniques consist of document
image processing methods based on water flow model, intelligent water drops algorithm
based on the dynamics of river systems, and water flow-like algorithm based on the
behavior of fluid flows. These techniques have been successfully applied to a variety of
optimization areas, such as pattern recognition, manufacturing and logistics.
2.3.1 Image Processing Methods Based on Water Flow Model
Kim et al. (2002) can be considered as one of the earliest authors who used the properties
of water flow to solve optimization problems. In the paper, the authors used a water flow
model to simulate an image processing problem. In particular, a grey level image is
considered as a three dimensional terrain including mountains and valleys, which
represent background and character regions in the image processing problem, respectively.
30
Start
Randomly generate a population of

P individuals
Calculate the fitness for each

individual in the initial population
Sort the individuals in ascending

order based on their fitness values
Select the top H individuals as the

elites to reproduce in next generation
Apply IPSO strategy for the S

better individuals and put the enhanced
individuals into the grafting population
No
Apply IGA for the remaining P-H-S

individuals together with the
individuals enhanced by IPSO
Choose the best P-H-S individuals

obtained by the IGA to complete the
population of next generation.
Update the best individual

from the population obtained
Satisfy stopping criteria?
Yes
End
Figure 2.10 Flow Chart of Botany-Grafting Inspired Algorithm
31
Based on the property of water flow always moving to lower regions, the authors
simulated with the pouring of water onto the terrain surface. In the field of image
processing, the objective often considered is to extract characters from backgrounds. Thus,
a threshold process is proposed to extract the valleys by the amount of filled water. A flow
chart of the whole method of Kim et al. (2002) is shown in Figure 2.11.
Start
w=0
Increase w by 1
w = wmax?
No Rain drop for each
pixel using a water
Yes
flow model
Label each pond based on
connectivity
1) Calculate average water

level for each pond
2) Assign the average value
for the water level
1) Extract the water from terrain

2)Threshold the water by Ostu’s
method
Stop
Figure 2.11 Flow Chart of the Method Proposed by Kim et al. (2002)
32
However, the method of Kim et al. (2002) requires long processing time since it does
not determine when to stop the iterative process. Also, characters on poor contrast
backgrounds in document images are often not separated successfully. Hence, Oh et al.
(2005) proposed an improved method which includes extraction of regions of interest
(ROI), an automatic stopping criterion, and hierarchical threshold process, to overcome
the drawbacks in the method of Kim et al. (2002). In the improved method, the input
image, known as terrain, is divided into regions of interest and desert regions. Then,
rainfall only occurs within the regions of interest. An automatic terminating criterion for
the iterative rainfall process was then provided to reduce the computational time needed
by the model. A flow chart of the method of Oh et al. (2005) is shown in Figure 2.12.
Figure 2.12 Flow Chart of the Improved Method Proposed by Oh et al. (2005)
33
Basu et al. (2007) continued to use this idea to deal with the problem of text line
extraction from optically scanned document images. Because of the appearance of multi-
skewed lines in text images, the problem becomes complex and difficult to solve
completely. To solve this problem, Basu et al. (2007) proposed a text line extraction
technique based on a water flow model. In this technique, hypothetical water flows, which
are required to move from both left and right sides of the image frame, are stuck by the
characters of text lines. Then, the boundaries of un-wetted areas on the image frame are
labeled to extract text lines. Brodic and Milivojevic (2009a and 2009b) also adopted the
method of Basu et al. (2007) for the problems of identifying and detecting reference text
line, respectively. However, these methods limited hypothetical water flows under a few
specified angles of document image frame. Thus, they failed in complex text examples in
which multi-skewed lines with large skew angle exist in the text images. Brodic and
Milivojevic (2010) proposed a modification of the water flow method for segmentation
and text parameters extraction of sample text at almost any skew angle. The modification
includes the extension of skew angles of the document image frame and the enlargement
of un-wetted image regions. The procedure of document text image identification
proposed by Brodic and Milivojevic (2010) is shown in Figure 2.13.
Figure 2.13 Procedure of Text Image Identification (Brodic and Milivojevic, 2010)
34
In summary, the works in this section only focus on combining the property of
water flow always moving down to lower regions and filling up valleys into their model or
technique for pattern recognition problems. For the document image processing problem
in which the objective is to extract characters from backgrounds, the valleys with the
amount of filled water are extracted. Otherwise for the problem of text line extraction
from optically scanned document images, un-wetted areas on the image frame are labeled
to be extracted. The models or techniques are simple and have not been established as
algorithmic methods.
2.3.2 Intelligent Water Drops Algorithm
The idea of constructing an optimization algorithm based on certain factors or processes
related to water flow was developed by Shah-Hosseini (2007). In this paper, the author
proposed an intelligent-water-drops (IWD) algorithm based on the dynamics of river
systems and the actions/reactions that happen between water drops and the environmental
changes in the flowing river. In this algorithm, artificial water drops are designed with two
properties of natural water drops, i.e., moving velocity and an amount of carrying soil. The
values of both properties may change when the water drops travel in the river since an
amount of removed soil from the river increases their velocity and volume. To travel from
the source to the destination, the water drops choose the paths with minimal soil. A flow
chart of the IWD algorithm is shown in Figure 2.14.
The IWD algorithm has been applied to solve optimization problems, such as the
traveling salesman problem (Shah-Hosseini, 2007), the multiple-knapsack problem (Shah-
Hosseini, 2008), and the n-queen puzzle (Shah-Hosseini, 2009). However, the properties
35
Edited by Foxit Reader
Chapter 2 Literature Review Copyright(C) by Foxit Software Company,2005-2007
For Evaluation Only.
of water flow and meteorological phenomena related to water drops have not been
explored fully in these papers. Also, this algorithm depends on the large number of
parameters defined by users, which may affect the performance of the algorithm.
Furthermore, the computational results obtained by the IWD algorithm for its applications
have not been good when compared with other algorithms.
add
step
Figure 2.14 Flow Chart of the IWD Algorithm (Shah-Hosseini, 2007)
36
2.3.3 Water Flow-Like Algorithm
An optimization algorithm based on the behaviors of water flow was introduced by Yang
and Wang (2007) for solving the bin packing problem, which is one of the well-known
discrete optimization problems. This algorithm is a multiple-agent-based optimization
method with an improvement idea of using a population of agents with size that is not
fixed. This algorithm may be described as follows: solution agents are modeled as water
flows which move on the terrain. Driven by the gravity and governed by the energy
conservation law, water flows will automatically move to lower altitudes on the terrain.
When moving from higher altitudes to lower ones on a rugged terrain, the water flow will
split into multiple sub-flows. Fluid momentum helps water flows adjust their compositions
and directions when they move through the rough terrains. When a number of water flows
move to the same position, they will be merged into a single water flow. In addition, water
flows may move upward to higher altitudes if their kinetic energy is larger than the
potential energy required. Furthermore, to explore search space, some water flows may
periodically evaporate and return to the terrain by precipitation.
In general, this is an evolutionary algorithm involving four operations: splitting and
moving, merging, evaporation, and precipitation. These operations are established and
controlled by some fundamental laws of physics, such as the law of energy conservation,
and the law of momentum conservation. Physics quantities such as mass, velocity, fluid
momentum, energy, and gravitational acceleration are also used as basic parameters to
construct this algorithm. A flow chart of the algorithm proposed by Yang and Wang
(2007) is shown in Figure 2.15.
37
Figure 2.15 Flow Chart of the Water Flow-Like Algorithm (Yang and Wang, 2007)
The experimental results and comparisons showed that this algorithm can perform well
for solving the bin packing problem. However, Yang and Wang (2007) only used two test
instances to evaluate and compare their algorithm with other algorithms, and may not be
sufficient to conclude that their algorithm is efficient for the bin packing problem. Also,
because of the dependence on the large number of controlled parameters defined by users,
this algorithm may spend a considerable amount of computational time for evaluating and
processing the operators.
Wu et al. (2010) adopted the logic of the algorithm proposed by Yang and Wang
(2007) to design a heuristic algorithm for solving the cell formation problem. This is one
38
of the important problems in cellular manufacturing when group technology is applied.
The proposed algorithm consists of two stages. In the first stage, the algorithm generates
initial feasible solutions in order to determine optimal cell size. In the second stage, the
optimal cell size obtained is used as a lower bound to search for optimal solutions. Since
the cell size may be a good lower bound for the solution process, the computational time
needed by the algorithm is significantly reduced, especially for large-sized instances.
The experimental results and comparisons for a set of 37 test instances from the
literature showed that the proposed algorithm of Wu et al. (2010) has performed better
than other algorithms, especially for large-sized problems. However, the large number of
controlled parameters still present in this algorithm may limit the performance of the
algorithm.
From the success obtained for solving optimization problems, such as the traveling
salesman problem, the bin packing problem, the multiple-knapsack problem, the n-queen
puzzle and the cell formulation problem, we can see that the algorithms inspired by the
behavior of water flows in nature indicate a promising optimization approach.
2.4 Conclusions and Possible Nature-Inspired Algorithm
A summary of the nature-inspired algorithms described in this chapter is shown in Table
2.1. In addition to the introduction of the basic ideas of constructing the algorithms, we
also provide the most important applications for which each algorithm has worked best.
This helps researchers have a condensed but ample overall view of the real-world
application capability of the nature-inspired algorithms.
39
Also, we present the timeline of all the nature-inspired algorithms presented in this
chapter (see Figure 2.16). From this figure, we can see that the optimization approaches
have recently developed significantly in terms of quantity as well as quality. Many
problem-solving models from nature are inspired by the design of optimization algorithms.
Successful applications of the algorithms for continuous and discrete optimization
problems in engineering have demonstrated the efficiency of these algorithms. Although
water flow inspired algorithms are developed recently, the algorithms have showed their
efficiency and effectiveness for solving some well-known optimization problems, such as
traveling salesman problem and bin packing problem. However, the potential of this
algorithm has not been fully exploited since many characteristics as well as behaviors of
water flow have not been considered. Hence, constructing an optimization algorithm based
on water flow models is still a promising research area.
A possible nature-inspired algorithm will be to integrate many additional features of
water flow in nature. Natural phenomena, such as the hydrological cycle and erosion
process, which have not been explored by other algorithms, will also be simulated in such
a proposed algorithm. Two major phases of the algorithm, the solution exploration and
exploitation phases which are inspired by the hydrological cycle and the erosion process
respectively, help to achieve a balance between solution diversification and intensification
capabilities to search for optimal solutions in reasonable computation time.
Moreover, this algorithm has only a small number of parameters defined by users,
which may limit the performance degradation of the algorithm due to the tuning of the
parameters. In addition, this algorithm can be applied to solve different optimization
40
Table 2.1 Summary of Applications of the Nature-Inspired Algorithms
Groups Sub-groups Algorithms Ideas Problems
Biologically Evolutionary Genetic algorithm The evolutionism Permutation flow shop

inspired algorithms of Darwin scheduling
algorithms (Reeves, 1995)
Memetic algorithm Genetic algorithm Traveling salesman
+ local search problem (Moscato &
Norman, 1992)
Shuffled frog- Memetic algorithm Water distribution
leaping algorithm + the social network (Eusuff &
behavior of frogs Lansey, 2003)
Stigmergic Termite algorithm Process of nest Mobile wireless ad-hoc
optimization reconstruction of networks (Roth &
algorithms termites Wicker, 2006)
Ant colony The foraging Traveling salesman
optimization procedure of ants problem (Dorigo &
Gambardella, 1997)
Bee colony The behavior of Telecommunication
optimization honey bees in network (Farooq, 2006)
foraging/mating
Swarm-based Particle swarm The social behavior Flow shop scheduling
optimization optimization of a school of fish (Liao et al., 2007)
algorithms or a flock of bird
Firefly algorithm The social behavior Continuous constrained
of fireflies optimization (Lukasik
& Zak, 2009)
Bat algorithm The echolocation Continuous constrained
behavior of bats optimization
(Yang, 2010)
Botanically An invasive weed The ecological Electromagnetic
inspired optimization process of weed designing problem
algorithms colonization and (Karimkashi & Kishk,
distribution 2010)
A botany-grafting The idea of grafting Spherical conformal
inspired algorithm in botany array (Li et al., 2010)
Water flow Intelligent water The dynamics and Traveling salesman
inspired drops algorithm reactions of river problem (Shah-
techniques systems Hosseini, 2007)
Water flow-like The energy Cell formation problem
algorithm conservation law (Wu et al., 2010)
41
problems without much change in the structure of the algorithm. These are features found
lacking in most existing water flow inspired algorithm.
Figure 2.16 Timeline of Nature-Inspired Algorithms
42
Chapter 3 A General Water Flow Algorithm
CHAPTER 3
A GENERAL WATER FLOW ALGORITHM
In this chapter, we introduce a novel nature-inspired algorithm for solving combinatorial
optimization problems. This algorithm simulates the hydrological cycle in meteorology
and the erosion phenomenon of water flow in nature, which represent the solution
exploration and exploitation capabilities of the algorithm, respectively. Being inspired by
the behaviors and characteristics of water flow in meteorology and nature, this algorithm
is named water flow algorithm (WFA). A detailed description of the natural phenomena
imitated to construct this algorithm is provided in the chapter. In Section 3.1, we present
the hydrological cycle in meteorology together with the components that constitute the
hydrological cycle, such as evaporation, condensation, transportation, precipitation,
transpiration, groundwater and run-off. The erosion phenomenon of water flow in nature
is described in Section 3.2. Finally, a detailed description of the general WFA for solving
combinatorial optimization problems is presented in Section 3.3.
43
3.1 Hydrological Cycle in Meteorology
Water is the source of all life on Earth. Hence, water plays an essential role in the living of
creatures as well as human beings. In nature, water exists in three states: solid, liquid and
gas. These states are affected by changes in the temperature of the environment. When
temperature increases, water can change from the solid state to the liquid state due to
melting, or from the liquid state to the gaseous state due to evaporation. Otherwise, when
temperature decreases, water can change from the liquid state to the solid state due to
freezing, or from the gaseous state to the liquid state due to condensation. When
temperature drops below the freezing point, water molecules can change directly from the
gaseous state to the solid state due to sublimation. Such changing of the states of water is a
fundamental concept in the explanation of many meteorological phenomena.
One of the well-known meteorological phenomena is the hydrological cycle. The
hydrological cycle reflects the circulation and conservation of water on Earth. Though
water in ocean, river, cloud and rain is often in a state of change, the total amount of water
on Earth is not changed. Due to the conservation of water, the circulation of water is
established. The hydrological cycle consists of the following main components.
(1). Evaporation: a process that transfers water from the liquid state in ocean to the
gaseous state in atmosphere.
(2). Condensation: a process that transfers water in atmosphere from the gaseous
state to the liquid state.
(3). Transportation: a movement of water vapor from ocean to land through
atmosphere due to wind.
44
(4). Precipitation: a process that transports water from atmosphere to the Earth’s
surface, also called rain.
(5). Transpiration: a process that transfers water from ground to atmosphere by the
evaporation of plants and vegetation.
(6). Groundwater: the amount of water that infiltrates through the Earth’s surface to
return to the ocean.
(7). Run-off: the rest of the amount of water on the Earth’s surface in rivers, lakes,
ponds or streams, after the transpiration and infiltration processes occur.
Figure 3.1 illustrates these seven basic components of the hydrological cycle. The
exploration phase of WFA proposed in this thesis is inspired by the hydrological cycle.
The phase is designed based on the basic components.
(4) Precipitation (3) Transportation

{Wind}
(2) Condensation
{Cloud}
(5) Transpiration
(7) Run-off
(1) Evaporation
(6) Groundwater
(Source: http://ww2010.atmos.uiuc.edu/(Gh)/guides/mtr/hyd/smry.rxml)
Figure 3.1 Basic Components of the Hydrological Cycle

45
A detailed description of the meteorological phenomenon is provided in Goudie
(1993). Here, we briefly present the relationship among the main components mentioned
in the hydrological cycle.
Firstly, sea water from the surface of the oceans is evaporated to become water vapor
in the atmosphere. When the water vapor is lifted and encounters cool air, it is condensed
to form clouds with moisture. Then, the moisture in clouds is transported around by wind
until it returns to the Earth’s surface under drops of water through precipitation. When the
drops of water fall down to the ground, they are subjected to two possible processes. In the
first process, some of the drops of water may be evaporated back into the atmosphere
through transpiration. In the second process, some of the drops of water may penetrate
through the Earth’s surface to be groundwater. The drops of water returning to the
atmosphere continue to join the clouds and move around, while the groundwater moves
back to the oceans.
The rest of the water on the Earth's surface is called runoff. Governed by the gravity
force, the runoffs move from higher altitudes to lower ones on the ground. Lakes, rivers or
streams are formed from the runoffs if they are held in valleys or depressions. When the
amount of precipitation increases dramatically, water in the lakes, rivers or streams may
flow over their barriers and move back to the oceans, where the hydrological cycle begins
again. This overflowing phenomenon is known as erosion process of water flow in nature,
which will be described in the next section.
46
3.2 Erosion Process of Water Flow in Nature
In nature, there are many phenomena affecting and changing the Earth’s surface. Erosion
process is one of such natural phenomena. In essence, erosion is a natural process of
weathering to the Earth’s surface, which detaches solids from the surface and deposits
them elsewhere (Holy 1982). This process is caused by movements of wind and water in
natural environment. In addition, the gravity force of Earth and the operations of living
creatures are the causes of the erosion process. However, since water covers over 70% of
the Earth’s surface area (Michael, 2006), we can see that it is the major cause of the
erosion process. Figure 3.2 shows an illustration of the erosion by water flow.
Figure 3.2
Erosion from Water Flow
(Source: http://www.salomart.com/images/erosion-2.jpg)
47
Normally, the consequence of erosion process caused by water is not serious to human
beings since it occurs slowly and gently in rivers or streams. However, when an amount of
water of the rivers or streams increases dramatically due to heavy precipitation or tide, the
erosion process may cause a flood. The flood can sweep away everything on the path
through which it passes.
The most important property of the erosion process caused by water flow is erosion
rate, or known as erosion capability, which is also a factor to recognize and control the
flood. The erosion rate is determined by an amount of water in watercourse, the velocity
of water streaming, as well as the state of land. In nature, an amount of water in
watercourse is affected by the amount of precipitation, intensity of precipitation and flood-
tide; while the velocity of water streaming is mainly affected by the slope of landscape (or
gravity force) and the amount of water in the watercourse. Here, the state of land merely
means the hardness of land. On the other hand, the erosion rate depends on three major
groups of factors, i.e., climatic factors, geologic factors and biological factors (Holy,
1982). The climatic factors include the amount and intensity of precipitation as mentioned
above. In addition, average temperature, wind speed, as well as storm frequency may be
considered as climatic factors affecting the erosion rate. The geologic factors consist of
the types of sediment or rock, their porosity and permeability, the slope of landscape, as
well as the shape of rocks. The biological factors are the ground cover area of vegetation,
the type of organisms living in the region, and the purpose of using land.
A different classification of factors that involves the impact of erosion capability is
provided in Figure 3.3. Here, these factors are divided into two main groups, i.e., natural
factors and human factors. Despite the classifications, it can be seen that there are many
48
factors affecting the erosion capability. Hence, it is not easy to control the erosion
rate/capability and flooding in general.
(Source: http://www.dartmoor-npa.gov.uk/learningabout/lab-printableresources/lab-
factsheetshome/lab-erosion)
Figure 3.3 Factors Affecting Erosion Capability
There are also many studies on the relationship between erosion process and factors
affecting this process. Such studies show that when the amount and intensity of
precipitation increase, erosion rate will increase. On the other hand, erosion rate is also
expected to adjust in response to climate changes due to many reasons. The most direct
reason is the change in the erosive power of precipitation. Other direct reasons include:
(1). Because of shifting precipitation regime and evaporation/transpiration rate, land
moisture is also changed. It affects infiltration and runoff ratios.
(2). Because of reducing concentration of land, land erodibility increases. Also, it
leads to increase in the amount of runoff due to appearance of rifts on the surface.
49
(3). Due to increasing temperatures, winter precipitation is shifted from non-erosive
snow to erosive rainfall.
(4). Due to melting of permafrost, a non-erodible land state may be shifted to be an
erodible land state.
All these reasons help to understand clearly about the close relationships of erosion rate
and other factors. We may then propose methods to control the erosion rate, or at least to
limit the loss caused by increasing the erosion rate.
In summary, erosion process is the displacement of solids usually caused by the water
currents, as well as the down-slope movement of the solids in response to gravity. The
erosion process is able to affect the Earth’s surface and change the direction of water
current. In addition, erosion process is able to smooth out obstacles and keep a state of
smoothed terrain for a long time period. Figure 3.4 shows an illustration of the terrain
smoothed by the erosion of water flow. Erosion rate, an important parameter of the
erosion process, is affected by many factors. The factors consist of the amount and the
intensity of precipitation, the texture of soil, the gradient of slope, etc. However, only two
major factors have been considered to design the capability of erosion process in the WFA.
These factors are the amount of precipitation as well as the texture of soil, including the
hardness of soil and the geographical shape of the surface. How to simulate the
hydrological cycle as well as to integrate the erosion process of water flow in constructing
an optimization algorithm is presented in the next section.
50
(Source: http://newscenter.lbl.gov/feature-stories/2007/04/22/damaged-land-buried-
carbon/)
Figure 3.4 Illustration for Smoothed Terrain by Erosion Process
3.3 General Water Flow Algorithm
In this section, we present a general water flow algorithm for combinatorial optimization
problems. The framework of the WFA can be customized and used to solve various types
of optimization problems.
The WFA mimics the hydrological cycle in meteorology and the erosion phenomenon
in nature, representing a balance between solution exploration and exploitation capabilities
51
in an optimization algorithm, respectively. The proposed algorithm is based on the
simulation of spreading raindrops into many places on the ground, the property of water
flow always moving from higher positions to lower positions, and the erosion capability of
water flow on the ground. The following descriptions further illustrate some of the
relationships between the terms used in the WFA and the concepts often used in other
well-known meta-heuristics, such as tabu search and genetic algorithm:
(1). The hydrological cycle reflects the circulation and conservation of water on the
Earth. It consists of many stages, i.e., evaporation, condensation, transportation,
precipitation, transpiration, groundwater and run-off (Goudie 1993). In the WFA,
we use the circulation of water to regenerate a population of new positions for
drops of water (DOWs) after each cloud is generated. It is similar to the
diversification procedure of restarting an initial solution in tabu search if the
current solution is not improved after a maximum number of iterations.
(2). The precipitation is the transfer of water from the atmosphere to land, usually
called rain (Goudie 1993). When there is rain, an amount of raindrops is
generated and it drops to the ground. In the WFA, the precipitation involves the
number of DOWs generated. It can be considered as generating a population of
solutions in a genetic algorithm.
(3). The concept of erosion is the detachment of soil particles caused by the impact of
water flow or raindrops (Holy 1982). The role of erosion in the WFA is to help
the search to intensify the promising local optimal positions and overcome the
obstacles to find better positions. It is similar to intensification in tabu search.
52
(4). The rate or intensity of erosion in nature depends on many factors, such as the
amount of precipitation and the soil type, etc. (Holy 1982). In the WFA, this
erosion capability is represented by the maximum number of iterations before
erosion process stops eroding at the local optimal position being considered. It is
similar to the concept of the maximum number of iterations without improvement
used in tabu search.
We now describe the general procedure of the WFA. Firstly, a cloud representing an
iteration randomly generates a set of DOWs onto some positions on the ground, which
represent solutions of the optimization problem. Next, due to the gravity force of the Earth
represented by a heuristic algorithm, the DOWs automatically move to local optimal
positions. They are held at these positions until the erosion condition is satisfied before
performing the erosion process. Then, depending on the amount of precipitation, the
falling force of precipitation and soil hardness at the local optimal positions, the erosion
process helps the DOWs overcome the local optimal positions to find better or global ones.
Almost all the ideas of constructing this procedure are novel, except for the idea of water
flow always moving to lower positions, which is similar to the previous works.
A flow chart of the general WFA for solving optimization problems is shown in Figure
3.5, while the pseudocode of the WFA is presented in Figure 3.6. The detailed
descriptions of the erosion condition, erosion capability, erosion process and other
operations of the general WFA for optimization problems are provided in the next
subsections.
53
Algorithm representation Hydrological cycle process
Start
Initialization
Generate a new cloud When it rains, the

DOWs are generated
and they fall to the
Randomly generate a population of ground
MaxPop DOWs with the positions
Exploration phase h
out of E-list
Apply a local search algorithm to

the generated DOWs to find their
local optimal positions
Due to gravity force, the
DOWs automatically move
Evaluate the soil hardness based on the from higher positions to
objective value of the problem and identify the lower positions (local
number of DOWs at the local optimal positions optimal positions) and they
are held there
Update P0-list and UE-list

No
Pick one position in UE-list

Erosion process happens
when the erosion
Perform erosion process condition is satisfied at
Satisfy erosion based on erosion capability the local optimal
No Yes
condition? and topology parameter positions
Exploitation phase h
determined
No
Update UE-list and E-list to

Have all positions in
continue with performing The terrain at eroded
UE-list been considered?
erosion process positions is smoothed
by the erosion process
Yes
Is MaxCloud reached?
Yes
Output the best positions in P0-list
End
Figure 3.5 Flow Chart of the General WFA
3.3.1 Encoding Scheme
For any optimization algorithm, encoding scheme, also known as solution representation,
is the first important step in the implementation of the algorithm. In the WFA, we have
54
Step 1:
Initialize the controlled parameters, i.e., the maximum number of clouds generated
(MaxCloud), the maximum number of DOWs that each cloud is allowed to generate
(MaxPop), the minimum number of DOWs in which the erosion process starts to
perform (MinEro), the maximum iterations that the erosion process will move to next
erosion direction if the position considering is not improved (MaxUIE);
Step 2:
For i = 1 to MaxCloud
Randomly generate a population of MaxPop positions for DOWs;
Determine the number of DOWs at each individual position;
Find the local optimal position for each individual position;
Update the best positions in P0-list;
Update the local optimal positions found and the number of DOWs into UE-list;
For each position in UE-list
If (satisfying the erosion condition) Then
Perform erosion process;
Update UE-list, P0-list, and E-list;
End if
Next each
Next i
Step 3:
Return the best positions found in P0-list.
Figure 3.6 Pseudocode of the General WFA

55
used DOWs to represent solutions of an optimization problem. Then, all the operators of
the WFA are performed on the DOWs. Depending on the structure of a certain problem,
DOWs are appropriately encoded to solve the problem efficiently.
In nature, the positional information of a DOW is shown by its longitude, latitude and
altitude on the ground. In an optimization problem, a DOW is associated with a feasible
solution of the problem. We consider the feasible solution and its objective value as
providing the longitude, latitude and altitude information for the position of the DOW on
the ground. For example, when applied to solving flow shop scheduling problems, a job
permutation will provide the longitude and latitude information for a DOW; while the
objective value, i.e., the completion time of jobs obtained by the corresponding job
permutation, will provide the altitude information for the DOW.
3.3.2 Memory Lists
In the WFA, three memory lists are used to support the search for global optimal positions
of DOWs. Firstly, the best position list, called the P0-list, stores the best optimal positions
found so far. This list is almost used in all optimization algorithms to output the best
solutions obtained by the algorithms. Secondly, the un-eroded position list, called the UE-
list, is used to store the local optimal positions which have not been eroded because of not
satisfying the erosion condition. Its purpose is to record the potential positions for the
subsequent erosion process of the WFA. The reasons for using local optimal positions as
potential positions/regions will be described and discussed later. Thirdly, the eroded
position list, called the E-list, is used to store eroded local optimal positions. The E-list is
inspired by terrain smoothing of erosion process. It plays an important role of preventing
56
DOWs from being regenerated to the eroded positions in the subsequent precipitations. It
would help to reduce the computation time needed by the algorithm.
3.3.3 Exploration Phase
The exploration phase of the WFA is inspired by the hydrological cycle in meteorology
presented in Section 3.1. The number of basic components of the hydrological cycle used
in the WFA depends on the optimization problem to be solved. For example, when solving
single-objective optimization problems, only transportation, precipitation and run-off of
the hydrological cycle are utilized in the exploration phase. However, when solving multi-
objective optimization problems, the exploration phase may include evaporation/
transpiration, condensation and groundwater. Despite the number of basic components
that are completely used, the general procedure does not undergo much change.
The exploration phase can be described as follows: at each cloud, we generate
randomly a set of positions for DOWs with size MaxPop. This is to simulate the spreading
of raindrops into many places on the ground. Based on the problem to be solved, we can
use an efficient constructive method to generate a set of seed positions for DOWs at the
first cloud, which may improve the performance of the WFA. After generating a
population of positions for the DOWs, a steepest descent hill sliding algorithm is used to
search for local optimal positions from these initial positions of the DOWs. In particular,
the hill sliding algorithm searches for the best improved position within the initial
position’s neighbors in terms of objective value. Then, this process is iteratively
performed in the same manner until no other improved position is found. Depending on
the problem, an appropriate neighborhood structure is proposed. The idea of determining
57
local optimal positions of DOWs is inspired by the property of water flow always moving
to lower positions.
In general, the exploration phase of the WFA results in a set of local optimal positions
of DOWs. The local optimal positions and the number of DOWs at these positions are
updated in the UE-list to be considered for performing the erosion process in the next
exploitation phase.
3.3.4 Exploitation Phase
Exploitation phase of the WFA is inspired by the erosion phenomenon of water flow in
nature presented in Section 3.2. The properties of the erosion phenomenon are simulated
in this phase of the WFA to guide DOWs to overcome the local optimal positions and
search for better or optimal positions. In the next subsection, we present the conditions to
perform erosion process for the local optimal positions in the UE-list. Also, the
calculations of erosion capability for the DOWs at the positions satisfying the conditions,
as well as the basic operational mechanism of the erosion process are described.
Depending on the nature of the problem, the corresponding operators of the erosion
process will be implemented to solve the problem efficiently.
3.3.4.1 Erosion Condition and Capability
Although many factors have effects on the erosion process of water flow in nature as
described in Section 3.2, only two major factors affect the erosion process significantly.
They are the hardness of soil and the amount of precipitation. In nature, the hardness of
soil varies according to the land area considered. However, when solving an optimization
58
problem representing the topography of the Earth, the landscape of the problem does not
change over time and space. Thus, we assume that the soil hardness value at every
position in the problem is the same. Thus, we are only concerned about the amount of
precipitation for determining the erosion condition to perform the erosion process in the
WFA. If the amount of precipitation at some local optimal position increases up to
MinEro, the erosion process would happen at the local optimal position.
Also, many factors affect the erosion rate/capability of water flow as presented in
Section 3.2. However, for simplicity, we only consider the capability of erosion process
based on two main factors: the amount of precipitation and its falling force. In
optimization problems, the amount of precipitation is represented by the number of DOWs
at the eroding local optimal position; while its falling force is represented by the objective
value of the corresponding problem at the position. We will next explain why the
objective value at the corresponding position is considered as the falling force of
precipitation.
In the WFA, the altitude information of a DOW is determined by the corresponding
objective value of a solution in an optimization problem. For the problem of minimization,
the DOW with the smaller altitude leads to the better quality solution. Moreover, the
falling force of raindrops depends on the distance between the cloud and the position of
the raindrops in which they fall on the ground. Hence, if we assume that the clouds move
at the same altitude, the impact of falling force at the lower positions is larger than that at
the higher positions. In other words, the erosion capability at the lower positions is larger
than that at the higher positions. This is the reason why the falling force of precipitation is
represented by the objective value at the corresponding position. Figure 3.7 illustrates the
59
erosion capability at two positions with two different altitudes. In Figure 3.7, the erosion
capability at point B is larger than that at point A.
hA hB
Figure 3.7 Illustration for Effect of Altitude on Erosion Capability
With the assumption that clouds are at the same altitude, the falling force of
precipitation to lower local optimal positions will cause erosion more easily than that of
higher ones. Then, we can infer that the erosion capability becomes stronger for local
optimal positions with the larger amount of precipitation and the lower objective values. It
may create a flexible operation scheme for the erosion capability of DOWs in the WFA.
Also, it helps the erosion process focus on exploiting promising regions strongly while
ignoring regions with poor performance. In particular, the relationship between these two
factors and the control parameter of erosion capability, MaxUIE, is expressed as follows:
MaxUIE = ϕ1Q (π * ) + ϕ 2 LB / z , (3.1)
60
where ϕ1 and ϕ2 are parameters representing the effect of precipitation and its falling
force, respectively. Also, Q(π * ) is the number of DOWs at the local optimal position π * ,
LB is a known lower bound, and z is the objective value at the eroding local optimal
position. Here, LB can be any lower bound obtained from the literature, or by solving a
relaxation of the problem in terms of the corresponding objective function.
In addition, we also propose a simple determination of erosion capability for solving
optimization problems with less complex structure. In this process, the erosion capability
does not depend on the amount of precipitation and its falling force. Here, we use a
constant value pre-specified for the erosion capability. When the amount of precipitation
at some local optimal position increases up to MinEro, the erosion process with the pre-
specified erosion capability would happen at the local optimal position. It is helpful for
optimization problems with less complex structure since the computation time may
decrease without affecting the solution quality obtained.
3.3.4.2 Erosion Process
Erosion process will be performed when the erosion condition at some local optimal
position is satisfied. Depending on the structure of the problem, the erosion capability at
the local optimal position may either be constant or determined from equation (3.1). The
strategy of erosion process is based on a topological parameter representing the
geographical surface, and whether an erosion direction is blocked.
The topological parameter Δd is defined as the difference between the objective value
of local optimal position and that of its neighboring position. We first choose the smallest
61
Δd as the erosion direction. If the erosion process for that direction does not improve
after MaxUIE iterations, we say that the direction is blocked. It means that water flow
cannot move in that direction and searching in that direction stops. This is followed by
backtracking, in which we restart the search from the local optimal position using the
direction with the next smallest Δd . If all directions for considering the local optimal
position are blocked, we call that position fully blocked and move it into the E-list, i.e., we
do not consider that position in the subsequent clouds/iterations. Otherwise, if there is a
direction with improvement when compared with the current local optimal position, we
will choose that direction to erode permanently for the local optimal position. Here, the
improvement means that the erosion process finds a new local optimal position, whose
objective value is smaller than that of the current local optimal position. This new local
optimal position is updated in the UE-list to continue with performing the erosion process.
The erosion process simulated in the WFA is thus close to the natural behavior of water
flow in the erosion phenomenon. In nature, if no unexpected change occurs, water flow
always moves by following a fixed stream. The entire erosion process can be presented in
pseudocode in Figure 3.8.
To perform the erosion process efficiently, we have used short-term memory in the
process, called track-list. Since the nature of water flow is to never move upstream, the
track-list is used to prevent the DOWs in the erosion process from moving back. The list
only temporarily records the previously passed positions of the DOWs that are determined
in the corresponding erosion direction. Thus, each erosion direction has a corresponding
track-list. The track-list becomes empty immediately after the erosion direction is
completely considered.
62
Procedure Erosion Process;
Begin
Do loop
Choose an un-eroded direction with the smallest Δd to erode;
Do loop
Apply steepest descent hill-sliding algorithm for the erosion direction chosen;
Until (new optimal position is found or no improvement after MaxUIE iterations)
If (the new optimal position is better than the eroding optimal position) Then
Update it into the UE-list to continue performing the erosion process;
Update the E-list;
End if
Until (all erosion directions are considered)
End.
Figure 3.8 Pseudocode for General Erosion Process of the WFA
The entire process of both exploration and exploitation terminates when the maximum
number of allowed iterations (MaxCloud) is reached.
Finally, we present the reasons why the erosion process only happens at the local
optimal positions with the large number of DOWs. The first reason is because the process
is inspired by the erosion phenomenon of water flow in nature. This phenomenon occurs
when the amount of water in watercourse increases dramatically as described in Section
3.2. The second reason is because the local optimal positions are considered as very
63
promising positions to search for global optimal positions. Intuitively, we can see that
because of using random number generation for the initial population of DOWs, the
probability of the DOWs generated onto every position is the same. Then, the center
positions of “big valleys” may be found more easily than the ones of “small valleys” (see
Figure 3.9). Also, it means that the number of DOWs at the center positions of “big
valleys” may increase faster than that at “small valleys”.
Whole search space 100%
10% 50% 10% 30%
“Small valley”
“Big valley”
“Local optimal”
Figure 3.9 Illustration for “Big Valley” and “Small Valley”
Generally, good properties of a solution may appear many times throughout the
iterations. Hence, local optimal positions found many times in the WFA can be considered
as promising positions or regions to find better local optimal positions by the erosion
process.
64
Chapter 4 WFA for Permutation Flow Shop Scheduling
CHAPTER 4
WFA FOR PERMUTATION FLOW SHOP SCHEDULING
In this chapter, we present a water flow algorithm (WFA) for solving the permutation flow
shop scheduling problem (PFSP), which is one of the well-known scheduling problems in
production. The proposed algorithm has been developed based on the general framework
of the WFA described in Chapter 3. The exploration phase of this algorithm is inspired by
the following basic components of the hydrological cycle, i.e., transportation, precipitation
and run-off; while the exploitation phase of this algorithm has used a pre-specified
constant value for determining the erosion capability of erosion process.
The structure of this chapter is organized as follows. In Section 4.1, we introduce the
PFSP and briefly describe a literature review of solution methods for the scheduling
problem. Then, the formulation of the PFSP is presented in Section 4.2. The customization
of the WFA for solving this scheduling problem is described in Section 4.3. The
computational results and comparisons among the WFA and other algorithms are shown
in Section 4.4. Finally, some conclusions are presented in Section 4.5.
65
4.1 Introduction
In production, the PFSP is one of the well-known scheduling problems in which n jobs
have to be processed by m machines with the same order of jobs on machines. The most
common objective of the scheduling problem is to minimize the completion time of jobs,
also known as makespan (Cmax), by specifying the sequence of jobs. In addition to this
objective, there are other desired measures of performance for this scheduling problem,
such as minimizing total flow time of jobs or minimizing total tardiness time of jobs.
Generally, the makespan criterion is often used because the minimization of makespan
ensures that production gets a high throughput (Pinedo, 2002).
The PFSP with makespan minimization has been proven to be an NP-hard problem
when the number of machines m > 3 (Rinnooy Kan, 1976), and is thus considered a
challenging optimization problem. Many research works in the literature have addressed
this scheduling problem, and they also proposed various optimization methods to solve it.
We can classify the methods according to two approaches: exact and heuristic algorithms.
Exact techniques are computationally effective for problems with small size (often less
than 20 jobs). However, for problems with large size, these techniques are
computationally intensive. On the other hand, as heuristic algorithms often find a good
solution rapidly, it is efficient to use them to solve problems with large size.
The heuristic algorithms can be classified into three major categories: constructive
algorithms, improvement algorithms and meta-heuristics. Some constructive algorithms
for the PFSP include Johnson (1954), Nawaz et al. (1983) and Koulamas (1998). Among
these algorithms, the algorithm by Nawaz et al. (1983) (NEH) is one of the effective
66
polynomial-time heuristic algorithms for the PFSP. Unlike constructive algorithms,
improvement algorithms start from an initial solution, and try to improve it through some
iterative procedures to obtain better solutions. Normally, the search for better solutions is
based on a predetermined neighborhood structure. Taillard (1990), Ho and Chang (1991),
and Suliman (2000) provide examples of improvement algorithms. Applying these
algorithms for solving the PFSP has led to some success, but the quality of solutions
obtained is not good for certain problems with large size. Recently, meta-heuristic
algorithms, which are inspired by the behavior of natural systems, have been extensively
developed and successfully applied to solve the PFSP. Some examples of the meta-
heuristic algorithms are ant colony optimization (Rajendran and Ziegler, 2004), particle
swarm optimization (Tasgetiren et al., 2007), and genetic algorithm (Nagano et al., 2008).
Hence, we propose a WFA for solving the PFSP. The proposed algorithm is developed
based on the general WFA presented in Chapter 3. In this algorithm, the exploration phase
simulates the basic components of the hydrological cycle, such as transportation,
precipitation and run-off. The exploitation phase uses a pre-specified constant value for
determining the erosion capability. Several well-known benchmark problem sets are used
to evaluate the performance of this algorithm. The best known values of the benchmark
problems in the literature are used to compare with the results obtained by the WFA. In
addition, we also used these results to compare with that of other efficient algorithms,
such as an adaptive learning approach combined with the NEH constructive heuristic
(NEH-ALA) proposed by Agarwal et al. (2006), or a constructive GA combined with
local search (CGALS) described by Nagano et al. (2008), etc.
67
4.2 Formulation of the PFSP
In this section, we describe and present the formulation of the PFSP. In this scheduling
problem, n jobs have to be processed by m machines with the same order of jobs on
machines. Let pji denote the processing time of job j (j = 1, 2,..., n) on machine i (i = 1,
2,..., m), and C(σj, m) denote the completion time of job σj on machine m. Then, the PFSP
aims to search for the best permutation of jobs processed through all machines.
Given the job permutation π = (σ 1 , σ 2 ,..., σ n ) , the completion time of the n-job m-
machine problem is calculated as follows:
C (σ 1 ,1) = pσ1 ,1 , (4.1)
C (σ j ,1) = C (σ j −1 ,1) + pσ j ,1 , j = 2,..., n, (4.2)
C (σ 1 , i ) = C (σ 1 , i − 1) + pσ1 ,i , i = 2,..., m, (4.3)
{ }
C (σ j , i ) = max C (σ j −1 , i ), C (σ j , i − 1) + pσ j ,i , j = 2,..., n; i = 2,..., m. (4.4)
Then, the makespan can be defined as:
Cmax (π ) = C (σ n , m). (4.5)
Hence, the PFSP with the makespan criterion aims to search for an optimal job
permutation π* in the set of all possible job permutations П, such that:
Cmax (π * ) ≤ C (σ n , m) ∀π ∈ Π. (4.6)
68
4.3 WFA for the PFSP
In this section, we present how the proposed WFA can be applied to solve the PFSP. The
basic components of the WFA described are the encoding scheme, memory lists,
exploration phase, and exploitation phase.
In the PFSP, a DOW is associated with a job permutation. We consider the job
permutation and its objective value as providing the longitude, latitude and altitude
information for the position of DOW on the ground. Given a job permutation
π = (σ 1 , σ 2 ,..., σ n ) , we define:
longitude(π ) = (σ 1 ,..., σ ⎢ n ⎥ ) (4.7)

⎢⎣ 2 ⎥⎦
latitude(π ) = (σ ⎢ n ⎥ ,..., σ n ) , (4.8)

⎢ 2 ⎥ +1
⎣ ⎦
where ⎢⎣ x ⎥⎦ is the largest integer less than or equal to x. The altitude of DOW is defined as
the corresponding makespan value of the given job permutation. The objective value is
calculated by equation (4.5). Figure 4.1 shows an illustrative example of a DOW and its
positional vector components for the PFSP with 6 jobs.
Job permutation Objective value

(makespan)
DOW 1 2 5 3 4 6 27
Altitude
Longitude Latitude
Figure 4.1 An Example of Solution Representation in the WFA for the PFSP
69
4.3.2 Memory Lists
In the WFA, we use three memory lists to support the search for global optimal positions.
Firstly, we use one memory list to store the best positions found so far, called the P0-list.
For the PFSP, the P0-list stores the job permutations with minimum makespan. This list is
updated when a new local optimal job permutation is found. Secondly, the un-eroded
position list, called the UE-list, is used to store local optimal job permutations which have
not been eroded due to not satisfying erosion condition. Its purpose is to record potential
job permutations for the subsequent erosion process. Thirdly, the eroded position list,
called the E-list, is used to store eroded local optimal job permutations. This list aims to
prevent next clouds from regenerating DOWs to the eroded job permutations. It would
help to save the computation time needed by the algorithm. Both UE-list and E-list are
updated after performing the erosion process. In addition, the UE-list is updated when a
new local optimal job permutation is found in the erosion process.
At each cloud (or iteration), after randomly generating a population of initial job
permutations for DOWs, the steepest descent hill sliding algorithm is used to search for
local optimal job permutations from the initial job permutations. In particular, the hill
sliding algorithm searches for the best improved job permutation within the initial job
permutation’s neighbors in terms of makespan value. Then, this process is performed
iteratively in the same manner until no other improved job permutation is found. In this
algorithm, a perturbation scheme based on a systematic pair-wise job exchange, a variant
of 2-opt algorithm, is used to construct the neighboring job permutations. In general, the
70
exploration phase in the WFA for the PFSP will result in a set of local optimal job
permutations. They will be updated in the UE-list to be considered for performing the
erosion process in the exploitation phase.
In the WFA for the PFSP, we perform the erosion process based only on the amount of
precipitation. If the amount of precipitation at some local optimal job permutation in the
UE-list increases up to MinEro, the erosion process will happen at the local optimal job
permutation.
In the current WFA, we consider the capability of erosion process based on a pre-
specified constant value. It means that the erosion capacity is independent of two factors,
i.e., the amount of precipitation and its falling force. In particular, when the erosion
process happens, the erosion capacity is determined by a constant MaxUIE, the maximum
number of iterations for the erosion process to move to the next erosion direction if the job
permutation is not improved.
In the erosion process of the WFA for the PFSP, the topological parameter Δd
representing geographical surface is calculated as the difference between the makespan
value of the local optimum job permutation and that of its neighboring job permutations.
Here, we still use the same neighborhood structure as in the exploration phase. Firstly, we
choose the smallest Δd to be the erosion direction. If the erosion process for that direction
does not improve after MaxUIE iterations, we say that the direction is blocked. In other
words, water flow cannot move in that direction and searching in that direction stops. This
is followed by backtracking, in which we restart the search from the local optimal job
71
permutation using the direction with the next smallest Δd . If all directions for considering
the local optimal job permutation are blocked, we call that job permutation fully blocked
and move it into the E-list. Then, we do not consider that job permutation in the following
clouds or iterations. Otherwise, if there is a direction with improvement when compared
with the current local optimal job permutation, we will choose that direction to erode
permanently for the local optimal job permutation. The improved local optimal job
permutation is updated in the UE-list to continue with performing the erosion process.
The entire process of both exploration and exploitation terminates when the maximum
number of allowed clouds or iterations (MaxCloud) is reached.
A flow chart of the WFA for solving the PFSP is shown in Figure 4.2. This flow chart
is a detailed extension of the flow chart presented in the Chapter 3.
4.3.5 A Numerical Example for Erosion Mechanism
In this section, a numerical example is used to illustrate for implementation of the erosion
mechanism in the exploitation phase of WFA. Data of the PFSP instance consists of:
The number of jobs: n = 5, and the number of machines: m = 5.
The processing time matrix of jobs on machines:
⎡2 3 4 9 7⎤
⎢8 2 6 5 1 ⎥⎥
⎢
Pij = ⎢ 4 2 7 8 5⎥ , for i = 1,.., n and j = 1,.., m.
⎢ ⎥
⎢2 4 5 6 3⎥
⎢⎣ 2 1 3 6 5 ⎥⎦
72
Start
Initialization
Generate a new cloud
Generate a new DOW with a

No
random job permutation
When it rains, the
Yes DOWs are generated
and they fall to the
j = j +1 Belong to E-list? ground
Exploration phase h
No
No
j = MaxPop?
Yes
Apply a local search algorithm to

the generated DOWs to find their
from higher positions to
Identify the number of DOWs at the
No lower positions (local
optimal positions) and they
are held there
Pick one job permutation

in UE-list Erosion process happens
when the erosion
Perform erosion process condition is satisfied at
Satisfy erosion based on erosion capability the local optimal
No Yes
condition? and topology parameter positions
determined
No
Update UE-list and E-list to

Have all job permutations The terrain at eroded
continue with performing
in UE-list been considered? positions is smoothed
erosion process
by the erosion process
Yes
Yes
Output the best job permutations in P0-list
End
Figure 4.2 Flow Chart of the WFA for the PFSP
73
For the WFA, the initialization of the parameter values is given as follows:
MaxCloud = 2, MaxPop =7, MinEro = 3, MaxUIE = 3.
In this illustration, we do not mention to the exploration phase of the WFA, thus we
assume that after doing the exploration phase, UE-list contains a set of two local optimal
job permutations as follows:
Table 4.1 UE-list of Local Optimal Job Permutations
UE-list Job permutations Cmax The number of DOWs
#1 [2, 1, 4, 3, 5] 47 5
#2 [1, 4, 5, 2, 3] 46 2
We see that only the first local optimal job permutation in the UE-list has satisfied the
erosion condition (the number of DOWs at the local optimal job permutation is greater
than or equals to MinEro). Hence, the erosion process will be performed at this job
permutation. Firstly, the neighboring job permutations of this local optimal job
permutation will be determined and ranked in the descend order with the objective
function value Cmax. In the WFA, only the neighbors of the eroding local optimal job
permutation are ranked to consider one after another. For the neighboring job
permutations belonging to the search path of the direction chosen, we do not need to rank
the job permutations. As a result, ranking neighbors of the eroding local optimal job
permutation [2, 1, 4, 3, 5] is shown in Table 4.2.
According to the erosion process of water flow described in the WFA, the smallest
topology (or neighboring job permutation) around the eroding local optimal job
permutation will be chosen to perform erosion first. Here, we choose the erosion direction
74
with respect to the neighboring job permutation [1, 2, 4, 3, 5] to perform erosion since its
objective value Cmax = 47 is smallest. In the case that there are many neighbors with the
same best objective value, we can choose randomly one of them to perform erosion.
Table 4.2 Possible Erosion Directions at the Local Optimal Job Permutation
Direction Neighbors Cmax
1 [1, 2, 4, 3, 5] 47
2 [2, 1, 4, 5, 3] 47
3 [5, 1, 4, 3, 2] 48
4 [2, 3, 4, 1, 5] 48
5 [3, 1, 4, 2, 5] 49
6 [2, 5, 4, 3, 1] 49
7 [4, 1, 2, 3, 5] 50
8 [2, 1, 5, 3, 4] 50
9 [2, 4, 1, 3, 5] 51
10 [2, 1, 3, 4, 5] 51
Next, we update the eroding local optimal job permutation [2, 1, 4, 3, 5] into the track-
list to prevent the DOWs move back this position. For the erosion direction chosen, if we
cannot find a better job permutation after MaxUIE = 3 iterations, we will consider the
erosion direction with the next smallest objective value of the neighboring job
permutations, i.e., job permutation [2, 1, 4, 5, 3] in Table 4.2. Then, the track-list will be
refreshed, and the first erosion direction is known as blocked direction. This procedure is
iterated until any improvement job permutation (i.e., the objective value of the job
permutation just found is better than that of the eroding local optimal job permutation [2, 1,
75
4, 3, 5]) is found in a direction. If an improvement job permutation is found, the steepest
hill-sliding algorithm is performed to obtain better new local optimal job permutation
from the improvement job permutation just found. Otherwise, if there is no improvement
through all erosion directions, the eroding local optimal job permutation is known as fully
blocked position and recorded into the E-list. Then, the erosion process will consider to
the next local optimal job permutation in the UE-list.
In this example, since we may find an improvement job permutation with respect to
the first erosion direction [1, 2, 4, 3, 5], this direction will be considered as the permanent
erosion direction of the DOWs at the local optimal job permutation [2, 1, 4, 3, 5]. The
steps to determine the improvement job permutation are shown in Table 4.3.
Table 4.3 Steps of Finding Improvement Job Permutation
Step 2 Step 3
No. Neighbors Cmax Remarks Neighbors Cmax Remarks
1 [2, 1, 4, 3, 5] 47 in track-list [4, 1, 2, 3, 5] 50
2 [4, 2, 1, 3, 5] 50 [2, 4, 1, 3, 5] 51
3 [3, 2, 4, 1, 5] 48 [3, 4, 2, 1, 5] 46
4 [5, 2, 4, 3, 1] 49 [5, 4, 2, 3, 1] 49
5 [1, 4, 2, 3, 5] 47 chosen [1, 2, 4, 3, 5] 47 in track-list
6 [1, 3, 4, 2, 5] 49 [1, 3, 2, 4, 5] 50
7 [1, 5, 4, 3, 2] 48 [1, 5, 2, 3, 4] 50
8 [1, 2, 3, 4, 5] 50 [1, 4, 3, 2, 5] 48
9 [1, 2, 5, 3, 4] 50 [1, 4, 5, 3, 2] 46 chosen
10 [1, 2, 4, 5, 3] 47 [1, 4, 2, 5, 3] 46
76
In particular, we will determine the neighboring job permutations and choose the best
job permutation among the neighbors without belonging to the track-list in order to
continue searching improvement job permutation. Through steps 2 and 3 in Table 4.3, we
find three improvement job permutations. Here, we choose randomly the job permutation
[1, 4, 5, 3, 2] to start searching better new local optimal job permutation. When applying
the steepest hill-sliding algorithm, we find the local optimal job permutation [1, 4, 3, 5, 2]
whose objective value Cmax = 45. Then, we update it into the UE-list to continue
performing the erosion process as shown in Table 4.4, and also update the eroded local
optimal job permutation [2, 1, 4, 3, 5] into the E-list.
Table 4.4 Updating the UE-list
UE-list Job permutations Cmax The number of DOWs
#1 [1, 4, 3, 5, 2] 45 5
#2 [1, 4, 5, 2, 3] 46 2
This erosion process only terminates when fully blocked state is reached. Then, the
next job permutation in the UE-list will be considered to perform the erosion process. If
all job permutations in the UE-list are considered, next cloud will be generated to explore
new search space. In this example, the entire process of both exploration and exploitation
terminates when MaxCloud = 2 is reached. Then, the best job permutation found in the P0-
list is displayed as the output of the WFA.
As for the case that the local optimal job permutation is considered as a fully blocked
position, we can intuitively observe as shown in Figure 4.3. Then, we cannot find any
improvement job permutation through all erosion directions.
77
MaxUIE = 3
Old “local optimal”

New “local optimal”
Figure 4.3 The Case of Fully Blocked Position
4.4 Computational Experiments and Comparisons
4.4.1 Benchmark Problem Sets
To evaluate the performance of WFA for the PFSP, we performed experiments for four
benchmark problem sets, i.e., that of Carlier, Heller, Reeves and Taillard. These are well-
known problem sets in the PFSP taken from the OR Library. In this chapter, a total of 121
instances comprising of 8 instances of Carlier, 2 instances of Heller, 21 instances of
Reeves (odd number instances) and 90 instances of Taillard are used. The best known
upper bounds of these problem sets from the literature were used to compare with the
results obtained by the WFA. Also, the results obtained by Agarwal et al. (2006) and
meta-heuristic algorithms in Nagano et al. (2008) were used to compare with our results.
Agarwal et al. (2006) performed empirically to select the best parameters, but they did not
mention the number of runs of their algorithm for solving the instances; while Nagano et
al. (2008) used the design-of-experiment method to tune their parameters and ran 5
independent replicates for each instance.

78
4.4.2 Platform and Parameters
The WFA has been coded using Visual Basic 6.0, and all experiments were performed on
an Intel Centrino Duo 1.60 GHz CPU with 1.5 GB of RAM running on Windows XP
Operating System. The computational complexity of the WFA for the PFSP is determined
based on the neighborhood structure used and the erosion process of this algorithm. In
particular, the WFA used 2-opt neighborhood structure, and the worst possibility of the
erosion process is to find through all n directions. Thus, the computational complexity of
the WFA may be estimated to be O(n3).
The choice of reasonable parameters for the WFA was determined by design-of-
experiment methods and the values are shown in Table 4.5. With the parameter sets, 5
independent replicates were used for each instance. The average or best results obtained
were used to evaluate the performance of the WFA and to compare with other algorithms.
4.4.3 Performance Measure
For comparison of objective values, we used the following average relative percentage
increase in objective value:
K
⎛ Heuristicsol (i ) − BN sol ⎞
∑⎜ BN sol
⎟
Δ= ⎝ ⎠ ×100
i =1
(4.9)
K
where Heuristicsol(i) and BNsol denote the makespan value obtained by the algorithm for the
ith replicate and the best known value from the literature, respectively; and K denotes the
number of replicates. If we only use the replicate with the best objective value to compute
the above Δ , then we denote the relative percentage increase in objective value by Δ Best .
79
Table 4.5 Parameter Sets for Benchmark Problem Sets
Value of Parameters
Benchmark
Problem Sets
MaxCloud MaxPop MinEro MaxUIE
Carlier 5 5 2 5
Heller 5 10 2 5
20J-05M 10 10 2 5
20J-10M 20 20 2 10
20J-15M 50 20 2 10
Reeves
30J-10M 20 20 2 10
30J-15M 30 10 2 10
50J-10M 10 10 2 10
75J-20M 10 10 2 10
20J-05M 5 10 2 10
20J-10M 20 20 2 10
20J-20M 30 20 2 10
50J-05M 20 10 2 10
Taillard
50J-10M 10 10 2 10
50J-20M 15 10 2 10
100J-05M 5 10 2 5
100J-10M 5 10 2 5
100J-20M 10 10 2 5
The best known values from the literature are used to be the reference points for the
evaluation and comparison of the WFA and other algorithms in this chapter. These values
may be the objective values of the optimal solutions of the benchmark instances used, or
those of the best solutions found by some algorithm so far. Hence, the comparison results
obtained may be negative values if new algorithm finds a better solution. Since we have
used the best known values published and summarized in the leading journal papers, the
comparison results obtain a high reliability.
80
4.4.4 Computational Results
From Table 4.6, it can be seen that the WFA outperforms the NEH-ALA although the
CPU running time of the WFA in some instances is larger. In particular, 55 out of 90
Taillard’s instances solved by the WFA have better results than that solved by the NEH-
ALA. In the rest of the instances, the NEH-ALA only performs better than the WFA in 11
instances, which are mainly 100job-5machine instances. For the instances with smaller
size, i.e., 20 jobs, we obtained results similar to that of the NEH-ALA, but the CPU
running time is smaller. Similar results are also obtained by the WFA when solving
Carlier’s, Heller’s, and Reeves’ benchmark problem sets. The WFA obtained the best
known upper bound for 10 out of 21 instances of Reeves, while the NEH-ALA only
obtained the best known upper bound for 2 instances. Moreover, the WFA has the
average relative percentage increase of 0.77%, while it is 1.51% for the NEH-ALA. For
the Heller benchmark problem sets, the best known results for these instances are provided
in Agarwal et al. (2006), which showed the best makespan value of 136 for the 20job-
10machine instance and 516 for the 100job-10machine instance. When applying the WFA
to solve the 20job-10machine problem, a new upper bound value of 135 was found.
Also, we compared the results obtained by the WFA with other efficient meta-
heuristic algorithms in Nagano et al. (2008) in Table 4.7 and Figure 4.4. From the table
and figure, it can be seen that the WFA dominates the performance of the simulated
annealing algorithm of Osman and Potts (SAOP), tabu search of Widmer and Hertz
(SPIRIT), differential evolutionary method of Onwubolu and Davendra (DE), NEH-ALA
of Agarwal et al., and constructive genetic algorithm of Nagano et al. (CGA). When
compared with the robust genetic algorithm of Ruiz et al. (GARMA) and CGALS of
81
Table 4.6 Comparison Results between the WFA and the NEH-ALA
No. of WFA NEH-ALA

Benchmarks
instances Δ Best (%) Time (s) Δ Best (%) Time (s)
Car11J-5M 1 0 0.17 0 6.6
Car13J-4M 1 0 0.22 0 2.6
Car12J-5M 1 0 0.25 0 8.6
Car14J-4M 1 0 0.23 0 10.7
Car10J-6M 1 0 0.66 0 6.0
Car8J-9M 1 0 0.25 0 6.4
Car7J-7M 1 0 0.11 0 2.8
Car8J-8M 1 0 0.19 0 4.5
Average 0 0
Hel20J-10M 1 -0.74 23 0 94
Hel100J-10M 1 0 2752 0 2565
Average -0.37 0
Rec20J-5M 3 0 81 0.11 16.5
Rec20J-10M 3 0 218 0.41 34.6
Rec20J-15M 3 0 3732 0.78 49.6
Rec30J-10M 3 0.23 5024 1.54 94.9
Rec30J-15M 3 0.34 4661 1.78 144.3
Rec50J-10M 3 0.27 5223 1.13 423.7
Rec75J-20M 3 4.56 5714 4.81 2876
Average 0.77 1.51
Tai20J-5M 10 0 30 0 34.2
Tai20J-10M 10 0 150 0.39 74.1
Tai20J-20M 10 0 854 0.38 152.3
Tai50J-5M 10 0 2173 -0.01 487.8
Tai50J-10M 10 1.42 2837 2.58 983
Tai50J-20M 10 4.30 3877 5.29 2750.3
Tai100J-5M 10 0.29 6842 0.11 3736
Tai100J-10M 10 0.98 7538 1.03 3837
Tai100J-20M 10 4.22 6836 4.97 8853.5
Average 1.25 1.64
82
Table 4.7 Average Relative Percentage Increase over the Best Known Solution for
Taillard’s Benchmarks Obtained by Meta-heuristic Algorithms
Instances WFA SAOP SPIRIT GARMA DE NEH-ALA CGA CGALS
20x5 0.00 0.93 4.01 0.29 3.98 1.38 1.33 0.05

20x10 0.00 2.59 5.65 0.63 5.86 2.22 2.42 0.19
20x20 0.00 2.33 4.84 0.41 4.53 1.78 2.08 0.08
50x5 0.00 0.48 1.90 0.06 4.28 0.46 0.32 0.02
50x10 1.49 3.34 5.84 1.76 11.48 3.44 3.72 1.65
50x20 4.35 4.47 7.46 2.62 14.73 4.66 4.98 2.67
100x5 0.37 0.28 0.93 0.07 4.27 0.46 0.21 0.02
100x10 1.11 1.53 2.96 0.60 10.42 1.54 1.46 0.60
100x20 4.38 4.68 6.26 2.52 16.08 4.49 4.52 2.84
Average 1.30 2.29 4.43 1.00 8.40 2.27 2.34 0.90
Figure 4.4 Means Plot for Comparing the WFA and Meta-heuristic Algorithms
83
Nagano et al., the WFA performs best on the small and medium instances (from 20job-
5machine to 50job-10machine problems), but for the large instances it performs closer to
the average of the group of algorithms, and the SAOP, NEH-ALA and CGA methods than
CGALS and GARMA.
4.5 Conclusions
In this chapter, we developed the WFA for solving the PFSP. In the algorithm, only three
basic components of the hydrological cycle, i.e., transportation, precipitation, and run-off,
were used in the exploration phase. For the exploitation phase, the erosion process with a
constant erosion capability was used. The WFA is tested and compared with other meta-
heuristic algorithms on the PFSP benchmark problem sets taken from the literature. The
results show that the algorithm is able to obtain good solutions to the benchmark problem
sets. Also, a new best known solution of a Heller benchmark instance is found by the
WFA.
84
Chapter 5 WFA for Flexible Flow Shop Scheduling
CHAPTER 5
WFA FOR FLEXIBLE FLOW SHOP SCHEDULING
In this chapter, we construct the WFA for solving the flexible flow shop scheduling
problem (FFSP), which is one of the NP-hard scheduling problems often encountered in
production environment. Here, we investigate the FFSP with limited or unlimited
intermediate buffers. A common objective of this problem is to find a production schedule
that minimizes the completion time of jobs. Other objectives that we have also considered
are minimizing the total weighted flow time of jobs and minimizing the total weighted
tardiness time of jobs.
While the proposed WFA is inspired by the hydrological cycle in meteorology and the
erosion phenomenon in nature, we have also combined the amount of precipitation and its
falling force to form a flexible erosion capability in this algorithm. This helps the erosion
process of the WFA to focus on exploiting promising regions strongly. Moreover, to
initiate the algorithm, we have used a constructive procedure to obtain a seed job
permutation. We have also proposed an improvement procedure for constructing a
complete schedule from a job permutation that represents the sequence of jobs in the first
85
stage of the FFSP. To evaluate the WFA for this scheduling problem, we have used
benchmark instances taken from the literature and randomly generated instances of the
problem. The computational results demonstrate the efficacy of this algorithm. Also, we
have obtained several improved solutions for the benchmark instances using the proposed
algorithm. We further illustrate the algorithm’s capability for solving problems in practical
applications by applying it to a maltose syrup production problem.
The structure of this chapter is organized as follows. In Section 5.1, we introduce the
FFSP and its important applications in modern production. A brief literature review of
problem classification and solution methods for the problem is also provided in this
section. Next, the details of the FFSP with limited or unlimited buffers are described in
Section 5.2. A full description of the proposed WFA for the FFSP with intermediate
buffers is provided in Section 5.3. An example of the FFSP with limited buffers in maltose
syrup production is presented in Section 5.4. Computational experiments and comparisons
based on benchmark instances of the FFSP with limited or unlimited buffers, randomly
generated instances, and the maltose syrup production problem are shown in Section 5.5.
Finally, some conclusions of this chapter are presented in Section 5.6.
5.1 Introduction
In production, flexible flow shop scheduling is one of the well-known NP-hard scheduling
problems. In this chapter, we focus on the FFSP with limited intermediate buffers. The
problem involves a set of jobs processed through several consecutive operation stages
with parallel identical machines in each stage, and there are limited intermediate buffers
between consecutive stages. The primary objective of this problem is to find a production
86
schedule to minimize the completion time of jobs, known as makespan (Cmax). There are
also other important objectives of this problem, such as to minimize the total weighted
flow time of jobs and to minimize the total weighted tardiness time of jobs. These
objectives help to achieve a high throughput for production. The FFSP with limited
buffers has been encountered in both traditional and modern manufacturing systems, such
as the electrics manufacturing (Wittrock, 1988), the paper production industry (Sherali et
al., 1990), the building industry (Grabowski and Pempera, 2000), the printed circuit board
assembly line in electronics industry (Sawik, 2001), and the continuous casting-hot charge
rolling production in steel industry (Tang and Xuan, 2006). It is also a special case of the
FFSP with unlimited intermediate buffers (Wardono, 2001).
Although many researchers have developed optimization techniques for solving a
variety of flow shop scheduling problems, only a few of them dealt with the FFSP with
limited buffers. A detailed review of the development of scheduling algorithms, as well as
the classification of FFSP, is given in Quadt and Kuhn (2007) and Ribas et al. (2010).
Quadt and Kuhn (2007) proposed the taxonomy of l-stage flexible flow line scheduling
procedures. The taxonomy focuses mainly on heuristic procedures that are split into
holistic and decomposition approaches. The decomposition approaches are further
classified into stage-oriented decomposition, job-oriented decomposition, and problem-
oriented decomposition approaches. Ribas et al. (2010) introduced a new classification for
published papers on FFSP. From a production perspective, this classification is based on
machine and job characteristics, relevant constraints, and objective functions. From a
solution perspective, this classification involves grouping the references into exact
approaches, heuristic procedures, hybrid approaches, and simulation/decision support
87
system procedures. Among these approaches, heuristic and hybrid approaches have
recently received considerable attention by many researchers (Ruiz and Vazquez-
Rodriguez, 2010). Wardono and Fathi (2004) developed a tabu search algorithm together
with a procedure for constructing a complete schedule to solve the FFSP with limited
buffers that minimizes job completion time. This algorithm is based on the stage-oriented
decomposition approach. Tavakkoli-Moghaddam et al. (2009) proposed a memetic
algorithm (MA), which involves a combination of genetic algorithm and nested variable
neighborhood search, for solving the flexible flow line scheduling problem with processor
blocking and without intermediate buffers. The MA obtained some promising results and
it can be considered as an efficient algorithm for solving the FFSP with no available
buffer space. However, there is no formal procedure for constructing a complete schedule
in the MA. Also, the quality of solutions obtained by these algorithms may not be good for
problems with large size.
In this chapter, we construct a water flow algorithm for solving the FFSP with limited
or unlimited intermediate buffers. To evaluate the performance of the WFA, we have
tested it on many instances of FFSP with intermediate buffers from the literature.
Moreover, we also compare the performance of the WFA with that of the tabu search
algorithm of Wardono and Fathi (2004) and MA of Tavakkoli-Moghaddam et al. (2009).
In addition, we introduce a problem encountered in the maltose syrup production industry
and use it to evaluate the efficiency of the WFA for solving problems arising in practical
applications.
88
5.2 FFSP with Intermediate Buffers
The FFSP with limited intermediate buffers is an NP-hard combinatorial optimization
problem (Wardono and Fathi, 2004) that may be formulated as follows. A set of N jobs is
processed on S consecutive production stages with ml parallel identical machines in each
stage l (l = 1,…,S). There are limited buffers between these consecutive stages (refer to
Figure 5.1) and we denote Bl as the capacity of the buffer at stage l. In this problem, jobs
have to be processed successively through all S stages. One machine in each stage can
only process one job at a time, and each job can only be processed on at most one machine
in each stage at the same time. In addition, each job is processed without preemption on
one machine in each stage. Moreover, a job can skip one or more stages but is unable to
go back to a previous stage. Here, we only consider processing time, weight, and due date
of jobs in this problem. Thus, we do not consider other characteristics, such as the
breakdown time of machines and set-up time of jobs in this model.
Buffer 1 Buffer S-1
1 1 1 1
2 2 2 2
N jobs
m1 m2 mS-1 mS
Stage 1 Stage 2 Stage S-1 Stage S

Figure 5.1 The Schematic of FFSP with Limited Intermediate Buffers
For the FFSP with limited buffers, we can divide it into two cases. The first one is the
case of no available buffer space for completed jobs between consecutive stages. It means
89
that after job j is finished by machine i in stage l, if there is no idle machine in the
subsequent stage l+1, then job j must wait on machine i in stage l until there is at least an
idle machine in stage l+1 to start processing job j in this stage. The job j is known as a
blocked job, and the corresponding machine i is known as a blocked machine. This case is
often encountered in maltose syrup production for the confectionery and sugar industries
(Hull, 2010), and the continuous casting-hot charge rolling production for the steel
industry (Tang and Xuan, 2006). The second case is the FFSP with finite buffer capacities
between consecutive stages. In this second case, after a job j is completed on machine i in
stage l, it can either be processed on an available machine in stage l+1 or wait in a
following buffer if there is no available machine in stage l+1. If there is no available
capacity in the following buffer, job j remains on blocked machine i in stage l until there is
available capacity in the following buffer or an available machine in stage l+1. In addition
to the two cases mentioned above, we also consider the FFSP with unlimited buffer
capacities between consecutive stages. In the FFSP with unlimited intermediate buffers,
after a job j is completed by machine i in stage l, the machine i is immediately available to
process awaiting jobs in the buffers of previous stage l-1 even though there is no available
machine in stage l+1.
A Gantt chart illustration of the FFSP with intermediate buffers is shown in Figure 5.2.
The data for this example is displayed in Table 5.1. In this example, we consider the FFSP
with 3 stages in which there is a machine in stage 1 and stage 3, and 3 machines in stage 2.
5 jobs are processed through these stages, in the order of job 1 to job 5 at stage 1. In the
first case of the FFSP with limited buffers (Figure 5.2a), when job 4 is completed on
machine 1 in stage 1, it has to wait in machine 1 until there is at least an idle machine in
90
stage 2. Thus, machine 1 in stage 1 is blocked. Hence, job 5 can only be processed on
machine 1 in stage 1 from time 5 when machine 1 in stage 2 is available. In the second
case of the FFSP with limited buffers (Figure 5.2b), there are buffers with capacity 1 after
stage 1 and stage 2, and they are denoted by B1 and B2 respectively. With B1, we can see
that machine 1 in stage 1 is not blocked from time 4 to 5 as in the first case. This is
because after job 4 is completed on machine 1 in stage 1, it is delivered to buffer 1. Then,
job 5 can be processed on machine 1 in stage 1 from time 4. Also, in Figure 5.2b, job 5 is
blocked on machine 2 in stage 2 from time 10 to 11, since buffer 2 is full and there is no
available machine in stage 3. However, for the FFSP with unlimited buffers (Figure 5.2c),
job 5 is not blocked on machine 2 in stage 2 from time 10 to 11 because it is delivered to
buffer 2 with infinite capacity. This shows a typical difference between the FFSP with
limited buffers and unlimited buffers. With infinite buffer capacities, blocked jobs or
machines do not exist in the problem.
To solve the FFSP with intermediate buffers, we first construct a WFA for solving the
FFSP with no available buffer space. Then, we apply the WFA with a modification of the
input data to solve the FFSP with finite buffer capacities between consecutive stages. The
aim of this modification is to convert the FFSP with finite buffers to one with no available
buffer space (McCormick et al., 1989). To do so, we consider a buffer with a capacity of C
as a stage with C parallel identical machines in which the processing time of jobs through
the machines is zero. We also assume that every job must be processed through all stages,
including the buffer stages. We illustrate the conversion process by an example in Table
5.1. As for the FFSP with unlimited buffers, we also perform a modification procedure of
constructing a complete schedule that is appropriate with the structure of the problem.
91
Stage Machine
1 1 1 2 3 4 b 5 b
2 1 1 4 b
2 2 5 b
3 3 b
3 1 1 2 4 5 3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
time, t
(a) The FFSP with no available buffer space between consecutive stages
b Blocked Job/Machine
Stage Machine Buffer
Job at
Buffer
1 1 1 2 3 4 5
B1= 1 4 5
2 1 1 4
2 2 5 b
3 3
B2 = 1 4 5 3
3 1 1 2 4 5 3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
time, t
(b) The FFSP with finite buffer capacities between consecutive stages
Stage Machine Buffer

1 1 1 2 3 4 5
B1= ∞ 4 5
2 1 1 4
2 2 5
3 3
B2 = ∞ 4 3
5
3 1 1 2 4 5 3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
time, t
(c) The FFSP with unlimited buffer capacities between consecutive stages
Figure 5.2 A Gantt Chart Illustration of the FFSP with Intermediate Buffers
92
Table 5.1 An Example of Converting FFSP with Finite Buffers to FFSP with No
Available Buffer
Processing Stage 1 Stage 2 Stage 3 Stage 4 Stage 5
time (buffer stage) (buffer stage)
Job 1 1 0 4 0 2
Job 2 1 0 6 0 3
Job 3 1 0 10 0 1
Job 4 1 0 4 0 1
Job 5 2 0 2 0 2
5.3 WFA for the FFSP with Intermediate Buffers
In this section, we present the operational mechanism of the WFA for solving the FFSP
with intermediate buffers. The proposed algorithm is based on the simulation of spreading
of raindrops into many places on the ground, as well as the property of water flow always
moving from higher positions to lower positions, and the erosion capability of water flow
on the ground.
Firstly, a cloud representing an iteration randomly generates a set of drops of water
(DOWs) onto some positions on the ground, which represent solutions of the FFSP. Next,
due to the gravity force of Earth represented by a heuristic algorithm, the DOWs
automatically move to local optimal positions. They are held at these positions until the
erosion condition is satisfied before performing the erosion process. Then, depending on
the amount of precipitation, the falling force of precipitation and soil hardness at the local
optimal positions, the erosion process helps the DOWs overcome the local optimal
positions to find better or global positions. A flow chart of the WFA for the FFSP with
intermediate buffers is shown in Figure 5.3. The details of the erosion condition, erosion
capability, erosion process, and other operations of the WFA for the FFSP with
intermediate buffers are described in the following subsections.

93
Initialization
Generate a new cloud (iteration) Construct a seed job

permutation using the
FFLL algorithm
Yes
Generate a new DOW with a
random job permutation using Apply the insertion
an encoding scheme scheme for the seed job When it rains, the DOWs are
permutation to generated and they fall to
construct a set of job the ground
j=j+1 Belong to E-list? permutations
Exploration phase h
No
Generate a set of
No DOWs from the set of
j = MaxPop? job permutations using
an encoding scheme
Yes
Apply the hill-sliding algorithm to the generated job permutations

with a 2-opt neighborhood structure and the H1-variant procedure to
find their local optimal job permutations Due to gravity force, the
Evaluate the soil hardness based on the objective value lower positions (local
obtained by the H1-variant procedure and identify the optimal positions) and they
number of DOWs at the local optimal job permutations are held there
No
Evaluate topology
Identify Choose the un-eroded
parameter around the job
erosion direction with the
Pick one job permutation permutation using the
capability at smallest topology
in UE-list 2-opt neighborhood
the job parameter to perform Erosion process happens
No structure and the
permutation erosion process when the erosion condition
Yes H1-variant procedure
is satisfied at the local
optimal positions
Satisfy erosion
condition? Update UE-list to No
continue with Is a better local optimal
Yes
No performing erosion job permutation found?
process
Have all job permutations No

Update E-list The terrain at eroded
in UE-list been considered?
positions is smoothed by the
Yes Have all directions been erosion process
Yes
chosen?
Yes
Output the best job permutations in P0-list
Figure 5.3 Flow Chart of the WFA for the FFSP with Intermediate Buffers
In the FFSP with intermediate buffers, a DOW is associated with a job permutation. We
consider the job permutations and their single objective value as providing the longitude,
94
latitude, and altitude information for the position of the DOWs on the ground. Given a job
permutation π = (σ 1 ,..., σ N ) , we define:
longitude(π ) = (σ 1 ,..., σ ⎢ N ⎥ ) (5.1)

⎢2⎥
⎣ ⎦
latitude(π ) = (σ ⎢ N ⎥ ,..., σ N ) , (5.2)

⎢⎣ 2 ⎥⎦ +1
where ⎣⎢ x ⎥⎦ is the largest integer less than or equal to x. The objective value could be the
makespan, total weighted flow time of jobs, or total weighted tardiness time of jobs.
However, we do not use all the possible objectives at the same time. Depending on the
single objective function adopted, the altitude of DOW is defined as the corresponding
objective value of the given job permutation. Figure 5.4 shows an illustrative example of a
DOW and its positional vector components for the FFSP with 6 jobs.
Job permutation Objective value

(makespan)
DOW 2 1 5 3 4 6 127
Altitude
Longitude Latitude
Figure 5.4 An Example of Solution Representation in the WFA for the FFSP
Note that the job permutations considered here are denoted by vectors of size N, which
represent the sequence of the given set of jobs performed in the first stage. As such vectors
cannot fully determine the schedule of the jobs through all the stages, a procedure for
constructing a complete schedule from such vectors is required to obtain the objective
value. Here, we propose an improvement over the procedure H1 of Wardono and Fathi
95
(2004), called the H1-variant procedure, to construct a complete schedule associated with
a given job permutation. Our proposed procedure starts from stage 2 and involves
choosing a job waiting in the preceding buffer for an available machine in the current
stage. The criteria for choosing the job, in order of priority, are job weight, due date, total
remaining processing time, processing time at current stage, and length of time at the
buffer. Thus, the job with the largest weight has the highest priority to be chosen. If two or
more jobs have the same largest weight, then the job with the earliest due date will be
chosen first. If there are jobs having the same maximum weight and earliest due date, then
the job with maximum total remaining processing time will be chosen. If there is still a tie,
the job whose processing time at the current stage is largest will be chosen. Any further
ties will be broken by choosing the job that is at the buffer for the longest time. This
procedure is unlike the procedure H1 of Wardono and Fathi (2004), in which only the last
criterion is used to choose the job to be processed. Moreover, the proposed improvement
is applicable to the FFSP with no available buffer space and the FFSP with unlimited
buffers. In the former case, we consider the completed jobs which are blocked at the
preceding stage as the available jobs waiting in buffer. In the latter case, as the buffer
capacities are infinite, the completed jobs are immediately transferred to the buffers to
wait for an available machine in the next stage by following the above mentioned rules of
choosing the jobs.
In addition, another improvement for assigning available machines based on the job
schedule is proposed. The assignment is mainly based on the first available machine
(FAM) rule of Wardono and Fathi (2004). However, for the available machines, we give
higher priority to the machines that have been used before. An example is shown in Figure
96
5.5. According to the modified FAM rule, job 1 is processed on machine 1 instead of
machine 2 at stage 2, even though machine 2 is available earlier and would have been
chosen based on the original FAM rule. This modified FAM rule will help to reduce the
number of machines used. Consequently, it has the advantage of possibly reducing the
resources used in the design problem.
Stage Machine
1 1 1 1
2 2 3 2 3
2 1 2 3 2 1 3
2 1
1 2 3 4 5 6 7 8 9 1011 1 2 3 4 5 6 7 8 9 1011
time, t time, t
(a) Procedure H1 (b) Proposed Procedure
Figure 5.5 A Comparison Between the FAM Rule and the Modified FAM Rule
Figure 5.6 shows a comparison of the overall performance between the two procedures
for an example taken from Wardono and Fathi (2004). In this example, since the weight
and due date of jobs are not given, we assume that they are the same for all jobs. Then,
choosing the appropriate job for the idle machines is based mainly on the total remaining
processing time of available jobs. From Figure 5.6, the completion time of all jobs is at
time 31 for the proposed procedure, while the completion time for the schedule of
Wardono and Fathi (2004) is at time 32. This shows a better performance of our proposed
constructive procedure over that of Wardono and Fathi (2004).
97
Vector representation: {2,6,9,8,7,10,3,4,1,5}

B1 = B2 = 2
Stage Machine b = Machine is blocked
1 1 2 9 8 10 b 4 5 b
2 6 7 3 b 1 b
2 1 2 6 8 3 b 5
2 9 7 10 4 b 1
3 1 2 7 10 5
2 9 8 4
3 6 3 1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
time, t
(a) Complete schedule obtained by the procedure H1
Stage Machine
1 1 2 9 8 10 b 4 b 5 b
2 6 7 3 b 1
2 1 2 6 10 1 4 b
2 9 8 3 7 b 5 b
3 1 2 3 1 7
2 9 10 4
3 6 8 5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
time, t
(b) Complete schedule obtained by the proposed H1-variant procedure
Figure 5.6 A Comparison Between the Procedure H1 and the H1-Variant Procedure
Based on the constructed complete schedule, the corresponding objective value can
then be determined. We thus incorporate the H1-variant procedure of constructing the
complete schedule in the determination of the objective value to simplify the
computational procedure of the WFA. This simplification is significant because the search
space of the problem is now limited to a set of possible permutations of N jobs and we do
98
not need to look at the number of machines in each stage. Although using the job
permutation representation may not cover the global optimal job permutation of the
scheduling problem, the best job permutation obtained by the WFA or any algorithm with
this representation is very near the global optimal job permutation. Furthermore, the job
permutation representation can be more easily integrated into metaheuristic algorithms
than the matrix representation, which cover the entire solution space but need expensive
computation time to determine whether a solution given in this representation is feasible
or not (Wardono and Fathi, 2004).
5.3.2 Memory Lists
Three memory lists of the WFA to support the search for global optimal positions are used
in the FFSP with intermediate buffers. Firstly, the best positions list, called the P0-list,
stores the job permutations with minimum makespan, minimum total weighted flow time
of jobs, or minimum total weighted tardiness time of jobs. Secondly, the un-eroded list,
called the UE-list, is used to store local optimal job permutations which have not been
eroded because of not satisfying the erosion condition. Its purpose is to record the
potential job permutations for the subsequent erosion process. Thirdly, the eroded list,
called the E-list, is used to store eroded local optimal job permutations. Among these lists,
the E-list plays an important role of preventing DOWs from being regenerated to the
eroded job permutations in the subsequent iterations. It would help to reduce the
computation time needed by the algorithm. The lists are updated in a similar manner as
that in Section 4.3.2.
99
In the first iteration of this phase, we generate a seed job permutation using the flexible
flow line loading (FFLL) algorithm (Pinedo, 2005). This is an efficient constructive
algorithm for flexible manufacturing systems with cyclic paths. It consists of three phases:
the machine allocation phase, the sequencing phase, and the release timing phase. The
objective of this algorithm is to minimize the work-in-process so as to reduce blocking
probabilities. Then, we use an insertion scheme to generate a set of job permutations for
the subsequent erosion process. Here, the insertion scheme is based on the seed job
permutation. This scheme is performed by removing a job from its present position and
inserting it at a different position, and then shifting the position of jobs between these two
positions by a unit accordingly. Although the exploration phase of the WFA is mainly
used to explore job permutation search space, the major objective of this phase is to
determine potential regions for performing erosion process in the exploitation phase.
Hence, we have used the generation of the set of initial seed job permutations to improve
the computation time of determining the potential regions. This may also improve the
convergence rate of the WFA, although the best job permutations may not come directly
from these initial seed job permutations.
In the following iterations of this phase, we no longer use the FFLL algorithm. Instead,
only randomly generated job permutations with population size MaxPop are used. In all
the iterations, after generating a population of job permutations for DOWs, a steepest
descent hill sliding algorithm is used to search for local optimal job permutations from
these initial job permutations. The hill sliding algorithm is similar as that described in
Section 4.3.3. In the hill sliding algorithm, a perturbation scheme based on a variant of 2-
100
opt algorithm is used to construct the neighboring job permutations. The 2-opt
neighborhood structure determines the set of all neighboring job permutations that can be
obtained from a current job permutation by exchanging positions of two jobs in the current
job permutation. In particular, if π ' is the job permutation obtained by swapping the
positions of two jobs σ i and σ j in a job permutation π , we can determine π ' by:
π ' [σ i ] = π ⎡⎣σ j ⎤⎦ , π ' ⎡⎣σ j ⎤⎦ = π [σ i ] ,

π ' [σ k ] = π [σ k ] for k ∈ N \ {i, j} , (5.3)
where π[σ] and π’[σ] denote the positions of job σ in the job permutation π and its
neighboring job permutation π’ respectively. The number of neighboring job permutations
obtained by the neighborhood structure is N(N-1)/2. The hill sliding algorithm with this 2-
opt neighborhood structure is used for all three objective functions considered here.
In general, the exploration phase in the WFA for the FFSP results in a set of local
optimal job permutations. They are updated in the UE-list to be considered for performing
the erosion process in the next exploitation phase.
In the WFA for the FFSP with intermediate buffers, we perform the erosion process based
on the amount of precipitation. If the amount of precipitation at some local optimal job
permutation increases up to MinEro (the minimum number of DOWs allowed to start the
erosion process), the erosion process would happen at the local optimal job permutation.
101
In the current WFA, we consider the capability of erosion process based on two main
factors, the amount of precipitation and its falling force. In the FFSP with intermediate
buffers, the amount of precipitation is represented by the number of DOWs at the eroding
local optimal job permutation, while its falling force is represented by the objective value
at the job permutation. We assume that clouds are at the same altitude, and the falling
force of precipitation to lower local optimal positions will cause erosion more easily than
that of higher ones, i.e., the erosion capability becomes stronger for local optimal job
permutations with larger amount of precipitation and lower objective values. It creates a
flexible operation scheme for the erosion capability of DOWs in the algorithm, and helps
the erosion process focus on exploiting promising regions strongly while ignoring regions
with poor performance. In particular, the relationship between these two factors and the
control parameter of erosion capability, MaxUIE (the maximum number of iterations for
the erosion process to move to the next erosion direction if the job permutation is not
improved), is expressed as:
MaxUIE = ϕ1Q (π * ) + ϕ 2 LB / z , (5.4)
where ϕ1 and ϕ2 are parameters representing the effect of precipitation and its falling
force respectively. Also, Q(π * ) is the number of DOWs at the local optimal job
permutation π * , LB is a known lower bound, and z is the objective value at the eroding
local optimal job permutation. Here, LB can be any lower bound obtained from the
literature, or by solving a relaxation of the problem. For the case of makespan
minimization, we have used the lower bound proposed by Sawik (2000) and Wardono and
Fathi (2004) for LB. As for the cases of minimizing the total weighted flow time of jobs
102
and total weighted tardiness time of jobs, we have used the lower bounds in Azizoglu et al.
(2001) and Akturk and Yildirim (1998) respectively.
In real-life, the falling force of precipitation does not affect the erosion capability as
much as the amount of precipitation. Hence, we formulate the relationship between these
two factors and the erosion capability as shown in equation (5.4). While there could be
other functions to express the property, from the computational experiments, this has
worked well on the scheduling problem. In particular, the effect of the falling force of
LB
precipitation is in the range [1, ϕ2 ] by using the function ϕ2 LB / z , due to ≤ 1 . In
z
addition, when running experiments we set ϕ2 to vary in [1, 4]. As for the effect of the
amount of precipitation, we set ϕ1 to vary in [2, 4], since the erosion process is often
performed at the local optimal job permutation with Q(π * ) ≥ 2 . Then, the falling force of
precipitation will have less effect on MaxUIE than the amount of precipitation. Although
parameters ϕ1 and ϕ2 may be considered as relative weights, we do not require the sum of
the weights to be equal to one.
Based on the preliminary computational experiments by using the design-of-
experiment method and setting other parameters to be constant, i.e., MaxCloud = 10,
MaxPop = 20, and MinEro = 3, the best values for ϕ1 and ϕ2 are determined to be 2 and 3
respectively when the WFA is applied to solve the FFSP with intermediate buffers.
103
The erosion process will be performed when the erosion condition is satisfied. The erosion
capability in this process at local optimal job permutations depends on equation (5.4). The
strategy of erosion process is based on a topological parameter representing the
geographical surface, and whether an erosion direction is blocked.
For the FFSP, the topological parameter Δd is calculated in the same manner as that
shown in Section 4.3.4. The backtracking strategy of the erosion process presented in
Section 4.3.4 is also used for solving the FFSP by the WFA. Here, we still use the 2-opt
neighborhood structure as in the exploration phase of the algorithm.
The entire process for both exploration and exploitation terminates when the
maximum number of allowed iterations (MaxCloud) is reached.
5.4 An Example of the FFSP in Maltose Syrup Production
In developing countries, confectionery and sugar industries are of great economic
importance. The main raw material used for such industries is maltose syrup, and its
production is thus crucial for these industries (Pedersen and Vang-Hendriksen, 2001). The
production of maltose syrup can be modeled as a FFSP with no available buffer space
between consecutive stages for completed jobs. An example of this problem involves nine
jobs, with each job handling one type of maltose syrup. They are processed through six
consecutive stages representing the six phases in the maltose syrup production process
(see Figure 5.7). Here, the process is continuous with connections through pipes and tanks.
104
The number of parallel identical machines in each stage, processing time of the jobs at
these stages, weight of the jobs, and due date of the jobs are shown in Table 5.2.
The jobs processed at the various stages of maltose syrup production not only depend
on the busy/idle state of consecutive machines as in the standard FFSP with limited
buffers, but also depend on the status of other machines. For example, jobs cannot go to
stage 1 until there is at least one idle machine in stage 2, even if the machine in stage 1 is
idle. Moreover, since the product is a liquid and stages 2 and 3 are connected directly, the
completion time of the jobs in stages 2 and 3 would be the same. As an example, Figure
5.8 illustrates a Gantt chart of the production problem with its data from Table 5.1. Note
that job 4 could have been processed at time 3 for the standard FFSP with no available
buffer space. However, due to some special requirement in maltose syrup production, job
4 can only be processed from time 7 when there is an idle machine in stage 2. As such, we
can consider this as a FFSP with controlled and limited buffers.
Table 5.2 Problem Data for Maltose Syrup Production
No. of Processing time of jobs (hours)

Stage
machines Job 1 Job 2 Job 3 Job 4 Job 5 Job 6 Job 7 Job 8 Job 9
1 1 6 6 6 6 6 6 6 8 8
2 4 49 49 49 49 49 23 23 11 11
3 1 7 7 7 7 7 7 7 7 7
4 1 2 2 2 2 2 2 2 2 2
5 1 7 5 5 5 5 5 5 7 5
6 1 1 1 1 1 1 1 1 1 1
Due date 120 87 87 172 172 72 72 120 72
Weight 0.10 0.10 0.10 0.05 0.05 0.20 0.20 0.05 0.15
105
Cereal grain mixing

phase
Glucose 1 Glucose 2 Glucose 3 Glucose 4
Filtering and cleansing

phase
Resin
Ceramic filtering
phase
Solidifying phase
Maltose tank Maltose tank

1 2
Figure 5.7 Maltose Syrup Production Process
Stage Machine
1 1 1 2 3 b 4 b 5
2 1 1 4
2 2 5
3 3
3 1 1 2 4 3 5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
time, t
Figure 5.8 Illustration of the FFSP with Controlled and Limited Buffers
106
Since the special restrictions in the production model only affect the H1-variant
procedure of constructing a complete schedule, we do not need to adjust the WFA to solve
the maltose syrup production problem. Thus, we consider the restrictions as constraints in
the H1-variant procedure for releasing the selected jobs to stage 1 and freeing the
machines at stage 2. For the former constraint, the selected jobs can only be processed at
stage 1 when at least one machine in stage 1 and one machine in stage 2 are idle. For the
latter constraint, the machine finishing a job at stage 2 is free only if the job at stage 3 is
also finished. The other rules of selecting and assigning jobs/machines in the H1-variant
procedure, as well as the procedure of determining the objective value are unchanged.
In this section, we present the results of computational experiments carried out on the
benchmark instances of Wittrock (1988), the randomly generated instances based on
Wardono and Fathi (2004), the randomly generated instances based on Tavakkoli-
Moghaddam et al. (2009), and an instance of the maltose syrup production problem
mentioned in the previous section. These experiments are used to evaluate the
performance of the WFA for solving the FFSP with limited and unlimited buffers.
5.5.1 Benchmark Instances and Randomly Generated Instances
Here, the benchmark instances of Wittrock (1988) shown in Table 5.3 are modified in a
similar manner as Wardono and Fathi (2004). Thus, where applicable, the transport time
(one minute) for the Wittrock instances is added to the processing time of the jobs at
stages 2 and 3 respectively. In these instances, the number of machines at stages 1, 2, and
107
3 are 2, 3, and 3 respectively. When solving the instances with finite buffers, we set the
buffer capacity to 3 for all stages.
For the instances generated based on Wardono and Fathi (2004), we use the procedure
of generating the type II dataset in that paper. These instances, called the TS instances,
were constructed with the number of jobs N = 20, 30, 40, 50; the number of stages S = 2, 3,
4; and the number of machines ml = 2, 4, 6 for all stages l. For each job, its processing
time is generated randomly as an integer from a uniform distribution [1, 100]. For each set
of parameter values, we construct and solve five instances. When solving the instances
with finite buffers, we set the buffer capacity to 1 for all stages. We compare the results
obtained by the WFA to the lower bounds proposed by Wardono and Fathi (2004). The
results for the Wittrock benchmark instances and the type II dataset obtained by the tabu
search algorithm of Wardono and Fathi (2004) are used to compare against the results
obtained by the WFA. In the comparison, the tabu search algorithms are called TS-H1 and
TS-Z3 for the FFSP with limited and unlimited buffers respectively.
Table 5.3 Problem Data for the Instances in Wittrock (1988)

Job Processing time (min) Production requirement of each instance
type Stage 1 Stage 2 Stage 3 1 2 3 4 5 6
A 39 11 14 12 - - - - -
B 13 28 54 1 - - - - -
C 22 56 60 26 - - 14 23 20
D 234 39 0 - - - 2 - -
E 39 25 80 - 6 7 4 - 1
F 13 70 54 - 14 20 16 - -
G 143 66 0 1 4 - - 3 1
H 0 28 14 7 - - - - -
I 26 39 74 - 6 - - - 5
J 18 59 34 - 4 - - - -
K 22 70 40 4 - - - - -
L 13 70 54 - 4 - - - -
M 61 46 34 - - 11 - 14 3
Total number of jobs 51 38 38 36 40 30
108
The instances generated based on Tavakkoli-Moghaddam et al. (2009) are called the
MA instances. They are constructed with the number of jobs N = 10, 20, 30, 40; and the
number of stages S = 2, 3, 4, 5, 6. Here, the number of machines in every stage is the same,
and the total number of machines in all stages is equal to twice the number of stages. For
each job, its integer processing time is generated by a uniform distribution [1, 10]. With
the combination of the above parameter values, we generate 20 different problem
instances. We compare the results obtained by the WFA to the lower bounds proposed by
Sawik (2000), as well as to the results obtained by Tavakkoli-Moghaddam et al. (2009),
which describes an algorithm for solving the FFSP with no available buffer space.
In addition to comparing with the results reported in Wardono and Fathi (2004) and
Tavakkoli-Moghaddam et al. (2009), we perform computational experiments on the TS
and MA instances with the same CPU and platform for all algorithms, i.e. WFA, TS-
H1/Z3, and MA. As the instances used in Wardono and Fathi (2004) and Tavakkoli-
Moghaddam et al. (2009) do not include the weight and due date of jobs, we also do not
consider these job characteristics in the randomly generated instances, even though the
WFA is capable of solving the FFSP with given weight and due date of jobs.
The WFA has been coded using Visual Basic 6.0, and all experiments have been
performed on an Intel Centrino Duo 1.60 GHz CPU with 1.5 GB of RAM running on
Windows XP Operating System. The computational complexity of the WFA for the FFSP
is determined based on the neighborhood structure used and the erosion process of this
algorithm. In particular, the WFA used 2-opt neighborhood structure, and the worst
109
possibility of the erosion process is to find for all n directions. Hence, the computational
complexity of the WFA is estimated to be O(n3).
The choice of reasonable parameters for the WFA is determined by design-of-
experiment methods. When implementing the design-of-experiment method, we may use
the independent level of values for the parameters, or the dependent ratio among the
parameters, e.g., the ratio of MaxPop and MinEro, to determine which values are best for
solving the scheduling problem. Here, we used the independent level of values for the
parameters since we have solved many types of FFSP data structures this way. The
independent levels of values for parameters were used as follows: MaxCloud = 5, 10, 15,
20; MaxPop = 5, 10, 15, 20; and MinEro = 2, 3, 4, 5. From the preliminary results, the
best parameter sets are summarized in Table 5.4. With these parameter sets, 20
independent replicates are used for each instance of Wittrock (1988), the maltose syrup
production example, and the MA instances as in the case of Tavakkoli-Moghaddam et al.
(2009). However, we only use 1 replicate for the TS instances as in the case of Wardono
and Fathi (2004). Here, Wardono and Fathi (2004), and Tavakkoli-Moghaddam et al.
(2009) performed a preliminary study to determine the best parameter values for TS-
H1/Z3, and MA respectively.
Table 5.4 Parameter Sets for Benchmark Instances
Value of parameters
Instances
MaxCloud MaxPop MinEro
Modified Wittrock
Instance 1 10 20 3
Instance 2, 3, 5 10 15 3
Instance 4, 6 10 10 4
Wardono & Fathi 10 20 3
Tavakkoli-Moghaddam 10 20 3
Maltose syrup production 5 10 5
110
5.5.3 Performance Measures
For a comparison of objective values, we have used the following average relative
percentage increase in objective value:
K
⎛ Heuristicsol (i ) − LBsol ⎞
∑⎜
i =1 ⎝ LBsol
⎟
Δ= ⎠ ×100. (5.5)
K
Here, Heuristicsol(i) and LBsol denote the objective value obtained by the algorithm for
the ith replicate and the lower bound value or the best known objective value from the
literature respectively, while K denotes the number of replicates. If we only use the
replicate with the best objective value to compute Δ , the relative percentage increase in
objective value is denoted by Δ Best . The smaller the value of Δ , the better the performance
of the algorithm. In this chapter, the lower bounds are used in all instances and the optimal
solution is only used in the maltose syrup production problem, so that Δ = 0 implies that
the algorithm has obtained the optimal solution.
The lower bound value or the best known objective value from the literature is used as
the reference values for the evaluation and comparison of the WFA and other algorithms
in this chapter. The best known objective value may be obtained from the optimal
solutions of the benchmark instances used, or the best solutions found by some algorithm
so far. Hence, the comparison results may be negative values if the proposed algorithm
finds a better solution.
For demonstrating the effect of problem size on the speed of different algorithms, we
use a CPU time ratio measure, denoted simply as ratio. This is the ratio of CPU time for
111
solving an instance under consideration to CPU time for solving the instance with the
smallest size, using the same number of iterations. If the value of ratio of an algorithm
increases quickly with respect to an increase in the problem size, it means that the speed of
the algorithm depends significantly on the problem size.
To compare the algorithms, we performed computational experiments on two sets of
generated TS and MA instances, using the same CPU and platform mentioned in Section
5.5.2 for all the algorithms, i.e., WFA, TS-H1/Z3, and MA. Instead of using the maximum
number of iterations in the experiments, CPU time limits with respect to the number of
jobs in the instances are imposed on all the algorithms. Table 5.5 shows the results of the
WFA, TS-H1/Z3, and MA on the TS instances. Here, we use the lower bound obtained by
Wardono and Fathi (2004) for LBsol in equation (5.5). The CPU time limits for the case of
FFSP with unlimited buffers are equal to one-sixth of the CPU time limits for the other
cases. This is because the procedure for constructing a complete schedule in the FFSP
with unlimited buffers requires a shorter computation time. As the MA is only designed
for the FFSP with no available buffer, we only run experiments for MA on TS instances
with Bl = 0. Table 5.6 shows the results of the WFA, TS-H1, and MA on the MA instances
for the case of FFSP with no available buffer. In this table, SD (Cmax) and SDOFV
represent the standard deviation of objective values obtained by the WFA/TS-H1 and the
MA respectively. Here, we use the lower bound obtained by Sawik (2000) for LBsol in
equation (5.5). From the results in Tables 5.5 and 5.6, we can see that the solution quality
obtained by the WFA outperforms the other algorithms for most instances. In a few
112
Table 5.5 Comparison Results of WFA, TS-H1/Z3, and MA with CPU Time Limit
for the TS Instances
CPU WFA TS-H1/Z3 MA

Test problem time
limit Δ Δ Δ
No. N × S × ml (s) Bl = 0 Bl = 1 Bl = ∞ Bl = 0 Bl = 1 Bl = ∞ Bl = 0
1 20× 2 × 2 0.16 0.04 0.04 0.52 0.04 0.10 5.09
2 20 × 2 × 4 1.14 0.71 1.11 2.51 1.79 2.69 18.58
3 20 × 2 × 6 4.02 4.01 5.39 7.49 6.49 8.11 26.74
4 20 × 3 × 2 300 if 0.57 0.25 0.21 2.49 0.29 0.67 8.75
Bl = 0, 1;
5 20 × 3 × 4 2.04 1.37 1.71 6.37 4.04 2.81 14.94
50 if
6 20 × 3 × 6 Bl = ∞ 1.15 0.99 1.61 4.74 2.27 2.96 19.25
7 20 × 4 × 2 4.25 1.99 1.89 4.80 1.94 1.29 17.45
8 20 × 4 × 4 7.48 7.10 6.82 11.81 8.96 7.03 24.77
9 20 × 4 × 6 8.00 6.36 7.60 12.29 10.33 9.09 22.32
10 30 × 2 × 2 0.83 0.07 0.15 1.63 0.27 0.09 5.57
11 30 × 2 × 4 1.77 0.78 0.94 2.69 2.38 1.30 9.28
12 30 × 2 × 6 3.16 2.49 2.52 6.67 3.10 3.78 17.69
13 30 × 3 × 2 600 if 1.95 0.09 0.06 2.03 0.04 0.09 15.33
Bl = 0, 1;
14 30 × 3 × 4 3.72 1.56 1.35 5.19 2.33 1.49 22.79
100 if
15 30 × 3 × 6 Bl = ∞ 5.36 5.00 4.16 11.16 8.34 5.97 24.65
16 30 × 4 × 2 4.48 1.57 1.42 4.77 1.97 1.42 21.00
17 30 × 4 × 4 7.18 6.10 6.27 9.01 7.34 6.17 31.51
18 30 × 4 × 6 10.69 9.88 9.61 10.63 9.01 8.68 30.34
19 40 × 2 × 2 0.68 0.05 0.15 1.04 0.12 0.13 4.44
20 40 × 2 × 4 2.04 0.71 0.71 2.24 0.95 1.12 11.79
21 40 × 2 × 6 2.90 1.68 2.04 3.67 2.74 2.02 18.81
22 40 × 3 × 2 1200 if 5.22 0.37 0.31 6.62 0.29 0.27 22.02
Bl = 0, 1;
23 40 × 3 × 4 4.57 1.68 1.94 4.64 1.93 1.42 26.49
200 if
24 40 × 3 × 6 Bl = ∞ 6.35 4.86 4.51 6.95 5.86 4.24 32.24
25 40 × 4 × 2 6.57 1.51 1.34 7.27 1.67 0.95 32.09
26 40 × 4 × 4 6.17 4.11 2.92 7.84 4.69 2.53 44.95
27 40 × 4 × 6 8.27 7.36 7.68 8.71 7.34 6.09 40.84
28 50 × 2 × 2 0.32 0.02 0.04 1.49 0.02 0.07 10.52
29 50 × 2 × 4 0.99 0.29 0.26 1.58 0.86 0.42 14.23
30 50 × 2 × 6 2.07 0.98 0.98 3.06 1.07 2.10 19.76
31 50 × 3 × 2 1800 if 4.36 0.43 0.34 4.69 0.38 0.31 19.87
Bl = 0, 1;
32 50 × 3 × 4 3.68 1.64 0.86 3.37 1.39 1.06 27.98
300 if
33 50 × 3 × 6 Bl = ∞ 6.39 3.78 3.40 7.72 4.26 3.71 34.31
34 50 × 4 × 2 9.39 1.71 0.69 10.01 2.04 0.68 33.63
35 50 × 4 × 4 7.46 3.97 2.26 9.10 4.12 2.43 47.71
36 50 × 4 × 6 6.91 5.04 3.02 7.39 6.26 3.20 48.98
Average 4.23 2.51 2.40 5.67 3.25 2.68 22.96
113
Table 5.6 Comparison Results of WFA, TS-H1, and MA with CPU Time Limit for
the MA Instances
Test problem WFA TS-H1 MA
CPU time
limit (s)
No. N × S × ml Δ SD (Cmax) Δ SD (Cmax) Δ SDOFV
1 10 × 2 × 2 0.00 0.00 0.00 0.00 0.00 0.00

2 10 × 3 × 2 5.00 0.00 5.00 0.00 8.87 0.60
3 10 × 4 × 2 120 21.35 0.00 21.35 0.00 24.83 0.83
4 10 × 5 × 2 25.30 0.00 25.30 0.00 33.98 0.82
5 10 × 6 × 2 33.33 0.00 33.33 0.00 40.25 1.51
6 20 × 2 × 2 0.00 0.00 0.00 0.00 4.15 0.69
7 20 × 3 × 2 6.75 0.37 6.67 0.41 20.44 1.09
8 20 × 4 × 2 300 17.26 0.44 17.61 0.51 32.39 0.52
9 20 × 5 × 2 17.90 0.64 18.23 0.66 34.11 2.62
10 20 × 6 × 2 21.97 0.51 22.55 0.51 43.65 2.01
11 30 × 2 × 2 1.92 0.44 3.90 0.60 11.98 0.45
12 30 × 3 × 2 1.95 0.47 3.30 0.60 17.62 0.62
13 30 × 4 × 2 600 11.80 1.03 13.33 1.03 32.08 1.04
14 30 × 5 × 2 20.22 0.68 21.09 0.75 45.76 3.51
15 30 × 6 × 2 22.30 0.79 23.55 0.89 53.11 1.94
16 40 × 2 × 2 0.42 0.00 0.46 0.22 7.85 0.83
17 40 × 3 × 2 5.10 0.67 6.91 1.02 20.60 1.84
18 40 × 4 × 2 1200 11.12 1.00 13.18 1.15 38.26 1.42
19 40 × 5 × 2 15.85 0.83 17.67 0.93 47.71 1.04
20 40 × 6 × 2 13.90 1.09 15.51 1.13 45.20 2.01
Average 12.67 0.45 13.45 0.52 28.14 1.27
4.00
3.50
Standard Deviation
3.00
2.50 WFA
2.00 TS-H1
1.50 MA
1.00
0.50
0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Test Problem Number

Figure 5.9 Standard Deviation of the Objective Values Obtained by the WFA, TS-
H1 and MA
114
instances, TS-H1/Z3 obtained slightly better results than WFA, especially in the case of
FFSP with unlimited buffers. However, from Figure 5.9, we can see that the standard
deviation values of WFA are smaller than that of the other algorithms in all instances. This
shows the robustness of the WFA when compared to the other algorithms.
In addition, we also investigate the tendency of the algorithms, i.e., WFA, TS-H1 and
MA, to find better solutions over computation time. The computational experiments are
carried out on the MA instance with 40 jobs, 6 stages, and 2 machines at each stage, as
well as CPU time limit of 1200 seconds for all algorithms. From Figure 5.10, we see that
the WFA may find a very good solution more quickly than TS-H1 and MA. Then, the
WFA improves this solution slowly to obtain a better solution or global optimal solution.
The final best solution obtained by the WFA is better than that of TS-H1 and MA.
Figure 5.10 Trajectory of Solution Improvement of the WFA, TS-H1, and MA
115
A further comparison between the results of the WFA using the maximum number of
allowed iterations and the results of the tabu search algorithms reported in Wardono and
Fathi (2004) on the randomly generated TS instances is performed in Table 5.7. We use
the lower bound obtained by Wardono and Fathi (2004) for LBsol in equation (5.5). From
Table 5.7, we can see that the WFA is more efficient than TS-H1/Z3 when solving all
three FFSP models. For the FFSP with no available buffer, the average relative percentage
increase of the WFA is 3.9% less than that of TS-H1. For the FFSP with finite buffers, the
average relative percentage increase of the WFA is 2.22% less than that of TS-H1. For the
FFSP with unlimited buffers, the average relative percentage increase of the WFA is 1.7%
less than that of TS-Z3. However, for some of the generated instances, the CPU time ratio
of the WFA is more than that of TS-H1/Z3.
A further comparison between the results of the WFA using the maximum number of
allowed iterations and the MA’s results reported in Tavakkoli-Moghaddam et al. (2009)
for solving the randomly generated MA instances is performed in Table 5.8. We use the
lower bound obtained by Sawik (2000) for LBsol in equation (5.5). From the results in
Table 5.8, we can see that the WFA outperforms MA in solution quality for all the
instances. The overall average relative percentage increase of the WFA is 14.63% less
than that of MA. The WFA is also more robust than MA as its average standard deviation
is less than that of MA. However, the CPU time ratio of WFA is larger than that of MA
from instance 12 onwards.
116
Table 5.7 Comparison Results between WFA and TS-H1/Z3 on the Randomly
Generated TS Instances
WFA TS-H1/Z3
Test Problem
Δ ratio Δ ratio
No. N×S×ml Bl = 0 Bl = 1 Bl = ∞ Bl = 0 Bl = 1 Bl = ∞ Bl = 0 Bl = 1 Bl = ∞ Bl = 0 Bl = 1 Bl = ∞
1 20× 2 × 2 0.31 0.10 0.10 1.00 1.00 1.00 1.45 0.47 0.15 1.00 1.00 1.00
2 20 × 2 × 4 1.00 0.20 0.20 1.54 1.76 1.58 4.78 2.41 1.99 1.09 1.50 1.91
3 20 × 2 × 6 6.56 5.40 4.82 2.16 2.47 2.56 8.72 7.79 8.00 1.59 1.43 2.18
4 20 × 3 × 2 0.08 0.08 0.08 2.57 2.64 2.68 5.85 0.79 0.49 2.62 1.81 2.67
5 20 × 3 × 4 0.43 0.14 0.14 2.91 3.65 2.61 10.24 9.16 8.11 2.67 2.12 2.55
6 20 × 3 × 6 0.00 0.00 0.00 3.46 5.04 3.53 1.52 1.58 1.44 2.23 2.13 2.39
7 20 × 4 × 2 3.91 0.00 0.00 4.08 6.95 5.05 10.56 3.18 2.81 3.04 4.84 3.85
8 20 × 4 × 4 2.27 1.20 1.20 4.68 6.64 4.37 11.88 10.26 10.24 3.41 4.87 5.03
9 20 × 4 × 6 5.29 5.29 5.29 5.27 8.06 5.97 10.35 10.04 9.17 3.80 3.00 4.33
10 30 × 2 × 2 0.15 0.00 0.00 4.55 3.84 3.26 3.84 0.33 0.19 3.34 3.04 2.61
11 30 × 2 × 4 0.43 0.43 0.14 4.93 4.22 4.91 3.09 1.30 0.66 3.46 3.05 3.70
12 30 × 2 × 6 1.21 1.21 0.78 5.22 5.41 8.75 5.13 3.37 2.76 4.23 4.13 6.91
13 30 × 3 × 2 2.29 0.00 0.00 6.13 7.14 9.69 7.93 0.80 0.46 4.86 5.71 5.39
14 30 × 3 × 4 3.86 1.49 0.53 9.72 10.25 12.41 6.38 4.68 3.23 8.03 6.86 8.12
15 30 × 3 × 6 4.95 4.26 4.26 10.55 14.49 15.78 7.30 6.01 4.95 7.89 8.99 9.82
16 30 × 4 × 2 1.31 0.24 0.00 11.03 11.64 11.51 11.51 3.38 2.31 10.15 11.29 8.76
17 30 × 4 × 4 5.73 5.73 5.53 12.44 13.90 15.19 10.33 7.58 5.70 12.10 11.51 13.76
18 30 × 4 × 6 7.55 7.55 6.09 15.69 17.82 18.21 8.97 9.08 8.05 13.52 12.57 12.85
19 40 × 2 × 2 0.76 0.10 0.00 6.20 6.72 8.57 3.30 0.28 0.07 5.07 5.10 5.21
20 40 × 2 × 4 1.18 0.43 0.30 7.97 8.57 10.81 2.69 0.96 0.33 7.16 6.79 8.33
21 40 × 2 × 6 2.09 0.98 0.98 9.48 13.66 12.40 2.94 1.79 1.16 10.57 10.32 11.15
22 40 × 3 × 2 5.19 0.05 0.05 17.50 17.09 15.71 6.41 1.18 0.14 13.93 12.44 11.15
23 40 × 3 × 4 5.21 1.51 0.34 19.93 21.81 23.20 6.12 2.46 1.28 14.80 13.88 15.82
24 40 × 3 × 6 2.40 2.40 2.40 22.89 22.84 25.71 5.72 5.86 2.86 20.44 16.36 26.64
25 40 × 4 × 2 6.42 0.74 0.65 22.65 29.14 25.43 12.63 2.32 1.05 25.48 29.47 18.94
26 40 × 4 × 4 2.82 1.35 1.02 24.70 30.45 31.11 9.90 5.20 2.97 23.86 33.02 30.00
27 40 × 4 × 6 3.89 3.43 3.43 29.63 33.51 31.39 9.06 7.06 6.49 35.04 28.37 26.85
28 50 × 2 × 2 0.81 0.00 0.00 7.34 11.63 11.94 3.39 0.59 0.12 9.96 9.03 7.97
29 50 × 2 × 4 1.65 0.52 0.20 9.11 15.16 12.94 3.73 1.01 1.23 12.56 13.57 13.00
30 50 × 2 × 6 2.04 1.32 1.08 10.07 14.19 14.80 3.71 1.89 1.39 13.58 12.62 19.73
31 50 × 3 × 2 1.84 0.03 0.03 22.66 31.38 22.49 8.30 1.78 0.88 25.04 21.02 12.55
32 50 × 3 × 4 2.08 0.31 0.03 26.03 37.71 32.29 6.93 2.27 1.40 28.71 31.48 25.64
33 50 × 3 × 6 3.90 2.50 1.90 34.29 43.93 44.29 6.88 3.95 3.24 33.67 30.41 36.61
34 50 × 4 × 2 8.76 1.84 0.49 34.99 57.14 35.71 12.80 3.64 1.22 36.50 50.00 25.97
35 50 × 4 × 4 6.75 2.07 0.87 43.49 58.57 58.74 9.39 5.49 3.89 55.44 49.99 51.39
36 50 × 4 × 6 5.49 3.22 1.70 42.22 59.43 61.43 7.04 6.24 5.31 55.42 51.05 55.18
Average 3.07 1.56 1.24 6.97 3.78 2.94
117
Table 5.8 Comparison Results between WFA and MA on the Randomly Generated
MA Instances
Test problem WFA MA

No. N × S × ml Δ SD (Cmax) ratio Δ SDOFV ratio
1 10 × 2 × 2 0.00 0.00 1.00 0.00 0.00 1.00
2 10 × 3 × 2 5.00 0.00 8.91 11.76 0.00 18.92
3 10 × 4 × 2 21.35 0.00 18.36 39.20 0.60 206.25
4 10 × 5 × 2 25.30 0.00 29.55 44.38 0.90 254.17
5 10 × 6 × 2 33.33 0.00 40.27 55.53 1.30 504.17
6 20 × 2 × 2 0.00 0.00 56.36 13.78 0.56 102.08
7 20 × 3 × 2 6.67 0.16 303.64 14.29 0.94 631.25
8 20 × 4 × 2 16.99 0.09 535.45 20.16 1.40 958.33
9 20 × 5 × 2 16.53 0.39 812.73 29.85 1.30 1650.00
10 20 × 6 × 2 21.02 0.09 1269.09 45.33 2.10 1887.50
11 30 × 2 × 2 1.76 0.24 233.64 5.13 0.90 256.25
12 30 × 3 × 2 1.35 0.19 829.09 15.34 2.30 583.33
13 30 × 4 × 2 10.66 0.19 1426.36 30.83 1.96 702.08
14 30 × 5 × 2 18.80 0.21 2806.36 41.15 3.00 1477.08
15 30 × 6 × 2 20.71 0.75 6539.09 44.36 1.80 2997.92
16 40 × 2 × 2 0.42 0.00 613.64 0.95 0.00 270.83
17 40 × 3 × 2 5.22 0.50 1677.27 14.61 1.30 1250.00
18 40 × 4 × 2 10.83 0.69 3523.64 34.07 2.20 2777.08
19 40 × 5 × 2 14.43 0.69 6541.82 34.40 2.20 2908.33
20 40 × 6 × 2 12.80 1.19 9286.36 40.73 2.50 6995.83
Average 12.16 0.27 26.79 1.36
The results of comparison between the WFA and the tabu search algorithms of
Wardono and Fathi (2004) on the modified Wittrock benchmark problems are shown in
Tables 5.9, 5.10, and 5.11. We use the lower bound obtained by Wardono and Fathi
(2004) for LBsol in equation (5.5). From these tables, it can be seen that the WFA
outperforms TS-H1 and TS-Z3 both in solution quality and CPU time ratio for most
benchmark instances. Thus, the WFA has obtained improved solutions than that reported
in Wardono and Fathi (2004), except for instance 1 in Table 5.11. For the FFSP with no
available buffer space, the average relative percentage increase in objective value of the
WFA is 0.95% less than that of TS-H1 (see Table 5.9). For the FFSP with finite buffers,
118
the average relative percentage increase of the WFA is 0.94% less than that of TS-H1 (see
Table 5.10). For the FFSP with unlimited buffers, the average relative percentage increase
of the WFA is 0.81% less than that of TS-Z3 (see Table 5.11). When solving instances
with different sizes, it is also observed that the CPU time ratio of the WFA is less than that
of TS-H1 and TS-Z3 for most instances.
Table 5.9 Comparison Results between WFA and TS-H1 for the FFSP with No
Available Buffer Space
WFA TS-H1
Problem Instance LB
Cmax Δ Best ratio Cmax Δ Best ratio
Modified Wittrock 1 746.3 811 8.67 5.73 822 10.14 6.47
2 758.0 826 8.97 2.89 839 10.69 3.29
3 758.7 815 7.42 2.25 822 8.35 2.29
4 755.3 820 8.57 1.46 825 9.22 4.30
5 961.5 971 0.99 2.51 974 1.30 2.69
6 666.7 682 2.29 1.00 686 2.90 1.00
Average 6.15 7.10
Table 5.10 Comparison Results between WFA and TS-H1 for the FFSP with Finite
Buffer Capacities
WFA TS-H1
Problem Instance LB
2 758.0 767 1.19 2.82 774 2.11 3.26
3 758.7 770 1.49 2.25 777 2.42 2.37
4 755.3 772 2.21 1.62 775 2.60 1.80
5 961.5 962 0.05 1.45 969 0.78 1.40
6 666.7 669 0.34 1.00 675 1.25 1.00
Average 1.25 2.19
119
Table 5.11 Comparison Results between WFA and TS-Z3 for the FFSP with
Unlimited Buffers
WFA TS-Z3
Problem Instance LB
2 758.0 764 0.79 2.88 773 1.98 2.82
3 758.7 770 1.49 2.05 776 2.28 1.86
4 755.3 769 1.81 1.98 776 2.74 2.40
5 961.5 962 0.05 1.43 969 0.78 1.45
6 666.7 669 0.34 1.00 677 1.55 1.00
Average 1.05 1.86
In addition, we also compared the results obtained by the WFA for the maltose syrup
production problem to the optimal values obtained by using enumeration. The results are
shown in Table 5.12 with two different LB values being used for the WFA. The first LB is
based on using the optimal objective value, while the second LB is based on using the
lower bounds derived from Wardono and Fathi (2004), Azizoglu et al. (2001) and Akturk
and Yildirim (1998) respectively for the three different objective functions. As for LBsol in
equation (5.5), we only use the optimal objective value. Note that the objective values
obtained by both the WFA and the enumeration method in Table 5.12 correspond to three
different solutions according to the three respective objective functions. The results in this
table show that the WFA is able to obtain the optimal solutions regardless of the two LB
values used. Moreover, the WFA obtained these optimal solutions with significantly
smaller computation time than that of the enumeration method.
From the experimental results for the benchmark instances, the randomly generated
instances, and the instance from maltose syrup production, we conclude that the WFA is
an efficient meta-heuristic algorithm for solving the FFSP with limited as well as
unlimited buffers.
120
Table 5.12 Computational Results of WFA for Maltose Syrup Production Problem
Optimal solution Solution obtained by Solution obtained by WFA

obtained by WFA when optimal when lower bound from
Objectives enumeration value is used for LB literature is used for LB
used
Optimal
value
Time (s) Δ Time (s) LB Δ Time (s)
Makespan 139.00 130 0.00 1.6 109.00a 0.00 1.8

b
Total weighted flow time 75.95 106 0.00 1.8 63.10 0.00 2.1
c
Total weighted tardiness 0.10 105 0.00 1.9 0.00 0.00 1.9
a
Lower bound derived from Wardono and Fathi (2004).
b
Lower bound derived from Azizoglu et al. (2001).
c
Lower bound derived from Akturk and Yildirim (1998).
5.6 Conclusions
In this chapter, we propose a WFA for solving the FFSP. It involves using an erosion
capability relationship function between the amount of precipitation and its falling force to
create a flexible operation scheme for the erosion process. This helps the erosion process
to focus on exploiting promising regions strongly. We show how this algorithm can be
applied to solve the FFSP with limited as well as unlimited buffers. In addition, we also
propose an improved procedure from that of Wardono and Fathi (2004) for constructing a
complete schedule of the FFSP problem. Computational experiments and comparisons
were carried out to show the performance of the proposed algorithm. The results show that
the WFA is a promising algorithm not only for solving benchmark instances and randomly
generated instances but also for solving problems arising in practical applications.
Improved solutions to benchmark problems are also found by our proposed algorithm.
Some preliminary results were first reported in Tran and Ng (2009), and the full results of
this chapter were then reported in Tran and Ng (2010).
121
Chapter 6 MOWFA for Multi-Objective Scheduling
CHAPTER 6
MOWFA FOR MULTI-OBJECTIVE SCHEDULING
In this chapter, we construct a multi-objective water flow algorithm (MOWFA) for
solving multi-objective scheduling problems. In particular, we investigate the multi-
objective flexible flow shop scheduling problem (MOFFSP) with limited intermediate
buffers. Two objectives of this scheduling problem are the minimization of the
completion time of jobs and the minimization of the total tardiness time of jobs. In the
MOWFA, landscape analysis is performed to determine the weights of objective
functions, which guide the drops of water to exploit potential regions and move
towards the optimal Pareto optimal solution set. We also include the evaporation and
precipitation processes in this algorithm to enhance the solution exploitation capability
of the algorithm in potential neighboring regions. In addition, we propose an
improvement process for reinforcing the final Pareto solution set obtained. The
performance of the MOWFA is tested with benchmark instances taken from the
literature and randomly generated instances. The computational results and
comparisons demonstrate the effectiveness and efficiency of the proposed algorithm.
Chapter 6 is organized as follows. In Section 6.1, we introduce the MOFFSP with
limited intermediate buffers and describe its applications. A brief literature review of
122
research works on the multi-objective scheduling problem is also presented in this
section. Section 6.2 describes the details of the MOFFSP with limited intermediate
buffers. Then, the proposed MOWFA for solving the multi-objective scheduling
problem is discussed in detail in Section 6.3. Computational results and comparisons
based on the benchmark instances taken from the literature and randomly generated
instances are shown in Section 6.4. Finally, some conclusions of this chapter are
presented in Section 6.5.
6.1 Introduction
Flexible flow shop scheduling (FFSP) with limited buffers is one of the well-known
scheduling problems due to its important applications in both traditional and modern
manufacturing systems. A brief description of the practical applications of this
scheduling problem can be found in Section 5.1. The FFSP with limited buffers is
known to be an NP-hard problem (Wardono and Fathi, 2004). It becomes even more
complex when we need to solve the problem in real-life production environment where
several conflicting objectives are simultaneously considered. Two primary objectives
of the problem investigated in this chapter are minimization of the completion time of
jobs (makespan or Cmax) and minimization of the total tardiness time of jobs.
Quadt and Kuhn (2007) and Ribas et al. (2010) presented a detailed review of
single-objective scheduling algorithms and the classification of FFSP. Although there
have been a lot of research works on the FFSP, only a few of them dealt with the FFSP
with limited buffers. There are also limited research works on the multi-objective
scheduling problem. An extensive review of multi-objective scheduling problems in
the past 13 years is presented by Lei (2009). From this review, only Wei et al. (2006)
123
proposed an evolutionary algorithm for solving the MOFFSP. However, the authors
did not consider the FFSP with limited buffers. Qian et al. (2009) presented a hybrid
differential evolution algorithm for solving the multi-objective permutation flow shop
scheduling problem with limited buffers, in which there is one machine at each stage.
Recently, Rashidi et al. (2010) proposed a hybrid parallel genetic algorithm for the
MOFFSP. In their paper, the authors investigated the FFSP with unrelated parallel
machines, sequence-dependent setup times, and processor blocking to minimize the
makespan and the maximum tardiness. The proposed solution procedure consists of
independent parallel genetic algorithms in which each genetic algorithm searches for
optimal solutions in different directions based on different assigned weights for each
subpopulation. The computational results show that it is an efficient algorithm for the
MOFFSP that was considered. However, assigning different weights to subpopulations
may not be efficient as one individual chromosome could be assigned with more than a
pair of weights. Moreover, the algorithm is dependent on a large parameter set defined
by the user. In addition, the solution representation is based on random keys, which
increases the computation time due to the need for decoding.
In this chapter, we propose the MOWFA for solving the MOFFSP with limited
intermediate buffers. This algorithm integrates several search procedures for solving
the multi-objective scheduling problem efficiently. In particular, in the exploration
phase of the MOWFA, the FFSP is divided into many scheduling sub-problems. Drops
of water (DOWs) in the sub-problems are assigned suitable weights based on landscape
analysis to search for optimal solutions in the corresponding directions. In the
exploitation phase of the MOWFA, the erosion process with local and global
neighborhood structures guides DOWs to overcome obstacles to search for better
124
optimal solutions in the region in which they are generated, and in the neighboring
regions, respectively. Also, we include an evaporation process in the algorithm to
enhance the solution exploitation capability of this algorithm in potential neighboring
regions. In addition, we propose an improvement search process for reinforcing the
final Pareto solution set obtained. The Wittrock benchmark instances taken from the
literature, as well as randomly generated instances, are used to evaluate the
performance of the proposed algorithm. The computational results and comparisons
show that the MOWFA is an efficient nature-inspired algorithm for solving the
MOFFSP with limited intermediate buffers.
6.2 MOFFSP with Intermediate Buffers
A detailed description of the FFSP with intermediate buffers has been provided in
Section 5.2. In this chapter, we only consider two cases of the FFSP with intermediate
buffers. The first case is the FFSP with no buffer between consecutive stages. Thus, if
there is no idle machine in the subsequent stage l + 1, a job completed on a machine in
stage l must then wait until at least a machine in stage l + 1 is available. The second
case is the FFSP with finite buffer capacities between consecutive stages. In this case,
if there is no available machine in the subsequent stage l + 1, then a job completed on a
machine in the previous stage l may wait in a following buffer l. If there is no available
buffer space, it remains on the blocked machine in stage l until there is an available
buffer space or an idle machine in stage l + 1.
Since intermediate buffers can be considered as machines with zero processing
time, the FFSP with finite buffers can be converted to one with no available buffer
space (McCormick et al., 1989). As a result, we only need to construct an MOWFA for
125
solving the MOFFSP with no buffer. Then, we can apply the MOWFA for solving the
MOFFSP with finite buffers by a modification of the input data described in Section
5.2. In the transformed problem, a buffer at stage l with a storage capacity Bl is
considered as a stage with Bl identical parallel machines (see Figure 6.1). The
processing time of jobs on the machines in the buffer stages is zero. Then, the total
number of stages in this problem becomes 2S – 1. Every job must be processed
through all stages, including the buffer stages.
Stage 1 Stage 2 Stage S
1 1 1 1 1
N jobs
2 2 2 2 2
M1 B1 M2 BS-1 MS
Buffer 1 Buffer S-1
Figure 6.1 FFSP with Operation Stages Including Intermediate Buffers
Let π denote a job permutation, Cjl denote the completion time of job j at stage l, dj
denote the deadline of job j, and Tj denote the tardiness time of job j. The tardiness
time of job j can be defined as T j = max {C jS − d j , 0} . Then, two objective functions
used in the multi-objective scheduling problem, namely minimizing the completion
time of jobs f1(π) and minimizing the total tardiness time of jobs f2(π), can be
determined by:
min f1 (π ) = max {C jS } , (6.1)

j∈N
N
min f 2 (π ) = ∑ T j . (6.2)
j =1
These objectives help to achieve a high throughput for production.

126
The general MOFFSP with the two objectives mentioned earlier can be described as
follows:
min f(π) = [f1(π), f2(π)], (6.3)
subject to π ∈ Π , (6.4)
where Π denotes the set of possible job permutations.
To solve the multi-objective optimization problem, we use the concepts of Pareto
optimality described below in terms of a minimization problem:
(a). Pareto dominance: a job permutation π1 dominates another job permutation π2,
denoted as π 1 f π 2 , if and only if we have:
f h (π 1 ) ≤ f h (π 2 ), ∀h ∈ {1, 2} and ∃k ∈ {1, 2}, f k (π 1 ) < f k (π 2 ) .
(b). Pareto optimal job permutation: a job permutation π1 is considered to be a
Pareto optimal job permutation if and only if there is no job permutation
π 2 ∈ Π that dominates π1.
(c). Pareto optimal set: Pareto optimal set is a set of all Pareto optimal job
permutations.
(d). Pareto optimal front: Pareto optimal front is the set of all objective values
corresponding to the job permutations in the Pareto optimal set.
(e). Non-dominated job permutation: a job permutation π is said to be non-
dominated with respect to a given set of job permutations if and only if π is not
dominated by any job permutation in the given set.
127
In the Pareto optimal set, the job permutations cannot be improved in any objective
function without degrading the value in at least one other objective function. Hence,
there is no job permutation that is the best for all objectives. As a result, multi-
objective optimization algorithms often need to find a set of good non-dominated job
permutations. To compare the performance of multi-objective optimization algorithms,
three main aspects are considered: the number of non-dominated job permutations in
the Pareto set obtained, the distance of the Pareto front obtained to the Pareto optimal
front, and the spread of non-dominated job permutations in the Pareto set obtained
(Silva et al., 2004).
In addition, we use a single-objective optimization approach to solve the multi-
objective scheduling problem. In this approach, the multi-objective optimization
problem is transformed to a single-objective optimization problem by combining
objective functions and their weights linearly. The main drawback of this approach is
that the weights of the objective functions are subjectively provided. To overcome this
drawback, we have performed landscape analysis of the multi-objective optimization
problem to determine correlation among objective functions. Then, weights are
automatically updated and assigned to corresponding objective functions. The single-
objective optimization approach is used to search for local optimal job permutations of
the multi-objective problem in the MOWFA. The details of performing the procedure
will be described in the next section.
6.3 MOWFA for the MOFFSP with Intermediate Buffers
In this section, we describe the procedure of MOWFA for solving the MOFFSP with
intermediate buffers, including the solution representation, procedure for constructing
128
a complete schedule, erosion condition, erosion capability, and erosion process. Other
important operators of the MOWFA for this problem, such as landscape analysis,
evaporation process, and improvement process, are also described.
In the MOFFSP with intermediate buffers, a DOW is associated with a job
permutation. The job permutation provides the longitude and latitude information for
the position of DOW on the ground. Its corresponding objective values provide the
altitude information for the position of the DOW. Given a job permutation
π = (δ1 ,..., δ N ) , we define:
longitude(π ) = (δ1 ,..., δ ⎢ N ⎥ ), and latitude(π ) = (δ ⎢ N ⎥ ,...,δ N ), (6.5)

⎢2⎥ ⎢ 2 ⎥ +1
⎣ ⎦ ⎣ ⎦
where ⎣⎢ q ⎦⎥ is the largest integer less than or equal to q. We also define the altitude of
the DOW as comprising of the two objective functions, makespan and total tardiness
time of the job permutation π. In Figure 6.2, we illustrate an example of a DOW and its
positional vector components for the MOFFSP with 8 jobs.
Job permutation π Objective values
N
Cmax ∑T j =1
j
DOW 1 7 8 3 6 5 4 2 117 827
Longitude(π) Latitude(π) Altitude(π)
Figure 6.2 An Example of Solution Representation in MOWFA for the MOFFSP
129
In the MOWFA, the job permutations providing the longitude and latitude
information represent the sequence of N jobs processed in the first stage. However, the
job permutation cannot fully determine a complete schedule for the jobs going through
all the stages, and thus the corresponding objective values are also not determined. As
such, we use the H1-variant procedure described in Section 5.3.1 to construct a
complete schedule associated with a given job permutation. This procedure has the
advantage that it is applicable to the FFSP with no buffer and the FFSP with unlimited
buffers. Also, with this procedure, the search space of the problem is limited to a set of
possible permutations of N jobs. Then, we do not need to monitor the number of
machines in each stage, and the computational procedure of the MOWFA becomes
simplified. As in the case of the single-objective FFSP in Chapter 5, although using the
job permutation representation for the multi-objective scheduling problem may not
include the global optimal job permutation of this problem, the best job permutation
found by the WFA or any algorithm with this representation is very near the global
optimal job permutation. Moreover, the job permutation representation is more easily
integrated into metaheuristic algorithms than the matrix representation, which can
cover the entire solution space but requires computation time to determine whether a
solution given in this representation is feasible (Wardono and Fathi, 2004).
6.3.2 Memory Lists
In the MOWFA, we use three memory lists in the search for the Pareto optimal set.
The first list, denoted as the Pareto-list, stores the best non-dominated job
permutations. The second list, denoted as the UE-list, is used to store local optimal job
permutations which have not been eroded due to the erosion condition not being
satisfied. The list aims to record the potential job permutations to be considered for
130
performing the erosion process. The final list, denoted as the E-list, is used to store
local optimal job permutations eroded. The E-list aims to prevent regenerating DOWs
into the job permutations that are eroded in the next iterations and would thus help to
save the computational time of the algorithm.
In our approach, multiple objective functions are linearly combined into a single
objective function so that the search for the local optimal job permutations is based on
a scalar objective function. Updating the UE-list and the E-list is also based on the
local optimal job permutations with respect to the scalar objective function. However,
updating the Pareto-list is based on the property of Pareto dominance in multi-
objective optimization.
Here, we describe the operational mechanism of the components in the exploration
phase of MOWFA. This phase aims to spread the DOWs to many places on the ground
to increase the solution diversification capability of the algorithm. It also helps the
DOWs to select suitable directions to search for the Pareto optimal set by analyzing the
landscape, which represents a set of objective values of feasible solutions.
6.3.3.1 Distinct Regions
In the exploration phase of this algorithm, we divide the MOFFSP with intermediate
buffers into multiple scheduling sub-problems so that the solution space of the problem
is divided into N distinct regions, where N is the instance size. This is achieved by
fixing the first position of a job in a job permutation from 1 to N when generating the
131
job permutations of DOWs. The rest of the positions in a job permutation are assigned
randomly. This can be represented with the following notation:
Ω = {Ω1 , Ω 2 ,..., Ω N }, with Ωi = { π | π [1] = i } for i = 1, 2,.., N , (6.6)
where Ω denotes the solution space of MOFFSP, and Ωi denotes ith distinct region.
6.3.3.2 Landscape Analysis
Landscape analysis for the distinct regions is performed to determine suitable
searching directions for the DOWs generated in the regions. Here, the searching
direction is defined by a pair of weights (c1, c2) that is used to combine the multiple
objective functions into a single objective function. Thus, the scalar objective function
f(π) is defined as:
f (π ) = c1 f1 (π ) + c2 f 2 (π ) , (6.7)
where c1 and c2 are the weights of objective function f1(π) and f2(π), respectively.
Based on the pair of weights, DOWs will search for local optimal job permutations in
the corresponding direction.
In this chapter, we use the landscape analysis approach proposed by Tantar et al.
(2008). At an iteration, we randomly generate a sample of MaxPop job permutations
for each distinct region Ωi . An approximation of the enclosing ellipse for the feasible
solution space in the distinct region is then obtained by such a sampling as described in
Tantar et al. (2008). Here, a complete ellipse is defined by the center coordinates (x0,
y0), the orientation angle α, and the major axis a and minor axis b. The orientation
angle provides the correlation degree that exists between the objective functions (see
Figure 6.3). Hence, finding non-dominated job permutations with an appropriate
132
orientation angle will lead to the Pareto optimal set. Given the orientation angle α, a
pair of corresponding weights (c1, c2) is determined by:
c1 = cos 2 (α ), (6.8)
c2 = sin 2 (α ). (6.9)
In the subsequent iterations, the approximation of the enclosing ellipse for region i
is determined based on the cumulative sample set of job permutations generated in the
region. When the sample size increases up to 100, we use the pair of weights (c1, c2)
obtained at the iteration for the corresponding region in the remaining iterations. We
do not need to recalculate the enclosing ellipse since the difference between the 100-
sample set and any larger sample set is almost unnoticeable (Tantar et al., 2008). This
will help to reduce computation time for performing the MOWFA.
f2(π) c2
c1
α
y0
0 x0 f1(π)
= Pareto optimal solutions
= Initial solutions generated
Figure 6.3 An Illustration for Finding the Pareto Set Based on the Orientation
Angle in MOWFA
133
6.3.3.3 Seed Job Permutations
To improve the performance of the MOWFA, we generate a seed job permutation in
the initial iteration by using the flexible flow line loading (FFLL) algorithm (Pinedo,
2005). A brief description of this algorithm is provided in Section 5.3.3. In order to
generate a set of N job permutations corresponding to N distinct regions, we use a
swap scheme based on the seed job permutation. This scheme is performed by
interchanging only the job at the first position with a job at a different position in the
seed job permutation. The FFLL algorithm is only used in the initial iteration. For
subsequent iterations, the FFLL algorithm is no longer used and instead, we use
randomly generated job permutations for each distinct region.
6.3.3.4 Hill-Sliding Algorithm
For the sample set of job permutations generated, including the seed job permutations
obtained by the swap scheme, we apply the single-objective optimization approach to
search for local optimal job permutations. Here, the steepest descent hill-sliding
algorithm described in Section 4.3.3 is used to guide DOWs to reach the local optimal
positions corresponding to the scalar objective function. Note that the pair of weights
(c1, c2), determined by approximating an enclosing ellipse for feasible solution space in
the corresponding distinct region, is used for transforming the multi-objective FFSP
into a single-objective FFSP.
These local optimal job permutations and the number of DOWs at the job
permutations are updated in the UE-list to be considered for performing the erosion
process in the subsequent exploitation phase.
134
6.3.4 Neighborhood Structures
In this chapter, the neighborhood structures used consist of the swap scheme described
earlier and the 2-opt neighborhood structure. The 2-opt neighborhood structure
determines all neighboring job permutations that can be obtained from a current job
permutation by exchanging positions of two jobs in the current job permutation. For
example, if π′ denotes the job permutation obtained by exchanging the positions of two
jobs δi and δj in a job permutation π, then π′ can be expressed as:
π ' [δ i ] = π ⎡⎣δ j ⎤⎦ , π ' ⎡⎣δ j ⎤⎦ = π [δ i ] ,

π ' [δ k ] = π [δ k ] for k ∈ N \ {i, j} , (6.10)
where π[δ] and π′[δ] denote the position of job δ in the job permutation π and its
neighboring job permutation π′, respectively.
In the MOWFA, we divide the 2-opt neighborhood structure into two types,
namely, local and global 2-opt neighborhood. In the local 2-opt neighborhood, we do
not exchange the job in the first position with a job in any other positions in a job
permutation. The neighborhood structure is used for finding local optimal job
permutations in distinct regions at the exploration phase. It is also used in the erosion
process. The neighborhood structure aims to find the Pareto solution set in a specific
distinct region. With the local 2-opt neighborhood, DOWs only move in the distinct
region in which they are generated. The number of neighboring job permutations
obtained by the neighborhood structure is (N–1)(N–2)/2. Otherwise for the global 2-opt
neighborhood, we can exchange the job in the first position with a job in any other
position of a job permutation. Thus, the global 2-opt neighborhood aims to guide the
DOWs to explore and exploit potential neighboring regions (see Figure 6.4). The
135
number of neighboring job permutations obtained by the neighborhood structure is
N(N–1)/2. It is used in the erosion process and the improvement phase.
Region 1 Region N ( Ω N )
Region i
Global 2-opt neighborhood

Local 2-opt neighborhood
Figure 6.4 An Illustration of the Search Direction of DOWs on Two Types of

2-opt Neighborhood
In this section, we present the conditions to perform erosion process for the job
permutations in the UE-list, the determination of erosion capability of DOWs at the job
permutations satisfying the conditions, and two schemes of erosion process. The
erosion process is the main operator of the exploitation phase which guides the DOWs
to overcome the local optimal positions and obtain better or optimal positions.
In the MOWFA for the MOFFSP with intermediate buffers, the condition to perform
erosion process is determined based on the amount of precipitation as described in
136
Section 4.3.4. Let MinEro denote the minimum amount of precipitation allowed to
start the erosion process. Then, the erosion process will happen at a local optimal job
permutation in the UE-list in which the amount of precipitation reaches MinEro. In the
MOWFA, the amount of precipitation is represented by the number of DOWs at the
eroding local optimal job permutation.
In the MOWFA, the erosion capability is based on two main factors, the amount of
precipitation and its falling force. The amount of precipitation is represented by the
number of DOWs as mentioned earlier, while its falling force depends on the altitude
of positions of DOWs. The falling force of precipitation to lower positions is stronger
than that of higher positions and would thus erode more easily for lower positions.
Hence, the erosion capability becomes stronger for local optimal job permutations with
larger number of DOWs and lower objective values.
Let MaxUIE denote the maximum number of iterations for the erosion process. We
use the following relationship between the erosion capability and the above two factors
as in Section 5.3.4.1:
*
MaxUIE = ϕ1Q (π * ) + ϕ 2 LB / f (π ) , (6.11)
where ϕ1 and ϕ2 represent the impact of precipitation and its falling force,
respectively. Based on preliminary experimentation, we set the values for ϕ1 and ϕ2 to
2 and 3, respectively. Here, Q(π * ) is the number of DOWs at the local optimal job
permutation π * , f(π*) is the scalar objective value of the local optimal job permutation,
LB is the scalar lower bound that is defined as:
LB = c1 LB1 + c2 LB2 , (6.12)
where (c1, c2) is the corresponding pair of weights used to determine f(π*), and LB1 and
LB2 represent the lower bounds of makespan and total tardiness time of jobs,
137
respectively. We have used the lower bound proposed by Wardono and Fathi (2004)
for LB1 and the lower bound proposed by Akturk and Yildirim (1998) for LB2.
In the MOWFA, when the erosion condition at some local optimal job permutation is
satisfied, the erosion process will be carried out. Before performing the erosion
process, the erosion capability of DOWs at the local optimal job permutation is
determined using equation (6.11). Then, the erosion process is performed following an
erosion strategy based on a topological parameter representing the geographical
surface of the optimization problem. For the MOFFSP, the topological parameter Δd h
is defined as the difference between the scalar objective value of the local optimal job
permutation π* and that of its neighboring job permutation π h* :
Δd h = f (π h* ) − f (π * ) . (6.13)
In the MOWFA for the MOFFSP, we use the two neighborhood structures
described in Section 6.3.4, i.e., the local and global 2-opt neighborhoods. With each
neighborhood structure, we have a specific erosion scheme for finding better local
optimal job permutations. The erosion scheme I, based on the local 2-opt
neighborhood, only searches for better local optimal job permutations in a given
distinct region, so that the DOWs only exploit a distinct region from which they are
generated (see Figure 6.5). The erosion scheme II, based on the global 2-opt
neighborhood, can search for better local optimal job permutations in other regions, so
that the DOWs can move and exploit neighboring regions different from the region that
they are generated (see Figure 6.6).
138
f2(π)
α
y0
0 x0 f1(π)
= Pareto optimal solutions
= Initial solution generated
= Corresponding local optimal solutions
= Intermediate solution with at least one worse objective value
Figure 6.5 An Illustration for Scheme I of the Erosion Process with the Local
2-Opt Neighborhood
f2(π)
αj
αi
0 f1(π)
= Initial solution generated in region i
= Local optimal solution in region i
= Solution found in region j after using the global 2-opt
= Local optimal solution in region j
Figure 6.6 An Illustration for Scheme II of the Erosion Process with the Global
2-Opt Neighborhood
139
In this erosion process, the E-list will also prevent the DOWs from moving to
eroded job permutations in other regions. Hence, the DOWs will only move to un-
eroded regions where their scalar objective value is improved.
6.3.6 Evaporation and Precipitation
In the MOWFA, we include an evaporation process in the algorithm to remove poor
local optimal job permutations in the UE-list. A precipitation process is also used to
regenerate the evaporated DOWs into neighboring regions, which reinforces the
erosion possibility in the regions.
In the evaporation process, the local optimal job permutations with one DOW, held
in the UE-list for a pre-specified number of iterations T, will be deleted from the list.
Then, in the next iteration, a precipitation process will generate the number of deleted
DOWs into new positions in neighboring regions. For example, if we remove one
DOW in region i from the UE-list, we will randomly generate DOWs in regions i – 1
and i + 1. In the case of the DOW being deleted from region 1 or N, it will be randomly
generated in regions N and 2, or regions N – 1 and 1, respectively.
6.3.7 Improvement Phase
The two schemes in the erosion process can be applied to a scalar objective function
and they aim to exploit strongly promising regions as well as potential neighboring
regions. However, since such an approach has been restricted to a scalar objective
function, its performance may be limited. Hence, we propose an improvement process
for the Pareto set obtained. This process is still based on the global 2-opt
neighborhood. However, it is performed by directly comparing two job permutations
140
based on the original objective functions in a similar manner as Ravindran et al.
(2005), which used the difference between the objective values obtained. To illustrate,
suppose we have two job permutations π and π′. The makespan and total tardiness time
of jobs for these two job permutations are denoted by Cmax, ST, and C′max, ST′,
respectively. Then, we define the following comparison values R and R′ to determine
which job permutation is better:
R = [(Cmax–Min(Cmax,C′max))/Min(Cmax,C′max)]+[(ST–Min(ST,ST′))/Min(ST,ST′)]
(6.14)
R′= [(C′max–Min(Cmax,C′max))/Min(Cmax,C′max)]+[(ST′–Min(ST,ST′))/Min(ST,ST′)]
(6.15)
If R is less then R′, then the best job permutation is π. Otherwise, the best job
permutation is π′.
In summary, after the maximum number of allowed iterations (MaxCloud) has been
reached, the exploration and exploitation phases terminate. Then, the final Pareto set
obtained will be improved by the improvement process described above. The
improvement process will terminate after ⎣⎢ N / 3⎦⎥ iterations.
A flow chart of the MOWFA for the MOFFSP with intermediate buffers is shown
in Figure 6.7.
6.4.1 Generation of Test Problems and Benchmark Problem Set
Computational experiments have been conducted to evaluate the performance of the
proposed MOWFA. The data required for the MOFFSP with intermediate buffers
141
MOWFA representation Hydrological cycle process
Start
Initialization
(i = 1)
Divide the problem into

N subproblems (distinct regions)
Construct a seed job permutation

using the FFLL algorithm
Use the swap scheme for the seed When it rains, the DOWs are
job permutation to construct generated and they fall to
a set of N job permutations the ground
Distribute a set of DOWs

Exploration phase h
corresponding to the set of job

permutations into N distinct regions
In each distinct region, a cloud randomly generates

MaxPop different positions of DOWs that do not
belong to the E-list
Determine a pair of weights (c1, c2)

for constructing a scalar objective function
corresponding to each distinct region
In each distinct region, apply the hill-sliding algorithm to the job permutations of from higher positions to
generated DOWs with a local 2-opt neighborhood structure and the H1-variant procedure lower positions (local
for scalar objective function f(x) to find their local optimal job permutations optimal positions) and they
are held there
Evaluate the falling force of precipitation based on the scalar objective
function value f(x) obtained by the H1-variant procedure and identify the
number of DOWs at the local optimal job permutations
i = i +1
Update the Pareto-list and the UE-list

Pick one job Choose the un-eroded when the erosion condition
permutation in UE-list Evaluate topology around
direction with the smallest is satisfied at the local
Determine erosion the job permutation using
Perform erosion topology parameter to optimal positions
capability at the the corresponding 2-opt
schemes I and II perform erosion process
job permutation neighborhood structure and
corresponding to each
weights (c1,c2)
erosion scheme
Satisfy erosion
Yes
No
condition?
Update the UE-list to Is a better
No
Exploitation phaseh
continue with local optimal

Update the Pareto-list Yes
performing erosion job permutation
process found?
No
The terrain at eroded
No
positions is smoothed by the
erosion process
No
Have all job
permutations
Update the E-list Yes Have all directions been
in UE-list been
chosen?
considered?
Yes
After a period of time,

Perform evaporation process for DOWs are evaporated into
Satisfy evaporation Yes
DOWs in UE-list the atmosphere. They are
condition?
generated onto the ground
by precipitation
No
i = MaxCloud ? Perform precipitation process

Improvement phase h
Perform the improvement phase

Yes for non-dominated job
permutations in the Pareto-list
Update the Pareto-list
End
Figure 6.7 Flow Chart of the MOWFA for the MOFFSP
142
consist of the number of jobs, the number of stages, the number of machines at each
stage, the number of buffers between consecutive stages, the processing times of jobs,
and the due date of jobs. We generated the data of the test problems based on the
procedure of generating datasets in Rashidi et al. (2010). We constructed the instances
with the number of jobs N = 20, 40, 60, 80, 100; the number of stages S = 2, 4; and the
number of machines in each stage is uniformly distributed in the interval [1, 4]. For
each job, its processing time is uniformly distributed in the interval [10, 100]. When
solving the instances with finite buffers, the buffer capacity for all stages in all
instances is set to be 3.The due date of a job is set as follows:
⎧S ⎫ LB
d j = max ⎨∑ p jl ⎬ + 1 × U , (6.16)
j∈N
⎩ l =1 ⎭ 10
where pjl denotes the processing time of job j in stage l, and LB1 denotes the lower
bound of makespan obtained by Wardono and Fathi (2004). Here, U denotes a uniform
random number between 0 and 1. This results in a tight due date for the jobs. For each
combination of jobs and stages, we have generated five different instances.
The benchmark problem set of Wittrock (1988), modified by Wardono and Fathi
(2004), was also used to test the performance of the MOWFA. In these instances, the
transport time (one minute) for the Wittrock instances is added to the processing time
of jobs at stages 2 and 3, respectively, where applicable. The number of machines at
stages 1, 2, and 3 are 2, 3, and 3, respectively. When solving the instances with finite
buffers, we set the buffer capacity to be 3 for all stages.
In this chapter, we have compared the results obtained by the MOWFA with those
obtained by the improved hybrid multi-objective parallel genetic algorithm
(IHMOPGA) of Rashidi et al. (2010) for the generated test problems and the
143
benchmark instances. The IHMOPGA is designed for solving the FFSP with unrelated
parallel machines, sequence-dependent setup times, and processor blocking, which is
one specific scenario of the general FFSP with intermediate buffers that is considered
in this chapter. Hence when compared with the MOWFA in the multi-objective
problem, the procedure of IHMOPGA is still the same as in Rashidi et al. (2010). The
only change in the IHMOPGA is in the input data. In particular, parallel machines at
each stage are identical, and the setup time is zero for all jobs at all machines in each
stage. Moreover, we also use the approach of converting the FFSP with finite buffers
to one with no available buffer space as described in the MOWFA for the IHMOPGA,
as the algorithm is used to solve the FFSP with finite buffers.
All the computational experiments described in this chapter were performed on an
Intel Centrino Duo 1.60 GHz CPU with 1.5 GB of RAM running on Windows XP
Operating System. The MOWFA and IHMOPGA have been coded using Microsoft
Visual Basic 6.0. Here, the computational complexity of the MOWFA for MOFFSP is
determined based on the neighborhood structure used and the erosion process of this
algorithm. In particular, the MOWFA used 2-opt neighborhood structure, and the
worst possibility of the erosion process is to find for all n directions. Thus, the
computational complexity of the MOWFA is estimated to be O(n3).
The choice of parameters for the MOWFA was determined by design-of-experiment
methods. In particular, we carried out several simulations that test the MOWFA on
instances with various values for the controlled parameters, MaxCloud, MaxPop,
MinEro, and T, and we chose the best values for each instance. These best parameter
144
values are shown in Table 6.1. For the IHMOPGA, we used the same parameter values
as Rashidi et al. (2010). In their paper, the authors did not mention exactly what
method is used to choose their parameters.
When comparing with the IHMOPGA, we used the CPU time limit stopping
criterion of N × M × S × 25 as presented in Rashidi et al. (2010). Here, M denotes the
average number of machines in S stages. For the Wittrock instances, we ran the
MOWFA and IHMOPFA with 5 independent replicates for each instance. The best
results obtained from these replicates were used to compare the performance of
MOWFA and IHMOPGA. For the randomly generated instances, we have used the
average results of 5 different instances corresponding to a combination of jobs and
stages to compare the performance of MOWFA and IHMOPGA. Rashidi et al. (2010)
also used 5 different replicates for each experiment when they ran the IHMOPGA.
Table 6.1 Parameter Sets of the MOWFA for Instances of the MOFFSP
Parameter values
Instances
MaxCloud MaxPop MinEro T
Modified Wittrock 10 5 3 3
Generated Dataset 10 10 3 3
6.4.3 Performance Metrics
In multi-objective optimization problems, proper comparison of algorithms is complex
and is often based on the non-dominated solution set obtained by the algorithms. Three
main aspects that are usually considered for evaluating the non-dominated solution set
obtained by the algorithms are the number of non-dominated solutions in the Pareto set
obtained, the distance of the Pareto front obtained to the Pareto optimal front, and the
145
diversity of non-dominated solutions in the Pareto set obtained (Silva et al., 2004).
Comparison metrics can also be based on other aspects, such as a direct comparison
between two algorithms based on the number of non-dominated solutions obtained by
the algorithms or a comparison based on a union of non-dominated solutions obtained
by the algorithms.
In this chapter, we use four comparison metrics representing all the above aspects.
The first metric, denoted by Ψ1 , is the number of distinct non-dominated solutions
obtained by each algorithm. The second metric, denoted by Ψ 2 , is the fraction of the
number of solutions in a non-dominated set obtained by an algorithm that are
dominated by the non-dominated solutions obtained by another algorithm. The third
metric, denoted by Ψ 3 , is the percentage of non-dominated solutions obtained by each
algorithm in the Pareto optimal set. The fourth metric, denoted by Ψ 4 , is the distance
of the non-dominated front obtained to the Pareto optimal front. As the Pareto optimal
set and Pareto optimal front are usually not known in advance, we would choose
distinct non-dominated solutions from a union of the non-dominated solutions obtained
by all the compared algorithms to form the Pareto optimal set and Pareto optimal
front, which are also the reference points in this comparison. This choice of the
reference points may affect the reliability of the comparison results if the algorithm
used to compare with the WFA has poor performance. Hence, to obtain reliable
comparison results we only choose the IHMOPGA which has been evaluated to have
good performance for solving the multi-objective FFSP.
Let P denote the union of non-dominated solutions obtained by the algorithms.
Since we only compare the MOWFA with the IHMOPGA, P is thus a non-dominated
solution set combined by P1 and P2, which denote the non-dominated solution sets
146
obtained by the MOWFA and the IHMOPGA, respectively. According to Van
Veldhuizen (1999), the first metric is defined as |P1| for the MOWFA and |P2| for the
IHMOPGA, respectively. Let D1 and D2 denote the number of solutions in the non-
dominated solution set obtained by the MOWFA and the IHMOPGA that are
dominated by the solutions in the IHMOPGA and the MOWFA, respectively. Then,
the second metric is defined as:
Di
Ψ i2 = for i = 1, 2. (6.17)
Pi
Let ND1 and ND2 denote the number of solutions in P1 and P2 which are not dominated
by any other solutions in P, respectively. In particular, ND1 and ND2 are determined as
follows:
NDi = Pi − {π i ∈ Pi | ∃π ∈ P : π f π i } for i = 1, 2. (6.18)
Then, the third metric is defined as:
NDi
Ψ i3 = for i = 1, 2. (6.19)
P
For the fourth metric, we use the following distance metric from Knowles and Corne
(2002):
1
Ψ i4 =
P
∑ min {dπ π | π
π ∈P
i, i ∈ Pi } for i = 1, 2, (6.20)
where dπ i ,π denotes the distance between a solution πi in the corresponding non-
dominated solution set Pi and a reference solution π in the non-dominated solution set
P. This is determined by:
dπ i ,π = ( f1* (π ) − f1* (π i ))2 + ( f 2* (π ) − f 2* (π i ))2 for i = 1, 2, (6.21)
147
where f i* (π ) denotes the ith objective normalized by using the reference solution set
P in Ishibuchi et al. (2003).
Among the performance metrics, the first and third ones focus on the diversity of
the non-dominated solution sets obtained, while the second and fourth ones focus on
the closeness of the non-dominated set to the Pareto optimal set. Here, we want to
maximize the diversity and minimize the closeness of the non-dominated solution set.
In this section, we provide the comparison results between the MOWFA and the
IHMOPGA for the MOFFSP with intermediate buffers. Table 6.2 shows the results of
MOWFA and IHMOPGA for the Wittrock instances with no available buffer capacity,
while Table 6.3 shows the results for the instances with finite buffers.
From Tables 6.2 and 6.3, we see that the MOWFA outperforms IHMOPGA in all
instances when we consider the metrics Ψ2, Ψ3, and Ψ4. For metric Ψ1, MOWFA is
worse than IHMOPGA in instance 1 for the case of no buffer (Table 6.2), and in
instances 1 and 6 for the case of finite buffer (Table 6.3). However, metric Ψ1 is not
significant when it is evaluated alone. It can be seen that at least 87.5% of the non-
dominated solutions obtained by the IHMOPGA in instances 1 and 6 are dominated by
the MOWFA (see the Ψ2 metric column of IHMOPGA in Tables 6.2 and 6.3).
Moreover, in instances 1 and 6, 100% of the non-dominated solutions obtained by the
MOWFA cover the final Pareto set (see the corresponding instances in the Ψ3 metric
column of MOWFA in Tables 6.2 and 6.3). From the average values of metrics Ψ1, Ψ2,
Ψ3, and Ψ4 in Tables 6.2 and 6.3, it can also be seen that the MOWFA obtained more
148
diversified and competitive Pareto sets than IHMOPGA in both cases of no buffer and
finite buffers.
Table 6.2 Comparison of MOWFA and IHMOPGA for the Wittrock

Benchmarks with No Buffer
Instance N×S MOWFA IHMOPGA

Ψ1 Ψ2 Ψ3 Ψ4 Ψ1 Ψ2 Ψ3 Ψ4
1 51×3 8.00 0.00 100.00 0.00 9.00 100.00 0.00 13.67
2 38×3 6.00 0.00 100.00 0.00 5.00 100.00 0.00 24.79
3 38×3 7.00 0.00 100.00 0.00 6.00 100.00 0.00 25.85
4 36×3 6.00 0.00 100.00 0.00 5.00 80.00 16.67 18.26
5 40×3 9.00 0.00 100.00 0.00 9.00 100.00 0.00 10.17
6 30×3 8.00 12.50 87.50 2.24 7.00 85.71 12.50 10.87
Average 7.33 2.08 97.92 0.37 6.83 94.29 4.86 17.27
Table 6.3 Comparison of MOWFA and IHMOPGA for the Wittrock

Benchmarks with Finite Buffers

Ψ1 Ψ2 Ψ3 Ψ4 Ψ1 Ψ2 Ψ3 Ψ4
1 51×3 7.00 0.00 100.00 0.00 8.00 87.50 14.29 16.46
2 38×3 14.00 28.57 76.92 0.86 11.00 72.73 23.08 6.65
3 38×3 8.00 12.50 70.00 1.78 8.00 62.50 30.00 5.23
4 36×3 9.00 0.00 90.00 1.43 9.00 88.89 10.00 8.36
5 40×3 4.00 0.00 100.00 0.00 4.00 75.00 25.00 12.5
6 30×3 7.00 0.00 100.00 0.00 8.00 87.50 14.29 9.38
Average 8.17 6.85 89.49 0.68 8.00 79.02 19.44 9.76
We further illustrate the comparison results between MOWFA and IHMOPGA for
Wittrock instances with no buffer and finite buffers in Figures 6.8 and 6.9,
respectively. It can be seen that the solutions obtained by the MOWFA cover more
area of the Pareto set than those obtained by the IHMOPGA. Also, the solutions of the
MOWFA are closer to the Pareto set than that of IHMOPGA.

149
Pareto Solution Set

Pareto Solution Set Pareto Solution Set
11400
13000 9400
12500 11350
9350
T o ta l ta r d i n e s s
12000
Total tardiness
9300 11300
11500
MOWFA MOWFA MOWFA
11000 9250 11250
IHMOPGA IHMOPGA IHMOPGA
10500 9200
11200
10000
9500 9150
11150
9000 9100
800 810 820 830 840 850 860 825 830 835 840 845 850 11100
810 815 820 825 830 835 840 845
Cmax Cmax
Cmax
a) Instance 1 b) Instance 2 c) Instance 3

Pareto Solution Set
5550
12000
7000
11800 5500
6950
11600
6900
T o t a l ta r d i n e s s
5450
T o ta l ta rd in e s s
11400
Total tardiness
6850
MOWFA 11200 MOWFA
5400
MOWFA
6800
IHMOPGA 11000 IHMOPGA 5350
IHMOPGA
6750
10800
6700 5300
6650 10600
10400 5250
6600
818 820 822 824 826 828 830 10200 5200
960 970 980 990 1000 1010 1020 1030 1040 680 685 690 695 700 705 710
Cmax
Cmax Cmax
d) Instance 4 e) Instance 5 f) Instance 6
Figure 6.8 Plot of Pareto Fronts Obtained by MOWFA and IHMOPGA on Wittrock Instances with No Buffer
150

Pareto Solution Set
9200 11200
10400
9100
10200 11100
10000 9000
11000
Total tardiness
T o tal tard in ess

9800 8900
10900
9600 MOWFA 8800 MOWFA
MOWFA
9400 10800
9200 IHMOPGA 8700 IHMOPGA IHMOPGA
10700
9000 8600
8800 10600
8500
8600 10500
8400
8400
8300 10400
760 765 770 775 780 785 790 795
760 770 780 790 800 810 820 830 765 770 775 780 785 790 795 800 805
Cmax Cmax Cmax
a) Instance 1 b) Instance 2 c) Instance 3

Pareto Solution Set
6800 5300
9545 5250
6700
9540 5200
6600
T o ta l ta rd in e s s
9535 5150
Total tardiness
Total tardiness
6500
MOWFA 9530 MOWFA 5100 MOWFA
6400
IHMOPGA 9525 IHMOPGA 5050 IHMOPGA
6300
9520 5000
6200
9515 4950
6100 4900
9510
6000 960 965 970 975 980 985 990 4850
770 780 790 800 810 820 668 670 672 674 676 678 680
Cmax
Cmax Cmax
d) Instance 4 e) Instance 5 f) Instance 6
Figure 6.9 Plot of Pareto Fronts Obtained by MOWFA and IHMOPGA on Wittrock Instances with Finite Buffers
151
The comparison results for the randomly generated instances are shown in Tables
6.4 and 6.5. Here, Table 6.4 shows the results of MOWFA and IHMOPGA for the
generated instances with no available buffer capacity, while Table 6.5 shows the
results for the generated instances with finite buffers.
From Tables 6.4 and 6.5, we can see that the MOWFA still outperforms IHMOPGA
in all instances. The average values of the metrics Ψ1, Ψ2, Ψ3 and Ψ4 in Tables 6.4 and
6.5 show that the MOWFA can obtain more diversified and competitive Pareto sets
than the IHMOPGA. For the case of no buffer in Table 6.4, although the number of
non-dominated solutions of MOWFA is less than that of IHMOPGA, they cover
96.8% of the area of the Pareto set, while the solutions of IHMOPGA only cover 9.6%
of the area of the Pareto set. Moreover, 88.72% of the solutions obtained by the
IHMOPGA are dominated by that of MOWFA, while the metric value of MOWFA is
only 2.55%. In addition, the distance of the solutions obtained by MOWFA to the
Pareto set is much closer than that of IHMOPGA, i.e., 0.48 and 11.93 for MOWFA
and IHMOPGA, respectively.
For the case of finite buffer in Table 6.5, although the quality of non-dominated
solutions obtained by the IHMOPGA has improved, they are still outperformed by the
non-dominated solutions obtained by the MOWFA. In particular, the dominated
solutions of the IHMOPGA have reduced to 75.72%, and they covered 17.66% of the
area of the Pareto set. They are thus worse than the non-dominated solutions obtained
by the MOWFA.
Through these comparison results, it can be concluded that the MOWFA
outperforms IHMOPGA for the instances tested, and the MOWFA is thus an efficient
algorithm for solving the MOFFSP with intermediate buffers.
152
Table 6.4 Comparison of MOWFA and IHMOPGA for the Randomly

Generated Instances with No Buffer

Ψ1 Ψ2 Ψ3 Ψ4 Ψ1 Ψ2 Ψ3 Ψ4
1 20×2 8.80 0.00 100.00 0.00 10.80 86.91 11.16 4.56
2 40×2 13.80 0.00 100.00 0.00 13.20 88.02 11.65 5.66
3 60×2 3.40 0.00 100.00 0.00 4.80 96.00 5.00 11.22
4 80×2 4.80 0.00 100.00 0.00 4.80 96.00 5.00 13.23
5 100×2 5.20 0.00 100.00 0.00 4.40 96.00 4.00 27.88
6 20×4 11.20 0.00 100.00 0.00 12.40 90.23 10.76 8.06
7 40×4 11.00 25.45 82.00 1.66 11.80 74.55 18.00 6.41
8 60×4 12.20 0.00 100.00 0.00 11.20 82.12 16.41 10.31
9 80×4 7.20 0.00 86.00 3.13 5.20 77.33 14.00 22.71
10 100×4 5.20 0.00 100.00 0.00 6.80 100.00 0.00 9.26
Average 8.28 2.55 96.80 0.48 8.54 88.72 9.60 11.93
Table 6.5 Comparison of MOWFA and IHMOPGA for the Randomly

Generated Instances with Finite Buffers

Ψ1 Ψ2 Ψ3 Ψ4 Ψ1 Ψ2 Ψ3 Ψ4
1 20×2 7.80 0.00 100.00 0.00 7.20 60.15 22.86 15.45
2 40×2 7.80 0.00 100.00 0.00 6.20 100.00 0.00 7.66
3 60×2 3.20 0.00 100.00 0.00 3.00 33.33 63.34 0.55
4 80×2 4.80 0.00 100.00 0.00 2.40 93.33 5.00 49.35
5 100×2 2.20 0.00 100.00 0.00 2.00 100.00 0.00 74.79
6 20×4 6.20 0.00 86.07 0.69 7.20 68.72 13.93 12.08
7 40×4 7.20 13.93 75.56 1.64 5.00 60.00 24.44 17.32
8 60×4 7.80 0.00 78.00 1.30 6.00 66.67 22.00 10.98
9 80×4 4.00 0.00 100.00 0.00 4.00 75.00 25.00 20.20
10 100×4 3.00 0.00 100.00 0.00 4.00 100.00 0.00 51.95
Average 5.40 1.39 93.96 0.36 4.70 75.72 17.66 26.03
153
6.5 Conclusions
In this chapter, a MOWFA is proposed for solving the MOFFSP with intermediate
buffers. To solve the problem by the proposed algorithm, we divide the feasible
solution set into multiple distinct regions. Then landscape analysis is performed for
each region to determine the weights for the objectives, which guide the DOWs to
search for local optimal solutions in the corresponding region. Two erosion schemes
are proposed in the MOWFA. The first scheme focuses on exploiting the current
region, while the second scheme helps DOWs exploit other potential regions. We also
incorporate an evaporation process in the MOWFA to enhance the exploitation
capability in potential neighboring regions. The Pareto solution set obtained is
reinforced by an improvement process.
We used the Wittrock benchmark instances and the randomly generated instances to
evaluate the performance of the MOWFA. The proposed algorithm is compared with
other algorithms that can be applied to solve the MOFFSP problem. The comparison
results show that the MOWFA outperforms the other algorithms for the test instances.
These results have been reported in Tran and Ng (2011a).
The MOWFA can be applied to solve the MOFFSP with a large number of
objectives. However, if we use the ellipsoid approximation method to determine the
relative weights of the objectives, the MOWFA can only successfully solve the
MOFFSP with at most three objectives. For the MOFFSP in which the number of
objectives is greater than three, we need to use a linear regression method to determine
the weights.
154
Chapter 7 WFA for Other Combinatorial Optimization Problems
CHAPTER 7
WFA FOR OTHER COMBINATORIAL
In this chapter, WFA is used to solve two other well-known combinatorial optimization
problems which are the quadratic assignment problem (QAP) and vehicle routing problem
(VRP). The QAP and VRP are NP-hard optimization problems often encountered in
facility layout design (Sahni, 1976) and the field of logistics and supply chain
management (Toth and Vigo, 2002), respectively.
For the QAP, a systematic precipitation generating scheme is included in the WFA for
spreading raindrop positions on the ground to increase the solution exploration capability
of the algorithm. Efficient local search methods are also used to enhance the solution
exploitation capability of this algorithm. The performance of the proposed algorithm is
tested with the benchmark instances taken from the QAPLIB (Burkard et al., 1997).
Computational results and comparisons show that the WFA is able to obtain good quality
or optimal solutions to the QAP instances in reasonable computation time.
For the VRP, we have developed a two-level WFA (2LWFA). The first level of the
proposed algorithm is to solve the mathematical programming model of the VRP with the
155
relaxation of the integrality constraints. At the second level, a modified WFA is then
applied to search for optimal solutions from the initial solutions obtained from the first
level. Here, we illustrate the performance of the 2LWFA with the capacitated vehicle
routing problem (CVRP). Some preliminary computational results show the efficiency of
this algorithm for the VRP.
Chapter 7 is organized as follows. In Section 7.1, we present the WFA proposed for
solving the QAP. In particular, we introduce the QAP and its important applications. A
literature review of solution methods for the problem is also described. Then, the
implementation of WFA for solving the QAP is presented. Computational experiments
and comparisons based on the QAP benchmark instances are shown in this section. In
Section 7.2, we present the 2LWFA proposed for solving the CVRP. The CVRP is first
introduced and a description of the 2LWFA is then provided. The results of preliminary
experiments carried out on the CVRP benchmark instances from the literature are also
shown. Finally, some conclusions of this chapter are provided in Section 7.3.
7.1 Quadratic Assignment Problem
7.1.1 Introduction
The QAP is one of the well-known NP-hard optimization problems (Sahni, 1976) due to
its important applications in practice, such as parallel and distributed computing (Bokhari,
1987), statistical data analysis (Hubert, 1987), testing of electronic devices (Eschermann
and Wunderlich, 1990), plant layout design (Rossin et al., 1999), data visualization
(Abbiw-Jackson et al, 2006), printed circuit board assembly process (Duman and Or,
156
2007), and website structure improvement (Qahri Saremi et al., 2008). The problem is
often described as follows: Given a set of facilities and a set of locations with the same
size n, assign the facilities to the locations such that the total cost of assignment is
minimized. The total cost w is calculated using the distance between locations and the
flow between facilities. The QAP can then be expressed as the problem of finding a
permutation π of n facilities as follows:
n n
min w (π ) = ∑ ∑ f ij d π [ i ]π [ j ] , (7.1)
π ∈Π n
i =1 j =1
where Π n denotes the set of possible permutations of N = {1, 2, …, n}, while π [i ]
and π [ j ] denote the location of facilities i and j in the permutation π, respectively.
Furthermore, fij is the flow between facilities i and j, and dπ [i]π [ j ] is the distance between
locations π [i ] and π [ j ] .
While there are some well solved special cases of the QAP (see for example,
Demidenko et al. (2006)), the QAP is generally difficult to solve (Taillard, 1995), and so
researchers have explored various possible solution methods. These solution methods can
be broadly classified into exact methods and heuristic methods. Exact methods aim to find
the optimal solution to the QAP but they often run into computational difficulties with
large QAP instances (Burkard et al., 1996; Cela, 1998; Hahn et al., 2001; Blanchard et al.,
2003; and Gasimov and Ustun, 2007). Heuristic methods on the other hand can be further
divided into constructive heuristics and improvement heuristics. Constructive heuristics
often construct a feasible solution for the QAP by assigning each facility to a location
according to some principles. Such constructive methods for solving the QAP can be seen
157
in Buffa et al. (1964), Arkin et al. (2001), and Gutin and Yeo (2002). Unlike constructive
heuristics, improvement heuristics attempt to improve an existing solution through some
iterative procedures. Some of the well-known algorithms belonging to this category are
the meta-heuristic algorithms, such as genetic algorithms (Ahuja et al., 2000; and Lim et
al., 2002), ant colony optimization algorithms (Maniezzo and Colorni , 1999; Ramkumar
et al., 2009; and Wong and See, 2010), greedy randomized adaptive search procedure (Li
et al., 1994), memetic algorithm (Merz and Freisleben, 2000), iterated fast local search
algorithm (Ramkumar et al., 2008), simulated annealing algorithm (Singh and Sharma,
2008), and DNA algorithm (Yang et al., 2008).
In this chapter, we develop the WFA for solving the QAP. It uses distributed drops of
water (DOWs) to represent the permutations in the QAP. Benchmark problem instances
drawn from the QAPLIB (Burkard et al., 1997) are used to evaluate the performance of
the proposed algorithm. The computational results and comparisons show that the WFA is
able to obtain good quality or optimal solutions to these instances and is a promising
nature-inspired algorithm for solving QAP problem instances.
7.1.2 WFA for the QAP
In this section, we describe the procedure of solving the QAP by the WFA in detail. The
main phases in the WFA include the exploration phase, the exploitation phase, and the
improvement phase, as well as a systematic precipitation generating scheme. We first
begin with a description of the basic components for the WFA.
158
7.1.2.1 Encoding Scheme and Memory Lists
For the QAP, we consider the permutation π of n facilities in the problem as the longitude
and latitude in the position of DOW on the ground, while the total cost of flow between
the facilities is encoded as the altitude. Given a permutation of n facilities π = (σ 1 ,..., σ n ) ,
we thus define:
longitude(π ) = (σ 1 ,..., σ ⎢ n ⎥ ), (7.2)

⎢⎣ 2 ⎥⎦
latitude(π ) = (σ ⎢ n ⎥ ,..., σ n ), (7.3)

⎢ 2 ⎥ +1
⎣ ⎦
where ⎢⎣ q ⎥⎦ is the largest integer less than or equal to q. Figure 7.1 shows an illustrative
example of a DOW and its positional vector components for the QAP with n = 8 facilities.
Permutation π
Total cost
σ1 σ2 σ3 σ4 σ5 σ6 σ7 σ8 w
DOW 7 8 5 3 1 2 6 4 9428
Altitude
Longitude Latitude
Figure 7.1 An Example of Solution Representation in the WFA for the QAP
With this encoding scheme, the neighborhood structure used in the WFA for the QAP
is mainly based on exchanging elements in a permutation. An example is the 2-opt
neighborhood structure used in the traveling salesman problem as well as in the QAP
159
(Merz and Freisleben, 2000). Thus if π' is the permutation obtained by exchanging
positions of two facilities i and j in the permutation π, we can determine π' by:
π ' [k ] = π [k ] for k ∈ N \ {i, j} ,

π ' [i ] = π [ j ] ,
π ' [ j ] = π [i ] . (7.4)
In addition, we also consider the neighborhood structure of an extended 2-opt
algorithm called the 2-opt mirror. In this 2-opt mirror algorithm, other than the usual 2-opt
neighboring permutations, we also consider the reflected permutation and its
corresponding 2-opt neighboring permutations. The reflected permutation π r of a
permutation π can be determined as follows:
π r [i ] = π [ n − i + 1] for i =1,..,n .
(7.5)
The neighborhood structures are used for both exploration and exploitation phases.
To support the search for global optimal permutations, three sets of memory lists,
namely the best permutations list (P0-list), the un-eroded permutations list (UE-list), and
the eroded permutations list (E-list), have been adopted to solve the QAP. Here the
functions of the lists and the procedure of updating these lists are similar as that described
in Chapter 4.
Let π cBest be the best optimal permutation found so far by the WFA at cloud c. Assume
that we have found a local optimal permutation π c*+1 at cloud c + 1. Updating the P0-list at
the cloud can then be described as follows:
160
⎧ π c*+1 if w(π c*+1 ) < w(π cBest ),

P0 -list = ⎨ Best
⎩ πc otherwise.
(7.6)
Updating of the UE-list includes two phases, the phase of removing the local optimal
permutations eroded and the phase of adding the ones just found, which are done in
succession. Updating of the E-list is only to add the local optimal permutations eroded.
They are shown as follows:
UE-list c +1 = ( UE-list c ∪ Π c2 ) \ Π1c , (7.7)
E-list c +1 = E-list c ∪ Π1c .

(7.8)
where Π1c and Π c2 denote the set of local optimal permutations eroded and the set of local
optimal permutations just found from initial permutations generated at each iteration,
respectively.
7.1.2.2 Exploration Phase
Two schemes for generating the initial population of DOW positions in the exploration
phase are proposed below.
The first scheme is the random permutation generator scheme. In this scheme, a
population Ω of permutations of DOWs is generated randomly for each cloud and the
number of such permutations is the maximum population size allowed (MaxPop). Since
the WFA mimics the property of water flow always moving from higher positions to
lower positions due to Earth’s gravity, a steepest descent hill sliding algorithm is then
161
applied to search for local optimal permutations from these initial permutations. In
particular, from an initial solution, the hill sliding algorithm searches for the best
improved solution within the initial solution’s neighbors in terms of objective value. Then,
this process continues to be performed iteratively for the improved solution obtained until
no other improved solution is found.
Due to the random nature of this scheme, the efficiency of WFA for solving the QAP
may fluctuate in instances with large size. To resolve this drawback of the first scheme as
well as to improve the solution exploration capability of WFA for QAP instances with
large size, a second scheme is proposed. This scheme is a systematic DOW generator
scheme that aims to distribute DOWs evenly into divided regions of the solution space.
We first divide the solution space into n regions, where n is the instance size. Then at each
cloud, the WFA generates n DOWs and each DOW is assigned to only one distinct region.
This means that a cloud would consist of n different permutations, which is achieved by
fixing the first position of a facility in a permutation from 1 to n when generating the
permutations of DOW. The rest of the positions in a permutation are assigned randomly.
This can be represented with the following notation:
{
Ω c = ⎡⎣π c1 ,..., π ci ,..., π cn ⎤⎦
T
}
π ci [1] = i , (i = 1,..., n ) , ( c = 1,..., MaxCloud ). (7.9)
A steepest descent hill sliding algorithm is also used in the second scheme to search for
local optimal permutations from these initial permutations.
At each cloud, the number of DOWs at the initial permutations is updated as follows:
162
π ci +1 ⎧Qcπ + 1 if π ci +1 = π ∈ Π 3c ,
Qc +1 =⎨ for i = 1,..., MaxPop,
⎩ 1 otherwise,
(7.10)
where Qcπ is the number of DOWs in permutation π at cloud c, and π ci +1 denotes the
permutation of the ith DOW generated at cloud c + 1, while the set of initial permutations
generated or local optimal permutations found through clouds 1 to c is denoted by Π 3c .
After the steepest descent hill sliding algorithm has been applied to find the local
optimal permutation π c*+1 from the initial permutation π ci +1 , we also update the number of
DOWs at the optimal permutation according to the following equation.
⎧⎪Qπ c+1 + Qcπ+c1+1

i
π c*+1 ∈ UE-list ,
*
π c*+1 if
Qc +1 =⎨ c i
π
⎪⎩ Qc +c1+1 otherwise.
(7.11)
In general, the solution exploration phase of the WFA for the QAP results in a set of
local optimal permutations. They are updated in the UE-list to be considered for the
erosion process in the next exploitation phase.
7.1.2.3 Exploitation Phase
The exploitation phase involves applying the erosion process to overcome the local
optimal permutations found in the exploration phase. Before describing the erosion
process, the erosion condition and capability are first described below:
163
7.1.2.3.1 Erosion Condition and Capability
The erosion process is triggered by the amount of precipitation and so if the number of
DOWs at a local optimal permutation increases to the threshold MinEro, the erosion
process is performed at this local optimal permutation.
Next, we consider the capability L of the erosion process. For the QAP, the erosion
capability is based on a factor, which is the number of DOWs at the eroding local optimal
permutation. In particular, the relationship between the erosion capability and this factor is
a nonlinear function. However, to simplify the computations in the WFA, we have set the
erosion capability to a constant value MaxUIE, so that the relationship can be described as
follows:
⎧⎪ MaxUIE
( ) if Qcπ c ≥ MinEro ,
*
π c*
L Qc =⎨ (7.12)
⎪⎩0 otherwise.
If the erosion process cannot find any improved permutation after MaxUIE search
steps, the erosion process stops and other permutations in the UE-list are considered for
performing the next erosion process.
7.1.2.3.2 Erosion Process
The erosion process is the main operator in the exploitation phase of WFA for solving the
QAP. Its task is to help the DOWs overcome local optimal permutations and obtain better
local optimal or global optimal permutations. In the QAP, the erosion process depends on
a topological parameter Δd h representing the geographical shape of the surface. It is
164
defined as the difference of total cost between the local optimal permutation and its hth
neighboring permutation:
( ) n(n − 1)
n n
Δd h = w (π h* ) − w (π * ) = ∑∑ fij dπ * [i ]π * [ j ] − dπ * [i ]π * [ j ] for h = 1,..., ,
i =1 j =1
h h
2
(7.13)
where π * and π h* denote the local optimal permutations and its hth neighboring
permutation respectively. In the case of the 2-opt mirror neighborhood, the number of
directions is n(n − 1) + 1 .
The aim of computing Δd h is to help the erosion process choose the most suitable
direction to perform erosion. The erosion process will choose the smallest Δd h to be the
first erosion direction. Searching for this direction will stop if the erosion process for the
erosion direction cannot find a better permutation after MaxUIE steps. Then, the erosion
process will be restarted with another erosion direction with the next smallest Δd h . If the
erosion process cannot find a better permutation for all the directions, we move the eroded
local optimal permutation into the E-list so that it will not be considered for erosion
process in the next clouds. On the other hand, if the erosion process is able to find a better
permutation than the eroding local optimal permutation, that erosion direction is chosen to
erode the local optimal permutation permanently. Then the new local optimal permutation
is updated in the UE-list to continue with performing the erosion process.
165
7.1.2.4 Improvement Phase
The exploration and exploitation phases of the WFA for the QAP terminate when the
maximum number of clouds (MaxCloud) has been generated. To improve the solutions
obtained by the WFA, we can apply the 3-opt algorithm at a final improvement phase.
This 3-opt algorithm would still use the steepest descent hill sliding method mentioned in
the previous section. A 3-opt list is also used to save the solutions obtained by this
improvement algorithm.
In summary, we have the following variants of WFA:
(1). Random 2-opt WFA: WFA with the random permutation generator scheme and
the 2-opt neighborhood structure.
(2). Systematic generator 2-opt WFA: WFA with the systematic DOW generator
scheme and the 2-opt neighborhood structure.
(3). Random 2-opt mirror WFA: WFA with the random permutation generator
scheme and the 2-opt mirror neighborhood structure.
A flow chart of the WFA for solving the QAP is shown in Figure 7.2.
7.1.3 Computational Experiments and Comparisons
To test the performance of the proposed WFA, we have used the random 2-opt WFA for
solving all the benchmark instances from the QAPLIB (Burkard et al., 1997). The
systematic generator 2-opt WFA and the random 2-opt mirror WFA are used to solve the
benchmark instances when the random 2-opt WFA has not obtained the best known value.
166
Initialization
Generate a new DOW

with permutation in
region i
Yes When it rains, the DOWs are
generated and they fall to
i=i+1 Belong to E-list? the ground
Exploration phase h
No
No
Spread to all n regions?
Yes
Apply the steepest ascent hill sliding algorithm

to search for the local optimal permutations of
DOWs generated in n regions
Determine the number of DOWs at the local lower positions (local
optimal permutations optimal positions) and they
are held there
No
Choose one permutation in

UE-list

Evaluate topology Perform erosion process for when the erosion condition
Satisfy erosion
Yes parameter around the un-eroded direction with the is satisfied at the local
condition?
permutation smallest topology parameter optimal positions
No
Update into UE-list to

Is a better local
No continue performing Yes
optimal permutation No
erosion process found?
No
The terrain at eroded

Have all permutations in UE-list Update E-list Yes Have all directions positions is smoothed by the
been chosen? been chosen?
erosion process
Yes
Update P0-list
Yes
Choose one best permutation in

P0-list
Improvement phase h
Apply 3-opt algorithm for the

permutation
No
Find better permutation? Yes Update 3-opt list
No
Have all permutations Output the best permutations

Yes
in P0-list been chosen? in 3-opt list
Figure 7.2 Flow Chart of the WFA for the QAP

167
7.1.3.1 Benchmark Problem Sets
The 134 instances drawn from the QAPLIB (Burkard et al., 1997) are well-known
benchmark problem sets in QAP with size ranging from 12 to 256. The best known upper
bounds of these problem sets obtained from the literature were used to compare with the
best results obtained by the WFA. In addition, we also compared these results with those
obtained by greedy randomized adaptive search procedure (GRASP) of Li et al. (1994), by
ant system (ANT) of Maniezzo and Colorni (1999), by greedy genetic algorithm (GGA) of
Ahuja et al. (2000), by hybrid genetic algorithm with partial local search (PGA) of
Lim et al. (2002), by iterated fast local search algorithm (IFLS) of Ramkumar et al. (2008),
by two-level modified simulated annealing based approach (MSA) of Singh and Sharma
(2008), and by population-based hybrid ant system (PHAS) of Ramkumar et al. (2009), all
of which are some of the most efficient meta-heuristic algorithms for solving the QAP
instances. Almost all the algorithms used the design-of-experiment method or trial runs to
select their best parameter values. However, the number of runs of each algorithm can be
different, such as 5 runs used for the GRASP, ANT, and PHAS, 7 runs used for GGA, and
10 runs for PGA. Also, some algorithms did not provide information on the number of
runs used, such as IFLS and MSA. Thus, to have a fair comparison, we used the smallest
number of runs that several algorithms have used, i.e., 5 runs.
7.1.3.2 Platform and Parameters
All the computational experiments were performed on an Intel Centrino Duo 1.60 GHz
CPU with 1.5 GB of RAM. The WFA has been coded using Microsoft Visual Basic 6.0.
Here, the computational complexity of the WFA for the QAP is determined based on the
168
neighborhood structures used and the erosion process of this algorithm. In particular, the
WFA used 2-opt neighborhood structure, and the worst possibility of the erosion process
is to find for all n directions. Hence, the computational complexity of the WFA is
estimated to be O(n3). Although the WFA for the QAP used 3-opt neighborhood structure
at the improvement phase, since the 3-opt algorithm is only used once, the computational
complexity of the WFA for the QAP is still O(n3).
The choice of parameters for WFA was determined by the design-of-experiment
method. In particular, we have carried out several simulations that test the WFA on all
types of QAP with various values for the controlled parameters, i.e., the exploration
parameters MaxCloud and MaxPop, and the exploitation parameters MaxUIE and MinEro.
The aim is to determine the best parameters of WFA for QAP that would achieve a
balance between solution exploration and exploitation capabilities in finding the best
solutions within reasonable computation time. Thus, smaller values may be used by the
parameters for larger problem instances. The values that were used are as follows:
MaxCloud = 2, 5, 10, 15, 20; MaxPop = 5, 10, 15, 20; MaxUIE = 5, 10, 15, 20; and
MinEro = 2, 3. From preliminary simulation results, the best parameter sets are shown in
Table 7.1. These parameter sets are then used for the random 2-opt WFA and the random
2-opt mirror WFA. For the systematic generator 2-opt WFA, we have used the parameter
sets (MaxCloud, MaxPop, MaxUIE, MinEro) = (50, n, 20, 2), (20, n, 10, 2), and (5, n, 5, 2)
for instances with size at most 50, more than 50 but at most 100, and more than 100
respectively, in order to allow a reasonable amount of erosion process to occur. With these
parameter sets, 5 independent replicates were used for each instance and the best results
were used to compare the WFA with other meta-heuristic algorithms.
169
Table 7.1 Parameter Sets of WFA for the QAP Benchmark Instances
Parameter values
Benchmarks n
MaxCloud MaxPop MaxUIE MinEro
Burkard 26 5 10 10 2
Christofides 12 – 20 10 10 10 2
22 15 10 10 2
25 20 10 10 2
Elshafei 19 10 10 10 2
Eschermann 16, 64 2 5 5 2
32 (a, b) 5 10 10 2
32 (c, e, g) 2 5 5 2
32 (d, h) 2 10 10 2
128 5 10 5 3
Hadley 12 – 20 10 10 10 2
Krarup 30, 32 20 10 10 2
Li & Pardalos 20, 30 10 10 10 2
40, 50, 60 10 20 10 2
70 10 20 15 2
80, 90 20 20 10 2
Nugent 12 – 28 10 10 5 2
30 20 10 10 2
Roucairol 12, 15 5 10 10 2
20 10 20 10 2
Scriabin 12, 15, 20 5 15 10 2
Skorin-Kapov 42 – 64 15 10 10 2
72, 81, 90 10 20 15 2
100 10 10 10 2
Steinberg 36 20 10 10 2
Taillard
(Taixxxa) 12 5 10 10 2
15, 17 5 20 10 2
20 – 35 20 20 10 2
40, 50 20 20 15 2
60, 80, 100 10 20 10 2
(Taixxxb) 12 – 20 5 5 10 2
25 10 10 10 2
30, 35, 40 20 10 10 2
50, 60, 80 10 10 15 2
100 5 20 10 2
150 5 10 5 2
(Taixxxc) 64 10 20 10 2
256 2 5 5 2
Thonemann 30 10 20 10 2
40 20 10 10 2
150 5 10 10 2
Wilhelm 50 15 20 10 2
100 10 10 10 2
170
7.1.3.3 Performance Measures
For comparison of objective values, we have used the following relative percentage
difference in objective value:
⎛ Heuristicsol − Optsol ⎞
Δ Best = ⎜ ⎟ × 100, (7.14)
⎝ Opt sol ⎠
where Heuristicsol and Optsol denote the best objective function value obtained by the
WFA and the best known value in the literature, respectively.
The best known values from the literature are used as the reference values for the
evaluation and comparison of the WFA and other algorithms for solving the QAP. These
values may be obtained from the optimal solutions of the benchmark instances used, or the
best solutions found by some algorithm so far. Thus, the comparison results may be
negative values if the WFA finds a better solution.
For evaluation and comparison of the overall performance of algorithms, we have used
criteria such as the average relative percentage difference for all instances solved and the
number of the best known solutions obtained in all instances solved. Thus, the algorithm
with a smaller average relative percentage difference and a larger number of the best
known solutions obtained would be considered to be a more effective optimization method.
171
7.1.3.4 Computational Results
The computational results and comparisons with other meta-heuristic algorithms, such as
GRASP, ANT, GGA, PGA, IFLS, MSA, and PHAS, are shown in Tables 7.2a to 7.2e,
while the improvement results obtained by the variants of WFA for the QAP benchmark
instances that have not been solved optimally by the 2-opt WFA are displayed in Table 7.3.
In Tables 7.2a to 7.2e, the column with the best results of WFA shows the best solutions
obtained by the random 2-opt WFA and the variants of WFA. Since the details of the
computation time of applying PHAS to the QAP instances were not shown in Ramkumar
et al. (2009), we did not include this information in Tables 7.2a to 7.2e. In addition, the
symbol “—” in the entries of these tables is used to indicate that the compared algorithms
have not solved the respective benchmark instances. For Table 7.3, the entries displayed in
italic font highlight the best results obtained by the respective variant of the WFA.
From Table 7.3, we can see that the systematic generator 2-opt WFA, the random 2-
opt mirror WFA and the WFA-3-opt have obtained better results than the random 2-opt
WFA for some of the instances when the random 2-opt WFA was unable to obtain the best
known solution. In particular, these variants improved the solution quality for 25 instances.
From the best results obtained by the WFA in Tables 7.2a to 7.2e, it can be seen that
out of the 134 instances from the QAPLIB (Burkard et al., 1997), the best known solutions
for 99 instances have been obtained by the WFA within reasonable computation time. The
WFA is also able to obtain solutions with a relative percentage difference of less than 2%
for all the remaining instances. The average relative percentage difference of WFA for all
the 134 instances is found to be 0.20%.
172
Table 7.2a Comparison Results of the WFA with Other Algorithms for Burkard’s and Christofides’ Instances
Best results of
Best Random 2-opt WFA GRASP ANT GGA PGA IFLS MSA PHAS
WFA
Instances known
Best Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time
value Best Time (s) Δ Best Best Best Best Best Best Δ Best
solution (s) (s) (s) (s) (s) (s) (s)
Bur26a 5426670 5426670 140.0 0 140.0 0 11.38 0 21.07 0 235 0 125 0 61.27 — — 0
Bur26b 3817852 3817852 108.0 0 108.0 0 59.45 0 35.03 0 225 0 9.5 0 60.27 — — 0
Bur26c 5426795 5426795 80.0 0 80.0 0 5.16 0 19.09 0 227 0 7.42 0 57.78 — — 0
Bur26d 3821225 3821225 126.0 0 126.0 0 15.12 0 19.4 0 213 0 8.42 0 61.27 — — 0
Bur26e 5386879 5386879 145.0 0 145 0 17.63 0 20.53 0 218 0 10.03 0 57.83 — — 0
Bur26f 3782044 3782044 197.0 0 197 0 5.05 0 11.23 0 204 0 6.68 0 59.19 — — 0
Bur26g 10117172 10117172 129.0 0 129 0 22.58 0 18.67 0 194 0 9.99 0 57.72 — — 0
Bur26h 7098658 7098658 97.0 0 97 0 37.58 0 5.67 0 204 0 6.82 0 57.47 — — 0
Chr12a 9552 9552 2.1 0 2.1 — — — — 0 19.6 0 0.54 0 1.09 0 40 0
Chr12b 9742 9742 1.8 0 1.8 — — — — 0 18.4 0 0.42 0 1.11 0 41 0
Chr12c 11156 11156 2.2 0 2.2 — — — — 0 20.2 0 1.29 0 1.02 0.26 38 0.27
Chr15a 9896 9896 11.4 0 11.4 — — — — 0.4 40.6 0 1.5 0 2.97 0 69 0
Chr15b 7990 7990 15.4 0 15.4 — — — — 0 41.8 0 1.31 0 3.08 2.7 72 0
Chr15c 9504 9504 14.3 0 14.3 — — — — 0 44 0 1.30 0 2.64 11.5 69 6.36
Chr18a 11098 11098 40.0 0 40 — — — — 0.4 79 0 2.11 5.14 7.23 1.71 103 14.25
Chr18b 1534 1534 37.0 0 37 — — — — 0 78.8 0 2.62 0 5.30 0 105 0
Chr20a 2192 2192 149.0 0 149 1.82 509 0 331 0 94.6 0.18 3.61 4.38 10.95 0 131 1.82
Chr20b 2298 2298 127.0 0 127 5.92 195 2.79 375 5.13 96.4 3.12 3.32 5.40 8.61 0 127 4.96
Chr20c 14142 14142 72.0 0 72 0.00 9.23 0 29.49 0 97.8 4.51 1.77 0 13.55 0 140 0
Chr22a 6156 6156 285.0 0 285 2.31 201 0 315 0.75 146 0 4.52 0.88 19.11 5.7 164 0.32
Chr22b 6194 6194 283.0 0 283 2.58 213 0.97 162 0 152 1.46 5.26 1.68 17.00 8.5 163 0
Chr25a 3796 3796 455.0 0 455 2.32 115 0 236 0 194 2.27 5.97 11.17 33.59 0 591 0

173
Table 7.2b Comparison Results of the WFA with Other Algorithms for Elshafei’s, Eschermann’s, Hadley’s, and Krarup’s
Instances
Best results
of WFA
Instances known
Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ
value Best solution Best Best Best Best Best Best Best Best
(s) (s) (s) (s) (s) (s) (s) (s)
Els19 17212548 17212548 67 0 67 — — — — 0 80.6 0 44.46 — — — — 0
Esc16a 68 68 0.1 0 0.1 — — — — 0 47.4 0 5.13 0 3.17 0 61 0
Esc16b 292 292 0.1 0 0.1 — — — — 0 48.2 0 0.19 0 2.75 0 60 0
Esc16c 160 160 0.1 0 0.1 — — — — 0 53.4 0 0.44 0 4.03 0 68 0
Esc16d 16 16 0.1 0 0.1 — — — — 0 53.2 0 0.5 0 3.98 — — 0
Esc16e 28 28 0.1 0 0.1 — — — — 0 46.8 0 0.32 0 2.28 — — 0
Esc16f 0 0 0.1 0 0.1 — — — — 0 46.0 — — 0 1.11 — — 0
Esc16g 26 26 0.1 0 0.1 — — — — 0 49.8 0 0.29 0 2.77 — — 0
Esc16h 996 996 0.1 0 0.1 — — — — 0 48.0 0 0.22 0 2.13 0 65 0
Esc16i 14 14 0.1 0 0.1 — — — — 0 51.6 0 0.16 0 2.05 — — 0
Esc16j 8 8 0.1 0 0.1 — — — — 0 402 0 0.32 0 2.91 — — 0
Esc32a 130 130 866 0 866 1.54 7.03 0 226 0 382 1.52 97.04 0 137 — — 0
Esc32b 168 168 258 0 258 0 2.80 0 40.59 0 400 0 33.61 0 110 — — 0
Esc32c 642 642 2.6 0 2.6 0 0.00 0 0.08 0 389 0 2.01 0 54.7 — — 0
Esc32d 200 200 39 0 39 0 1.92 0 2.13 0 353 0 2.76 0 74.3 — — 0
Esc32e 2 2 1 0 1 0 0.00 0 0.05 0 370 0 0.66 0 46.09 — — 0
Esc32g 6 6 1 0 1 0 0.00 0 0.07 0 371 0 1.27 0 28.41 — — 0
Esc32h 438 438 141 0 141 0 3.41 0 2.64 0 349 0 6.54 0 85.75 — — 0
Esc64a 116 116 47 0 47 — — — — 0 2631 0 194 0 1522 — — 0
Esc128 64 64 6976 0 6976 — — — — — — 0 1631 — — — — 0
Had12 1652 1652 1.1 0 1.1 — — — — — — 0 4.27 0 0.97 0 41 0
Had14 2724 2724 2.7 0 2.7 — — — — — — 0 10.25 0 1.97 0 64 0
Had16 3720 3720 7.4 0 7.4 — — — — — — 0 5.38 0.05 3.64 0 88 0
Had18 5358 5358 23 0 23 — — — — — — 0 18.54 0 6.52 0 118 0
Had20 6922 6922 48 0 48 0 2.8 0 159 — — 0 15.26 0 10.58 0 148 0
Kra30a 88900 88900 2040 0 2040 0.00 292 0 199 0 301 0.89 71 1.34 106 — — 0
Kra30b 91420 91420 2095 0 2095 0.32 268 0 140 0 331 0 123 0.13 102 — — 0.08
Kra32 88700 88700 2052 0 2052 — — — — — — — — 0 172 — — 0
174
Table 7.2c Comparison Results of the WFA with Other Algorithms for Li & Pardalos’ and Skorin-Kapov’s Instances
Best results of
WFA
Instances known
Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ
value Best solution Best Best Best Best Best Best Best Best
(s) (s) (s) (s) (s) (s) (s) (s)
Lipa20a 3683 3683 65 0 65 0 0.99 0 107 0 74.8 0 0.59 0 16.11 — — 0
Lipa20b 27076 27076 76 0 76 0 0.66 0 0.00 0 74.4 0 0.39 0 16.78 — — 0
Lipa30a 13178 13178 352 0 352 0 46.43 0 54.85 0 345 0 5.66 0 120 — — 0.99
Lipa30b 151426 151426 349 0 349 0 7.31 0 0.00 0 337 0 2.78 0 122 — — 0
Lipa40a 31538 31538 3065 0 3065 1.13 306.00 1.02 281 0.96 1022 0 19.46 0 490 — — —
Lipa40b 476581 476581 1981 0 1981 0 6.21 0 0 0 1026 0 9.52 0 486 — — —
Lipa50a 62093 62619 4099 0.80 4568 — — — — 0.95 1486 0.82 57.32 1.02 1556 — — 0
Lipa50b 1210244 1210244 3578 0 3578 — — — — 0 1509 0 39.96 0 1462 — — 0
Lipa60a 107218 108103 6025 0.79 12396 — — — — 0.77 3057 0.64 137 0.84 3668 — — 0.81
Lipa60b 2520135 2520135 4112 0 4112 — — — — 0 3047 0 86.13 0 3724 — — 0
Lipa70a 169755 170956 8057 0.71 8057 — — — — 0.71 6148 0.62 233 0.77 8067 — — —
Lipa70b 4603200 4603200 8503 0 8503 — — — — 0 6123 15.9 196 0 7762 — — —
Lipa80a 253195 254853 10144 0.63 17685 — — — — 0.61 9519 0.61 373 0.67 15220 — — —
Lipa80b 7763962 7763962 10800 0 10800 — — — — 0 9499 16.56 332 20.33 15965 — — —
Lipa90a 360630 362854 12812 0.57 26123 — — — — 0.58 12358 0.54 592 0.63 27909 — — —
Lipa90b 12490441 12490441 25677 0 25677 — — — — 0 12319 0 503 0 27788 — — —
Sko42 15812 15836 4691 0.03 6961 — — — — 0.25 1006 0.35 365 0.30 614 — — 0
Sko49 23386 23510 5051 0.32 8158 — — — — 0.21 1252 0.19 714 0.45 1318 — — 0.05
Sko56 34458 34568 4792 0.20 8537 — — — — 0.02 2976 0.06 907 0.47 2613 — — 0.12
Sko64 48498 48796 5624 0.31 14405 — — — — 0.22 3788 0.09 1399 0.25 4936 — — 0
Sko72 66256 66660 7454 0.47 19254 — — — — 0.29 5078 0.21 1987 0.73 8663 — — 0.03
Sko81 90998 91452 7449 0.50 7449 — — — — 0.20 10964 0.12 2680 0.43 16960 — — 0.05
Sko90 115534 116922 7640 0.91 21076 — — — — 0.27 12698 0.43 3822 0.45 28787 — — 0.02
Sko100a 152002 153426 8767 0.94 8767 — — — — 0.21 16608 0.22 1486 1.30 309 — — 0.19
Sko100b 153890 155288 6724 0.85 7574 — — — — 0.14 14729 0.30 1405 2.34 274 — — —
Sko100c 147862 149628 7476 1.15 8576 — — — — 0.20 20314 0.06 873 1.50 284 — — —
Sko100d 149576 151196 8335 0.97 19097 — — — — 0.17 20302 0.27 863 1.03 293 — — —
Sko100e 149150 151056 11632 0.90 26534 — — — — 0.24 21127 0.33 745 1.55 301 — — —
Sko100f 149036 150510 9172 0.91 10200 — — — — 0.29 21479 0.41 781 1.73 285 — — —

175
Table 7.2d Comparison Results of the WFA with Other Algorithms for Nugent’s, Roucairol’s, Scriabin’s, Steinberg’s,
Thonemann’s, and Wilhelm’s Instances
Random 2-opt Best results of

Best GRASP ANT GGA PGA IFLS MSA PHAS
WFA WFA
Instances known
Best Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Time Δ Best
value Best Best Best Best Best Best Best
solution (s) (s) (s) (s) (s) (s) (s) (s)
Nug12 578 578 1.6 0 1.6 — — — — 0 19 0 6.84 0 1.41 0 36 0
Nug14 1014 1014 2.0 0 2.0 — — — — — — 0 7.71 0.39 3.11 — — 0.2
Nug15 1150 1150 5.8 0 5.8 — — — — 0 41.4 0 8.3 0 4.02 0 73 0
Nug16a 1610 1610 16.3 0 16.3 — — — — — — 0 11.24 0 5.59 — — 0
Nug16b 1240 1240 18.1 0 18.1 — — — — — — 0 11.02 0 5.59 — — 0
Nug17 1732 1732 22.3 0 22.3 — — — — — — 0 11.95 0 7.31 — — 0
Nug18 1930 1930 31.1 0 31.1 — — — — — — 0.31 13.56 0 9.55 — — 0
Nug20 2570 2570 74.6 0 74.6 0 2.53 0 119 0 97.8 0 20.73 0 16.06 0 132 0
Nug21 2438 2438 135 0 135 — — — — — — 0 29.80 0 20.63 — — 0
Nug22 3596 3596 119 0 119 — — — — — — 0 43.82 0 25.84 — — 0
Nug24 3488 3488 183 0 183 — — — — — — 0 33.83 0 39.75 — — 0.06
Nug25 3744 3744 181 0 181 — — — — — — 0 42.10 0 47.66 — — 0
Nug27 5234 5234 225 0 225 — — — — — — — — 0 80.56 — — 0
Nug28 5166 5166 276 0 276 — — — — — — — — 0.12 98.33 — — 0
Nug30 6124 6124 644 0 644 0.42 523 0 181 0.07 354 0.42 109 2.12 117 0.06 887 0.07
Rou12 235528 235528 1.2 0 1.2 — — — — 0 19.6 0 0.30 0 1.06 0 35 0
Rou15 354210 354210 3.4 0 3.4 — — — — 0 34.6 0 0.56 0 2.95 0.71 71 0
Rou20 725522 725522 57.8 0 57.8 0 165 0 245 0.16 75.2 0 1.43 0.02 11.73 0.06 127 0
Scr12 31410 31410 1.3 0 1.3 — — — — 0 18.8 0 0.44 0 1.11 0 38 0
Scr15 51140 51140 4.6 0 4.6 — — — — 0 35.2 0 0.42 0 3.09 0 78 0
Scr20 110030 110030 23.4 0 23.4 0 157 0 46.1 0 79.6 0 1.57 0 12.69 2.13 137 0
Ste36a 9526 9526 3057 0 3057 1.81 276 0.76 295 0.27 710 0 221 0 204 — — —
Ste36b 15852 15852 3425 0 3425 0.92 180 0.25 213 — — 0 235 3.43 222 — — —
Ste36c 8239110 8239110 3696 0 3696 0.89 142 0.33 321 — — 0 24.07 — — — — —
Tho30 149936 149936 1379 0 1379 0.00 216 0 288 0 396 0 132 0.29 119 — — —
Tho40 240516 240620 3704 0.04 3704 1.17 184 0.66 312 0.32 958 0.05 344 0.53 502 — — —
Tho150 8133398 8238058 21600 1.29 21600 — — — — — — 0.41 729 — — — — —
Wil50 48816 48916 3933 0.06 11856 — — — — 0.07 2115 0 695 0.28 1499 — — —
Wil100 273038 274446 15320 0.34 23725 — — — — 0.2 20544 0.15 1252 0.27 51121 — — —
176
Table 7.2e Comparison Results of the WFA with Other Algorithms for Taillard’s Instances
Best results of
Random 2-opt WFA GRASP ANT GGA PGA IFLS MSA PHAS
Best known WFA
Instances
value Best Time Δ Best Time Δ Best Time Δ Best Time Δ Best Time Δ Best Time Δ Best Time Δ Best Time Δ Best
solution (s) (s) (s) (s) (s) (s) (s) (s)
Tai12a 224416 224416 0.7 0 0.7 — — — — — — 0 0.18 0 1.05 0 35 0
Tai12b 39464925 39464925 0.5 0 0.5 — — — — — — 0 0.62 0 1.03 0.03 53 0
Tai15a 388214 388214 8.1 0 8.1 — — — — — — 0 0.69 0 2.99 0.39 73 0.01
Tai15b 51765268 51765268 4.2 0 4.2 — — — — — — 0 0.7 0 3.05 0.47 77 0
Tai17a 491812 491812 18 0 18 — — — — — — 0.55 0.96 0.38 5.55 0 86 0.56
Tai20a 703482 703482 860 0 860 0 484 0 160 — — 0.84 1.38 0.47 11.42 0.21 124 0.8
Tai20b 122455319 122455319 22 0 22 — — — — — — 0 2.70 0 12.52 5.6 167 0
Tai25a 1167256 1169030 1157 0 7268 1.43 355 0.55 206 — — 0.77 3.27 2.00 33.03 — — 1.57
Tai25b 344355646 344355646 332 0 332 — — — — — — 0 3.77 5.59 35 — — 0
Tai30a 1818146 1832590 1969 0.57 2623 1.58 265 1.46 332 — — 1.34 6.72 1.11 83.06 — — 1.37
Tai30b 637117113 637117113 3169 0 3169 — — — — — — 0 14.11 2.22 81 — — 0
Tai35a 2422002 2436540 2477 0.59 6185 1.90 531 1.64 232 — — 1.29 12.09 1.24 177 — — 1.3
Tai35b 283315445 283315445 4749 0 4749 — — — — — — 0 23.9 3.54 186 — — 0.19
Tai40a 3139370 3160484 4612 0.67 4612 2.20 325 2.05 253 — — 1.08 18.37 1.85 354 — — 1.7
Tai40b 637250948 637250948 5355 0 5355 — — — — — — 0 36.95 5.60 328 — — 0
Tai50a 4938796 5031472 5354 1.46 11342 — — — — — — 1.31 58.21 2.25 1104 — — 2.48
Tai50b 458821517 458926056 7166 0.02 7166 — — — — — — 0 64.77 0.42 1032 — — 0
Tai60a 7205962 7342990 9962 1.55 14450 — — — — — — 1.79 104 2.75 2740 — — 2.37
Tai60b 608215054 612153786 8584 0.65 8584 — — — — — — 0 148 0.47 2621 — — 0.02
Tai64c 1855928 1855928 5834 0 5834 — — — — — — 0 28.96 0.03 237 — — 0
Tai80a 13511780 13821180 10084 1.87 27967 — — — — — — 1.41 360 2.46 11333 — — 2.37
Tai80b 818415043 827982667 15800 1.17 15800 — — — — — — 0.03 424 2.79 10533 — — 0
Tai100a 21052466 21538854 11274 1.76 36544 — — — — — — 1.29 785 2.33 35781 — — —
Tai100b 1185996137 1198498100 20268 1.05 20268 — — — — — — 0.32 855 0.52 34336 — — —
Tai150b 498896643 508566248 44677 1.94 44677 — — — — — — 0.2 3414 0.38 290186 — — —
Tai256c 44759294 44896638 35434 0.27 367250 — — — — — — 0.16 1956 0.27 73180 — — —
177
Table 7.3 Improved Results of the WFA Variants for the QAP Instances Not Optimally Solved by 2-Opt WFA
Systematic generator Random 2-opt

Best Random 2-opt WFA WFA-3-opt
2-opt WFA mirror WFA
Instances known
Best Best Best Best
value Time (s) Time (s) Time (s) Time (s)
solution solution solution solution
Lipa50a 62093 62619 4099 62703 5099 62666 6481 62593 4568
Lipa60a 107218 108103 6025 108172 10534 108070 12396 108070 12396
Lipa80a 253195 254853 10144 254840 17053 254800 17685 254800 17685
Lipa90a 360630 362854 12812 362906 19905 362673 26123 362673 26123
Sko42 15812 15836 4691 15816 6961 15830 8252 15816 6961
Sko49 23386 23510 5051 23460 8158 23474 8971 23460 8158
Sko56 34458 34568 4792 34528 8537 34558 11047 34528 8537
Sko64 48498 48796 5624 48648 14405 48758 13885 48648 14405
Sko72 66256 66660 7454 66570 19254 66820 15072 66570 19254
Sko90 115534 116922 7640 116968 18859 116632 20212 116590 21076
Sko100b 153890 155288 6724 155728 18431 155398 21747 155204 7574
Sko100c 147862 149628 7476 149862 17480 149990 19819 149564 8576
Sko100d 149576 151196 8335 151186 19884 151022 19097 151022 19097
Sko100e 149150 151056 11632 151140 19147 150588 25371 150498 26534
Sko100f 149036 150510 9172 151032 18680 150794 21531 150390 10200
Tai25a 1167256 1169030 1157 1169030 2453 1167256 7268 1167256 7268
Tai30a 1818146 1832590 1969 1828576 2623 1830918 10287 1828576 2623
Tai35a 2422002 2436540 2477 2436458 6185 2443826 9234 2436458 6185
Tai50a 4938796 5031472 5354 5010798 11342 5026322 15100 5010798 11342
Tai60a 7205962 7342990 9962 7317694 14450 7353798 19163 7317694 14450
Tai80a 13511780 13821180 10084 13790286 12708 13764720 27967 13764720 27967
Tai100a 21052466 21538854 11274 21577638 17538 21422344 36544 21422344 36544
Tai256c 44759294 44896638 35434 44879868 367250 44881948 103680 44879868 367250
Wil50 48816 48916 3933 48856 13800 48846 11856 48846 11856
Wil100 273038 274446 15320 274244 18825 274608 27522 273980 23725
178
WFA
Average Percentage Difference
1.0 0.99
GRASP
0.87
0.8 ANT
0.69
GGA
0.6 0.51
(%)
PGA
0.41
0.4 IFLS
0.28 0.24
0.17 0.20 MSA
0.2 0.14 0.13
0.04 0.04 PHAS
0.00
0.0
44 44 87 130 130 40 110 Number of
(a). Instances Solved
100 94 96 91
Number of Best Known
84
Solutions Obtained
80 80
72
66
60 58
40 40 40
40
2 33
26
20
0
44 44 87 130 130 40 110 Number of
(b). Instances Solved
Figure 7.3 Comparison of WFA with Other Algorithms on (a). Average Percentage Difference; and (b). Number of Best Known Solutions Obtained
179
When compared with other meta-heuristic algorithms, namely GRASP, ANT, GGA,
PGA, IFLS, MSA, and PHAS, it can be seen from Tables 7.2a to 7.2e that WFA
outperforms GRASP, ANT, and MSA in all the instances, and also outperforms
GGA,PGA, IFLS, and PHAS in many of the instances. When comparing the overall
performance using the average relative percentage difference, as well as the number of the
best known solutions obtained, the WFA dominates these meta-heuristic algorithms as
shown in Figure 7.3, especially when compared with the GRASP, the ANT, the IFLS, and
the MSA. This shows that the WFA is able to obtain good results when compared to other
efficient meta-heuristic algorithms.
7.2 Vehicle Routing Problem
In this section, we present the two-level WFA for solving VRP. In this algorithm, the first
level focuses on solving the VRP with relaxation of integrality constraints. Then at the
second level, a modified WFA uses the initial solutions obtained by the first level to
search for the optimal solutions. Here, we illustrate the performance of the 2LWFA with
the CVRP. The performance of the 2LWFA has been tested on some CVRP benchmark
instances obtained from the literature. The experimental results obtained are compared
with the best known solutions found from the literature, and they demonstrate the
efficiency of the 2LWFA for solving the CVRP.
7.2.1 Capacitated Vehicle Routing Problem
In the field of logistics and supply chain management, how to arrange an appropriate
supplier-to-customer assignment and determine an efficient distribution routing is very
180
important. A good solution can not only improve the efficiency in operation, but also
significantly reduce operating costs. Many problems that originate from these concerns are
classified as VRPs, which are NP-hard problems (Toth and Vigo, 2002).
The CVRP can be considered as one of the typical VRPs. This problem consists of a
fleet of vehicles with pre-specified limited capacity, which has to serve a set of customers
with specific demands. In this thesis, we have focused on the CVRP with one depot.
Hence, the vehicles must start and end at the same depot. The CVRP aims to assign the
vehicles to the customers and to find the efficient routes of the vehicles so that total travel
distance (or time) is minimized. The problem has some important constraints, such as (1)
total demand of the customers served by a vehicle should not exceed the capacity of the
vehicle, (2) one customer can only be served by one vehicle, (3) all customers must be
served, (4) the number of vehicles used cannot exceed the number of given vehicles. In
this problem, the distance between the depot and customers, as well as between a
customer and other customers, are also given.
There are many research works related to the CVRP. Among them, the works that
formulate the CVRP as an integer linear programming problem have generally led to
effective exact solution approaches for this problem. In this research work, we have used
the integer linear programming formulation presented in Kulkarni and Bhave (1985) and
corrected in Imdat et al. (2004). The integrality constraints of the formulation are relaxed
to allow the first level of the 2LWFA to find good initial solutions.
181
7.2.2 Two-level WFA for the CVRP
A 2LWFA has been proposed for solving the CVRP. The first level of this algorithm is to
solve the mathematical programming model of the CVRP with the relaxation of the
integrality constraints. This is to reduce the computation time of obtaining good initial
solutions for the 2LWFA. Then, the obtained optimal solution is adjusted by a check-and-
fit procedure, and a perturbation scheme is applied to generate a set of initial feasible
solutions for the second level of this algorithm. At the second level, a modified WFA is
applied to search for optimal solutions from the initial seed solutions.
7.2.2.1 First Level
The efficiency and effectiveness of solving the CVRP based on the mathematical
programming formulation depends on the solvers used. Most solvers require large
amounts of computation time and may fail to return solutions for problems with large size.
To avoid this problem and reduce the computation effort required, we relax the integrality
constraints of the CVRP. In particular, we focus on the decision variable Xij in the
mathematical programming formulation of the CVRP. It is a binary variable that is equal
to 1 if and only if customers (or depot) i and j are connected. We relax these variables to
real numbers in [0, 1] and solve the resulting relaxed model with the commercial
optimization package LINGO 5.0. This relaxed CVRP model can be described by the
equations from (7.15) to (7.23). In these equations, n is the number of customers, m is the
number of routes, and Cij represents the cost or distance between customers i and j.
Variables Xij or Xji exist only if qi + q j ≤ Q , where qi and qj represent the demand of
customers i and j respectively, and Q is the maximum capacity of the vehicle. A variable
182
ui is associated with each customer i, which is used to formulate equation (7.20) to ensure
that the solution contains no sub-tours disconnected from the depot. Other equations
represent the common constraints of the CVRP described in Section 7.2.1.
Min ∑C
i≠ j
ij X ij (7.15)
n
S.t.: ∑X
j =1
0j = m, (7.16)
∑X
i =1
i0 = m, (7.17)
n
∑X
j =0
ij = 1 for i = 1,..., n, (7.18)
j ≠i
n
∑X
i =0
ij = 1 for j = 1,..., n, (7.19)
i≠ j
ui − u j + QX ij ≤ Q − q j for i, j = 1,..., n and i ≠ j , (7.20)
qi ≤ ui ≤ Q for i = 1,..., n, (7.21)
X ij ∈ [0;1] for i, j = 0,..., n and i ≠ j , (7.22)
m ≥ 1 and m ∈ Z + , (7.23)
For the resulting optimal solution, we arrange the values of decision variables in a
descending order. Those Xij variables with large values can be considered to be potential
connections of customers (or depot) i and j. As such, we will assign these Xij to be 1. Also,
we check the feasibility of the assignment based on the constraints of the CVRP, and fit in
the most appropriate assignment. This check-and-fit procedure is performed iteratively
until a feasible solution of the CVRP is determined. Then, a perturbation scheme based on
the cross-exchange procedure in Taillard et al. (1997) is used to generate a set of seed
solutions from the feasible solution. This cross-exchange procedure can be illustrated by
183
Figure 7.4. In this figure, the square and circles represent the depot and the customers
respectively. Each route begins and ends at the same depot (or square). Firstly, we remove
two edges (A1, A1’) and (B1, B1’) from route 1, and two edges (A2, A2’) and (B2, B2’)
from route 2. Next, the segments (A1’, B1) and (A2’, B2) which may include some
customers are swapped to formulate the new routes, i.e., new route 1 contains the new
edges (A1, A2’) and (B2, B1’), while new route 2 contains (A2, A1’) and (B1, B2’).
These new routes are accepted only when they satisfy all constraints of the CVRP.
Route 1 Route 2 Route 1 Route 2
A1 A2 A1 A2
A1’ A2’ A1’ A2’
B1 B2 B1 B2
B1’ B2’ B1’ B2’
Figure 7.4 The Cross-Exchange Procedure (Taillard et al., 1997)
We now describe the check-and-fit procedure in greater detail and illustrate it with a
benchmark instance of Christofides and Eilon (1969) for the CVRP. The instance used is
184
E-n13-k4, which has 12 customers served by 4 vehicles. Firstly, we use LINGO 5.0 to
solve the problem, and the values of the decision variables obtained are as follows: X0,1 =
X0,2 = X0,6 = X0,9 = X1,0 = X2,0 = X3,0 = X6,0 = X9,12 = X12,3 = 1; X4,7 = X4,11 = X7,8 = X5,10 = X7,4
= X7,10 = X8,5 = X8,11 = X10,5 = X10,7 = X11,4 = X11,8 = 0.5; X2,7 = 0.28; and all the rest of the
variables Xi,j = 0. Here, index 0 denotes the depot. Secondly, we apply the check-and-fit
procedure to obtain a feasible solution. In particular, we consider the decision variables
whose value is 1. The corresponding customers in the decision variables are assigned in
the same route. For example, we have X0,1 = X1,0 = 1, and also we assign customer 1 into
route 1. Similarly, we obtain 4 partial routes, i.e., route 1: (1), route 2: (2), route 3: (6),
and route 4: (9, 12, 3). Route 4 consists of 3 customers since X0,9 = X9,12 = X12,3 = X3,0 = 1.
Next, we consider whether any customer is connected with a customer assigned in the
partial routes. In this case, we see that only X2,7 = 0.28, and thus we assign customer 7 into
route 2. In addition, since X7,4 = X4,11 = 0.5, we also assign customers 4 and 11 into route 2,
i.e., route 2: (2, 7, 4, 11). Although we still have X11,8 = 0.5, we do not assign customer 8
into route 2, since including customer 8 would violate the constraint of maximum capacity
of each vehicle. Moreover, we chose the customer 4 instead of customer 8 to assign into
route 2 after assigning customer 7. This is because we choose based on the minimum
distance between customers. Here, the distance between customers 7 and 4 is shorter than
the distance between customers 7 and 8. Since we do not have any nonzero decision
variable that may connect the unassigned customers with the last customer in the partial
routes, we consider and choose the customer with minimum distance between it and the
last customer in partial routes. Here, we will assign customer 10 into route 3. Because of
X10,5 = 0.5, we also assign customer 5 right after customer 10 in route 3. After applying the
check-and-fit procedure, we construct a feasible solution with 4 complete routes, i.e., route
185
1: (1, 8), route 2: (2, 7, 4, 11), route 3: (6, 10, 5), and route 4: (9, 12, 3). The total
corresponding travel distance is 307. Based on the cross-exchange procedure described
above, the feasible solution is used to generate a set of seed solutions that are then used as
initial solutions in the modified WFA.
A flow chart of the first level of the 2LWFA for the CVRP is shown in Figure 7.5.
Start the first level
Relax the integrality constraints of the

mathematical programming
formulation of the CVRP
Solve the relaxed formulation of the

CVRP
Apply the check-and-fit procedure to

obtain a feasible solution for the CVRP
Use a perturbation scheme to construct

a set of initial seed solutions from the
feasible solution obtained
Assign the number of DOWs

corresponding to the number of initial
seed solutions generated
End the first level
Figure 7.5 Flow Chart of the First Level of the 2LWFA for the CVRP
186
7.2.2.2 Second Level
After constructing a set of seed solutions, a modified WFA is applied to search for optimal
solutions from the seed solutions. Basically, this algorithm is similar to the WFA
described in Chapter 3. However, a new solution representation for the CVRP is being
used in the modified WFA. Also, a variable neighborhood structure based on the k-opt
algorithm is used in the erosion process to enhance the flexibility and efficiency of this
process. From the literature, the k-opt algorithm with k ≥ 5 has obtained solutions with
insignificant improved quality compared to the k-opt algorithm with other smaller values
of k, and yet requires a large amount of computation time. Hence, we chose the value of k
to be 3 and 4 for the variable neighborhood structure in the erosion process of the
modified WFA. The implementation of this variable neighborhood search is similar to the
method in Hansen and Mladenovic (2001). For the neighborhood structure in the
exploration phase of the modified WFA, the 2-opt algorithm is used.
The WFA encodes a feasible solution of an optimization problem and its objective
value into a DOW, which is a component of a cloud representing a pool of solutions. For
the CVRP, we consider the values of the coordinates (X, Y) as the longitude and latitude
in the position of DOW on the ground, while the total traveling distance/time is encoded as
the altitude. In some instances when the values of coordinates (X, Y) for the depot and
customers are not known, we use the solution representation of the CVRP with specific
routes for vehicles. Actually, this solution can also be expressed by the values of
coordinates (X, Y) with the corresponding sequences.
187
We use an example to illustrate the solution representation of the CVRP. Here, we
denote integer 0 to be the depot, and other integers from {1, 2,…, n} to be the customers.
For example, a solution representation may then be shown as [0, 2, 1, 3, 0, 5, 4, 6]. This
solution includes two vehicles: a vehicle starts from the depot to serve the customers with
routing 2, 1, and 3, and goes back to the depot; another vehicle starts from the depot to
serve the customers with routing 5, 4, and 6, and goes back to the depot. With this solution
representation, integer 0 can appear many times in the solution depending on the number
of vehicles given, but other integers can only appear once in the solution. This is suitable
for the constraints of the VRP presented in this chapter. Figure 7.6 shows an example of a
DOW and its positional vector components for the CVRP with n = 6 customers and 2
vehicles.
Solution to VRPs 0 2 1 3 0 5 4 6 DOW
Coordinate X 145 159 151 130 145 163 128 146 Longitude
Coordinate Y 215 261 264 254 215 247 252 246 Latitude
Objective value Total travel distance = 244 Altitude
Figure 7.6 An Example of Solution Representation in the 2LWFA for the CVRP
A flow chart of the modified WFA of 2LWFA for the CVRP is shown in Figure 7.7.
188
Randomly generate the number of

DOWs with feasible solutions of
Start the second level
the CVRP that do not belong to
the E-list
Find the local optimal positions of

the solutions generated by
applying the hill-sliding algorithm
Update the P0-list and the UE-list
Pick one solution in the UE-list
No
Evaluate topology Choose the un-eroded
No
parameter with a direction with the smallest
Satisfy erosion condition? Yes
corresponding variable topology parameter to
neighborhood structure perform erosion process
No
No
Update the UE-list to Is a better No
continue with performing Yes local optimal solution
the erosion process found?
No
Have all the solutions in

Update the E-list
the UE-list been considered?
Have all directions been
chosen?
Yes Yes
Is the maximum number Yes

Have all neighborhood
of clouds reached? structures been applied?
Yes
Output the best solutions of the CVRP

End the second level
in the P0-list
Figure 7.7 Flow Chart of the Modified WFA for the CVRP
7.2.3 Preliminary Experiments
The 2LWFA has been coded using Visual Basic 6.0 and linked with LINGO 5.0. All
preliminary experiments have been performed on an Intel Centrino Duo 1.60 GHz CPU
with 1.5 GB of RAM. Here, the computational complexity of the 2LWFA for the CVRP is
determined based on the neighborhood structure used and the erosion process of this
189
algorithm. In particular, the 2LWFA used k-opt neighborhood structure (k = 3 or 4), and
the worst possibility of the erosion process is to find for all n directions. Thus, the
computational complexity of the 2LWFA is estimated to be O(n5).
These experiments are carried out on the benchmark instances of Christofides and
Eilon (1969) and Fisher (1994) taken from the literature for the CVRP. The best known
values of the benchmark instances from the literature are also used as reference values for
the evaluation of the WFA for solving the CVRP. These values may be obtained from the
optimal solutions of the benchmark instances used, or the best solutions found by some
algorithm so far. The results obtained are shown in Table 7.4. The choice of parameters
for the 2LWFA was determined by design-of-experiment methods and the best values of
the parameters include MaxCloud = 20, MaxPop = 10, MinEro = 3, and MaxUIE = 5.
From Table 7.4, we see that the 2LWFA can obtain optimal solutions for the instances
with small and medium size within reasonable computation time. For instances with larger
size, i.e., F-n72-k4, E-n76-k7, and E-n101-k8, the proposed algorithm can also find
solutions with a relative deviation of 1.7%, 0.6%, and 4.4% over the optimal/best known
solution, respectively. The average relative percentage difference for all instances is only
0.6%.
The results from Table 7.4 also show that the solution obtained by 2LWFA achieved
an average improvement of 20.97% over the initial solution obtained by applying the
check-and-fit procedure to the solution found using LINGO. We find that these
improvements are more significant for the instances with larger size.
190
Table 7.4 Experimental Results for the CVRP
Initial solution CPU time (secs)

Optimal/Best after applying
Instance 2LWFA
known value check-and-fit
Lingo WFA
procedure
E-n13-k4 247 307 247 0.2 0.50
E-n22-k4 375 390 375 0.5 1.75
E-n23-k3 569 623 569 0.6 4.47
E-n30-k3 534 582 534 1.0 7.50
E-n33-k4 835 910 835 1.5 35.25
E-n51-k5 521 627 521 3.0 120.15
E-n76-k7 683 843 687 18.0 1261.72
E-n101-k8 817 1012 853 50.0 4840.14
F-n45-k4 724 839 724 2.6 102.66
F-n72-k4 237 425 241 20.5 1435.72
7.3 Conclusions
In this chapter, the WFA has been developed for solving other combinatorial optimization
problems, such as the QAP and the VRP. To solve the QAP by the proposed algorithm
with enhanced solution diversification and intensification capabilities, a systematic DOW
generator scheme to distribute the positions of DOW is applied, while neighborhood
structures, such as the 2-opt mirror and 3-opt, are used to focus on strong searching of
191
promising regions. The benchmark problem sets from the QAPLIB (Burkard et al., 1997)
are used to evaluate the performance of the WFA. The computational results show that the
WFA is able to generate optimal solutions for many benchmark problems of QAP, and
near-optimal solutions for the remaining problems. The proposed algorithm is also
compared with other meta-heuristic algorithms from the literature. The results of the
comparison show that the WFA compares favorably with other meta-heuristic algorithms
used to solve the QAP. These results have been reported in Ng and Tran (2011).
To solve the CVRP efficiently, we developed a two-level WFA. The first level of the
proposed algorithm is to solve the mathematical programming model of the CVRP with
the relaxation of the integrality constraints. At the second level, a modified WFA is
applied to search for optimal solutions from the initial solutions obtained from the first
level. The results of the preliminary experimentations show the potential of the 2LWFA to
solve the CVRP, as well as other types of VRPs. These results have been reported in Tran
and Ng (2011b).
192
Chapter 8 Conclusions and Future Research Work
CHAPTER 8
CONCLUSIONS AND FUTURE RESEARCH WORK
In this thesis, we have introduced a novel nature-inspired algorithm, known as the water
flow algorithm, for solving combinatorial optimization problems. The proposed algorithm
has simulated the hydrological cycle in meteorology and the erosion phenomenon in
nature, which represent the solution exploration and exploitation capabilities of the
algorithm, respectively. Many characteristics of water flow in nature, such as water always
moving to lower positions, distributing onto many places on the ground, and eroding
terrain, are imitated to design the operators of this algorithm. The WFA has been applied
to solve a variety of NP-hard combinatorial optimization problems, i.e., flow shop
scheduling problem, flexible flow shop scheduling problem with intermediate buffers,
quadratic assignment problem, and capacitated vehicle routing problem. Also, this
algorithm has been developed to solve multi-objective scheduling problem, which is an
NP-hard optimization problem with many practical applications in modern production
environment.
193
According to the literature review, there are limited meta-heuristic algorithms inspired
by the behaviors of water flow in nature. These works have not demonstrated the
efficiency of the algorithms for solving various optimization problems. Hence, with the
success of the WFAs for solving such single-objective and multi-objective problems as
illustrated in this thesis, we have put forth a new and effective nature-inspired
optimization approach. This approach may be applied to solve other optimization
problems by adjusting the components of the algorithms appropriately.
Chapter 8 is organized as follows. In Section 8.1, we present some conclusions of the
thesis. The contributions of this thesis are also highlighted in this section. Section 8.2
provides some possible future research works.
8.1 Conclusions
This thesis focuses on constructing the general WFA and the implementation of this
algorithm for solving NP-hard combinatorial optimization problems, such as the flow shop
scheduling problem, flexible flow shop scheduling problem with intermediate buffers,
quadratic assignment problem, and vehicle routing problem. Both single-objective and
multi-objective optimization approaches of the WFA were also developed to solve such
optimization problems. From the experimental results, we can conclude that the WFA is a
promising method that is able to obtain good quality solutions to the optimization
problems within reasonable computation time.
The main contributions of this thesis consist of six parts, and they can be outlined as
follows:
194
(1). A novel nature-inspired algorithm, known as the water flow algorithm, for solving
NP-hard optimization problems was constructed. The WFA has mostly simulated
the characteristics of water flow in nature and the components of the hydrological
cycle in meteorology. This algorithm consists of two major phases, the solution
exploration and exploitation phases, which are inspired by the hydrological cycle
and the erosion phenomenon, respectively. The WFA has achieved a balance
between the solution exploration and exploitation capabilities to search for optimal
solutions in reasonable computation time. In the algorithm, the number of
controlled parameters defined by users is small. Hence, the computation time
needed by this algorithm is significantly reduced, which helps to increase the
performance of the WFAs.
(2). With some modifications, the WFA could be developed to solve several single-
objective NP-hard optimization problems, such as the permutation flow shop
scheduling problem, flexible flow shop scheduling problem with intermediate
buffers, quadratic assignment problem, and capacitated vehicle routing problem.
For the PFSP, we have used a basic version of the WFA to solve this problem. The
algorithm obtained the best known solutions for almost all the benchmark
instances used, and a new best known solution of a Heller benchmark instance was
found by this algorithm. In addition, the WFA outperforms several meta-heuristic
algorithms used in the computational comparisons.
(3). For the FFSP with intermediate buffers, some components of the WFA were
modified to solve this scheduling problem. In particular, an improved procedure
for constructing a complete schedule of the FFSP was integrated into the algorithm.
195
The procedure helps the WFA to ignore the number of machines at each stage,
which decreases the computational complexity of the algorithm for this problem.
Moreover, we combined the amount of precipitation and its falling force to form a
flexible erosion capability in this algorithm. This helps the erosion process to focus
on exploiting promising regions strongly. The experimental results show that the
WFA is an efficient algorithm for solving benchmark instances from the literature
and randomly generated instances. Many improved solutions to the benchmark
problems were found by this algorithm. Also, the results demonstrate the potential
of the algorithm to solve real-world problems, such as in maltose syrup production.
Moreover, the comparison results show that the WFA outperforms other meta-
heuristic algorithms, such as the tabu search and memetic algorithm, for solving
the FFSP with intermediate buffers.
(4). The WFA is able to solve the QAP effectively. In this version of the WFA, a
systematic DOW generator scheme to distribute the positions of DOWs was
proposed to increase the exploration capability of the algorithm, while the
neighborhood structures, such as the 2-opt mirror and 3-opt, were integrated to
focus on strong searching of promising regions. The WFA obtained the best known
solutions for 99 out of the 134 instances from the QAPLIB within reasonable
computation time. The average relative percentage difference of the algorithm for
all the 134 instances was found to be 0.20%. In addition, the WFA outperforms all
the algorithms when compared in terms of average percentage difference and
number of the best known solutions obtained.
196
(5). A two-level WFA was developed to solve vehicle routing problems. The first level
of this algorithm is to solve the mathematical programming model of the VRPs
with the relaxation of the integrality constraints. At the second level, a modified
WFA is then applied to search for optimal solutions from the initial solutions
obtained from the first level. In the modified WFA, a variable neighborhood
structure based on the k-opt algorithm was used to enhance the flexibility and
efficiency of the erosion process. In this thesis we illustrated the performance of
the 2LWFA for solving the capacitated vehicle routing problem. The experimental
results show the potential of the 2LWFA to solve this problem as well as other
types of VRPs.
(6). In addition to the capability of solving single-objective optimization problems, the
WFA is also able to solve multi-objective optimization problems by modifying and
integrating some specialized components. In this thesis, the WFA was developed
to solve the multi-objective FFSP with intermediate buffers. In this algorithm,
landscape analysis was performed to determine the weights of objective functions,
which guide DOWs to exploit potential regions and move towards the optimal
Pareto solution set. Also, the evaporation and precipitation processes were
included into this algorithm to enhance the solution exploitation capability of the
algorithm in potential neighboring regions. Moreover, an improvement process for
reinforcing the final Pareto solution set obtained was proposed. The experimental
results, based on benchmark instances taken from the literature and randomly
generated instances, demonstrate the effectiveness and efficiency of the MOWFA.
197
The comparison results also show that the MOWFA outperforms other algorithms
for the test instances.
8.2 Future Research Work
In this section, we present some possible future research work and directions for the WFA.
These consist of improvement and application of the algorithms.
Firstly, the choice of parameters for the WFAs can be improved. While the current
choice of parameters used in the WFAs for single-objective and multi-objective
optimization problems has obtained good computational performance, further research on
how to improve the choice of parameters used in the WFAs can still be performed. This
would lead to greater efficiency for the WFAs when solving the optimization problems, as
well as the possibility of obtaining solutions with better quality by the algorithms.
Secondly, when we use the erosion process based on the lower bound of the
optimization problem, i.e. for the flexible flow shop scheduling problem with intermediate
buffers, the performance of this process depends on the quality of lower bound used. In
this thesis, we only used the lower bounds from the literature. Hence, obtaining a good
lower bound for the scheduling problem to be used in the WFA may be considered as
possible future research work.
Thirdly, we have illustrated the potential of the WFAs for solving other optimization
problems in this thesis. The WFA and the MOWFA can be used to solve other types of
single-objective and multi-objective optimization problems respectively, that are similar to
the optimization problems investigated in this thesis, by adapting some components of the
198
algorithms appropriately. Depending on the specific structure of optimization problems
considered, we can customize the WFAs to solve the problems efficiently and this would
certainly be part of future research work. Some examples of such potential optimization
problems include the job shop scheduling problem and traveling salesman problem.
In addition, we can apply the WFA to solve the scheduling problems with other
important objective functions, such as minimizing total flow time of jobs, minimax
tardiness, or minimizing total idle time of machines. This helps to allow a more
comprehensive evaluation of the performance of the WFA. We can even develop the
WFA for solving scheduling problems with data noise and uncertainty which are
commonly found in most practical single-objective and multi- objective problems.
Stochastic techniques for dealing with such problems can then be integrated into the
WFA.
Lastly, we can extend the WFA for solving continuous optimization problems by
designing appropriate neighborhood structures for continuous variables. Furthermore,
some smoothing functions or methods in the field of continuous optimization may be
integrated into the erosion process of the WFA to enhance the performance of the
algorithm.
199
References
REFERENCES
[1]. Abbiw-Jackson R., B. Golden, S. Raghavan, and E. Wasil. A Divide-and-Conquer

Local Search Heuristic for Data Visualization, Computers and Operations
Research, 33(11), pp. 3070–3087. 2006.
[2]. Abraham A., C. Grosan, and V. Ramos. Stigmergic Optimization. Netherlands:

Springer-Verlag Berlin Heidelberg. 2006.
[3]. Agarwal A., S. Colak, and E. Eryarsoy. Improvement Heuristic for the Flow Shop
Scheduling Problem: An Adaptive Learning Approach, European Journal of
Operational Research, 169(3), pp. 801–815. 2006.
[4]. Ahuja R.K., T.L. Magnanti, and J.B. Orlin. Network Flows: Theory, Algorithms,
and Applications. New Jersey, US: Prentice Hall. 1993.
[5]. Ahuja R.K., J.B. Orlin, and A. Tiwari. A Greedy Genetic Algorithm for The
Quadratic Assignment Problem, Computers and Operations Research, 27(10),
pp.917–934. 2000.
[6]. Akturk M.S, and M.B. Yildirim. A New Lower Bounding Scheme for the Total
Weighted Tardiness Problem, Computers and Operations Research, 25(4), pp.
265–278. 1998.
[7]. Applegate D., R.E. Bixby, V. Chvatal, and W. Cook. Finding Cuts in the TSP. A
Preliminary Report Distributed at the Mathematical Programming Symposium,
Ann Arbor, Michigan. 1994.
200
References
[8]. Arkin E.M., R. Hassin, and M. Sviridenko. Approximating the Maximum

Quadratic Assignment Problem, Information Processing Letters, 77(1), pp. 13–16.
2001.
[9]. Azizoglu M., E. Cakmak, and S. Kondakci. A Flexible Flow Shop Problem with
Total Flow Time Minimization, European Journal of Operational Research, 132(3),
pp. 528–538. 2001.
[10]. Banks A., J. Vincent, and C. Anyakoha. A Review of Particle Swarm Optimization
- Part I: Background and Development, Natural Computation, 6(4), pp.467–484.
2007.
[11]. Basak A., S. Pal, S. Das, A. Abraham, and V. Snasel. A Modified Invasive Weed
Optimization Algorithm for Time-Modulated Linear Antenna Array Synthesis,
IEEE Congress on Evolutionary Computation (CEC’10). 2010.
[12]. Basturk B., and D. Karabogo. An Artificial Bee Colony (ABC) Algorithm for
Numerical Function Optimization. In: IEEE Swarm Intelligence Symposium 2006,
May 12-24, Indianapolis, IN, USA. 2006.
[13]. Basu S., C. Chaudhuri, M. Kundu, M. Nasipuri, and D.K. Basu. Text Line
Extraction from Multi-Skewed Handwritten Documents, Pattern Recognition,
40(6), pp. 1825–1839. 2007.
[14]. Bitam S., M. Batouche, and E. Talbi. A Survey on Bee Colony Algorithms. In:
Proceedings of IEEE International Symposium on Parallel & Distributed
Processing (IPDPSW’ 10), pp. 1–8. 2010.
[15]. Blanchard A., S. Elloumi, A. Faye, and N. Wicker. A Cutting Algorithm for the
Quadratic Assignment Problem, INFOR, 41(1), pp. 35–49. 2003.
[16]. Bokhari S.H. Assignment Problems in Parallel and Distributed Computing. Kluwer
Academic Publishers, Boston, MA. 1987.
201
References
[17]. Brodic D., and Z. Milivojevic. Reference Text Line Identification Based on Water
Flow Algorithm. In: Proceedings of ICEST 2009, SP-2 Sect, Veliko Tarnovo,
Bulgaria. 2009a.
[18]. Brodic D., and Z. Milivojevic. Modified Water Flow Method for Reference Text
Line Detection. In: Proceedings of ICCS 2009, Sofia, Bulgaria. 2009b.
[19]. Brodic D., and Z. Milivojevic. An Approach to Modification of Water Flow

Algorithm for Segmentation and Text Parameters Extraction. In: Proceedings of
DoCEIS 2010, IFIP AICT 314, pp. 324–331. 2010.
[20]. Buffa E.S., G.C. Armour, and T.E. Vollmann. Allocating Facilities with CRAFT,
Harvard Business Review, 42(2), pp. 136–158. 1964.
[21]. Burkard R.E., E. Cela, G. Rote, and G.J. Woeginger. The Quadratic Assignment
Problem with A Monotone Anti-Monge and A Symmetric Toeplitz Matrix: Easy
and Hard Cases. In: Proceedings of the 5th International Conference on Integer
Programming and Combinatorial Optimization, Vancouver, British Columbia,
Canada, pp. 204–218. 1996.
[22]. Burkard R.E., S.E. Karisch, and F. Rendl. QAPLIB - A Quadratic Assignment
Problem Library, Journal of Global Optimization, 10(4), pp. 391–403. 1997.
[23]. Cela E. The Quadratic Assignment Problem: Theory and Algorithms. In: D.Z. Du,
P. Pardalos (Eds.), Combinatorial Optimization, Kluwer Academic Publishers,
London. 1998.
[24]. Chatterjee S., C. Carrera, and L. Lynch. Genetic Algorithms and Traveling
Salesman Problems, European Journal of Operational Research, 93(3), pp. 490–
510. 1996.
[25]. Christofides N., and S. Eilon. An Algorithm for the Vehicle Dispatching Problem,
Operational Research Quarterly, 20(3), pp. 309–318. 1969.
202
References
[26]. Christofides N., A. Mingozzi, P. Toth, and C. Sandi. Combinatorial Optimization.

John Wiley & Sons. 1979.
[27]. Dantzig G.B., and J.H. Ramser. The Truck Dispatching Problem, Management
Science, 6(1), pp. 80–91. 1959.
[28]. Darwin C. On the Origin of Species by Means of Natural Selection. John Murray,
London. 1859.
[29]. Dawkins R. The Selfish Gene. Oxford: Oxford University Press. 1976.
[30]. Demidenko V.M., G. Finke, and V.S. Gordon. Well Solvable Cases of the
Quadratic Assignment Problem with Monotone and Bimonotone Matrices, Journal
of Mathematical Modeling and Algorithms, 5(2), pp. 167–187. 2006.
[31]. Deneubourg J.L., S. Goss, J.M. Pasteels, D. Fresneau, and J.P. Lachaud. Self-
Organization Mechanisms in Ant Societies (II): Learning in Foraging and Division
of Labor. In: From Individual to Collective Behavior in Social Insects (J.M.
Pasteels and J.L. Deneubourg, Eds.), Experientia Supplementum, 54, pp. 177–196.
1987.
[32]. Dorigo M. Optimization, Learning and Natural Algorithms. Ph.D. thesis,

Politecnico di Milano, Italie. 1992.
[33]. Dorigo M., and L.M. Gambardella. Ant Colony System: A Cooperative Learning
Approach to the Traveling Salesman Problem, IEEE Transaction on Evolutionary
Computation, 1(1), pp.53–66. 1997.
[34]. Dorigo M., G.D. Caro, and L.M. Gambardella. Ant Algorithms for Discrete
Optimization, Artificial Life, 5(2), pp.137–172. 1999.
[35]. Duman E., and I. Or. The Quadratic Assignment Problem in the Context of the
Printed Circuit Board Assembly Process, Computers and Operations Research,
34(1), pp. 163–179. 2007.
203
References
[36]. Eschermann B., and H.J. Wunderlich. Optimized Synthesis of Self-Testable Finite
State Machines. In: Proceedings of the 20th International Symposium Fault-
Tolerant Computing (FTCS 20), Newcastle Upon Tyne, UK, pp. 390–397. 1990.
[37]. Eusuff M.M., and K.E. Lansey. Optimization of Water Distribution Network
Design Using the Shuffled Frog Leaping Algorithm, Journal of Water Resources
Planning and Management, 129(3), pp. 210–25. 2003.
[38]. Farooq M. From the Wisdom of the Hive to Intelligent Routing in

Telecommunication Networks: A Step towards Intelligent Network Management
through Natural Engineering, PhD Thesis, University of Dortmund, Germany.
2006.
[39]. Gasimov R.N., and O. Ustun. Solving the Quadratic Assignment Problem Using F-
MSG Algorithm, Journal of Industrial and Management Optimization, 3(2), pp.
173–191. 2007.
[40]. Goldberg D.E. Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley. 1989.
[41]. Goudie A. The Nature of the Environment. Oxford: Blackwell Science. 1993.
[42]. Grabowski J., and J. Pempera. Sequencing of Jobs in Some Production System,
European Journal of Operational Research, 125(3), pp. 535–550. 2000.
[43]. Grosan C., and A. Abraham. Stigmergic Optimization: Inspiration, Technologies

and Perspectives, Studies in Computational Intelligence (SCI), 31, pp. 1–24. 2006.
[44]. Gutin G., and A. Yeo. Polynomial Approximation Algorithms for TSP and QAP
with A Factorial Domination Number, Discrete Applied Mathematics, 119(1–2),
pp. 107–116. 2002.
[45]. Haddad O.B., A. Afshar, and M.A. Marino. Honey Bees Mating Optimization
Algorithm (HBMO): A New Heuristic Approach for Engineering Optimization. In:
204
References
Proceedings of the First International Conference on Modeling, Simulation and

Applied Optimization, Sharjah, UAE. 2005.
[46]. Hahn P.M., W.L. Hightower, T.A. Johnson, M. Guignard-Spielberg, and C.

Roucairol. Tree Elaboration Strategies in Branch and Bound Algorithms for
Solving the Quadratic Assignment Problem, Yugoslavian Journal of Operational
Research, 11(1), pp. 41–60. 2001.
[47]. Hansen P., and N. Mladenovic. Variable Neighborhood Search: Principles and
Applications, European Journal of Operational Research, 130(3), pp. 449–467.
2001.
[48]. Ho J.C., and Y.L. Chang. A New Heuristic for the N-Job, M-Machine Flow Shop
Problem, European Journal of Operational Research, 52(2), pp. 194–202. 1991.
[49]. Holland J.H. Adaptation in Natural and Artificial Systems. Ann Arbor: University
of Michigan Press. 1975.
[50]. Holy M. Erosion and Environment. Elmsford: Pergamon. 1982.
[51]. Hubert L.J. Assignment Methods in Combinatorial Data Analysis. Marcel Dekker
Inc., New York. 1987.
[52]. Hull P. Glucose Syrups: Technology and Applications. UK, Wiley-Blackwell.

2010.
[53]. Huynh T.H. A Modified Shuffled Frog Leaping Algorithm for Optimal Tuning of
Multivariable PID Controllers, IEEE International Conference on Industrial
Technology (ICIT’08), Chengdu, China, pp. 1–6. 2008.
[54]. Imdat K., L. Gilbert, and B. Tolga. A Note on the Lifted Miller–Tucker–Zemlin
Tour Elimination Constraints for the Capacitated Vehicle Routing Problem,
205
References
[55]. Ishibuchi H., T. Yoshida, and T. Murata. Balance between Genetic Search and
Local Search in Memetic Algorithms for Multiobjective Permutation Flowshop
Scheduling, IEEE Transactions on Evolutionary Computation, 7(2), pp. 204–223.
2003.
[56]. Johnson S.M. Optimal Two-And Three-Stage Production Schedules with Setup
Times Included, Naval Research Logistics Quarterly, 1(1), pp. 61–68. 1954.
[57]. Karimkashi S., and A.A. Kishk. Invasive Weed Optimization and Its Features in
Electromagnetics, IEEE Transactions on Antennas and Propagation, 58(4), pp.
1269–1278. 2010.
[58]. Kennedy J., and R.C Eberhart. Particle Swarm Optimization. In: Proceedings of
IEEE International Conference on Neural Networks, Piscataway, NJ, pp. 1942–
1948. 1995.
[59]. Kim I.K., D.W. Jung, and R.H. Park. Document Image Binarization Based on
Topographic Analysis Using A Water Flow Model, Pattern Recognition, 35(1), pp.
265–277. 2002.
[60]. Knowles J.D., and D.W. Corne. On Metrics for Comparing Nondominated Sets.
In: Proceedings of Congress on Evolutionary Computation, pp. 711–716. 2002.
[61]. Koulamas C. A New Constructive Heuristic for Flow Shop Scheduling Problem,
[62]. Koopmans T., and M.J. Beckmann. Assignment Problems and the Location of
Economics Activities, Econometric, 25(1), pp. 53–76. 1957.
[63]. Kulkarni R.V., and P.R. Bhave. Integer Programming Formulations of Vehicle
Routing Problems, European Journal of Operational Research, 20(1), pp. 58–67.
1985.
[64]. Lei D. Multi-Objective Production Scheduling: A Survey, International Journal of

Advance Manufacturing Technology, 43(9–10), pp. 926–938. 2009.
206
References
[65]. Li Y., P.M. Pardalos, and M.G.C. Resende. A Greedy Randomized Adaptive
Search Procedure for the Quadratic Assignment Problem. In: P.M. Pardalos, H.
Wolkowicz (Eds.), Quadratic Assignment and Related Problems, DIMACS Series
in Discrete Mathematics and Theoretical Computer Science 16, pp. 237–261. 1994.
[66]. Li W.T., X.W. Shi, Y.Q. Hei, S.F. Liu, and J. Zhu. A Hybrid Optimization
Algorithm and Its Application for Conformal Array Pattern Synthesis, IEEE
Transactions on Antennas and Propagation, 58(10), pp. 3401–3406. 2010.
[67]. Liao C.J., C.T. Tseng, and P. Luarn. A Discrete Version of Particle Swarm
Optimization for Flow Shop Scheduling Problems, Computers and Operations
Research, 34(10), pp. 3099–3111. 2007.
[68]. Lim M., Y. Yuan, and S. Omatu. Extensive Testing of A Hybrid Genetic
Algorithm for Solving Quadratic Assignment Problems, Computational
Optimization and Applications, 23(1), pp. 47–64. 2002.
[69]. Liong S.Y., and M.D. Atiquzzaman. Optimal Design of Water Distribution
Network Using Shuffled Complex Evolution, Journal of the Institution of
Engineers, Singapore, 44(1), pp. 93–107. 2004.
[70]. Lukasik S., and S. Zak. Firefly Algorithm for Continuous Constrained
Optimization Task, ICCCI 2009, Lecture Notes in Artificial Intelligence (Eds. N.T.
Nguyen, R. Kowalczyk, S.M. Chen), 5796, pp. 97–100. 2009.
[71]. Mallahzadeh A.R., H. Oraizi, and Z. Davoodi-Rad. Application of the Invasive

Weed Optimization Technique for Antenna Configuration, Progress in
Electromagnetics Research, 79, pp. 137–150. 2008.
[72]. Maniezzo V., and A. Colorni. The Ant System Applied to the Quadratic
Assignment Problem, IEEE Transactions on Knowledge and Data Engineering,
11(5), pp. 769–778. 1999.
[73]. Martello S., and P. Toth. Knapsack Problems: Algorithms and Computer
Implementations. New York: John Wiley and Sons. 1990.
207
References
[74]. McCormick S.T., M.L. Pinedo, S. Shenker, and B. Wolf. Sequencing in an

Assembly Line with Blocking to Minimize Cycle Time, Operations Research,
37(6), pp. 925–936. 1989.
[75]. Mehrabian A.R., and C. Lucas. A Novel Numerical Optimization Algorithm

Inspired from Weed Colonization, Ecological Informatics, 1(4), pp. 355–366. 2006.
[76]. Mehrabian A.R., and A.Y. Koma. Optimal Positioning of Piezoelectric Actuators
on A Smart Fin Using Bio-Inspired Algorithms, Aerospace science and technology,
11(2–3), pp.174–182. 2007.
[77]. Merz P., and B. Freisleben. A Genetic Local Search Approach to the Quadratic
Assignment Problem. In: Proceedings of the 7th International Conference on
Genetic Algorithms, San Diego, CA: Morgan Kaufmann, pp. 465–72. 1997.
[78]. Merz P., and B. Freisleben. Fitness Landscape Analysis and Memetic Algorithms
for the Quadratic Assignment Problem, IEEE Transactions on Evolutionary
Computation, 4(4), pp. 337–352. 2000.
[79]. Michael P. Fundamentals of Physical Geography (2nd Edition).

PhysicalGeography.net. Retrieved 2007-03-19. 2006.
[80]. Michalewicz Z. Genetic Algorithms + Data Structures = Evolution Programs.

Springer Verlag, Berlin. 1992.
[81]. Moscato P. On Evolution, Search, Optimization, Genetic Algorithms and Martial

Arts: Towards Memetic Algorithms. Technical Report Caltech Concurrent
Computation Program, Report 826, California Institute of Technology, Pasadena,
California, USA. 1989.
[82]. Moscato P., and C. Cotta. A Gentle Introduction to Memetic Algorithms. In:
Handbook of Metaheuristics (F. Glover and G. Kochenberger, eds.), pp. 105–144.
Kluwer Academic Publishers, Boston MA. 2003.
208
References
[83]. Moscato P., and M.G. Norman. A Memetic Approach for the Traveling Salesman
Problem—Implementation of A Computational Ecology for Combinatorial
Optimization on Message-Passing Systems. In: International Conference on
Parallel Computing and Transputer Application, Amsterdam, Holland: IOS Press,
pp. 177–86. 1992.
[84]. Mostafa S.A., R. Mahnaz, R.K. Ashkan, and C. Lucas. A Study of Electricity
Market Dynamics Using Invasive Weed Colonization Optimization, IEEE
Symposium on Computational Intelligence and Games (CIG’08). 2008.
[85]. Nagano M.S., R. Ruiz, and L.A.N. Lorena. A Constructive Genetic Algorithm for
Permutation Flowshop Scheduling, Computers and Industrial Engineering, 55(1),
pp. 195–207. 2008.
[86]. Nakrani S., and C. Tovey. On Honey Bees and Dynamic Server Allocation in
Internet Hosting Centers, Adaptive Behaviors, 12(3–4), pp. 223–240. 2004.
[87]. Nawaz M., E.E. Enscore Jr, and I. Ham. A Heuristic Algorithm for the M-Machine,
N-Job Flow-Shop Sequencing Problem, OMEGA: International Journal of
Management Science, 11(1), pp. 91–95. 1983.
[88]. Ng K.M., and T.H. Tran. A Water Flow Algorithm for Solving the Quadratic
Assignment Problem, Submitted to Applied Mathematics and Computation. 2011.
[89]. Oh H.H., K.T. Lim, and S.I. Chien. An Improved Binarization Algorithm Based on
A Water Flow Model for Document Image with Inhomogeneous Backgrounds,
Pattern Recognition, 38(12), pp. 2612–2625. 2005.
[90]. Onwubolu G., and D. Davendra. Scheduling Flow Shops Using Differential
Evolution Algorithm, European Journal of Operational Research, 171(2), pp. 674–
692. 2006.
[91]. Osman I.H., and C.N. Potts. Simulated Annealing for Permutation Flow Shop
Scheduling, OMEGA International Journal of Management Science, 17(6), pp.
551–557. 1989.
209
References
[92]. Otsu N. A Threshold Selection Method from Gray-Level Histograms, IEEE

Transaction on Systems, Man, and Cybernetic, 9(1), pp. 62–66. 1979.
[93]. Papadimitriou C.H., and K. Steiglitz. Combinatorial Optimization: Algorithms and

Complexity. Englewood Cliffs, N.J.: Prentice Hall. 1982.
[94]. Pedersen S., and H. Vang-Hendriksen. Method for Production of Maltose and/or
Enzymatically Modified Starch. World Intellectual Property Organization,
International Publication Number: WO 01/16349 A1. 2001.
[95]. Pinedo M. Scheduling: Theory, Algorithms, and Systems. Upper Saddle, N.J.:
Prentice Hall. 2002.
[96]. Pinedo M. Planning and Scheduling in Manufacturing and Services. New York:
Springer. 2005.
[97]. Qahri Saremi H., B. Abedin, and A. Meimand Kermani. Website Structure
Improvement: Quadratic Assignment Problem Approach and Ant Colony
Metaheuristic Technique, Applied Mathematics and Computation, 195(1), pp.
285–298. 2008.
[98]. Qian B., L. Wang, D.X. Huang, W.L. Wang, and X. Wang. An Effective Hybrid
De-Based Algorithm for Multi-Objective Flow Shop Scheduling with Limited
Buffers, Computers and Operations Research, 36(1), pp. 209–233. 2009.
[99]. Quadt D., and H. Kuhn. A Taxonomy of Flexible Flow Line Scheduling
Procedures, European Journal of Operational Research, 178(3), pp. 686–698. 2007.
[100]. Rajendran C., and H. Ziegler. Ant-Colony Algorithms for Permutation Flowshop
Scheduling to Minimize Makespan/Total Flowtime of Jobs, European Journal of
[101]. Rameshkumar K., R.K. Suresh, and K.M. Mohanasundaram. Discrete Particle
Swarm Optimization (DPSO) Algorithm for Permutation Flow Shop Scheduling to
210
References
Minimize Makespan. Lecture Notes in Computer Science, 3612, pp. 572–581.

2005.
[102]. Ramkumar A.S., S.G. Ponnambalam, and N. Jawahar. Iterated Fast Local Search
Algorithm for Solving Quadratic Assignment Problems, Robotics and Computer-
Integrated Manufacturing, 24(3), pp. 392–401. 2008.
[103]. Ramkumar A.S., S.G. Ponnambalam, and N. Jawahar. A Population-Based Hybrid

Ant System for Quadratic Assignment Formulations in Facility Layout Design,
International Journal of Advanced Manufacturing Technology, 44(5–6), pp. 548–
558. 2009.
[104]. Rashidi E., M. Jahandar, and M. Zandieh. An Improved Hybrid Multi-Objective

Parallel Genetic Algorithm for Hybrid Flow Shop Scheduling with Unrelated
Parallel Machines, International Journal of Advanced Manufacturing Technology,
49(9–12), pp. 1129–1139. 2010.
[105]. Ravindran D., A. Noorul Haq, S.J. Selvakuar, and R. Sivaraman. Flow Shop
Scheduling with Multiple Objective of Minimizing Makespan and Total Flow
Time, International Journal of Advanced Manufacturing Technology, 25(9–10), pp.
1007–1012. 2005.
[106]. Reeves C.R. A Genetic Algorithm for Flowshop Sequencing, Computers and
Operations Research, 22(1), pp.5–13. 1995.
[107]. Resnick M. Turtles, Termites, and Traffic Jams: Explorations in Massively Parallel
Microworlds. Bradford Books. 1997.
[108]. Ribas I., R. Leisten, and J.M. Framinan. Review and Classification of Hybrid Flow
Shop Scheduling Problems from a Production System and a Solutions Procedure
Perspective, Computers and Operations Research, 37(8), pp. 1439–1454. 2010.
[109]. Richardson P. Bats. Natural History Museum, London. 2008.
211
References
[110]. Rinnooy Kan A.H.G. Machine Scheduling Problems: Classification, Complexity

and Computations. The Hague: Nijhoff. 1976.
[111]. Rossin D.F., M.C. Springer, and B.D. Klein. New Complexity Measures for the
Facility Layout Problem: An Empirical Study Using Traditional and Neural
Network Analysis, Computers and Industrial Engineering, 36(3), pp. 585–602.
1999.
[112]. Roth M., and S. Wicker. Termite: A Swarm Intelligent Routing Algorithm for
Mobile Wireless Ad-Hoc Networks, Studies in Computational Intelligence (SCI),
31, pp. 155–184. 2006.
[113]. Ruiz R., and J.A. Vazquez-Rodriguez. The Hybrid Flow Shop Scheduling Problem,
[114]. Sahni S., and T. Gonzalez. P-complete Approximation Problems, Journal of the
ACM, 23(3), pp. 555–565. 1976.
[115]. Sawik T. Mixed Integer Programming for Scheduling Flexible Flow Lines with
Limited Intermediate Buffers, Mathematical and Computer Modeling, 31(13),
pp.39–52. 2000.
[116]. Sawik T. Mixed Integer Programming for Scheduling Surface Mount Technology
Lines, International Journal of Production Research, 39(14), pp. 3219–3235. 2001.
[117]. Sayadi M.K., R. Ramezanian, and N. Ghaffari-Nasab. A Discrete Firefly Meta-

Heuristic with Local Search for Makespan Minimization in Permutation Flow
Shop Scheduling Problems. International Journal of Industrial Engineering
Computations, 1, pp. 1–10. 2010.
[118]. Seeley T.D., S. Camazine, and J. Sneyd. Collective Decision Making in Honey
Bees: How Colonies Choose Among Nectar Sources, Behavioural Ecoloy and
Sociobiology, 28(4), pp. 277–290. 1991.
212
References
[119]. Sepehri-Rad H., and C. Lucas. A Recommender System Based on Invasive Weed
Optimization Algorithm. In: IEEE Congress on Evolutionary Computation (CEC
2007), pp. 4297–4304. 2007.
[120]. Shah-Hosseini H. Problem Solving By Intelligent Water Drops. In: Proceedings of

IEEE Congress on Evolutionary Computation (CEC 2007), pp. 3226–3231. 2007.
[121]. Shah-Hosseini H. Intelligent Water Drops Algorithm: A New Optimization

Method for Solving the Multiple Knapsack Problem, International Journal of
Intelligent Computing and Cybernetics, 1(2), pp. 193–212. 2008.
[122]. Shah-Hosseini H. The Intelligent Water Drops Algorithm: A Nature-Inspired

Swarm-Based Optimization Algorithm, International Journal of Bio-Inspired
Computation, 1(1–2), pp. 71–79. 2009.
[123]. Sherali H.D., S.C. Sarin, and M.S. Kodialam. Models and Algorithms for a Two-
Stage Production Process, Production Planning and Control, 1(1), pp. 27–39. 1990.
[124]. Silva J.D.L., E.K. Burke, and S. Petrovic. An Introduction to Multiobjective

Metaheuristics for Scheduling and Timetabling. In: Lecture Notes in Economics
and Mathematical Systems, 535, pp. 91–129. 2004.
[125]. Singh S.P., and R.R.K. Sharma. Two-Level Modified Simulated Annealing Based
Approach for Solving Facility Layout Problem, International Journal of Production
Research, 46(13), pp. 3563–3582. 2008.
[126]. Suliman S.M.A. A Two-Phase Heuristic Approach to the Permutation Flow-Shop

Scheduling Problem, International Journal of Production Economics, 64(1–3), pp.
143–152. 2000.
[127]. Taillard E. Some Efficient Heuristic Methods for the Flow Shop Sequencing
[128]. Taillard E. Comparison of Iterative Searches for the Quadratic Assignment

Problem, Location Science, 3(2), pp. 87–105. 1995.
213
References
[129]. Taillard E., P. Badeau, M. Gendreau, F. Guertin, and J.Y. Potvin. A Tabu Search
Heuristic for the Vehicle Routing Problem with Soft Time Windows,
Transportation Science, 31(2), pp. 170–186. 1997.
[130]. Tang L.X., and H. Xuan. Lagrangian Relaxation Algorithms for Real-Time Hybrid
Flowshop Scheduling with Finite Intermediate Buffers, Journal of the Operational
Research Society, 57(3), pp. 316–324. 2006.
[131]. Tantar E., C. Dhaenens, J.R. Figueira, and E. Talbi. A Priori Landscape Analysis
in Guiding Interactive Multi-Objective Metaheuristics. In: Proceedings of IEEE
Congress on Evolutionary Computation (CEC 2008), pp. 4104–4111. 2008.
[132]. Tasgetiren M.F., Y.C. Liang, M. Sevkli, and G. Gencyilmaz. A Particle Swarm
Optimization Algorithm for Makespan and Total Flowtime Minimization in the
Permutation Flowshop Sequencing Problem, European Journal of Operational
Research, 177(3), pp. 1930–1947. 2007.
[133]. Tavakkoli-Moghaddam R., N. Safaei, and F. Sassani. A Memetic Algorithm for

the Flexible Flow Line Scheduling Problem with Processor Blocking, Computers
and Operations Research, 36(2), pp. 402–414. 2009.
[134]. Teodorovic D., T. Davidovic, and M. Selmic. Bee Colony Optimization: The
Applications Survey, ACM Transactions on Computational Logic, pp. 1–20. 2011.
[135]. Toth P., and D. Vigo. Models, Relaxations and Exact Approaches for the
Capacitated Vehicle Routing Problem, Discrete Applied Mathematics, 123(1–3),
pp. 487–512. 2002.
[136]. Tran T.H., and K.M. Ng. A Water Flow Algorithm for Flexible Flow Shop
Scheduling with Limited Intermediate Buffers. In: Proceedings of the 4th
Multidisciplinary International Conference on Scheduling: Theory and
Applications (MISTA’09), Dublin, Ireland, pp. 606–616. 2009.
[137]. Tran T.H., and K.M. Ng. A Water-Flow Algorithm for Flexible Flow Shop
Scheduling with Intermediate Buffers, Journal of Scheduling, In Press. 2010.
214
References
[138]. Tran T.H., and K.M. Ng. A Water Flow Algorithm for Multi-Objective Flexible
Flow Shop Scheduling with Intermediate Buffers, Submitted to Computers and
Operations Research. 2011a.
[139]. Tran T.H., and K.M. Ng. Two-Level Water-Flow Algorithm for Vehicle Routing
Problems, Submitted to the 5th Multidisciplinary International Conference on
Scheduling: Theory and Applications (MISTA’11), Phoenix, Arizona, USA. 2011b.
[140]. Van Veldhuizen D.A. Multiobjective Evolutionary Algorithms: Classifications,

Analyses, and New Innovations. PhD dissertation, Department of Electrical and
Computer Engineering, Graduate School of Engineering, Air Force Institute of
Technology, Wright-Patterson AFB, Ohio, US. 1999.
[141]. Wang W., B. Wu, Y. Zhao, and D. Feng. Particle Swarm Optimization for Open
Vehicle Routing Problem, Lecture Notes in Computer Science, 4114, pp. 999–
1007. 2006.
[142]. Wardono B. Algorithms for the Multi-Stage Parallel Machine Problem with Buffer
Constraints. Ph.D. Dissertation, North Carolina State University. 2001.
[143]. Wardono B., and Y. Fathi. A Tabu Search Algorithm for the Multi-Stage Parallel
Machine Problem with Limited Buffer Capacities, European Journal of
[144]. Wei Z., X.F. Xu, and S.C. Deng. Evolutionary Algorithm for Solving Multi-
Objective Flow Shop Scheduling Problem, Computers Integrated Manufacturing
Systems, 12, pp. 1227–1234 (in Chinese). 2006.
[145]. Widmer M., and A. Hertz. A New Heuristic Method for the Flow Shop Sequencing
[146]. Wittrock R.J. An Adaptable Scheduling Algorithm for Flexible Flow Lines,
Operations Research, 36(3), pp. 445–453. 1988.
215
References
[147]. Wong K.Y., and P.C. See. A Hybrid Ant Colony Optimization Algorithm for
Solving Facility Layout Problems Formulated as Quadratic Assignment Problems,
Engineering Computations: International Journal for Computer-Aided Engineering
and Software, 27(1), pp. 117–128. 2010.
[148]. Wu T.H., S.H. Chung, and C.C. Chang. A Water Flow-Like Algorithm for
Manufacturing Cell Formation Problems, European Journal of Operational
Research, 205(2), pp. 346–360. 2010.
[149]. Yang F.C., and Y.P. Wang. Water Flow-Like Algorithm for Object Grouping
Problems, Journal of the Chinese of Industrial Engineers, 24(6), pp. 475–488. 2007.
[150]. Yang X., Q. Lu, C. Li, and X. Liao. Biological Computation of the Solution to the
Quadratic Assignment Problem, Applied Mathematics and Computation, 200(1),
pp. 369–377. 2008.
[151]. Yang X.S. Engineering Optimization via Nature-Inspired Virtual Bee Algorithms,
IWINAC 2005, Lecture Notes in Computer Science, 3562, pp. 317–323. 2005.
[152]. Yang X.S. Nature-Inspired Metaheuristic Algorithms. UK: Luniver. 2008.
[153]. Yang X.S. A New Metaheuristic Bat-Inspired Algorithm. In: Nature Inspired
Cooperative Strategies for Optimization (Eds. J.R. Gonzalez et al.), Studies in
Computational Intelligence, Springer Berlin, 284, Springer, pp. 65-74. 2010.
[154]. Yu G. Industrial Applications of Combinatorial Optimization. Boston: Kluwer

Academic Publishers. 1998.
[155]. Zhang X., Y. Wang, G. Cui, Y. Niu, and J. Xu. Application of A Novel IWO to the
Design of Encoding Sequences for DNA Computing, Computers and Mathematics
with Applications, 57(11-12), pp. 2001–2008. 2009.
[156]. Zhi X., X. Xing, Q. Wang, L. Zhang, X. Yang, C. Zhou, and Y. Liang. A Discrete
PSO Method for Generalized TSP Problem. In: Proceedings of International
Conference on Machine Learning and Cybernetics, pp. 2378–2383. 2004.
216

Waterflow

Uploaded by

Copyright:

Available Formats

Waterflow

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Waterflow

Uploaded by

Copyright:

Available Formats

CORE Metadata, citation and similar papers at core.ac.

A WATER FLOW ALGORITHM FOR

TRAN TRUNG HIEU

NATIONAL UNIVERSITY OF SINGAPORE

TRAN TRUNG HIEU

First of all, I would like to express my sincerest gratitude to my supervisor, Associate

Professor Ng Kien Ming, at the Department of Industrial and Systems Engineering,

National University of Singapore, for his encouragement and guidance throughout my

my research work as well as thesis.

extend my acknowledgement to the officers of the department for their assistance in

handling my administrative matters.

Tran Trung Hieu

TABLE OF CONTENTS .................................................................................................... ii

ABSTRACT ..................................................................................................................... viii

LIST OF TABLES ............................................................................................................. x

LIST OF FIGURES ........................................................................................................ xiii

GLOSSARY ................................................................................................................... xvi

CHAPTER 1 INTRODUCTION ................................................................................ 1

1.1 Combinatorial Optimization Problems ................................................................... 2

1.2 Nature Inspired Algotithms..................................................................................... 4

1.3 Motivation and Research Objectives ...................................................................... 7

1.4 Main Contributions of the Thesis ........................................................................... 8

1.5 Outline of the Thesis ............................................................................................. 10

CHAPTER 2 LITERATURE REVIEW .................................................................. 13

2.1 Biologically Inspired Algorithms ......................................................................... 14

2.1.1 Evolutionary Algorithms ............................................................................ 14

2.1.2 Stigmergic Optimization Algorithms .......................................................... 19

2.1.3 Swarm-Based Optimization Algorithms ..................................................... 22

2.2 Botanically Inspired Algorithms ........................................................................... 27

2.2.2 A Botany-Grafting Inspired Algorithm ...................................................... 30

2.3 Water Flow Inspired Techniques .......................................................................... 30

2.3.1 Image Processing Methods Based on Water Flow Model .......................... 30

2.3.2 Intelligent Water Drops Algorithm ............................................................. 35

2.3.3 Water Flow-Like Algorithm ....................................................................... 37

2.4 Conclusions and Possible Nature-Inspired Algorithm .......................................... 39

CHAPTER 3 A GENERAL WATER FLOW ALGORITHM .............................. 43

3.1 Hydrological Cycle in Meteorology ..................................................................... 44

3.2 Erosion Process of Water Flow in Nature............................................................. 47

3.3 General Water Flow Algorithm ............................................................................ 51

3.3.1 Encoding Scheme........................................................................................ 54

3.3.2 Memory Lists .............................................................................................. 56

3.3.3 Exploration Phase ....................................................................................... 57

3.3.4 Exploitation Phase ...................................................................................... 58

3.3.4.1 Erosion Condition and Capability .................................................. 58

3.3.4.2 Erosion Process .............................................................................. 61

CHAPTER 4 WFA FOR PERMUTATION FLOW SHOP SCHEDULING ....... 65

4.1 Introduction .......................................................................................................... 66

4.2 Formulation of the PFSP ....................................................................................... 68

4.3 WFA for the PFSP ................................................................................................ 69

4.3.2 Memory Lists .............................................................................................. 70

4.3.3 Exploration Phase ....................................................................................... 70

4.3.4 Exploitation Phase ...................................................................................... 71

4.3.5 A Numerical Example for Erosion Machenism .......................................... 72

4.4 Computational Experiments and Comparisons ..................................................... 78

4.4.1 Benchmark Problem Sets ............................................................................ 78

4.4.2 Platform and Parameters ............................................................................. 79

4.4.3 Performance Measure ................................................................................. 79

4.4.4 Computational Results ................................................................................ 81

4.5 Conclusions ........................................................................................................... 84

CHAPTER 5 WFA FOR FLEXIBLE FLOW SHOP SCHEDULING ................. 85