AN IMPROVED ANT SYSTEM ALGORITHM FOR UNEQUAL AREA FACILITY LAYOUT PROBLEMS

KOMARUDIN

A thesis submitted in fulfillment of the requirements for the award of the degree of Master of Engineering (Mechanical Engineering)

Faculty of Mechanical Engineering Universiti Teknologi Malaysia

OCTOBER 2009

iii

To my God, Allah 'azza wa jalla Then to my beloved parents, wife and daughter

iv

ACKNOWLEDGEMENTS

In preparing this thesis, I was in contact with many people, researchers, and academicians. They have contributed towards my understanding and thoughts. In particular, I wish to express my sincere appreciation to my main thesis supervisor, Dr. Wong Kuan Yew, for his encouragement, guidance, critics and friendship. I am also indebted to Universiti Teknologi Malaysia (UTM) for funding my Master study.

My fellow friends from the Indonesian Student Association (PPI) should also be recognized for their support. My sincere appreciation also extends to all my colleagues and others who have provided assistance at various occasions. Their views and tips are useful indeed. Unfortunately, it is not possible to list all of them in this limited space. I am grateful to all my family members.

ABSTRACT

To date, a formal Ant Colony Optimization (ACO) based metaheuristic has not been applied for solving Unequal Area Facility Layout Problems (UA-FLPs). This study proposes an Ant System (AS) algorithm for solving UA-FLPs using the Flexible Bay Structure (FBS) representation. In addition, this study proposes an improvement to the FBS representation when solving problems which have empty spaces. The proposed algorithm uses several types of local search to improve its search performance. It was extensively tested using 20 well-known problem instances taken from the literature. The proposed algorithm is effective and can produce all of the best FBS solutions (or even better) except for 2 problem instances. In addition, it can improve the best-known solution for 7 problem instances. The improvement gained by the proposed algorithm is up to 21.36% compared to previous research. Evidently, the proposed algorithm is also proven to be effective when solving large problem sets with 20, 25, and 30 departments. Furthermore, this study has implemented a Fuzzy Logic Controller (FLC) to automate the tuning of the AS algorithm. The experiments involved tuning four parameters individually, i.e. number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate, as well as tuning all of them at once. The results showed that FLC could be used to replace manual parameter tuning which is time consuming. The results also showed that instead of using static parameter values, FLC has the potential to help the AS algorithm to achieve better objective function values.

vi

ABSTRAK

Sehingga kini, metaheuristik Ant Colony Optimization (ACO) belum lagi digunakan untuk menyelesaikan Unequal Area Facility Layout Problems (UA-FLPs). Matlamat projek ini adalah untuk membangunkan satu algoritma Ant System (AS) untuk menyelesaikan UA-FLPs dengan menggunakan model Flexible Bay Structure (FBS). Selain itu, kajian ini juga membuat pembaikan kepada model FBS apabila menyelesaikan UA-FLPs yang mengandungi ruang kosong. Algoritma yang dibina tersebut menggunakan beberapa jenis pencarian tempatan (local search) untuk meningkatkan prestasi pencariannya. Ianya diuji secara luas menggunakan 20 masalah terkenal yang diambil daripada literatur. Algoritma tersebut adalah berkesan dan boleh menghasilkan semua solusi FBS terbaik (bahkan lebih baik) kecuali untuk 2 set masalah. Selain itu, algoritma tersebut juga boleh memperbaiki solusi terbaik untuk 7 set masalah. Pembaikan yang diperolehi mencapai 21.36% bila dibandingkan dengan kajian-kajian yang lepas. Jelasnya, algoritma ini juga berkesan untuk menyelesaikan masalah bersaiz besar yang mempunyai 20, 25, dan 30 jabatan. Tambahan lagi, projek ini telah menerapkan Fuzzy Logic Controller (FLC) untuk menala secara otomatik nilai parameter-parameter di dalam algoritma AS. Eksperimen dengan FLC membabitkan penalaan empat parameter secara berasingan, iaitu number of ants, pheromone information parameter, heuristic information parameter, dan evaporation rate, dan penalaan semua empat parameter secara sekaligus. Hasilnya menunjukkan bahawa FLC boleh digunakan untuk menggantikan penalaan parameter secara manual. Hasilnya juga menunjukkan bahawa FLC berpotensi untuk membantu algoritma AS untuk mencapai fungsi objektif yang lebih baik daripada menggunakan nilai parameter yang statik.

vii

TABLE OF CONTENTS

CHAPTER DECLARATION DEDICATION

TITLE

PAGE ii iii iv v vi vii x xi xii xv 1 1 2 3 4 4 5 6 7 7 7 8

ACKNOWLEDGEMENTS ABSTRACT ABSTRAK TABLE OF CONTENTS LIST OF TABLES LIST OF FIGURES LIST OF ABBREVIATIONS LIST OF APPENDICES 1 INTRODUCTION 1.1 1.2 1.3 1.4 1.5 1.6 1.7 2 Overview Background of Research Problem Statement Objective of Study Scope of Study Organization of Thesis Conclusions

LITERATURE REVIEW 2.1 2.2 Introduction Combinatorial Optimization Problem 2.2.1 Complexity of Combinatorial Optimization 2.2.2 Intensification and Diversification in

viii Metaheuristic 2.2.3 Parameter Tuning for Metaheuristics 2.3 2.4 Facility Layout Problems Unequal Area Facility Layout Problems 2.4.1 Exact procedures 2.4.2 Heuristic Procedures 2.4.3 Metaheuristic Algorithm 2.5 Ant Colony Optimization 2.5.1 Ant Colony Optimization Framework 2.5.2 ACO Applications in Facility Layout Problems 2.5.3 Parameter Tuning for Ant Colony Optimization 2.6 3 Conclusions 10 12 15 20 26 30 32 35 36 38 41 45 47 47 47 50 52 54 54 56 57 57 57

METHODOLOGY 3.1 3.2 Introduction Methodology 3.2.1 Ant System Development 3.2.2 Fuzzy Logic Controller (FLC) Development 3.3 3.4 3.5 Research Hypothesis Data Collection Conclusions

ANT SYSTEM FORMULATION 4.1 4.2 Introduction Structure of the Proposed Ant System (AS) Algorithm 4.2.1 Solution representation of Unequal Area Facility Layout Problems 4.2.2 Objective function 4.2.3 Ant Solutions Construction 4.2.4 Local Search Procedures 4.2.5 Pheromone update scheme 4.2.6 Stopping criteria 4.3 4.4 Implementation details of the algorithm Conclusions

59 62 63 66 67 68 68 72

ix 5 PARAMETER TUNING USING FUZZY LOGIC 5.1 5.2 5.3 5.4 5.5 6 Introduction Manual Parameter Tuning Fuzzy Logic Controller FLC-AS Formulation Conclusions 73 73 73 75 77 81 82 82

EVALUATION OF THE PROPOSED ALGORITHM 6.1 6.2 Introduction Problem Sets and Parameter settings used to evaluate the Proposed Algorithm 6.3 Evaluation of the Proposed Algorithm with Manual Parameter Tuning 6.4 Evaluation of the Proposed Algorithm with Fuzzy Logic Controller 6.5 Conclusions

82

85

88 93 94 94 95 97 98 110 111-136

CONCLUSIONS 7.1 7.2 7.3 Introduction Conclusions Future Work

REFERENCES PUBLICATIONS Appendices A-B

LIST OF TABLES

TABLE NO. Table 2.1 Table 2.2 Table 2.3 Table 2.4 Table 2.5 Table 3.1 Table 3.2 Table 5.1 Table 5.2

TITLE Comparisons of FLP classes Ant Colony Optimization general form Applications of ACO algorithms in FLPs ACO parameter values in various optimization problems Parameter Tuning for ACO algorithms UA-FLP problem sets used in this research Characteristics of problem sets used in previous research Manual parameter tuning results Rule bases for (a) number of ants and evaporation rate, and (b) pheromone information parameter and heuristic information parameter

PAGE 17 36 39 42 45 55 56 75

79 83

Table 6.1 Table 6.2

Problem set data Statistical data on results and computation time of the proposed algorithm

85

Table 6.3

Results and comparison of the proposed algorithm with previous research 86

Table 6.4

Results and comparison of the proposed algorithm with the best-FBS and best-known solutions 87

Table 6.5

Results and comparison of the proposed algorithm with Fuzzy Logic Controller 89

xi

LIST OF FIGURES

FIGURE NO. PAGE Figure 2.1

TITLE

Exponential complexity and its blow-up point (modified from Ibaraki, 1987) 9

Figure 2.2

An illustration of intensification and diversification in metaheuristics (adapted from Blum and Roli, 2003) 11 22

Figure 2.3 Figure 2.4

(a) Block Layout and (b) Detailed Layout

The solution of UA-FLP with continuous representation by Anjos and Vannelli (2006) 23 24 25 25

Figure 2.5 Figure 2.6 Figure 2.7 Figure 2.8

The solution of UA-FLP with QAP model by Hardin and Usher The solution of UA-FLP with FBS model in Shebanie II (2004) The solution of UA-FLP with STS model in Shebanie II (2004)

Demonstration of the shortest path finding capability of an ant colony (Blum, 2005a) 36 48 51 58

Figure 3.1 Figure 3.2 Figure 4.1 Figure 4.2 Figure 4.3 Figure 4.4

Research methodology main steps FLC-AS development methodology The proposed AS algorithm

An Example of Solution Representation in the Proposed Algorithm 59 Layout generated from the ant solution representation in Figure 4.2 60 An example of layout solution with (a) original FBS and (b) modifiedFBS 61 69 76 78

Figure 4.5 Figure 5.1 Figure 5.2 Figure 5.3

Details of the proposed AS algorithm Scheme for parameter tuning in metaheuristics The proposed FLC-AS algorithm

Graphs showing membership functions for (a) input parameters, and (b) output parameters 79

xii

LIST OF SYMBOLS

the total area of the facility

cij

the cost per unit distance per unit material from department i to department j

Cik

department i located in location candidate k (department i located in the kth element of the department sequence)

dij

the distance between departments i and j

f(s)

ant objective function

F(s)

quality function

fij

the number of material flow from departments i to department j

the height of the facility

hi

the height of department i

li

length of department i

lmin

minimum length of department

xiii mi sum of material flow from and to department i

the number of departments

N(s)

available departments which have not been used in the corresponding ant

p(Cik) -

probability to locate department i to location candidate k

pinf

the number of infeasible departments

Supd

the set of ants that is used for updating the pheromone

Vall

the best overall objective function value found

Vfeas

the best feasible objective function value found

weight for the corresponding ant solution s.

the width of the facility

wi

the width of department i the centroid of an area below fuzzy membership function c.

xk

x axis value for the center of location candidate k

xo

the crisp value

yk

y axis value for the center of location candidate k

xiv max k c ik pheromone information parameter

maximum aspect ratio

heuristic information parameter

the parameter for location candidate k

fuzzy membership function

evaporation rate

pheromone value associated with the solution to locate department i to location candidate k

xv

LIST OF APPENDICES

APPENDIX

TITLE

PAGE

A B

Data Sets for UA-FLPs Best layout obtained by the proposed AS algorithm

112 127

CHAPTER 1

INTRODUCTION

1.1

Overview

Determining the physical organization of a production system is defined as a Facility Layout Problem (FLP). The term production system does not only point to manufacturing, but it also concerns service and communication systems (Meller and Gau, 1996). The objective of an FLP is mainly to minimize the total material handling cost of the production system. It can also include or combine various objectives such as to maximize space utilization, maximize flexibility, and maximize employee satisfaction and safety (Muther, 1955). In general, FLPs can include various layouts such as manufacturing, hospital, office, and construction layouts.

One of the important FLPs is the Unequal Area Facility Layout Problem (UAFLP) which was originally formulated by Armour and Buffa (1963). In UA-FLPs, there is a fixed rectangular facility with dimensions H and W, where H is the height and W is the width. A number of departments which do not need to have the same area requirement must be arranged according to the following criteria: (1) all departments must be located inside the facility, (2) all departments must not overlap with each other, and (3) the final dimensions of the departments must fulfill some maximum ratio constraints and/or minimum value restrictions (Meller and Gau, 1996). The goal of the problem is to partition the facility into departments so as to minimize the total material movement cost. This objective is based on the material handling principle that material handling cost increases proportionally with the distance which must be traveled by the materials.

A UA-FLP is a complete Nondeterministic Polynomial (NP) problem and thus, recent research using exact algorithms can only optimally solve up to 11 departments (Meller et al., 2007). In recent years, researchers have proposed several metaheuristic approaches to obtain high quality solutions for large UA-FLPs (Meller and Gau, 1996; Singh and Sharma, 2005). Researchers have also proposed several UA-FLP representations such as Flexible Bay Structure (FBS), Slicing Tree Structure (STS), and Quadratic Assignment Problem (QAP) to reduce the complexity of UAFLPs (Meller and Gau, 1996).

1.2

Background of Research

As a metaheuristic, the Ant Colony Optimization (ACO) family was introduced by M. Dorigo and his colleagues in the early 1990s (Dorigo et al., 1996). The first ACO algorithm developed was Ant System (AS) and subsequently, many variants have been developed in order to improve the algorithm performance. Among other ACO variants are Elitist AS (EAS), Rank-based AS (RAS), MAX-MIN Ant System (MMAS), and Ant Colony System (ACS). ACO was initially tested for solving the Traveling Salesman Problem (TSP), and then it has been implemented to solve various kinds of optimization problems such as Facility Layout Problem (FLP), scheduling, vehicle routing, timetabling, data mining, bioinformatics problems, etc. (Blum, 2005a).

One area which has been successfully solved by ACO is FLP. It is a very hard optimization problem and can be applied in many application areas (Meller and Gau, 1996). These reasons have encouraged many researchers to study it. The use of ACO to solve FLPs is mainly based on two reasons (Sttzle and Dorigo, 1999). Firstly, ACO is suitable to solve discrete optimization problems because its solution structure uses a discrete solution representation. Secondly, ACO can use heuristic information to exploit the problems. This heuristic information can help ACO to search and examine promising solution regions.

3 Unfortunately, the application of ACO in solving FLPs is mainly limited to Quadratic Assignment Problems (QAPs) as will be discussed later in the literature review section. Only little research has been done to solve other classes of FLPs. For this reason, this research aims to implement ACO for solving another class of FLP, i.e. UA-FLPs.

When using metaheuristic algorithms, one of the important factors is the balance between intensification and diversification in the search space. To provide this balance, determining the right values for metaheuristic parameters is one of the critical issues (Eiben et al., 1999). To handle this matter, this research implements a Fuzzy Logic Controller (FLC) to automatically tune the ACOs parameters.

1.3

Problem Statement

UA-FLP as one of the well-known problems in facility planning has been studied for years. It is a very difficult problem because of the constraints contained. Conventional methods can only optimally solve UA-FLPs with up to 11 departments (Meller et al., 2007). Meanwhile, the size of the problem still grows and the largest problem currently has 35 departments (Liu and Meller, 2007). Therefore, researchers have proposed several metaheuristic approaches to solve UA-FLPs especially for the large problems (Tate and Smith, 1995; Liu and Meller, 2007; Scholz et al., 2009). Basically, metaheuristic approaches can only report the best-obtained solutions since they may not guarantee to obtain the optimum solutions. Therefore, there is an opportunity to use other metaheuristic approaches to improve the best-known solutions for UA-FLPs.

On the other hand, ACO has not been used for solving UA-FLPs although it is considered as one of the state of the art algorithms for solving FLPs, particularly QAPs. Therefore, it is advantageous to implement ACO for solving UA-FLPs in order to show that it can be used as an alternative to build a facility layout. Such an implementation may achieve better results compared to approaches proposed by

4 previous studies. This will then encourage the use of ACO for solving UA-FLPs, with more confidence.

The selection of parameter values for ACO is essential to its performance. Past methods which have been used for parameter tuning in ACO have been reviewed but no one has used FLC to tune ACOs parameter values. On the other hand, FLC performed well when it was implemented to tune the parameter values of other metaheuristic algorithms. Therefore, the use of FLC to automatically tune the ACOs parameters is considered necessary because it may benefit ACOs performance. If the results are as good as those obtained from manual tuning, then FLC can be used to replace manual tuning which is time consuming.

1.4

Objective of Study

The objectives of this research are to:

i.

Formulate an AS algorithm with FLC for solving UA-FLPs. Particularly, FLC is used to automatically tune the parameter values in AS.

ii. Evaluate the performance of the algorithm in solving UA-FLPs. The evaluation is based on comparing its performance, technically the best objective function values achieved with those obtained from previous studies.

1.5

Scope of Study

This study covers the design, development, and evaluation of the proposed algorithm for solving UA-FLPs. In addition, this research is restricted by the limitations listed below:

5 i. This study chooses the AS algorithm for the ACO implementation and it does not consider other ACO variants. AS is chosen because this research is an initial effort to use ACO for solving UA-FLPs and furthermore, AS is the most basic algorithm of ACO.

ii. In order to simplify UA-FLPs, this research uses the FBS model representation. The FBS model is selected because it has been thoroughly studied and thus, it provides a good basis for comparison.

iii. All the case problems are taken from previous studies which have been published in the literature. The smallest problem contains 7 departments whereas the largest one has 35 departments.

iv. This study is only interested in solving UA-FLPs which have a rectangular facility with fixed dimensions.

v. The total department areas must not exceed the facility area. In addition, this research will also solve UA-FLPs with empty spaces.

vi. All the departments are considered as rectangular and they have certain constraints that bound their dimensions with minimum and maximum values. In addition, this research does not take into account problems which have departments with fixed dimensions or locations.

vii. Rectilinear distance is used in material movement cost calculation.

1.6

Organization of Thesis

This thesis is structured as follows. Chapter 2 describes the literature review which lays down the fundamentals of UA-FLP and ACO. It provides a

6 comprehensive review on the previous approaches used to solve UA-FLPs and clearly highlights that ACO has not been used for solving them. Chapter 3 discusses the methodology used in conducting this research. Chapter 4 introduces the formulation of the proposed AS algorithm for solving UA-FLPs. Meanwhile, chapter 5 presents the FLC formulation applied to the AS algorithm. In chapter 6, evaluation results obtained from the computational experiment are presented and discussed. Finally, chapter 7 gives several conclusions for the research and directions for future work.

1.7

Conclusions

In this chapter, the background of the research, its objectives and scopes have been described. The idea of using AS for solving UA-FLPs is a relatively novel concept. It is indicated that the use of FLC may improve the performance of AS by balancing its search intensification and diversification. This research is expected to produce an AS algorithm for solving UA-FLPs. This research is also expected to become a new paradigm for tuning the parameters in AS by using FLC. In the next chapter, a thorough review of the related literature will be provided.

CHAPTER 2

LITERATURE REVIEW

2.1

Introduction

This chapter provides the basic knowledge of metaheuristics, Ant Colony Optimization (ACO), Unequal Area Facility Layout Problems (UA-FLPs) and parameter tuning for ACO. This chapter becomes the foundation for formulating an Ant System (AS) for solving UA-FLPs. This chapter also gives the starting point to develop a Fuzzy Logic Controller (FLC) to automatically tune the AS's parameters.

As will be described later, AS has not been used for solving UA-FLPs. Furthermore, it is also shown that FLC has not been used for automatically tuning the AS's parameters. These two considerations have motivated the author to conduct this research on using AS and FLC for solving UA-FLPs.

2.2

Combinatorial Optimization Problem

The term Combinatorial Optimization (CO) is composed of combinatorial and optimization. An optimization problem P is generally defined as follows (Ibaraki, 1987):

P: maximize (minimize) f (x ) subject to x S

(2.1)

where S X denotes a feasible region in the underlying space X. Namely, S is the set of feasible solutions satisfying the imposed constraints. The function f : S R , where R is the set of real number, is called the objective
function. A feasible solution x S is optimal if no other feasible solution y satisfies

f ( y ) > f ( x) in the case of maximization (or

f ( y ) < f ( x ) in the case of

minimization). Maximization can be expressed as minimization since maximizing

f (x ) is equal to minimizing f (x) , and vice versa. The sets X and S take a
variety of forms in application.

P is a CO problem if X and S are discrete or combinatorial (implying that X and S are discrete sets of finite elements or countable infinite elements) (Ibaraki,

1987). From the definition, it can be concluded that many problems such as, Linear Programming (LP), Traveling Salesman Problem (TSP), Quadratic Assignment Problem (QAP), Vehicle Routing Problem (VRP), and more importantly, Unequal Area Facility Layout Problem (UA-FLP) are in the CO domain.

2.2.1 Complexity of Combinatorial Optimization

The complexity of an optimization problem can be determined with two important measures, time complexity and space complexity (Ibaraki, 1987; Maffoli and Galbiati, 2000). Time complexity is related to the number of computation steps such as additions, multiplications, comparisons, and so forth which is required to solve the optimization problem, whereas space complexity is associated with the size of the required space when performing the computation, i.e., the input data space and the working space.

An optimization problem is regarded as easy if it can be solved by an algorithm with time complexity O( N k ) for a constant k, where N is the size of the problem. Such a complexity is said to have a polynomial order, and the algorithm is

9 called a polynomial time algorithm. On the other hand, if any algorithm for solving the problem requires a complexity not bounded by a polynomial N, it is considered to be difficult. Typical algorithms which do not have a polynomial order complexity are having a logarithmic order O( N log N ) or exponential order O (k N ) .

Unfortunately, many CO problems turn out to be difficult to solve. Many CO problems show typical behaviors of exponential functions; they blow-up at certain points as shown in Figure 2.1. In other words, algorithms with exponential complexity can be practical only when they are used for solving a problem with size
N smaller than its blow-up point. So it is more convenient to use approximation

algorithms for solving a problem with size higher than its blow-up point.

Computation Time

Blow up point

Problem Size

Figure 2.1 Exponential complexity and its blow-up point (modified from Ibaraki, 1987) Approximation algorithms do not guarantee optimal solutions but can find near-optimal solutions. One of the approximation algorithms which is extensively explored nowadays is metaheuristic. Osman and Laporte (1996) defined metaheuristic as an iterative generation process which guides a subordinate heuristic by combining intelligently different concepts for exploring and exploiting the search space, and learning strategies are used to structure information in order to find nearoptimal solutions efficiently.

10 For a broaden view, Gendreau and Potvin (2005) have proposed a unifying
view of metaheuristics. They argued that every metaheuristic can be divided, but not

always into a small number of algorithmic components. They are:

Construction. This is used by all metaheuristics for creating initial solution(s).

Recombination. This component generates new solutions from the current ones

through the recombination process.

Random Modification. This is used to modify current solution(s) through random

perturbations.

Improvement. This is used to explicitly improve the solutions(s); for example, by

selecting the best solution in the neighborhood, by applying a local search or by projecting the solution in the feasible region.

Memory Update. This component updates either the standard memories (Tabu

Search), pheromone trails (ACO), populations (Genetic Algorithm) or reference sets (Scatter Search).

Parameter/Neighborhood Update. This component adjusts the parameter values

or modifies the neighborhood structures.

2.2.2 Intensification and Diversification in Metaheuristic

The effectiveness of metaheuristics is measured by two considerations, i.e. quality of solution, and computation time needed. A good metaheuristic can produce a solution as close as possible to the optimal solution and still minimizing the computation time needed.

11 To deal with these two contradictive objectives, the terms intensification and
diversification have been introduced (Glover and Laguna, 1997). Occasionally these

terms are called in different names, such as exploitation and exploration (Eiben et
al., 1999), or local search and global search (Blum and Roli, 2003). These terms

help us to understand the nature of metaheuristic algorithms and at the same time, give guidelines to improve them.

A very good explanation about these terms has been delivered by Glover and Laguna (1997). According to them, intensification focuses on examining neighbors of elite solutions while diversification encourages the search process to examine unvisited regions and generates solutions that differ in significant ways from those which have been found before. These descriptions are illustrated in Figure 2.2.

Figure 2.2 An illustration of intensification and diversification in metaheuristics (adapted from Blum and Roli, 2003) As can be seen in Figure 2.2, during the intensification, the metaheuristic algorithm excessively tries to find a neighborhood solution which is better than its current elite solution. If the search only relies on the intensification component, it

12 will trap the algorithm in the local optimum since no other neighborhood solution is better than the local optimum. The search should avoid wasting excessive computing time in the local optimum region and thus diversification should be activated. So, the algorithm must also incorporate the diversification component to allow it to leave the local optimum by jumping the search process to an unvisited region. The notions neighbor, locality, and area (or region) are subjected to the characteristics of the solution space.

It can be concluded that the intensification and diversification components play an important role in the success of metaheuristic algorithms. In order to have the ability to control intensification or diversification, researchers have proposed several methods (Blum and Roli, 2003). One method which implicitly takes control of this role is parameter tuning (Eiben et al., 1999). Usually, this activity is done during the end of an iteration as described in the previous section (Gendreau and Potvin, 2005). This technique will guide the search process whether it should activate intensification or diversification.

2.2.3 Parameter Tuning for Metaheuristics

Parameter tuning can be identified as a process to find the correct combination of an algorithm's parameters for each individual problem (Bhrde et al., 2005). It should be noted that parameter tuning is different from parameter setting (Eiben et al., 1999). The reasons for parameter tuning are (Eiben et al., 1999; Wolpert and Macready, 1997):

Generally, the performances of all metaheuristic algorithms depend on their parameter values. Small size optimization problems are usually not too influenced by their parameter values.

13 There are no universal parameter values which can solve all optimization problems effectively and efficiently, and thus, parameter tuning is needed.

Research has shown that parameter tuning has several purposes. It can be used for one or more of these purposes (Bhrde et al., 2005; Dai et al., 2006; Hartono
et al., 2004; Rpke and Pisinger, 2006; Syarif et al., 2004):

To balance between intensification and diversification during the search process

To speed-up the algorithm by giving good operators more probabilities to occur

To maintain the solution feasibility by controlling the penalty function parameter

Parameter tuning methods can be differentiated by examining their conditions when they are used in a metaheuristic algorithm. If the parameter values being tuned are always same or not changing during the execution, the method can be considered as a static parameter tuning system. For this system, parameter values are tuned before the running of the algorithm and they remain static during the execution. On the other hand, a parameter tuning method which has its parameter values being changed during the running of the algorithm is called a dynamic parameter tuning
system. Apart from this two systems, there can be a mix system (Bhrde et al. 2005),

which tunes the parameters before running the algorithm, and the parameter values are dynamic during the execution of the algorithm.

By examining the properties of the methods, the author classifies the static parameter tuning system into two classes, experimentation method, and landscape
analysis method. Meanwhile a dynamic parameter tuning method can be categorized

into four classes; random value method, deterministic method, adaptive method, and
self-adaptive method.

Experimentation method or manual parameter tuning is the earliest method used in parameter tuning. Since the initial development of metaheuristic algorithms,

14 researchers have realized about the importance of parameter values. Different parameter values might lead to different optimum results and/or computation time. For this reason, researchers performed several experimentations to execute metaheuristic algorithms with various parameter values. Usually, the values which yielded the best result were claimed as the optimum parameter values. It could be observed that the works done by De Jong (1975) and Grefenstette (1986) were the earliest attempts to use manual parameter tuning. Although it has several drawbacks, this method is still used nowadays because of its simplicity.

Another method to tune parameter values in metaheuristic algorithms is landscape analysis method. This method is influenced by the No Free Lunch
Theorem (NFLT) (Wolpert and Macready, 1997). The theorem argues that one single

algorithm will not perform better than other algorithms for all kinds of optimization problems and their instances. This is because every optimization problem has its own characteristics and behaviors (Koppen, 2004). Within an optimization problem, it has many problem instances that may have different difficulties and properties. Therefore, a metaheuristic algorithm must be able to deal with different CO problems and their instances in a different way. Thus, Bhrde et al. (2005) proposed to classify the problem landscape before effective parameter adaptation may occur in a metaheuristic.

Another approach dealing with parameter tuning is using random parameter values. Mostly, researchers who called their method as parameter free algorithm did this (Sawai and Kizu, 1998). They argued that due to the robustness of their method, there is no need to tune the parameter values. While other researchers proposed the parameter values to be dynamic, this method fixes several parameters and changes the rest dynamically by using random functions. Usually they use a narrow scale of parameters to be randomized. For example, if the mutation rate in Genetic Algorithm (GA) has a value between 0 and 1, this method randomizes it between 0.6 and 0.8. This method has helped people to deal with parameter tuning, but its efficiency and effectiveness can be hard to prove.

15 Deterministic method changes the parameter values with deterministic rules. The parameters constantly change, either ascending or descending or change according to other deterministic functions. The change is not related to the algorithm performance; it just depends on the number of iterations. It can be argued that Simulated Annealing (SA) has a similar behavior to this method.

Adaptive method is extensively used in tuning the parameters of metaheuristic algorithms. In this method, the results from earlier iterations will be extracted to become useful information for guiding the parameter tuning process. This method tries to adapt the state of a running metaheuristic algorithm and takes an action which can balance intensification and diversification in that particular metaheuristic. Many studies have been done with this method (Ko et al., 1996; Hartono et al., 2004; Bhrde et al., 2005; Dai et al., 2006; Rpke and Pisinger 2006).

Another interesting method for parameter tuning is self-adaptive parameter tuning. This method tries to treat parameters as decision variables which must be solved. This method usually comes from the evolutionary computation family. In this method, parameters are encoded into chromosomes and processed with mutation and recombination. The better values of the encoded parameters will lead to better individuals, which in turn are more likely to survive and produce offsprings, and hence propagate these better parameter values. This method was shown to perform effectively in Hinterding (1997).

2.3

Facility Layout Problems

Determining the physical organization of a production system is defined as a Facility Layout Problem (FLP). The term production system does not only point to manufacturing but it also concerns service and communication systems (Meller and Gau, 1996). The objective of an FLP is mainly to minimize the total material handling cost of the production system. It can also include or combine various

16 objectives such as to maximize space utilization, maximize flexibility, and maximize employee satisfaction and safety (Muther, 1955).

Facility layout plays an important role in the manufacturing environment since it determines the physical relationships of manufacturing activities. It represents the locations of sub-facilities where materials will be moved from and to. Therefore, it determines the traveling distance of materials in manufacturing processes. Material movement can be considered as a non-value added process in manufacturing. Thus, facility layout must be considered as an approach to control material movement cost.

In manufacturing, material movement cost is considered as a part of material handling cost. In a typical manufacturing industry, material handling accounts for 25% of all employees, 55% of all factory space, and 87% of production time (Frazelle, 1986). In addition, material handling is estimated to represent between 15% and 70% of the total cost of a manufactured product (Tompkins et al., 2003). The decrease of material movement cost will reduce the material handling cost and will eventually cut down the manufacturing cost.

FLP is one of the old engineering problems and still an active area nowadays. Muther (1955) has developed a systematic procedure for designing plant layout in the 1950s. Research on facility layout still continues until nowadays, including innovative and developmental works to integrate all components in facilities planning. Two thorough literature reviews on FLPs were done by Meller and Gau (1996) and Singh and Sharma (2005). Meller and Gau (1996) reviewed nearly 100 papers published in 10 years before 1996, and Singh and Sharma (2005) reviewed nearly 140 papers published in 20 years before 2005. Both of them discussed recent trends in facility layout research in their respective time. They highlighted a tendency to integrate the design of a facility layout with the design of a material handling system. In addition, they also concluded the need for a stochastic facility layout rather than a static one.

17 Muther (1955) has listed the objectives of plant layout. Although the objectives are dedicated for plant layout, they are also relevant for facility layout. The objectives can be summarized as follows:

to reduce the traveling distances of materials.

to have a regular flow of the parts and products and to recucenot permitting bottleneck in the production.

to effectively utilize the space occupied by the facilities.

to enhance satisfaction and safety of the workers.

to obtain flexibility that can be easily readjusted for changing conditions.

FLP can be divided into three main categories, namely, (1) QAP, (2) UA-FLP, and (3) Machine Layout Problem (MLP). The comparisons of these three categories are summarized in Table 2.1

Table 2.1 Comparisons of FLP classes

Problem type QAP UA-FLP MLP

No 1 2 3

Department size Same size, fixed dimension or ignored Varied size, decision variables Varied size, fixed dimension

Location candidate Fixed location Decision variables Decision variables

Facility area Fixed or ignored Total department areas Total department areas or free dimension

QAP was introduced by Koopmans and Beckman (1957) to model the problem of locating interacting facilities of equal areas. The objective of QAP is to find an assignment of all interacting facilities to all locations such that the total cost of assignment is minimized. QAP is a special case of FLP in which all departments have equal areas and all locations are fixed and known.

18 UA-FLP was originally formulated by Armour and Buffa (1963). The goal is to partition a region into departmental sub-regions so as to minimize the total material movement costs. The department areas do not need to be the same but there are dimension ratio constraints or minimum dimension restrictions. Researchers have been proposing several representations to reduce the complexity of UA-FLP. Two successful representations are the Flexible Bay Structure (FBS) model and Slicing Tree Structure (STS) model (Meller and Gau, 1996). These allow researchers to use discrete optimization algorithms to solve UA-FLPs.

MLP has several fixed size machines and a number of products which need to be produced in specific processes. The objectives are to determine the location of machines in a given space while at the same time minimizing material handling (Tompkins et al., 2003).

In recent research, some publications have discussed special cases of FLP. Below are several special cases of FLP which have been published in the literature:

1. Multi-Floor Facility Layout Problem

The problem is to allocate sub-facilities in a certain facility which has many floors. In addition, there are interactions between floors causing material handling in vertical movement. The objectives of this problem are to determine the dimensions and locations of departments (including which floor they will be placed), as well as the vertical material handling systems. The algorithms for solving Multi-Floor FLP have been discussed in Johnson (1982), Bozer et al. (1994), Meller and Bozer (1996) and Meller and Bozer (1997).

2. Facility Layout Problem with Input/Output Points

Rather than using the centroid of each sub-facility to calculate distance, the algorithm must try to find the exact location of the flow in and out points. Usually the points are located on the boundaries of each sub-facility. In addition, research has been done to use the aisle structure to calculate actual distance rather than rectilinear

19 distance. The research on this subject can be referred in Montreuil and Ratliff (1989) and Kim and Kim (1999).

3. Dynamic Facility Layout Problem

This problem uses dynamic production demand as a replacement for static production demand. Facility layout is not optimized for a certain period only, but it is optimized for a number of given periods (or in stochastic demand conditions). The optimum facility layout will be based on the material handling cost in each period and also the re-arrangement cost between periods if there are changes in the facility layout. The research on this issue has appeared in Rosenblatt (1986) and Balakrishnan et al. (1992).

4. Facility Layout Problem in Flexible Manufacturing System (FMS)

Flexible Manufacturing System (FMS) is a production system in which a set of machines and flexible material-handling equipment like robots, Automated Guided Vehicles (AGVs), etc. are linked and controlled by a central computer. FMS is different from the classical manufacturing system due to the higher degree of automation, less number of machines, frequent setups, higher volume and flow of information, etc. The objective of this type of problem is to produce a layout of machines so that it can provide flexibility in FMS. The work of Solimanpur et al. (2005) has tried to solve a single row layout problem in the FMS environment.

5. Facility Layout Problem in Cellular Manufacturing System (CMS)

Cellular Manufacturing System (CMS) is a mix production system strategy between mass production and job shop manufacturing. In CMS, similar parts are classified into part families which will be assigned into an appropriate machine cell. This has a cost-effectiveness advantage because it reduces the part traveling distances since the manufacturing processes of each part are predominantly performed in one machine cell. It can still balance the machine loads because the loads are distributed among machine cells and products are not bounded in a certain

20 cell. On the other hand, this system has the flexibility to adjust production schedules because product families have similarity in processes so that the sequence will not change much. The objective of this type of layout problem is to configure/group machine cells and their locations in a facility. The works of Chen and Srivstava (1994) and Vakharia and Chang (1997) have attempted to solve this problem using metaheuristic approaches.

2.4

Unequal Area Facility Layout Problems

Unequal Area Facility Layout Problem (UA-FLP) is primarily used to model a manufacturing layout. The problem arises when there is a need to make or modify a facility layout. The need can be caused by (i) development of a new manufacturing facility, (ii) expansion of an old manufacturing facility, (iii) alteration of the production amounts and sequences, etc.

UA-FLP was originally formulated by Armour and Buffa (1963). They assumed that there is a given fixed rectangular region, or facility, with dimensions H and W, where H is the height and W is the width of the facility. There are also several departments which need to be assigned to the facility. The number of departments, the area of each department, and the material flow values associated with each pair of departments are assumed to be known. The objective function for minimizing the total material movement cost is given by:

Total Cost =
n

i =1 j =1,i j

d
n

ij

f ij c ij

(2.2)

the number of departments

fij

the total material flow between department i and department j, where

i, j = 1, 2, , n

21
dij

the distance between department i and department j, where i, j = 1, 2, , n

cij

the cost of moving one unit material per unit distance from department i to department j, where i, j = 1, 2, , n

The goal is to partition the facility into departmental sub-regions so as to minimize the total material movement cost. This objective is based on the material handling principle that material handling cost increases with traveling distance. The distance is measured in a variety of ways; here are two ways which are widely used:

Input-output (I/O) points distance: The distance is measured between

specified I/O points of two departments and in some cases, it is measured along the aisles of the facility (Tretheway and Foote, 1994). The major drawback of this accurate measure is that the locations of the I/O points and/or the aisles are still unknown until the detailed layout has been obtained.

Centroid-to-centroid (CTC) distance: The distance is calculated between the

centers (centroids) of two departments. The shortcoming of the CTC distance is department shapes can become very long and narrow if the shape restrictions are loose. This is because an algorithm based on the CTC distance attempts to align the department centroids as close as possible (Tate and Smith, 1995). Another weakness of the CTC distance is an L-shaped department may have a centroid that falls outside of the department (Francis and White, 1974).

Two calculation methods can be used to compute the two distance measures above. Rectilinear distance is the most common calculation method used. It computes distance using paths which are parallel to the perpendicular (orthogonal) axes. Armour and Buffa (1963), Tate and Smith (1995), and Liu and Meller (2007) are among the researchers who used rectilinear distance to compute the distance between two departments. The second calculation method is Euclidean distance. This method computes distance by using a straight-line path connecting two points. Euclidean distances were used in Tam and Li (1991) and van Camp et al. (1991).

22 UA-FLP is very applicable to model real-world manufacturing layout problems, especially those which use a process layout approach. Several case problems found in the literature were originally taken from real-world manufacturing problems (van Camp, 1989; Meller, 1992; Liu and Meller, 2007). In addition, Meller (1992) stated that he used his proposed algorithm to solve the Ohio lamp plant layout belonging to the General Electric Company.

Solving an UA-FLP will produce a block layout, which specifies the relative location of each department (see figure 2.3a). The result can be expanded to obtain a
detailed layout, which specifies the exact department locations, aisle structures,

input/output (I/O) point locations, and the layout within each department (see figure 2.3b).

2 1 2 1 3

(a)

(b)

Figure 2.3 (a) Block Layout and (b) Detailed Layout UA-FLP can be modeled using various representations. Each representation has its particular characteristics and complexities. There are four representation models which are categorized based on the characteristic of the final layout. They are the Continuous model, QAP model, Flexible Bay Structure (FBS) model, and Slicing Tree Structure (STS) model.

1. Continuous model

The Continuous model has the characteristic that the departments' placements and dimensions in the facility are not restricted to any rule other than the problem itself. This ensures the model has the possibility to generate all layout solutions

23 including the exact optimal solution. Nevertheless, this model is very hard to be solved when the solution space becomes large. This leads to the development of other models which are easier to be solved. Generally, the continous model has been used to solve UA-FLP using Mixed Integer Linear Programming (MILP) (Meller and Gau, 1996) and mathematical programming (Anjos and Vannelli, 2006). Figure 2.4 is a final layout example modeled as a continuous representation which cannot be produced by other representations.

Figure 2.4 The solution of UA-FLP with continuous representation by Anjos and Vannelli (2006) 2. Quadratic Assignment Model (QAP) model

Several researchers have used the QAP model for solving UA-FLP (Armour and Buffa, 1963; Hardin and Usher, 2005). All departments in the problem are broken into small grids which have the same area. Then the algorithm tries to gather the grids which come from the same department into an interconnected region. This can be achieved by adding large artificial flows among grids which come from similar departments. As shown in Figure 2.5, this model often produces non-

24 rectangular blocks in the final layout. In addition, Bozer and Meller (1997) showed that this model is ineffective because it implicitly adds department constraints.

Figure 2.5 The solution of UA-FLP with QAP model by Hardin and Usher (2005) 3. Flexible Bay Structure (FBS) model

In the FBS formulation, the placement of departments in an UA-FLP generates columns or bays. Each bay does not need to have the same width and same number of departments. The width of a bay is automatically adjusted by the number of departments it contained. Therefore, the problem becomes simpler and easier to be solved. The problem complexity is reduced into determining the departments' placement order and the total number of departments that each bay contains. FBS has an advantage in that the bays will become candidates for aisle structures and this facilitates users to transform the model into an actual facility design (Konak et al., 2006).

25

Figure 2.6 The solution of UA-FLP with FBS model in Shebanie II (2004) 4. Slicing Tree Structure (STS) model

In the STS model, the final UA-FLP layout can be described as a tree representation. The tree representation contains departments as nodes and also u, b, r, and l as connecting nodes. The connecting nodes act as the slicing operators in the formation of sub-facility regions. The four slicing operators are u for up cut, b for bottom cut, r for right cut, and l for left cut. The department nodes and the slicing operators define relative department locations and adjacencies, as well as the partition sequences. The STS model has been used by Scholz et al. (2009).

Figure 2.7 The solution of UA-FLP with STS model in Shebanie II (2004)

26 Besides developing models to represent an UA-FLP, many researchers have proposed their methodologies to solve it. These methodologies can be classified into exact procedures, heuristic and improvement procedures, and metaheuristic algorithms.

2.4.1 Exact procedures

Montreuil

(1990)

introduced

Mixed-Integer

Programming

(MIP)

formulation for solving UA-FLP. The formulation uses a continuous model to represent the UA-FLP layout. The notations and formulation are shown below.

Notations:

(The first six items of the notations are parameters, while the last four items are decisions variables.)

1. The building is L units in the x-direction and W units in the y-direction.

2. The indices i, j will be used for departments, where i, j = 1, 2, N.

3. Each department i has an area requirements ai.

4. The upper and lower limits on the length or the width of department i are denoted by ubi and lbi respectively.

5. The minimum and maximum perimeters of department i are denoted by pi and Pi respectively.

27 6. The set of positive flows is denoted by F = {f ij }. That is f ij > 0 i, j F and

furthermore F = M . The mth positive flow, fm, originates from department i(m) and terminates at department j(m).

7. The rectilinear distance between department i and department j is expressed as the sum of the distance in the x-direction, dijx, and the distance in the y-direction,
dijy. Note that for flow m, the distances in the x and y directions are denoted by dmx and dmy respectively.

8. The location of department i is indicated by its centroid, which is denoted by (xi,

yi).

9. Each department is rectangular shaped; department i has dimension 2li in the xdirection and dimension 2wi in the y-direction.

10. The relative location decision variables are denoted by zxij and zyij. In general, the
zij decision variables determine whether one department is to the north, south,

east, or west of another department in the layout.

Montreuil's formulation:
x 1 if i must be to the west (left ) of j z ij 0 if the above imposition is not enforced

y 1 if i must be to the north ( above) of j z ij 0 if the above imposition is not enforced

min

f (d
m m

x m

y + dm

)
m

(2.3)

s.t.

x d m xi(m) x j (m)

(2.4)

28
x d m x j ( m ) xi ( m )

(2.5)

y d m yi(m) y j ( m)

(2.6)

y d m y j (m) yi( m)

m i i i i

(2.7)

li x i L l i wi y i W wi

(2.8)

(2.9)

lbi 2l i ubi

(2.10)

lbi 2wi ubi

(2.11)

x y 2 z ij + z x + z ij + z y 3 ji ji

i, j; i < j i, j

(2.12)

x xi + li x j l j + Lz ij

(2.13)

x j + l j x i l i + Lz x i, j ji

(2.14)

y y i + wi y j w j + Wz ij

i, j i, j
i

(2.15)

y j + w j y i wi + Wz y ji pi 4(li + wi ) Pi

(2.16)

(2.17)

The objective (Equation 2.3) is based on the multiplication of the material flow cost and the rectilinear distance between department centroids. A standard linear

29 programming trick is used to linearize the absolute values in the distance function (see Eqs. 2.4-2.7). In Eqs. (2.8) and (2.9) each department is constrained to be within the facility, while in Eqs. (2.10) and (2.11) the maximum and minimum lengths of the department rectangles are constrained. The constraint set (Equation 2.12) ensures that the relative department location constraints are relaxed in two or three directions. In Eqs. (2.13) - (2.16), the relative location decision variables are utilized to ensure that departments do not overlap. Finally, in Equation (2.17), a bounded perimeter constraint is used as a surrogate area constraint because the actual area constraint, 4liwi = ai, is nonlinear. Although this MIP approach is powerful and promising, only problems with six or less departments can be solved optimally (Meller and Gau, 1996). This is due to the extensive computing time requirements for large-scale problems.

Meller et al. (1999) pointed out that Montreuils perimeter constraint is modeled such that an increase in the UA-FLP aspect ratio directly corresponds to the increases in errors between the actual and solved areas. The difference between the original and surrogate areas can be as much as 11%, 25%, 36%, and 44% for aspect ratios of 2, 3, 4, and 5, respectively. Thus Meller et al. (1999) proposed an improved surrogate perimeter constraint that provides a more realistic and effective implementation by forcing a department area to adhere rigorously to changes in its perimeter. They reported that their approach could reach optimal solutions for UAFLPs with up to eight departments.

Sherali et al. (2003) provided a similar approach that significantly reduces errors in department areas by using a polyhedral outer approximation of the area constraints and branching priorities. Using the polyhedral approximation and other innovative techniques, they reported optimal solutions for UA-FLPs with up to nine departments. Meller et al. (2007) have presented a new formulation for UA-FLPs based on the sequence-pair representation. They tightened the structure of the problems, and extended the solvable solution space from problem with nine departments to problem with eleven departments.

30 In addition, Konak et al. (2006) introduced an MIP approach for UA-FLPs with Flexible Bay Structure (FBS). In their method, the nonlinear department area constraints are modeled in a continuous plane without using any surrogate constraint. The department dimension is managed by controlling the width of the bay. The width of each bay is calculated as the total area of the departments assigned to that bay divided by the length of the facility in the y axis direction. Their method is extensively tested and it could find optimal solutions for problems with up to 15 departments.

2.4.2 Heuristic Procedures

The majority of UA-FLP algorithms can be identified as either construction heuristic algorithms or improvement heuristic algorithms. In the former category, the solutions are constructed from scratch. It is considered the simplest heuristic approach to solve UA-FLPs, but the quality of the solution produced is generally unsatisfactory. On the other hand, improvement heuristic algorithms use iterations to improve the initial solution. Improvement methods can be easily combined with construction methods. In general, several heuristic approaches have been published in the literature including CRAFT, SHAPE, NLT, and MULTIPLE (Meller and Gau, 1996).

CRAFT (Computerized Relative Allocation of Facilities Technique) is the oldest improvement-type approach developed by Armour and Buffa (1963). CRAFT works by improving the initial layout through the exchange of departments. It performs two-way or three-way exchanges of the centroids of non-fixed departments that are equal in area or adjacent. For each exchange, CRAFT calculates an estimated cost reduction and it chooses the exchange with the largest estimated cost reduction (steepest descent). It then performs the department exchanges and continues until no estimated cost reduction can be obtained. Constraining the department exchanges to those departments that are adjacent or equal in area is likely to affect the solution

31 quality. The procedure has also been criticized because it can produce departments with irregular shapes (Tompkins et al., 2003).

SHAPE developed by Hassan et al. (1986), is a construction algorithm that utilizes a discrete representation and an objective function based on rectilinear distances between department centroids. The department selection sequence is dependent on a ranking, which is based on each department's flow and a user-defined critical flow value. Department placement begins at the center of the layout. Subsequent department placement is based on the objective function value, with departments placed on each of the layout's four sides. The algorithm is easy to implement; however, because the department shape is controlled by the objective function, the shape of the facility may deteriorate towards the end.

NLT (Nonlinear Optimization Layout Technique) is a construction algorithm developed by van Camp et al. (1991). It is based on nonlinear programming techniques and it utilizes Euclidean distance to calculate distances between department centroids. In NLT, there are three constraints: departments cannot overlap, cannot be located outside the facility, and cannot be assigned to a less than required area. The constrained model is transformed into an unconstrained form by an exterior point quadratic penalty method. With a three-stage approach, a difficult problem is solved using the solution from the previous stage (as an initial solution point). The department shapes are all rectangular.

MULTIPLE (MULTI-floor Plant Layout Evaluation) is a single or multi-floor improvement-type algorithm developed by Bozer et al. (1994). MULTIPLE uses a discrete representation and enhances CRAFT by increasing the number of exchanges considered in each iteration. In addition, MULTIPLE can restrict the irregularity of department shape by using an irregularity measure based on the perimeter and area of each department. However, since it uses a discrete representation, the department shapes may not be rectangular.

Anjos

and

Vannelli

(2006)

introduced

two-phase

mathematical

programming method to solve UA-FLPs. The first phase used in their method is a

32 relaxation of UA-FLPs, which intends to find a good starting point for algorithm implementation. In the second phase, an iterative algorithm is used to solve the exact formulation of the problems as a non-convex Mathematical Program with Equilibrium Constraints (MPEC). Although this two-phase mathematical

programming method achieves better objective function values (as compared to those accomplished by other previous research) when solving UA-FLPs, it has been tested only on one UA-FLP instance introduced by Armour and Buffa (1963).

2.4.3 Metaheuristic Algorithm

Several metaheuristics such as SA, GA, Tabu Search (TS), and Particle Swarm Optimization (PSO) have been used to approximate the solutions for very large UA-FLPs. The SA algorithm originated from the theory of statistical mechanics and is based upon the analogy between the annealing of solid metals and solving optimization problems. GA iteratively searches for the global optimum by generating new solutions through processes similar to genetic reproduction approaches. PSO uses several particles to represent solutions and allows them to fly in a solution space so that they cooperate to converge to the optimum solution.

In terms of the application of metaheuristics, Tam (1992) used LOGIC (Layout Optimization using Guillotine-Induced Cut) which is an SA algorithm to solve UA-FLPs with STS representation. With a given slicing tree and department area values, the layout can be determined by recursively partitioning a rectangular area and placing the departments into the areas according to the four specific branching operators. Since this approach is likely to produce long and narrow department shapes, two shape constraints are added as penalties to the objective function. The algorithm uses two-way exchanges of branching operators as an attempt to find a better layout.

Tate and Smith (1995) proposed a GA for solving UA-FLPs using FBS representation. A dynamic or adaptive penalty function (as shown in Equation 2.18)

33 is used to guide the search into feasible solution regions. This method searches a solution space by varying department-to-bay assignments or by adding or removing bay breakpoints. The algorithm generates good layouts and it is shown to outperform CRAFT and NLT.

p(m) = (m) k (V feas Vall )

(2.18)

where:

p(m) = the penalty function

= the number of infeasible departments

= parameter that determines the severity of the penalty function

Vfeas

= the best feasible objective function value found

Vall

= the best overall objective function value found

Hardin and Usher (2005) used PSO to solve UA-FLPs with QAP representation. Each department is divided into grids/tiles with equal size. The solution construction is performed by managing tiles based on two different rules. The first rule attempts to move tiles to their own department centroid, and the second rule attempts to move tiles to the centroid of other departments which have high volume relationship. Then new solutions are constructed based on the two rules above. This method shows an improvement when compared to the CRAFT method.

Recently, Liu and Meller (2007) proposed an approach to solve UA-FLPs represented as sequence pairs, by using GA and MIP. Conceptually, Liu and Meller (2007) used GA to modify the solutions represented as sequence-pairs. Thereafter, the sequence-pair representation is transformed to become a feasible solution with MIP. The main purpose of using the sequence-pair representation is to eliminate all

34 infeasible binary variables which make large UA-FLPs difficult to solve. Liu and Meller (2007) claimed that their method could achieve optimal solutions for problems with up to eleven departments. Besides this, they have shown some improvements when solving problem instances with larger data sets.

More recently, Scholz et al. (2009) proposed a TS algorithm with STS representation for solving UA-FLPs. They incorporated a bounding curve for dealing with fixed departments and free dimension facility in UA-FLPs. Their algorithm incorporated four types of neighborhood moves to find better solutions. They compared their algorithm with previous research and showed large improvements.

In order to enhance the UA-FLP methods to produce a more realistic layout solution, a Maximum Aspect Ratio (MAR) has been introduced by van Camp et al. (1989), which is given by,

max{l i , wi } max = min {l , w } i i

(2.19)

Where:

max

= maximum aspect ratio

li

= length of department i, i = 1, 2, 3 ..n

wi

= width of department i, i = 1, 2, 3 ..n

The MAR ensures that the UA-FLP methods produce an acceptable physical layout solution. A normal MAR is usually between two and seven. As the MAR becomes smaller, a UA-FLP becomes highly constrained since the allowable department dimensions become more restricted. On the other hand, a MAR higher than seven requires significant computing time because there are more candidate solutions to be considered.

35 From the previous literature review, it can be said that UA-FLP is still an active area. It also shows that Ant Colony Optimization (ACO) or particularly Ant System (AS) has not been utilized to solve UA-FLPs. This encourages the author to apply AS for solving them.

2.5

Ant Colony Optimization

Ant Colony Optimization (ACO) was introduced by Marco Dorigo and his colleagues in the early 1990s (Blum, 2005a). ACO works by imitating the foraging behavior of ants when searching the shortest path to reach their food. Figure 2.8 shows an experiment that demonstrates the shortest path finding capability of ant colonies. In the initial condition, all ants are in their nest and there exists two paths of different lengths between the ants nest and the food source. When searching for food, ants initially explore both paths in equal probability. 50% of the ants take the short path and 50% of them take the long path to the food source. The ants that have taken the short path have arrived earlier at the food source. Therefore, the probability to take again this path is higher. The pheromone trail on the short path receives, in probability, a stronger reinforcement, and the probability to take this path grows. Finally, due to the pheromone evaporation on the long path, the whole colony will, in probability, use the short path.

ACO was initially developed for solving the TSPs. Thereafter, its application was extended to solve various kinds of discrete optimization problems including the QAPs, VRPs, Job-Shop Scheduling Problems (JSPs), etc.

ACO is currently among the state-of-the-art algorithms for solving, for example, the sequential ordering problem (Dorigo and Gambardella, 2000), resource constrained project scheduling problem (Merkle et al., 2002), open shop scheduling problem (Blum, 2005b), and 2D and 3D hydrophobic polar protein folding problem (Shmygelska and Hoos, 2005).

36

Figure 2.8 Demonstration of the shortest path finding capability of an ant colony (Blum, 2005a)

2.5.1 Ant Colony Optimization Procedure

After parameter initialization and problem data input, every ACO iteration consists of ant solutions construction, local search procedures (optional), and pheromone information update (see Table 2.2).

Table 2.2 Ant Colony Optimization procedure

Set parameters, initialize pheromone trails while termination conditions not met do ConstructAntSolutions ApplyLocalSearch UpdatePheromones end while {optional}

37 The explanations of these three components in ACO are as follows (Blum, 2005a):

ConstructAntSolutions(). A set of m artificial ants construct solutions from

elements of a finite set of available solution components. Solution construction starts with an empty partial solution. Then, at each construction step, the current partial solution is extended by adding a feasible solution component without violating any problem constraint. The choice of a solution component is selected probabilistically at each construction step. The rules for the probabilistic choice of solution components vary across different ACO variants but the common rule used in AS (Dorigo et al., 1996) is:

p(ci | s ) =

c i ( s )

[ i ] .[ (ci )] , [ i ] .[ (ci )]

ci (s )

(2.20)

(ci) is an optional weight value for each feasible solution component ci N(s).
Furthermore, and are positive parameters that determine the pheromone information parameter and the heuristic information parameter.

where i is the pheromone value associated with the solution component ci.

ApplyLocalSearch(). Once solutions have been constructed, and before

updating the pheromones, often some optional actions may be required. These are often called daemon actions, and can be used to implement problem specific and/or centralized actions, which are intended to improve the solutions obtained by the ants.

UpdatePheromones(). The aim of the pheromone update is to avoid a too

rapid convergence in sub-optimal regions and to increase the pheromone values associated with good or promising solutions. Usually, these are achieved by (i) decreasing all the pheromone values through pheromone evaporation, and (ii) increasing the pheromone values associated with a chosen set of solution(s) Supd:

38

i (1 ). i + . ws F ( s ) {sS upd |ci s }

(2.2)

parameter called evaporation rate, F: S R+ is called the quality function such that

where Supd is the set of solutions that is used for the update, (0, 1] is a

f (s) < f (s') F (s) F (s'), s s' S, and w is the weight for corresponding solution s.

To date, a few types of ACO variants exist (e.g. Ant System (AS), Elitist Ant System (EAS), Rank-based Ant System (RAS), Max-Min Ant System (MMAS) and Ant Colony System (ACS)) and there are little differences among them. In AS, pheromone values are updated by all the m ants that have built a solution. EAS updates the pheromone values by using all ants in the respective iteration, as well as the best-so-far ant (Dorigo et al., 1996). Instead of using all the ants, RAS only uses several high quality ants plus the best-so-far ant for updating the pheromone values (Bullnheimer et al. 1999). In MMAS, pheromone update only comes from the iteration best ant or the best-so-far ant, and the value of the pheromone is bounded (Sttzle and Hoos, 2000). ACS differs by introducing a parameter qo as the probability for choosing a solution component which maximizes [i].[(ci)] or using 1-qo as the probability to perform a construction step according to Equation 2.20 (Dorigo and Gambardella, 1997). In addition, ACS uses the best-so-far solution update rule and introduces local pheromone update.

2.5.2 ACO Applications in Facility Layout Problems

As mentioned earlier, ACO has successfully solved many CO problems (Blum, 2005a). ACO is suitable to be used for solving CO problems since it uses a discrete representation. ACO also has another advantage in using heuristic information to exploit the problems so that it can solve them in a shorter computation time. Moreover, ACO often uses local search to gain an intensification advantage just like other metaheuristics.

39 One area that has been solved successfully by ACO is FLP. FLP is a very hard optimization problem and can be applied in many areas. These reasons have encouraged many researchers to solve it. Unfortunately, the applications of ACO in solving FLPs are mainly limited to QAP. Only little research has been conducted to solve other FLP classes. Table 2.3 shows the applications of ACO in solving FLPs.

Table 2.3 Applications of ACO algorithms in FLPs

No 1 2 3 4 5 6 7 8 9 10 11 12 13 Reference Maniezzo et al. (1994) Gambardella et al. (1997) Maniezzo and Colorni (1999) Maniezzo (1999) Sttzle and Hoos (2000) Talbi et al . (2001) Corry and Kozan (2004) Solimanpur et al . (2004) AbdelRazig et al . (2005) Baykasoglu et al . (2005) Solimanpur et al . (2005) Demirel and Toksari (2006) McKendall Jr. and Shang (2006) FLP class QAP QAP QAP QAP QAP QAP MLP QAP QAP QAP QAP QAP QAP QAP QAP QAP QAP ACO variant AS-QAP HAS-QAP AS-QAP ANTS-QAP MMAS Parallel Ant Colonies Ant Colony System Modified ACO Modified ACO Modified ACO Modified ACO Hybrid ACO-SA Hybrid Ant System Ant Colony with stochastic local search Hybrid Ant Colony System ACO with guided local search MMAS

14 Mouhoub and Wang (2006) 15 Ramkumar and Ponnambalam (2006) 16 Hani et al . (2007) 17 Wong and See (2009)

ACO has been implemented to solve QAP since its early development. The first developed algorithm is Ant System (AS), and then several kinds of ACO variants have been reported to solve QAP such as ANTS-QAP, MAX-MIN Ant System (MMAS-QAP), and Hybrid Ant System (HAS-QAP). These algorithm share two similar characteristics: (1) the pheromone trail of ACO is concerned with the desirability to assign facility i to location j, and (2) local search procedures are used to improve the ant solutions.

The differences among the ACO variants above can be summarized as follows (Sttzle and Dorigo, 1999): In AS, the heuristic information of facility i to be located in location j is inversely proportional to the sum of distances from location j

40 to all other locations and the sum of flows from facility i to all other facilities. ANTS-QAP computes heuristic values dynamically by assigning lower bounds on the solution costs of the completed partial solutions. ANTS-QAP does not use pheromone evaporation, but it uses a mechanism to increase or decrease the pheromone values depending on the quality of the solutions. MMAS does not use heuristic information but it relies on local search to gain high quality solutions. In addition, in MMAS, the pheromone values are bounded to specific upper and lower bounds. The pheromone values are assigned to the upper and lower bound values if an operation causes them to go beyond these limits. HAS-QAP does not use pheromone trails to construct solutions, but it tries to modify the current solution by swapping two facilities which are chosen randomly or selected based on probabilities taken from pheromone trail values. In HAS-QAP, only the global best ant is allowed to update the pheromone values.

The computational results among the ACO variants show that ANTS-QAP is the best ACO variant on all the problem instances tested. In addition, computational results show that ACO based algorithm is currently among the best algorithms for solving real-life and structured QAP instances (Sttzle and Dorigo, 1999).

In addition, ACO has been used by Corry and Kozan (2004) for solving the dynamic machine layout problem. The solution representation used consists of three parts: locations, orientations, and the order in which machines are positioned. The locations are represented in grid blocks to provide a framework for depositing pheromone trail information. An artificial ant constructs a solution by choosing a machine and its location. The pheromone trail is recorded as the ant moves while constructing the solution. Then this information is used for updating the pheromone values. The model has shown better result than the heuristically reduced integer programming method.

41 2.5.3 Parameter Tuning for Ant Colony Optimization

In contrast to the development of ACO's structure, parameter tuning in ACO has only received little attentions. Only several papers have been reported to develop and improve this issue (e.g. Dorigo et al., 1996; Dorigo and Di Caro, 1999; Watanabe and Mitsui, 2003). Several researchers have tuned the various parameters in ACO, while others have experimented with tuning the ACO's parameters when it is combined with another metaheuristic such as GA and PSO.

For example, Dorigo et al. (1996) have experimented with parameter tuning in ACO to solve TSP. They gave guidelines on the boundary of the basic parameters in ACO: : relative importance of the pheromone information, 0 : heuristic information parameter, 0 : evaporation rate, 0 < 1

In another study, Dorigo and Di Caro (1999) experimentally concluded that good parameter values for ACO are; m = problem size (for TSP), = 1, = 5, and = 0.5.

As mentioned earlier, ACO has been successfully used to solve the sequential ordering problem, resource constrained project scheduling problem, open shop scheduling problem, and 2D and 3D hydrophobic polar protein folding problem. It is found that for the cases above, the optimum ACO parameters used are different and they are not always the same with Dorigo and Di Caro (1999)'s suggestion. The parameter values used for the cases above can be seen in Table 2.4. Note that parameter = 0 shows that the algorithm does not use heuristic information, and ant number m = 1 because of hybridization with beam search.

42

Table 2.4 ACO parameter values in various optimization problems

No 1 2 3 4 Optimization problem Sequential ordering problem Project scheduling problem Open shop scheduling problem Bioinformatic problem m 10 5 1 10-15 1 1 10 1 0 1 0 2 0.1 0.975 0.1 0.6 Reference Gambardella and Dorigo (2000) Merkle et al. (2002) Blum (2005b) Shmygelska and Hoos (2005)

From the fact that researchers tried to tune their parameter values in ACO and they concluded different parameter values as optimal values, two points can be concluded:

The performance of ACO is influenced by its parameter values. This is because they implicitly influence search intensification and diversification in the algorithm.

There are no universal parameter values for ACO that can be used to solve all optimization problems efficiently and effectively. The differences in 'the optimal parameter values' come from the differences in problem size, optimization problem type, and problem instance characteristic.

The mistakes in setting parameter values may lead the algorithm to bias and ruin its performance. This is because the algorithm cannot diversify enough to reach the optimum region and/or cannot intensify in the correct direction. Instead of using static parameter tuning, several researchers dynamically tune the parameter values in ACO (Watanabe and Mitsui, 2003; Qin et al., 2003; Favuzza et al., 2006). This is done to gain control of the intensification and diversification mechanisms in ACO. Basically, the parameter tuning methods in ACO can be classified into static parameter tuning, adaptive parameter tuning, and self-adaptive parameter tuning.

43 In terms of the first category, Dorigo and Gambardella (1997) have proposed the optimum number of ants in ACS. They showed that the optimum number of ants is influenced by the problem size dimension. Nevertheless, they concluded that ACS works well when m = 10 based on their experimental observation.

Gambardella and Dorigo (2000) proposed that the qo value (see section 2.5.1) is a function of the problem size, i.e. qo = 1 s/n. This makes qo dependent on the problem size n. s is the expected number of nodes selected with the probabilistic rule.

Pilat and White (2002) proposed two hybrid methods incorporating GA into ACS. The first algorithm uses a population of ants that encodes decision variables in ACS. In every iteration, four ants will be selected to execute the ACS and GA operations. Each selected ant will first run the ACS algorithm to generate a solution. Then through crossover and mutation, the two worst ants will be replaced by the new ones. They compared this hybrid algorithm with ACS and did not find significant improvement in finding optimum solutions. The second hybrid algorithm is used to find ACSs optimum parameter values (as alternative to Dorigo and Gambardella (1997)'s proposal). This algorithm is quite similar with the first one except that the population consists of one complete ACS algorithm rather than an ant. They concluded that the parameter values would influence the algorithms performance and no magic parameter values can be used for all optimization problems.

Watanabe and Matsui (2003) proposed a method to dynamically tune the size of the candidate set (ants) in ACS. The candidate set is used to limit the search space only to promising regions. They restricted the candidate set by only considering 90% accumulation of the sorted pheromone concentration. With this method, it is not necessary to set the size of the candidate set in advance. The computational results with several graph coloring problem instances indicated that the proposed control mechanism can potentially improve the efficiency of ACO algorithms, especially for large optimization problems.

Qin et al. (2006) used an adaptive ant colony algorithm to solve phelogenetic tree construction problem. The term adaptive in this algorithm refers to dynamically

44 tuning and which are based on pheromone distributions. At the initial stage of the algorithm, ants should select solution components based mainly on heuristic information . This can be achieved by assigning a relatively large value to . After some iterations, the pheromone values for the solution components are increased, thus their influence becomes more and more important. Therefore the value of should be increased so that the pheromone influence can be gained. Experimental results show that the proposed method has better performance than GA. It can converge faster and obtain better quality results than GA.

Hao et al. (2006) introduced the use of adaptive ACS by incorporating PSO. The role of PSO is to find the correct values for ACO parameters by formulating them as particles. PSO is executed only when the new best solution is found. The proposed solution was tested with TSP problem instances, and compared with pure ACS. The results showed that PSO-ACS performs better than ACS. In addition, Hao with different groups of researchers have also conducted dynamic parameter tuning on the heuristic information parameter (Huang et al., 2006) and pheromone trail evaporation rate (Hao et al., 2007). The results indicated improvement as

compared to the traditional ACO.

Favuzza et al. (2006) used a varying qo based on the number of unimproved iterations. If the algorithm reaches a local convergence indicated by a high number of unimproved iterations, then the value of qo will be decreased, thus allowing the algorithm to heighten the diversification process. As soon as the algorithm leaves the local convergence, the qo value will be increased again allowing the intensification process to happen. The proposed algorithm has proven to be robust in finding the optimal reinforcement strategy for distribution systems.

Randall (2004) proposed a near parameter free ACO. Particularly, the parameter search process is integrated within the running of the algorithm. In other classes of metaheuristics, this method is called self-adaptive parameter tuning (Eiben et al., 1999). The proposed method has comparable results to the standard implementation of ACO. In addition, this method removes the need to tune the parameters by hand.

45 In Table 2.5, the various parameter tuning methods for ACO together with their drawbacks are summarized. The significance of these methods is dependent on their implementation. More research is needed to examine and compare previous studies using the same optimization problem, problem size, and problem instances. As described in Table 2.5, major drawback of these methods is the incapability to control search intensification and diversification. In addition, there is no previous research which has used Fuzzy Logic Controller (FLC) to tune the parameters of ACO.

Table 2.5 Parameter Tuning for ACO algorithms

No Method Problem 1 Proposes a value for ant TSP number 2 Determines q o based on problem size 3 Genetic Algorithm Sequential ordering TSP Drawbacks Only for ant number, does not have the ability to balance exploitation and exploration Only changes the tuning need from q o to s Overheating the algorithm, does not show satisfied results Only useful for limiting diversification and increasing intensification, cannot control diversification Not useful for balancing diversification and intensification Diversification may become bias depending on the distribution of pheromone trail and heuristic information Does not have the ability to balance exploitation and exploration Does not give concern to solutions which are trapped in local optima Reference Dorigo and Gambardella (1997) Gambardella and Dorigo (2000) Pilat and White (2002) Watanabe and Matsui (2003)

4 Candidate Set

Graph Coloring

5 Integrates parameter search within metaheuristic 6 changes q o based on the number of unimproved iterations 7 PSO

TSP and QAP

Randall (2004)

Distribution System

Favuzza et al. (2006)

TSP

Hao et al. (2006)

8 Dynamic tuning based Phelogenetic on pheromone diversity Tree

Qin et al. (2006)

2.6

Conclusions

This chapter has discussed the importance of FLPs along with a review on the applications of ACO in FLPs. It is found that ACO has not been implemented to solve UA-FLPs which have been used for years to model manufacturing layout

46 problems. To alleviate the complexities of UA-FLPs, this chapter has described several layout representations. One of the important layout representations is FBS which will divide a facility into bays, and then departments will be assigned to them. In addition, FBS has an advantage because the bays may become aisle candidates in the final manufacturing layout.

In addition, balancing intensification and diversification in the search space is an important key element in metaheuristics. This can be achieved by dynamically tuning the parameter values. This chapter has reviewed the parameter tuning methods that have been used for ACO. It is found that the previous parameter tuning methods have not been able to effectively balance intensification and diversification.

CHAPTER 3

METHODOLOGY

3.1

Introduction

The methodology used in this research begins with problem definition. After an extensive literature review, several research hypotheses are generated. Then, the necessary data are collected before starting the developmental work. The development of the proposed algorithm covers the algorithm design, codification, and verification. Following this, the algorithm is tested using numerous case problems taken from the literature. Based on the results obtained from the evaluation, analysis will be carried out.

3.2

Methodology

The methodology used to complete this research consists of seven main steps, which are shown in Figure 3.1.

1. Problem definition

In this step, the problem definition is determined along with the research boundaries and research objectives. Firstly, the problem to be solved is identified in this step. In order to ensure the research is completed on schedule, research

48 boundaries are clearly determined. The research objectives are also determined because they act as the guidance for the rest of the steps.

Figure 3.1 Research methodology main steps 2. Literature review

In this step, an extensive literature review is conducted. Previous studies are reviewed especially in the areas of Unequal Area Facility Layout Problems (UAFLPs), Ant Colony Optimization (ACO), and parameter tuning in ACO. From the literature review, the current trend related to this research can be found. It also shows the research opportunity which should be tackled in this research.

49 3. Research hypothesis

Hypotheses are then determined based on the research objectives and literature review. These hypotheses represent preliminary conclusions that must be supported or rejected when completing the research.

4. Data collection

The next step is to determine and collect what are needed to conduct the experiment and test the hypotheses. The important data needed are test problems which have been published in the literature. Test problems act as inputs for algorithm design and testing. On the other hand, previous results found in the literature are used to evaluate the performance of the proposed algorithm.

5. Development of the ACO algorithm for UA-FLPs

The development of the algorithm is divided into two phases. Phase I is intended to design the basic ACO formulation for solving UA-FLPs and phase II is intended to construct the Fuzzy Logic Controller (FLC) which will dynamically tune the parameter values. The algorithm development uses the C++ programming language since it is fast and yet easy to be implemented.

6. Testing and evaluation of the algorithm

The final algorithm is tested using test problems taken from the literature. The test problems are chosen from those which are well-known and widely used in the literature. It is intended that the selected test problems cover the whole range of problem size (from small until large). In addition, they are tested using conditions which are similar to those from previous research. Then the results are evaluated by comparing them with previous research in terms of the best known solution and best Flexible Bay Structure (FBS) solution.

50 7. Result analysis and conclusion

Analysis is conducted to evaluate the performance of the algorithm. This includes analyzing its advantages and disadvantages. This step also highlights some important notes for developing the algorithm and gives some future directions for continuing the research. Finally, the hypotheses will be answered and conclusions will be drawn.

3.2.1 Ant System Development

This research uses the Ant System (AS) framework for implementing ACO. The detailed algorithm development methodology is shown in Figure 3.2 and the steps involved are discussed below:

1. Algorithm formulation

The algorithm is formulated using the Ant System (AS) framework because it is the most fundamental concept in ACO. The structure and implementation of AS is designed based on paradigms obtained from the literature. Although AS has been implemented to solve other classes of FLPs such as Quadratic Assignment Problems (QAPs), it still needs modification in solving UA-FLPs. Besides this, there are several optional components that can be considered in the implementation of the algorithm such as local search, global update rule, etc. In addition, to alleviate the complexity of UA-FLPs, Flexible Bay Structure (FBS) is used to model them.

2. Algorithm codification

Then, the AS algorithm is codified using the C++ language. C++ is chosen because it is proven superior in computation time and yet easy to be programmed. Due to maintenance reasons, the codification uses the structural programming

51 paradigm. In structural programming, one big program is divided into small programs in order to decrease redundancy and increase maintenance easiness.

Start Formulation of AS Algorithm Codification of Algorithm Verification of Algorithm Testing of Algorithm

Test problems from literature Manual parameter tuning

Formulation of FLC Integration of FLC with AS Codification of FLCAS Verification of Final Algorithm Test problems from literature Testing of Final Algorithm End

Figure 3.2 FLC-AS development methodology 3. Algorithm verification

In conjunction with the algorithm codification, the program will also be inspected for bugs and mistakes. Each part of the program will be tested whether there is any problem such as division by zero, variable not found, etc. After all parts

52 of the program can run properly, small test problems will be input to the program. The first iteration will be examined to check if there is any problem or bug. Several random testings will also be conducted to ensure that the program runs without any problem.

4. Algorithm testing and manual parameter tuning

After the program can be properly used, it will be tested using test problems taken from the literature. In the initial stage, the best parameter setting will be determined using manual parameter tuning on one test problem. Manual parameter tuning will be conducted on five different levels for several parameters. The parameter values which can achieve the best objective function will be used as the optimal parameter settings. These settings will be used for conducting the experiment for the rest of the test problems.

3.2.2 Fuzzy Logic Controller (FLC) Development

This section is intended to develop the FLC and integrate it with AS. To complete this part, this research uses the following steps:

1. FLC Formulation

The objective of this step is to design the FLC which consists of three main components; input parameters along with their fuzzy membership functions, output parameters along with their fuzzy membership functions, and fuzzy inference engine. Input parameters are associated with the convergence and/or divergence level of the algorithm. Meanwhile, output parameters are algorithm parameters that will control the searching process of the algorithm. Fuzzy inference engine will become a bridge, which converts the convergence/divergence level into a specific action that must be taken to adjust the convergence/divergence level.

53 2. Integration of FLC with AS

The function of this step is to integrate the FLC with the pure AS. This step includes determining where and when to put the FLC into the algorithm. It also decides the frequency of executing the FLC.

3. FLC-AS codification

In this step, the complete FLC-AS concept is transformed into the C++ language. In the same way as AS programming, the codification uses the structural programming paradigm. FLC becomes a new module and it will be inserted into the main AS algorithm.

4. FLC-AS verification

This step ensures that the FLC is functioning. It also examines the connection between FLC and AS. The FLC module will be tested whether there is any problem such as division by zero, variable not found, etc. After all parts of the program can run properly, small test problems will be input to the program. The first iteration will be examined to check if there is any problem or bug. Several random testings will also be conducted to ensure that the program runs without problems.

5. FLC-AS testing

After the final program can be properly used, it will be tested using test problems found in the literature. The initial parameter values for running the algorithm will be taken from the manual parameter tuning. For each test problem, five replications will be conducted.

54 3.3 Research Hypothesis

Solving UA-FLPs using AS is a complex research. The need for obtaining high quality solutions and balancing search intensification and diversification is a challenge. To address these problems, the following hypotheses are proposed:

High quality UA-FLP solutions can be achieved by implementing an initial AS algorithm with FBS representation.

FLC can be used to replace manual parameter tuning and to automatically tune the parameter values of AS.

3.4

Data Collection

The initial AS and FLC-AS algorithms are tested using numerous problem sets taken from the literature. The list of the problem sets can be found in Table 3.1. Note that max is the maximum aspect ratio constraint and lmin is the minimum side length constraint. The complete data for the problem sets can be found in Appendix A.

In order to allow the use of FBS representation on problem sets M11s, M11a, M15s, M15a and M25, the author have modified them so that they do not have a fixed size department. In addition, the fixed location constraint is changed into a sequence constraint. The modified fixed size department then becomes the last department that will be assigned in the FBS representation. As for BA12TS and BA14TS, they are the modified problem sets initially provided by Tate and Smith (1995) to change the empty space to dummy department(s).

55 Table 3.1 UA-FLP problem sets used in this research

Problem Number of Facility size Common shape Reference set departments Width Height constraint Meller et al. (1998) max = 4 1 O7 7 8.54 13.00 Meller et al. (1998) max = 5 2 FO7 7 8.54 13.00 Meller et al. (1998) max = 5 3 FO8 8 11.31 13.00 4 O9 5 V10s 6 V10a 7 M11s 8 M11a 9 M15s 10 M15a 11 M25 12 NUG12 13 NUG15 14 BA12 15 BA12TS 16 BA14 17 BA14TS 18 AB20 19 SC30 20 SC35 9 10 10 11 11 15 15 25 12 15 12 16 14 14 20 30 35 12.00 25.00 25.00 6.00 3.00 15.00 15.00 15.00 3.00 3.00 7.00 7.00 6.00 6.00 2.00 12.00 15.00 13.00 51.00 51.00 6.00 2.00 15.00 15.00 15.00 4.00 5.00 9.00 9.00 10.00 10.00 3.00 15.00 16.00 max = 5 lmin = 5 max = 5 lmin = 1 max = 5 lmin = 1 max = 5 max = 5 max = 5 max = 5 lmin = 1 lmin = 1 lmin = 1 lmin = 1 max = 1.75 max = 5 max = 4 Meller et al. (1998) van Camp (1989) van Camp (1989) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) van Camp (1989) van Camp (1989) van Camp (1989) Tate and Smith (1995) van Camp (1989) Tate and Smith (1995) Armour and Buffa (1963) Liu and Meller (2007) Liu and Meller (2007)

No

The test problems are chosen from those which are well-known and widely used. Table 3.2 can be used to compare the characteristics of this research's problem sets with those used in previous research. As can be seen, different researchers have applied different type and number of problem sets in their evaluation. Moreover, not all problem instances can be obtained because their data are not published. Nevertheless, by comparing Table 3.1 and Table 3.2, it can be said that the problem sets used in this study are adequate and comparable to previous research (20 problem instances are used as compared to a minimum of 1 and a maximum of 27). Most of the previous studies have only tested less than 20 problem sets. The smallest problem instance used in this study is just one department larger than the smallest problem instance found in the literature. In addition, the proposed algorithm will be tested using several large problem instances which have appeared in the literature.

56 Table 3.2 Characteristics of problem sets used in previous research

Number of problem sets tested 1 10 3 4 15 21 27 7 Minimum problem size 20 5 11 10 6 7 6 7 Maximum problem size 20 30 25 20 9 20 35 20

No 1 2 3 4 5 6 7 8

Literature Armour & Buffa (1963) van Camp (1989) Bozer et al. (1994) Tate and Smith (1995) Sherali et al. (2003) Konak et al. (2006) Liu and Meller (2007) Scholz et al . (2009)

3.5

Conclusions

This chapter has discussed about the methodology used in this research. The methodology consists of seven main steps, (i) problem definition, (ii) literature review, (iii) research hypothesis, (iv) data collection, (v) algorithm development, (vi) algorithm testing and evaluation, and (vii) analysis and conclusion. The development of the algorithm is achieved by two steps, (i) AS development, and (ii) FLC-AS development and integration. This chapter has also provided the research hypotheses which will be tested in this research. Finally, it has also provided the problem sets which will be used to evaluate the proposed algorithm.

CHAPTER 4

ANT SYSTEM FORMULATION

4.1

Introduction

This chapter presents the Ant Colony Optimization (ACO) formulation for solving Unequal Area Facility Layout Problems (UA-FLPs). As mentioned earlier, Ant System (AS) is used as the basic ACO formulation in this study. The structure of the AS algorithm comprises three main parts, i.e. construct ant solutions, apply local search, and update pheromones. Specifically, ant solutions are constructed based on both pheromone information and heuristic information. The proposed algorithm uses five types of local search to improve the search. In addition, the algorithm uses the iteration-best ant and global-best ant to update the pheromone values.

4.2

Structure of the Proposed Ant System (AS) Algorithm

The proposed AS algorithm follows the ACO general form, as presented in Table 2.2. In this implementation, the main looping of the proposed AS algorithm consists of (i) construct ant solutions, (ii) apply local search procedures and (iii) update pheromone values. The details are summarized in Figure 4.1.

The proposed AS algorithm uses the Flexible Bay Structure (FBS) representation to transform the continuous characteristic of UA-FLPs into a discretebased optimization problem. The advantages of the FBS representation are: (i) it

58 decreases the problem complexity and (ii) it generates aisle candidates in the final layout. The layout infeasibility issue is handled by adopting the adaptive penalty objective function used by Tate and Smith (1995).

Figure 4.1 The proposed AS algorithm The proposed algorithm constructs ant solutions based on pheromone information and heuristic information. The heuristic information is obtained from processing the material flow data and relative department locations. Meanwhile the pheromone information is collected during each iteration from the global-best ant and iteration-best ant.

59 4.2.1 Solution representation of Unequal Area Facility Layout Problems

The proposed AS algorithm uses FBS to represent an UA-FLP. Conceptually, FBS divides the facility into vertical or horizontal bays. The bay width is flexible, depending on the departments that it contained. In addition, the number of bays and the number of departments in each bay are also flexible. The proposed algorithm tries to find the optimum value for the number of bays, the number of departments contained in each bay, and the department-placement order that could minimize the objective function value.

Each ant solution consists of two parts. The first part represents the department sequence. This sequence is the order of n departments represented by integer numbers which will be placed into the facility. The second part represents break points which manage the number of bays and the number of departments contained in each bay. This second part is represented by n-1 binary numbers. 1 represents a bay break and 0 otherwise. The bay width is adjusted by dividing the sum of department areas contained by the facility length.

Figure 4.2 shows an example of an ant solution which is used in the proposed algorithm. This ant solution is for an UA-FLP with n = 7. The first part contains the order of seven departments (5-3-2-1-6-4-7) which will be placed into the facility. The second part contains six binary numbers (0-0-1-0-1-0) which represent break bay locations.

Figure 4.2 An Example of Solution Representation in the Proposed Algorithm This representation will be transformed into a layout solution in Figure 4.3. Specifically, the departments are placed from left to right, and bottom to top based on the sequence 5-3-2-1-6-4-7. For the bay breaks, number 1 appears two times, in the

60 3rd and 5th places. These show that the breaks are generated after the third and fifth departments. So there are 3 bays, the first bay contains departments {5, 3, 2}, the second bay contains departments {1, 6} and the third bay contains departments {4, 7}.

With FBS representation, the proposed algorithm only searches solutions that comprise a combination of department sequence and bay break sequence. Hence, the complexity of UA-FLPs using FBS representation is n!2n-1. This representation also has the advantage in that the bays will become aisle candidates and this facilitates users to transform the model into an actual layout.

Figure 4.3 Layout generated from the ant solution representation in Figure 4.2 In order to deal with problem sets with empty spaces, the author proposed two improvements for FBS representation: (1) extending the feasible solution space by using empty space to fulfill department constraints and (2) improving the objective function by recursively filling the bay with empty space. The author refers to this representation as modified-FBS (mFBS). An example of mFBS representation is shown in Figure 4.4.

The transformation from FBS representation into mFBS solution is based on two procedures:

61 1. Determine the bay height based on the total area of departments which are located in a bay. Whenever departments that violate department constraints are found, the bay height will be increased by adding an available empty space into the bay to satisfy the constraints. Whenever an empty space is added, it will be proportionally placed in the left and right side of the bay.

(a)

(b)

Figure 4.4 An example of layout solution with (a) original FBS and (b) modifiedFBS 2. After all department constraints in each bay are satisfied and if there is still an unoccupied empty space, select the middle bay and add the available empty space until the maximum bay height is fulfilled. Examine if this improves the objective function value. If there is an improvement, the modification will be accepted. Otherwise, the original layout will be restored. Whenever an empty space is added, it will be proportionally placed in the left and right side of the bay. This procedure will be repeated for all bays by selecting the nearest bay from the center point of the facility. In addition, this procedure will be executed only if the empty space is still available.

62 4.2.2 Objective function

The goal of the proposed algorithm is to minimize the total material handling cost. The cost (objective function) for an ant is calculated based on Equation 4.1.

Ant Objective Function =

n

i =1 j =1,i j

c
n

ij

f ij d ij + p inf (V feas Vall )

(4.1)

where:

the number of departments

cij

the cost of moving one unit material per unit distance from department i to department j, where i, j = 1, 2, , n

fij

the number of material flows from department i to department j, where i, j = 1, 2, , n

dij

the distance between departments i and j, where i, j = 1, 2, , n

pinf

the number of infeasible departments

Vfeas

the best feasible objective function value found

Vall

the best overall objective function value found

The distance is measured using rectilinear distance from centroid-to-centroid of each department. In addition, the objective function incorporates an adaptive penalty function (Tate and Smith, 1995) to guide the search process towards feasible solution regions. Specifically, an ant solution is given a penalty value which is proportional to the number of infeasible departments that it contains. A department is

63 considered as infeasible if it violates the department constraints, i.e. either the maximum aspect ratio or minimum side length.

4.2.3 Ant Solutions Construction

The proposed algorithm employs the pheromone information and heuristic information for constructing ant solutions. In this respect, the two parts of an ant solution are not generated concurrently. The bay breaks are generated first and their constructions are solely based on pheromone information. The generated bay breaks are transformed into a dummy layout which will be used to calculate the heuristic information for that particular ant solution. Following this, the department sequences are constructed based on pheromone information and heuristic information.

During execution, the proposed algorithm keeps the pheromone information for the two parts of the ant solution (department sequences and bay breaks). For each element of the department sequences, pheromone information is the relative probability to choose department 1 until department n. While for each element of the bay breaks, pheromone information corresponds to the desirability to select a bay break or not.

When adding a solution component, the proposed algorithm only chooses the solution component which has not appeared in that particular ant solution to avoid duplication. The proposed algorithm cannot guarantee that the generated ant solution is feasible because the feasibility can only be known after a complete layout solution is formed. Instead, it uses an adaptive penalty function to guide the search process to the feasible regions as mentioned in section 4.2.2.

After the bay breaks have been generated, the proposed algorithm creates a dummy layout solution assuming that all departments have the same area (obtained by dividing the facility area with the number of departments). The dummy layout specifies the location candidates which represent the relative locations of

64 departments in the final layout solution. The location candidates will be used as a basis for determining the heuristic information for department sequences construction.

The proposed algorithm uses the heuristic information to help achieves a higher quality solution in a shorter time. The heuristic information helps to guide the proposed algorithm to place a department with high material flow in a location candidate nearer to the center of the facility. In order to provide such heuristic information, the proposed algorithm introduces as a parameter for location candidate. Assume that point (0, 0) is located at the bottom left of the facility. The parameter for each location candidate is calculated based on Equation 4.2.

L L W W = |x |+ | y | k 2 k 2 k 2 2

(4.2)

where:

the parameter for location candidate k, where k = 1,2,, n

the facility length

the facility width

xk

x axis value for the center of location candidate k

yk

y axis value for the center of location candidate k

Equation 4.2 ensures that a location candidate which is nearer to the center of the facility will have a higher value. Apparently, the equation is a linear function since the proposed algorithm uses the rectilinear distance as the measuring parameter.

65 By using the pheromone information and heuristic information, the proposed algorithm generates the department sequences by randomly choosing them based on probabilities calculated using Equation 4.3. The ik acts as the pheromone information collected in previous search to guide the proposed algorithm to good solution regions that have been explored. The product of m and represents the heuristic information which heuristically guides the proposed algorithm to place a department with high material flow in a location candidate nearer to the center of the facility.

p(C ik | s ) =

cik ( s )

[ ik ] .[mi .k ] , [ ik ] .[mi .k ]

cik ( s )

(4.3)

where

p(Cik) =

probability to locate department i to location candidate k

Cik

department i located in location candidate k (department i located in the kth element of the department sequences)

ik

pheromone value associated with the solution to locate department i to location candidate k

mi

sum of material flow from and to department i

parameter for location candidate k

pheromone information parameter

heuristic information parameter

66 N(s) = available departments which have not been used in the corresponding ant

4.2.4 Local Search Procedures

In order to enhance the search performance of AS, the author incorporates five types of local search as improvement procedures into the algorithm. They are:

1. Swap between department sequence, which randomly chooses two different departments and exchanges their locations.

2. 1-insert procedure on department sequence, which randomly chooses one department and moves it to a new location in the department sequence.

3. 2-opt procedure on department sequence. This local search randomly chooses a subset chain of the department sequence and rearranges it in the opposite direction.

4. Bay break change procedure, which randomly changes the bay break (from 0 to 1 or vice versa).

5. Bay break swap procedure, which randomly swaps the bay break with its neighbor.

The local search procedures can be classified into two categories. Both are used to provide a effective search for the proposed algorithm. The first category (procedures 1-3) is a neighborhood search to find a good department sequence and the second category (procedures 4-5) is a neighborhood search to find a good bay layout. All the procedures are used because of the large solution space of certain problem instances. All of them have the same probability to occur since the solution

67 space of one problem instance may differ from those of other problem instances and thus, it is hard to predict the most effective local search procedure.

During implementation, the algorithm continuously probes the neighborhood of the solution search space (using local search procedures) to find a solution with a better quality. Whenever a better solution (refers to a solution that either (1) improves the solution feasibility (indicated by a decrease in the number of infeasible departments), or (2) improves the objective function value without degrading the solution feasibility) is found, the neighborhood that contains such a solution will be explored in the next local search step.

Although the proposed algorithm incorporates five types of local search, it randomly selects one of them in an implementation, and recursively repeats it until a stopping criterion is met. As in this case, the stopping criteria for the local search are specified as: (1) the maximum number of steps pre-specified by users, and (2) the number of steps where the local search does not improve the solution quality. Whenever one of these criteria is met, the local search will be terminated.

4.2.5 Pheromone update scheme

The proposed algorithm uses the standard pheromone updating strategy to avoid premature search convergence and to bias the search process into a solution search space that contains good or promising solutions. It uses the global-best and iteration-best ant solutions to update the pheromone value. The pheromone updating process is shown in Equation 4.4.

ik (1 ). ik + .

{sS upd |cik s }

s

w F ( s)

(4.4)

where:

68 Supd = the set of ants that is used for updating the pheromone evaporation rate, (0, 1) quality function such that f (s) < f (s') F (s) F (s'), s s' S, with f(s) is the objective function. The proposed algorithm uses F(s) = 1/f(s)

F(s)

weight for the corresponding ant solution s. The proposed algorithm uses w = 1 for all Supd.

4.2.6 Stopping criteria

The proposed algorithm uses two stopping criteria to terminate its search: (1) the maximum number of iterations, which sets an upper limit for the number of steps involved in a single implementation and, (2) the maximum number of iterations where AS could not further improve the best-so-far solution. The latter helps the algorithm to terminate the search if a better solution could not be found. Whichever stopping criterion is met first, the proposed algorithm will be stopped.

4.3

Implementation details of the algorithm

The detailed implementation of the proposed algorithm is shown in Figure 4.5. In short, the algorithm is modified from the general AS framework to accommodate the FBS representation, adaptive penalty function, ant solution construction, local search procedures and pheromone update scheme for UA-FLPs.

69
Start i = iteration number un = unimproved iteration number cp = computation time

i =0, un=0, cp=0

Set algorithm parameter

m = ant number = pheromone trail parameter = heuristic information parameter = evaporation rate L = facility length in x-direction W = facility width in y-direction n = department number ai = area for department i ubi = upper bound for the length and width of department i lbi = lower bound for the length and width of department i fij = material flow from department i to department j

imax , unmax, LSmax, unLSmax

Set algorithm and local search stopping criteria

m,

Set parameter values

Data : L, W, n, ai, ubi, lbi, fij

Input problem data

sfi = sum of flow from department i to all other departments.

Calculate partial heuristic information table sfi

ik

=pheromone value for placing department i to location k

Initialize pheromone value table ik

f(s)* = infinity S* = {}

Set global best objective function and solution

i=i+1

Update iteration counter

d=0

Initialize ant number counter d

f(i)* = infinity Si * = {}

Set iteration best objective function and solution

Figure 4.5 Details of the proposed AS algorithm

70

Figure 4.5 Details of the proposed AS algorithm (continued)

71
B

Check the best ant in iteration i, f(d-th) < f(i)* ? yes f(i)* = f(d-th) Si* = S(d-th)

no C

Ant number counter d = ant number m ?

no

yes Check the global best ant, f(i)* < f(s)* ? yes un = 0 f(s)* = f(i)* S* = Si*

no

un = un + 1

Select ant Supd, for pheromone update

Update pheromone value

i = imax ? OR un = unmax ?

no

yes End

Figure 4.5 Details of the proposed AS algorithm (continued)

72

4.4

Conclusions

This chapter has presented an AS algorithm with FBS representation for solving UA-FLPs. The main algorithm

consists of three parts, i.e. construct ant solutions, apply local search procedures, and update pheromones. In addition, this chapter has also proposed an improvement to the FBS representation to solve UA-FLPs with empty spaces.

CHAPTER 5

PARAMETER TUNING USING FUZZY LOGIC

5.1

Introduction

This chapter discusses the parameter setting of the proposed algorithm. This chapter mentions the manual parameter tuning conducted on the proposed algorithm. In addition, it also explains the formulation of the Fuzzy Logic Controller (FLC) to automatically tune the parameters of the proposed algorithm. FLC is chosen because it can handle imprecise propositions which represent the state of the running proposed algorithm. This facilitates the proposed algorithm to balance its intensification and diversification efforts. By balancing these efforts, it ensures that the proposed algorithm does not waste too much computation effort in an unattractive region and has the ability to escape from local optima. In addition, FLC can be used as an alternative for manual parameter tuning which is time consuming.

5.2

Manual Parameter Tuning

In the case of manual parameter tuning, the parameter values are determined based on a randomly selected test problem AB20. This tuning is conducted on five different levels for the four main parameters of the proposed algorithm. Initially, these parameters are assigned with the following values recommended by Dorigo et al. (1999):

74

Number of ants m = problem size, Pheromone information parameter = 1, Heuristic information parameter = 5, Evaporation rate = 0.5.

Then each parameter value is varied on five different levels. This is done by changing one parameter at a time while the others are held constant. Since the tuning involves four parameters, 20 experiments are needed. The parameter values which achieve the best objective function (minimum objective function) will be used as the optimal parameter settings.

The results of the manual parameter tuning are shown in Table 5.1. From the table, it can be seen that, the best parameter values are: number of ants = n (n = number of departments), pheromone information parameter = 1, heuristic information parameter = 5, and evaporation rate = 0.1. These values will be used for testing the rest of the problem sets.

Having conducted the manual parameter tuning, the next section will discuss about the use of FLC to dynamically tune the parameter values.

75 Table 5.1 Manual parameter tuning results

No Number of Pheromone Heuristic information Evaporation ants information parameter parameter rate 1 5 1 5 0.5 2 10 1 5 0.5 3 20 1 5 0.5 4 30 1 5 0.5 5 35 1 5 0.5 6 20 1 5 0.5 7 20 3 5 0.5 8 20 5 5 0.5 9 20 10 5 0.5 10 20 15 5 0.5 11 20 1 1 0.5 12 20 1 3 0.5 13 20 1 5 0.5 14 20 1 10 0.5 15 20 1 15 0.5 16 20 1 5 0.1 17 20 1 5 0.3 18 20 1 5 0.5 19 20 1 5 0.7 20 20 1 5 0.9 Minimum AS best 5699.20 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 5677.83 AS worst 6065.68 6065.68 5699.20 6065.68 5804.28 5699.20 6065.68 6065.68 6065.68 5699.20 5699.20 6065.68 6065.68 6065.68 5699.20 5699.20 5699.20 6065.68 5699.20 6065.68 5699.20 AS average 5866.81 5784.96 5686.38 5832.97 5703.12 5686.38 5755.40 5910.54 5755.40 5686.38 5686.38 5759.67 5837.24 5759.67 5686.38 5682.10 5690.65 5837.24 5686.38 5755.40 5682.10

5.3

Fuzzy Logic Controller

Zadeh (1988) introduced the fuzzy membership function to handle imprecise propositions. Unlike the traditional logic which only has true (1) or false (0) value, the fuzzy membership function may have values between the interval [0, 1]. For example, temperature would be a linguistic variable in the control of a heater. It can take on linguistic values such as high, low, quite low, etc. A temperature value of 40oC may have a membership value of 0.8 for hot because it is above the normal room temperature, and it could get higher. These imprecise propositions can be used to represent the state of a running algorithm (whether intensifying or diversifying). The fuzzy logic for tuning metaheuristic parameters can be regarded as a control system. Thus, the implementation of this Fuzzy Logic Controller (FLC) can be depicted in Figure 5.1. To implement it, several things must be decided; input parameters along with their membership function, output parameters along with their membership function, and the fuzzy rule base which specifies the relationships between the input and output parameters.

76

Figure 5.1 Scheme for parameter tuning in metaheuristics There are several papers which have used FLC to tune parameter values in metaheuristics. Most of them are implemented to tune evolutionary algorithms. For example, Wang et al. (1996) proposed a fuzzy logic controlled genetic algorithm (FCGA) to adaptively adjust the crossover rate and mutation rate during the optimization process. This algorithm was tested using a power economic dispatch problem. The result showed that FCGA has much better performance than the conventional genetic algorithm. Xue et al. (2005) proposed the use of FLC to dynamically tune parameters in multi-objective differential evolution (MODE). They used two inputs for FLC, i.e. population diversity (PD) and generation percentage (GP) already performed. For the outputs, they used greediness and perturbation factor of the reproduction operator. These two parameters control explicitly the exploitation and exploration of the evolutionary algorithm. They showed that FLCMODE obtained better results in 80% of the testing examples compared to the conventional MODE.

The significance of parameter values in metaheuristics has been laid down by Eiben et al. (1999). Although their discussion was aimed at the evolutionary algorithm, their work was also applicable to AS since both metaheuristics are stochastic based algorithms. They stated that parameter dependencies, time consuming efforts, and unnecessarily optimal parameter values are the technical drawbacks of manual parameter tuning. In brief, they suggested the use of dynamic parameter setting.

77 Based on these reasons, this research studies the effect of FLC on dynamically tuning the parameters of AS. This study involves four basic parameters of AS; number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate.

5.4

FLC-AS Formulation

Basically, FLC is added to automatically tune the proposed AS algorithms parameter values. FLC updates the parameter values in every iteration of the algorithm. It takes the convergence and divergence states of the running algorithm as inputs, and then feedbacks the parameter values changes as outputs into the main AS loop. The proposed FLC-AS algorithm can be seen in Figure 5.2.

The role of FLC is to balance the search intensification and diversification of the algorithm. The input parameters for the FLC are the standard deviation of solutions and the number of unimproved iterations. The former can represent how solutions in one iteration differ from those in another iteration. Meanwhile the latter represents the degree of which the current search is trapped in local optima. In other words, these two input parameters help to indicate the state of intensification and diversification of the algorithm. Before they are transformed into fuzzy sets, they are normalized by dividing them with their respective maximum value. The membership function for the input parameters is shown in Figure 5.3(a).

78

Figure 5.2 The proposed FLC-AS algorithm The proposed algorithm utilizes four output parameters which are aimed to influence its search intensification and diversification. These parameters are the number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate. In this research, there are four experiments which tune the parameters individually and one experiment which tunes all of them at once. The membership function for the output parameters is shown in Figure 5.3(b).

The input parameters will be processed to produce the appropriate action by applying the fuzzy rule base. These rule bases usually takes the form of IF-THEN

79 statements. The rule bases for the number of ants and evaporation rate is shown in Table 5.2(a), whereas the one for the pheromone information parameter and heuristic information parameter is shown in Table 5.2(b). The abbreviations for the input parameter values are VL = very low, L = low, M = medium, H = high, and VH = very high, and those for the output parameter values are NL = negative large, N = negative, MV = maintain value, P = positive, and PL = positive large.

Membership value

Very Low Low Medium High Very High

Membership value

Negative Large Negative Maintain V alue Positive Positive large

0 0 0.25 0.5 0.75 1

-1 -0.5

0 0 0.5 1

Norm alized input param eter value

Norm alized output param eter value

(a)

(b)

Figure 5.3 Graphs showing membership functions for (a) input parameters, and (b) output parameters Table 5.2 Rule bases for (a) number of ants and evaporation rate, and (b) pheromone information parameter and heuristic information parameter
Standard deviation of solutions Output VL Number of unimproved iterations VL L M H VH PL PL P P MV L PL P P MV N M P P MV N N H P MV N N NL VH
Number of unimproved iterations Output VL VL L M H VH NL NL N N MV L NL N N MV P M N N MV P P H N MV P P PL VH MV MV P P PL Standard deviation of solutions

MV N N NL NL

(a)

(b)

80 The rule bases are designed to control the search intensification and diversification of the proposed algorithm. It is made to provide the algorithm with an ability to jump out from local optima. In addition, the rule bases are intended to limit the search when it is performed in a too wide region. The rule bases above can be interpreted as follows:

When the algorithm is trapped in local optima (standard deviation of solutions is low), the FLC will try to widen the algorithm's search by increasing the number of ants.

When the algorithm searches in a very wide region (the number of unimproved iterations is low), then the FLC will help to focus the search by increasing the evaporation rate.

If the current search does not help the algorithm to escape from local optima, then the FLC will try to change the direction of the search. The direction is altered by modifying the pheromone information parameter and/or heuristic information parameter.

After applying the fuzzy rule base, the fuzzy outputs will be defuzzified to be transformed into crisp values. This research uses the centroid method (Mendel, 1995) as the defuzzification technique. The centroid method can be seen in Equation 5.1 in which xo is the crisp value, and x is the centroid of the area below the fuzzy membership function c. The changes in the output parameters are limited to 10% of the gap between their respective maximum and minimum value. The FLC is only executed after 10 iterations to allow the algorithm to sample the distribution of the input parameter values.

xo =

( x).x.dx ( x).dx
C C

(5.1)

81

5.5

Conclusions

This chapter has described the parameter setting for the proposed AS algorithm. Initially, this chapter discusses the implementation of manual parameter tuning. Then it describes the implementation of the Fuzzy Logic Controller (FLC) to automatically tune the proposed algorithm. The input parameters for the FLC are the standard deviation of solutions and the number of unimproved iterations. In addition, the proposed algorithm utilizes four output parameters which are aimed to influence its search intensification and diversification. These parameters are the number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate.

CHAPTER 6

EVALUATION OF THE PROPOSED ALGORITHM

6.1

Introduction

This chapter presents the evaluation of the proposed algorithm. At first, the proposed algorithm is tested using the parameter values obtained from manual parameter tuning (refer to section 5.2). After that, the proposed algorithm is tested together with the Fuzzy Logic Controller (FLC) as described in section 5.4. Four basic parameters of the AS algorithm are tuned with FLC, i.e. number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate. Four experiments are conducted to tune each of the parameters individually and one experiment is conducted to tune all of them at once. Finally, this chapter provides a discussion about the results obtained.

6.2

Problem Sets and Parameter settings used to evaluate the Proposed Algorithm

The proposed algorithm is tested using numerous problem sets taken from the literature. The list of problem sets can be found in Table 6.1. Note that max is the maximum aspect ratio constraint, lmin is the minimum side length constraint and both best-FBS solution and best-known solutions are in cost unit.

83

In order to allow the use of FBS representation on problem sets M11s, M11a, M15s, M15a and M25, the author has modified them so that they do not have a fixed size department. In addition, the fixed location constraint is changed into a sequence constraint. The modified fixed size department then becomes the last department that will be assigned in the FBS representation. As for BA12TS and BA14TS, they are the modified problem sets initially provided by Tate and Smith (1995) to change the empty space to dummy department(s).

Table 6.1 Problem set data

Problem Depart- Common shape set ments constraint max = 4 1 O7 7 max = 5 2 FO7 7 3 FO8 4 O9 5 V10s 6 V10a 7 M11s 8 M11a 9 M15s 10 M15a 11 M25 12 NUG12 13 NUG15 14 BA12 15 BA12TS 16 BA14 17 BA14TS 18 AB20 19 SC30 20 SC35 8 9 10 10 11 11 15 15 25 12 15 12 16 14 14 20 30 35 max = 5 max = 5 lmin = 5 max = 5 lmin = 1 max = 5 lmin = 1 max = 5 max = 5 max = 5 max = 5 lmin = 1 lmin = 1 lmin = 1 lmin = 1 max = 1.75 max = 5 max = 4 Best FBS solution* Best known Problem data solution* 131.63 Meller et al. (1998) 23.12 20.95 Meller et al. (1998) 22.39 22.39 Meller et al. (1998) 241.06 235.95 Meller et al. (1998) 19,994.10 van Camp (1989) 21,463.07 van Camp (1989) 1,317.79 Bozer et al. (1994) 1,185.20 Bozer et al. (1994) 27,781.95 Bozer et al. (1994) 29,157.60 Bozer et al. (1994) 1,588.37 Bozer et al. (1994) 265.60 van Camp (1989) 526.75 van Camp (1989) 8,180.00 van Camp (1989) 8,600.33 Tate and Smith (1995) 4,712.33 van Camp (1989) 4,927.69 Tate and Smith (1995) 3,707.00 Liu and Meller (2007) 3,604.00 Liu and Meller (2007) Reference Best FBS solution Best known solution Sherali et al. (2003) Konak et al. (2006) Sherali et al. (2003) Konak et al. (2006) Sherali et al. (2003) Konak et al. (2006) Sherali et al. (2003) Konak et al. (2006) Scholz et al . (2009) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Gau and Meller (1999) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Gau and Meller (1999) Gau and Meller (1999) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Scholz et al . (2009) Konak et al. (2006) Konak et al. (2006) Konak et al. (2006) Scholz et al. (2009) Konak et al. (2006) Konak et al. (2006) Liu and Meller (2007) Liu and Meller (2007)

No

22,899.65 21,463.07 1,317.79 1,225.00 27,781.95 31,779.09 265.60 526.75 8,801.33 8,600.33 5,004.55 4,927.69 6,890.82 -

6,890.82 Armour and Buffa (1963) Konak et al. (2006) Konak et al. (2006)

* Values in the 5th and 6th columns represent cost It can be said that the problem sets used can represent the UA-FLP domain. The problem size varies from 7 to 35 departments. To date, the largest problem set known is SC35 (with 35 departments) which was originated from a recent research done by Liu and Meller (2007). Twenty problem sets are used in this research whereas the largest number of problem sets which have been used in previous research is 27 (as shown in Table 3.2). This study can only test less than 27 problem sets because some of them cannot be obtained (their data are not published).

Most of the problem sets do not have a too restrictive department constraint. Most of them have a maximum aspect ratio of 4 or 5, or a minimum side length of 1.

84 These constraints bound the department solution space so that the departments will not be too long and narrow. One problem set which has a restrictive constraint is AB20. This problem set has a maximum aspect ratio of 1.75. This constraint generates a little solution space because it restricts the department dimensions and allows little combinations of departments to be placed in one bay.

Table 6.1 also provides the best FBS solutions and best known solutions. These can be used to compare the performances of the proposed algorithm with those from previous research. This comparison is also important to examine whether the proposed algorithm can achieve the optimal solutions or not.

All of the best FBS solutions are obtained from Konak et al. (2006). Their research used Mixed Integer Programming (MIP) to solve UA-FLPs with FBS representation. They claimed that their method could generate optimal solutions for problems with up to 14 departments. However, it could not be used to solve larger problems.

On the other hand, the best known solutions are obtained from various research. They are originated from the MIP approach (Meller et al., 1999; Sherali et al., 2003), continuous representation approach (Liu and Meller, 2007), Slicing Tree Structure (STS) representation approach (Scholz et al., 2009) and FBS representation approach (Konak et al., 2006).

The author uses a maximum number of local search steps = 1000n, and a maximum number of local search steps spent when the solution quality could not be improved = 100n. The proposed algorithm is replicated five times with a maximum number of iterations = 1000 and a maximum number of unimproved iterations = 500. The algorithm is coded with C++ and compiled using GCC 4.3.0. It is tested using an Intel Centrino Duo processor (1.7 GHz) and a Linux operating system.

85 6.3 Evaluation of the Proposed Algorithm with Manual Parameter Tuning

In this experiment, the proposed algorithm uses the parameter values obtained from manual parameter tuning; number of artificial ants = n (n = number of departments), pheromone information parameter = 1, heuristic information parameter = 5, and evaporation rate = 0.1.

The statistical results obtained for the proposed algorithm are summarized in Table 6.2. From the table, it can be shown that the proposed algorithm is effective since the difference between the best and worst solutions is relatively low. It can also be seen that the computational time of the proposed algorithm is acceptable for facility layout planning purposes.

Table 6.2 Statistical data on results and computation time of the proposed algorithm
Objective function value No Problem set 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 O7 FO7 FO8 O9 V10s V10a M11s M11a M15s M15a M25 NUG12 NUG15 BA12 BA12* BA12TS BA14 BA14* BA14TS AB20 SC30* SC35* AS Best 136.58 23.12 22.39 241.06 22,899.64 21,463.10 1,321.35 1,204.15 23,197.80 27,545.30 1,496.42 262.00 536.75 8,786.00 8,299.50 8,587.05 5,004.55 4,913.22 4,927.69 5,677.83 3,679.85 3,962.72 AS mean 136.58 23.12 22.39 241.06 22,899.64 21,463.10 1,321.35 1,204.15 23,232.12 27,682.00 1,508.11 262.00 536.75 8,786.00 8,299.50 8,587.05 5,004.55 4,913.22 4,932.78 5,690.65 3,716.56 4,061.89 AS worst 136.58 23.12 22.39 241.06 22,899.64 21,463.10 1,321.35 1,204.15 23,369.40 27,809.20 1,525.19 262.00 536.75 8,786.00 8,299.50 8,587.05 5,004.55 4,913.22 4,936.18 5,699.20 3,749.46 4,148.29 Computation time (second) Best time Total time (1 replication) (5 replications) 38 194 17 97 26 134 44 228 57 316 61 309 83 457 77 396 210 1,330 259 1,604 2,215 9,567 126 637 328 1,792 164 890 301 1,960 445 3,121 313 2,009 498 2,913 401 1,836 1,055 7,640 17,935 74,533 28,671 148,616

*the problem is solved using mFBS

86 The comparisons of the proposed algorithm with previous research can be examined in Table 6.3 and Table 6.4. The algorithm performs better than the approach proposed by Tate and Smith (1995). Although both methods use FBS representation, the proposed algorithm utilizes a high number of local search procedures which help it to focus on finding good results. In general, the proposed algorithm produces the same or better results compared to the method proposed by Konak et al. (2006) except for problem sets M11s and NUG15. This is probably because they used Mixed Integer Programming which is weak in solving large problem instances. The proposed algorithm can produce better results than the method used by Liu and Meller (2007) on BA12*, BA14*, and SC30* because of the use of mFBS. On the other hand, the proposed algorithm cannot outperform the method proposed by Scholz et al. (2009) since they use an STS representation which can produce more solution candidates.

Table 6.3 Results and comparison of the proposed algorithm with previous research
No 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Problem Set O7 FO7 FO8 O9 V10s V10a M11s M11a M15s M15a M25 NUG12 NUG15 BA12 BA12* BA12TS BA14 BA14* BA14TS AB20 SC30* SC35* Tate and Smith (1995) 8,861.00 5,080.10 7,205.40 Konak et al. (2006) 23.12 22.39 241.06 22,899.65 21,463.07 1,317.79 1,225.00 27,781.95 31,779.09 265.60 526.75 8,801.33 8,801.33 8,600.33 5,004.55 5,004.55 4,927.69 6,890.82 Liu and Meller (2007) 131.63 20.73 22.31 235.95 19,997.00 8,702.00 8,702.00 5,004.00 5,004.00 3,706.83 3,604.12 Scholz et al. (2009) 132.00 239.07 19,994.10 8,264.00 8,264.00 4,712.33 4,712.33 AS Best objective function value 136.58 23.12 22.39 241.06 22,899.64 21,463.10 1,321.35 1,204.15 23,197.80 27,545.30 1,496.42 262.00 536.75 8,786.00 8,299.50 8,587.05 5,004.55 4,913.22 4,927.69 5,677.83 3,679.85 3,962.72

*the proposed algorithm uses mFBS for solving the problems In addition, as shown in Table 6.4, the proposed algorithm is able to produce the best FBS solution (or even better) for all the problem sets except only for M11s

87 and NUG15. These outstanding results can be achieved because of the use of the right representation in AS and the contribution of local search procedures. Meanwhile, when a comparison with the best known solution is made, the proposed algorithm can improve 7 problem sets, i.e. M15s, M15a, M25, NUG12, BA12TS, AB20, and SC30. Despite this, it is unable to improve the previous best known solution for 10 problem sets. This is due to the nature of the FBS representation that restricts the solution space (Konak et al., 2006).

Table 6.4 Results and comparison of the proposed algorithm with the best-FBS and best-known solutions
Objective function value AS Best 136.58 23.12 22.39 241.06 22,899.64 21,463.10 1,321.35 1,204.15 23,197.80 27,545.30 1,496.42 262.00 536.75 8,786.00 8,299.50 8,587.05 5,004.55 4,913.22 4,927.69 5,677.83 3,679.85 3,962.72 Percent Difference with Best FBS** Best known** -3.62% 0.00% -9.39% 0.00% 0.00% 0.00% -2.12% 0.00% -12.69% 0.00% 0.00% -0.27% -0.27% 1.73% -1.57% 19.76% 19.76% 15.37% 5.85% 6.14% 1.37% 1.37% -1.86% -1.86% 0.17% -6.90% 6.05% -1.44% 0.15% 0.15% 0.00% -5.84% 1.86% -4.09% 0.00% 0.00% 21.36% 21.36% 0.74% -9.05%

No 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

Problem Set O7 FO7 FO8 O9 V10s V10a M11s M11a M15s M15a M25 NUG12 NUG15 BA12 BA12* BA12TS BA14 BA14* BA14TS AB20 SC30* SC35*

*the proposed algorithm uses mFBS for solving the problems **positive values in the last two columns represent improvements Furthermore, the results between BA12 and BA12*, and BA14 and BA14* can be compared to evaluate the effectiveness of mFBS. The results show that the mFBS (which is used in BA12* and BA14*) performs better than the original FBS (which is used in BA12 and BA14). As stated in the previous section, mFBS can

88 extend the solution space by using free space to help the layout to fulfill department constraints. It also tries to improve the solution by adding empty space to the bays.

The proposed algorithm is effective when solving large problem sets. As shown, it performs better on problem sets M25, AB20, and SC30 which have 25, 20, and 30 departments respectively. The results reflect the possibility to improve other UA-FLP representation techniques when solving these problems. When the proposed algorithm is used for solving SC35, it is unable to improve the best known solution. Although the problem is similar to SC30, SC35 has a more restricted problem constraint (the maximum aspect ratio is = 4). This problem constraint restricts the algorithm to find a collection of departments which can be placed in one bay, thus simultaneously decreases the objective function.

6.4

Evaluation of the Proposed Algorithm with Fuzzy Logic Controller

In this experiment, the proposed algorithm uses FLC to tune the parameter values. For the initial parameter values, this research uses the following settings: number of artificial ants = n (n = number of departments), pheromone information parameter = 1, heuristic information parameter = 5, and evaporation rate = 0.1. When applying FLC, the following limits for the parameter values are used: number of ants - between 5 and 2n-5, pheromone information parameter and heuristic information parameter - between 1 and 15, and evaporation rate - between 0.1 and 0.9.

The results obtained from testing the AS algorithm with fuzzy logic are shown in Table 6.5. In the third column, there are six types of implementations. FLC0 means AS without FLC (manual parameter tuning), FLC1 means AS with FLC on number of ants, FLC2 means AS with FLC on pheromone information parameter, FLC3 means AS with FLC on heuristic information parameter, FLC4 means AS with FLC on evaporation rate and FLC5 means AS with FLC on all the four parameters at once. The computation times needed for these implementations are not shown because they do not differ much from those given in Table 6.2

89 Table 6.5 Results and comparison of the proposed algorithm with Fuzzy Logic Controller
Problem set Objective function value AS best AS mean AS worst 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 136.58 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 23.12 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 22.39 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 241.06 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 22,899.64 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 21,463.10 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,321.35 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 1,204.15 Percent difference with FLC0 AS best* AS mean* AS worst* 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

No

Method FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5

O7

FO7

FO8

O9

V10s

V10a

M11s

M11a

* positive values in the last three columns represent improvements

90 Table 6.5 Results and comparison of the proposed algorithm with Fuzzy Logic Controller (continued)
Problem set Objective function value AS best AS mean AS worst 23,197.80 23,232.12 23,369.40 23,197.80 23,197.80 23,197.80 23,197.80 23,197.80 23,197.80 23,197.80 23,264.72 23,369.40 23,197.80 23,232.12 23,369.40 23,197.80 23,232.12 23,369.40 27,545.30 27,682.00 27,809.20 27,545.30 27,545.30 27,545.30 27,545.30 27,638.72 27,701.00 27,545.30 27,669.86 27,701.00 27,545.30 27,607.58 27,701.00 27,545.30 27,669.86 27,701.00 1,496.42 1,508.11 1,525.19 1,522.69 1,533.35 1,550.23 1,483.48 1,499.01 1,525.91 1,505.53 1,522.58 1,543.17 1,484.82 1,503.93 1,524.06 1,502.23 1,530.72 1,568.93 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 262.00 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 536.75 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,299.50 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,596.08 8,632.19 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 8,587.05 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 4,913.22 Percent difference with FLC0 AS best* AS mean* AS worst* 0.0% 0.1% 0.7% 0.0% 0.1% 0.7% 0.0% -0.1% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.5% 1.0% 0.0% 0.2% 0.4% 0.0% 0.0% 0.4% 0.0% 0.3% 0.4% 0.0% 0.0% 0.4% -1.7% -1.6% -1.6% 0.9% 0.6% 0.0% -0.6% -1.0% -1.2% 0.8% 0.3% 0.1% -0.4% -1.5% -2.8% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% -0.1% -0.5% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

No

Method FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5

M15s

10

M15a

11

M25

12 NUG12

13 NUG15

14

BA12

15 BA12TS

16

BA14

* positive values in the last three columns represent improvements

91 Table 6.5 Results and comparison of the proposed algorithm with Fuzzy Logic Controller (continued)
Problem set Objective function value AS best AS mean AS worst 4,927.69 4,932.78 4,936.18 4,927.69 4,929.39 4,936.18 4,927.69 4,927.69 4,927.69 4,927.69 4,929.39 4,936.18 4,927.69 4,932.78 4,936.18 4,927.69 4,927.69 4,927.69 5,677.83 5,690.65 5,699.20 5,677.83 5,857.68 6,167.87 5,677.83 5,832.97 6,065.68 5,677.83 5,677.83 5,677.83 5,677.83 5,682.10 5,699.20 5,677.83 5,910.54 6,065.68 3,679.85 3,716.56 3,749.46 3,654.75 3,737.90 3,795.27 3,700.19 3,746.88 3,820.45 3,649.41 3,707.04 3,762.74 3,727.14 3,746.45 3,762.34 3,662.67 3,726.04 3,804.61 3,962.72 4,061.89 4,148.29 3,941.12 4,052.00 4,170.52 4,006.41 4,069.06 4,135.07 3,965.57 4,047.76 4,136.80 3,962.43 4,067.36 4,132.52 4,009.60 4,104.60 4,174.42 Percent difference with FLC0 AS best* AS mean* AS worst* 0.0% 0.1% 0.0% 0.0% 0.1% 0.2% 0.0% 0.1% 0.0% 0.0% 0.0% 0.0% 0.0% 0.1% 0.2% 0.0% -2.9% -7.6% 0.0% -2.4% -6.0% 0.0% 0.2% 0.4% 0.0% 0.2% 0.0% 0.0% -3.7% -6.0% 0.7% -0.6% -1.2% -0.5% -0.8% -1.9% 0.8% 0.3% -0.4% -1.3% -0.8% -0.3% 0.5% -0.3% -1.4% 0.5% 0.2% -0.5% -1.1% -0.2% 0.3% -0.1% 0.3% 0.3% 0.0% -0.1% 0.4% -1.2% -1.0% -0.6%

No

Method FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5 FLC0 FLC1 FLC2 FLC3 FLC4 FLC5

17 BA14TS

18

AB20

19

SC30

20

SC35

* positive values in the last three columns represent improvements The table provides the best, the mean (average), and the worst objective function values gathered from 5 replications for each problem instance. The last three columns provide the differences between AS with FLC and without FLC (FLC0). In addition, modified FBS (mFBS) is used for solving problems with empty space (BA12, BA14, SC30 and SC35).

When comparing the results in the AS-best column, AS with FLC performs as good as AS with manual parameter tuning on 17 problem sets since all the implementations of FLC on these particular problems produce the same best objective function values as those generated by AS with manual parameter tuning. In general, all of these 17 problem sets do not have a too large number of departments. This concludes that the parameter values have a relatively low influence on small problem instances.

92 The difference starts when AS with FLC is used for solving M25, SC30, and SC35 which have 25, 30, and 35 number of departments respectively. Nevertheless, the results of using FLC are still encouraging because the differences are very small which are below 2%. The high number of departments means a wider solution space must be explored. Hence, it can be concluded that parameter values influence the performance of the AS algorithm when it is used for solving large problem instances.

The performances of the five types of FLC implementations are not too varied. When FLC is used to tune only one parameter (FLC1, FLC2, FLC3, and FLC4), the results are comparable with those obtained when it is not used (FLC0). These can be achieved because the values of the other fixed parameters are taken from manual parameter tuning. On the other hand, when FLC is used to tune the four parameters at once, good quality results are also attained. This proves that FLC can be utilized to produce good solutions for AS by automating the tuning of its parameters. This brings an advantage in using FLC because it omits the need to perform manual parameter tuning which is time consuming.

Additionally, the use of FLC for tuning AS's parameters resulted in both improvement and deterioration on M25, SC30, and SC35 problem instances. FLC can find better results on 6 cases whereas it performs poorer on 8 cases. The variability in performance of the AS algorithm is a basic characteristic of a metaheuristic. Although metaheuristic algorithms have their own strengths, they still rely on random numbers to conduct their search.

Furthermore, it cannot be decided which type of FLC implementation (FLC1, FLC2, FLC3, FLC4, FLC5) is the best since the results are varied. For example, FLC1 can improve problem sets SC30 and SC35 but it performs a little poorer when solving M25. However, the most useful FLC implementation is FLC5 since it does not need any parameter value from manual parameter tuning. FLC5 can certainly be used as a parameter tuning method for the AS algorithm.

93

6.5

Conclusions

This chapter has evaluated the proposed AS algorithm with FBS representation for solving UA-FLPs. The proposed algorithm is effective and can produce all of the best FBS solutions (or even better) except for M11s and NUG15. In addition, it can improve the best known solution for 7 problem instances, i.e. M15s, M15a, M25, NUG12, BA12TS, AB20, and SC30. This research has also proposed an improvement to the FBS representation by using free or empty space. The improvement obtained is justified on problem sets BA12 and BA14. Evidently, the proposed algorithm is also proven to be effective when solving large problem sets: M25, AB20, and SC30.

Furthermore, this chapter has described the testing of the FLC to automate the tuning of the proposed algorithm. The experiments involved tuning four parameters individually, i.e. number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate, as well as tuning all of them at once. The results show that FLC can be used to replace manual parameter tuning which is time consuming. The results also show that instead of using static parameter values, FLC has the potential to help the AS algorithm to achieve better objective function values.

CHAPTER 7

CONCLUSIONS

7.1

Introduction

This study attempts to investigate the performance of Ant System (AS) when solving Unequal Area Facility Layout Problems (UA-FLPs). To date, a formal Ant Colony Optimization (ACO) based metaheuristic has not been applied to solve them. Recent research has mainly used it to deal with Quadratic Assignment Problems (QAPs). Generally, ACO has been proven to perform outstandingly when solving QAPs, which are the backbone problems for UA-FLPs (Sttzle and Dorigo, 1999). This motivates the author to create a new AS for solving UA-FLPs, and to investigate its performance.

In addition, parameter values play important roles in a metaheuristic algorithm. The significance of parameter values in metaheuristics has been laid down by Eiben et al. (1999). They stated that parameter dependencies, time consuming efforts, and unnecessarily optimal parameter values are the technical drawbacks of manual parameter tuning. In brief, they suggested the use of dynamic parameter setting.

Based on these reasons, this study investigates the effect of Fuzzy Logic Controller (FLC) on dynamically tuning the parameters of AS. This study involves four basic parameters of AS; number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate.

95

The proposed algorithm with and without FLC are tested using numerous UA-FLPs which are represented as Flexible Bay Structures (FBSs). The problem instances comprehensively represent the UA-FLP domain since their sizes range from 7 to 35 departments. In addition, the problem instances chosen have been heavily used in previous research.

7.2

Conclusions

This study has involved an in-depth literature review on several domains, i.e. ACO, facility layout, and parameter tuning for metaheuristics. In addition, this study has formulated an AS algorithm for solving UA-FLPs. It has also designed a Fuzzy Logic implementation for tuning the algorithm's parameter values. Then the algorithm is tested using well-known problem instances from the literature. In short, the following key points can be made: 1. This research has formulated an AS algorithm with FBS representation for solving UA-FLPs. The formulation of the proposed algorithm follows the past characteristics of AS in solving QAPs. The main looping of the proposed algorithm consists of three main parts: (1) construct ant solutions, (2) apply local search procedures, and (3) update pheromone values. The FBS representation is represented by department sequence and bay break sequence. The proposed algorithm uses several local search procedures to enhance the search on department sequence and bay structure layout. In this study, the global-best solution and iteration-best solution are used to update the pheromone values in every iteration. In addition, the study proposes an improvement to the FBS structure (modified-FBS or mFBS) by using empty space to fulfill department constraints. 2. This study is the first that applies the AS concept for solving UA-FLPs. The ACO family has been used to solve Facility Layout Problems (FLPs) but its implementation was intended to solve QAPs and Machine Layout Problems (MLPs). This condition is shown in the literature review section (Table 2.3). In

96 addition, this study is the initial attempt to use FLC to automatically tune the parameter values of AS. Previous research has used dynamic parameter tuning but none has tried to utilize FLC as shown in Table 2.4. 3. This research has described the implementation of FLC to automate the tuning of the proposed algorithm. The input parameters for the FLC are the standard deviation of solutions and the number of unimproved iterations. The former can represent how solutions in one iteration differ from those in another iteration. Meanwhile the latter represents the degree of which the current search is trapped in local optima. The proposed algorithm utilizes four output parameters which are aimed to influence its search intensification and diversification. These parameters are the number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate. Then several rule bases are

designed to control the search intensification and diversification of the algorithm. They are made to provide the algorithm with an ability to jump out from local optima. In addition, the rule bases are intended to limit the search when it is performed in a too wide region. 4. This study has evaluated the proposed AS algorithm with FBS representation for solving UA-FLPs. The proposed algorithm is effective and can produce all of the best FBS solutions (or even better) except for M11s and NUG15. In addition, the proposed algorithm can improve the best known solution for 7 problem instances, i.e. M15s, M15a, M25, NUG12, BA12TS, AB20, and SC30. The improvement obtained by using the mFBS representation is justified on problem sets BA12 and BA14. Evidently, the proposed algorithm is also proven to be effective when solving large problem sets: M25, AB20, and SC30. 5. This study has described the testing of FLC to automate the tuning of the proposed algorithm. The experiments involved tuning four parameters individually, i.e. number of ants, pheromone information parameter, heuristic information parameter, and evaporation rate, as well as tuning all of them at once. The results show that FLC can be used to replace manual parameter tuning which is time consuming. The results also show that instead of using static parameter values, FLC has the potential to help the AS algorithm to achieve better objective function values.

97 7.3 Future Work

In terms of future work, this study can be expanded in the following ways:

1. It can be extended to include additional constraints of UA-FLPs such as fixed department location, fixed department dimension, and free facility dimension.

2. It can include special cases of UA-FLPs, such as multiple-floor UA-FLPs, multiperiod UA-FLPs, stochastic UA-FLPs, or UA-FLPs with input-output points.

3. The proposed AS algorithm can be substituted with other ACO variants such as Max-Min Ant System (MMAS), Ant Colony System (ACS), etc. The substitution with other ACO variants is expected because they have been proposed by researchers to complement and improve the performance of the basic AS algorithm.

Finally, based on the results obtained, it is believed that this research will bring a significant advancement to the FLP domain.

98

REFERENCES

AbdelRazig, Y., El-Gafy, M., and Ghanem, A. (2006). Dynamic Construction Site Layout Using Ant Colony Optimization. Transportation Research Board (TRB) Annual Meeting. 22-26 January. Washington, D.C., USA: TRB.

Anjos, M.F., and Vannelli, A. (2006). A New Mathematical-Programming Framework for Facility Layout Problem. INFORMS Journal on Computing 18(1), 111-118.

Armour, G..C., and Buffa, E.S. (1963). A Heuristic Algorithm and Simulation Approach to Relative Allocation of Facilities. Management Science. 9(2), 294-300.

Balakrishnan, J., Jacobs, F. R., and Venkataramanan, M. A. (1992). Solutions for The Constrained Dynamic Facility Layout Problem. European Journal of Operational Research. 57(2), 280-286.

Baykasoglu, A., Dereli, T., and Sabuncu, I. (2006). An Ant Colony Algorithm for Solving Budget Constrained and Unconstrained Dynamic Facility Layout Problems. The International Journal of Management Science. 34(4), 385-396.

Bhrde, F. M. G., McGinnity, T. M., and McDaid, L. J. (2005). Landscape Classification and Problem Specific Reasoning for Genetic Algorithms. International Journal of Systems and Cybernetics. 34(9/10), 1469-1495.

99 Blum, C. (2005a). Ant Colony Optimization: Introduction and recent trends. Physics of Live Reviews. 2(4), 353-373.

Blum, C. (2005b). Beam-ACO: Hybridizing ant colony optimization with beam search: An application to open shop scheduling. Computers and Operations Research 32(6), 15651591.

Blum, C., and Roli, A. (2003). Metaheuristics in Combinatorial Optimization: Overview and Conceptual Comparison. ACM Computing Surveys. 35(3), 268308.

Bozer, Y.A., and Meller, R.D. (1997). A Reexamination of the Distance-Based Facility Layout Problem. IIE Transactions. 29(7), 549-560.

Bozer, Y.A., Meller, R.D., and Erlebacher, S.J. (1994). An Improvement-Type layout Algorithm for Single and Multiple Floor facilities. Management Science. 40(7), 918-932.

Bullnheimer, B., Hartl, R.F., and Strauss, C. (1999). A new rank-based version of the Ant System: A computational study. Central European Journal for Operations Research and Economics. 7(1), 25-38.

Chen, W. H., and Srivstava, B. (1994). Simulated annealing procedures for forming machine cell in-group technology. European Journal of Operational Research. 75(1), 100111.

Corry, P., and Kozan, E. (2004). Ant Colony Optimisation for Machine Layout Problems. Computational Optimization and Applications. 28(3), 287-310.

Dai, C. H., Zhu, Y.F. and Chen, W.R. (2006). Adaptive Probabilities of Crossover and Mutation in Genetic Algorithms Based on Cloud Model. Information Theory Workshop. 22-26 October. Chengdu, China: IEEE, 710-713.

100 De Jong, K.A. (1975). An analysis of the behavior of a class of genetic adaptive systems. Doctor Philosophy. Department of Computer Science, University of Michigan, Ann Arbor: USA.

Demirel N.C. and Toksari M.D. (2006) Optimization of the Quadratic Assignment Problem using an Ant Colony Algorithm. Applied Mathematics and Computation. 183(1), 427-435.

Dorigo, M., and Di Caro, G. (1999). Ant colony optimization: A new meta-heuristic. Proceeding of the 1999 Congress on Evolutionary Computation. 6-9 July. Washington, DC, USA: IEEE, 1470-1477.

Dorigo, M., and Gambardella, L.M. (1997). Ant Colony System: A cooperative learning Approach to The Traveling Salesman Problem. IEEE Transactions on Evolutionary Computation. 1(1), 53-66.

Dorigo, M., Maniezzo, V., and Colorni, A. (1996). Ant System: Optimization by a colony of cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics Part B. 26(1), 29-41.

Eiben, A.E., Hinterding, R. and Michalewicz, Z. (1999). Parameter Control in Evolutionary Algorithms. IEEE Transactions on Evolutionary Computation. 3(2), 124-141.

Favuzza, S., Graditi, G. and Sanseverino, E.R. (2006). Adaptive and Dynamic Ant Colony Search Algorithm for Optimal Distribution Systems Reinforcement Strategy. Applied Intelligence. 24(1), 31-42.

Francis, R.L., and White, J.A. (1974). Facility Layout and Location: An Analytical Approach. Englewood Cliffs, NJ: Prentice-Hall.

101 Frazelle, E.H. (1986). Material Handling: A Technology for Industrial

Competitiveness. Material Handling Research Center Technical Report, Georgia Institute of Technology, USA.

Gambardella, L.M. and Dorigo, M. (2000). Ant Colony System hybridized with a new local search for the sequential ordering problem. INFORMS Journal on Computing. 12(3), 237255.

Gambardella, L.M., Taillard, E.D. and Dorigo. M. (1997). Ant colonies for the QAP. Technical Report 4-97, IDSIA, Lugano, Switzerland.

Gendreau, M., and Potvin, J.-Y. (2005). Metaheuristics: A bibliography. Annals of Operations Research. 140(1), 189-213.

Glover, F., and Laguna, M. (1997). Tabu Search. Boston: Kluwer.

Grefenstette, J.J. (1986). Optimization of Control Parameters for Genetic Algorithms. IEEE Transactions on Systems, Man, and Cybernetics. 16(1), 122-128.

Hani, Y., Amodeo, F., Yalaoui, F., and Chen, H. (2007). Ant Colony Optimization for Solving an Industrial Layout Problem. European Journal of Operational Research. 183(2), 633-642.

Hao, Z., Cai, R.C., and Huang, H. (2006). An Adaptive Parameter Control Strategy for ACO. Proceeding of International Conference on Machine Learning and Cybernetics. August 13-16. Dalian, China: 203-206.

Hao, Z., Huang, H., Qin, Y., and Cai, R. (2007). An ACO Algorithm with Adaptive Volatility Rate of Pheromone Trail. Proceeding of International Conference on Computational Science. 27-30May. Beijing, China: 1167-1170.

102 Hartono, P., Hashimoto, S. and Wahde, M. (2004). Labeled-GA with Adaptive Mutation Rate. Proceeding of IEEE Congress on Evolutionary Computation. 20-23 June. Portland, Oregon: 1851-1858.

Hassan, M.M.D., Hogg, G.L., and Smith, D.R. (1986). Shape: A construction algorithm for area placement evaluation. International Journal of Production Research. 24(5), 1283-1295.

Hinterding, R. (1997). Self-adaptation using Multi-chromosomes. Proceeding of IEEE International Conference on Evolutionary Computation.13-16 April. Indianapolis, USA: 87-91.

Huang, H., Yang, X., Hao, Z., and Cai, R. (2006). A Novel ACO Algorithm with Adaptive Parameter. Proceeding of International Conference on Intelligent Computing. 16-19 August. Yunnan, China: 12-21.

Ibaraki, T. (1987). Combinatorial optimization problems and their complexity. In Enumerative Approaches to Combinatorial Optimization - Part I. Annals of Operations Research. 10, 1-49.

Johnson, R. V. (1982). Spacecraft for multi-floor layout planning. Management Science. 28(40), 407-417.

Kim, J.-G., and Kim, Y.-D. (1999). A branch and bound algorithm for locating input and output points of departments on the block layout. Journal of the Operational Research Society. 50(5), 517-525

Ko, M-S., Tang, T-W. and Hwang, C-S. (1996). Adaptive Crossover Operator based on Locality and Convergence. Proceedings of the 1996 IEEE International Joint Symposia on Intelligence and System. 4-5 November. Rockville, USA: IEEE, 18-22.

103 Konak, A., Kulturel-Konak, S., Norman, B.A., and Smith, A.E. (2006). A New Mixed Integer Formulation for Optimal Facility Layout design. Operation Research Letters. 34(6), 660-672.

Koopmans, T. C., and Beckman, M. (1957). Assignment problems and the location of economic activities. Econometrica. 25(2), 53-76.

Koppen, M. (2004). No-Free-Lunch theorems and the diversity of algorithms. Congress on Evolutionary Computation. 1, 235-241.

Liu, Q., and Meller, R.D. (2007). A sequence-pair representation and MIP-modelbased heuristic for the facility layout problem with rectangular departments. IIE Transactions. 39(4), 377-394.

Maffoli, F. and Galbiati, G. (2000). Approximability of hard combinatorial optimization problems: an introduction. Annals of Operations research. 96(14), 221-236.

Maniezzo, V. (1999). Exact and approximate nondeterministic tree-search procedures for the quadratic assignment problem. INFORMS Journal on Computing. 11(4), 358-369.

Maniezzo, V., and Colorni, A. (1999). The Ant System applied to the quadratic assignment problem. IEEE Transaction of Knowledge and Data Engineering. 11(5),769-778.

Maniezzo, V., Colorni, A., and Dorigo, M. (1994). The ant system applied to the quadratic assignment problem. Technical Report IRIDIA/94-28, Universit Libre de Bruxelles, Belgium.

104 McKendall Jr, A.R., and Shang, J. (2006). Hybrid Ant Systems for the Dynamic Facility Layout Problem. Computers and Operations Research. 33 (3), 790803.

Meller, R.D. (1992). Layout Algorithms for Single and Multiple Floor Facilities. Doctor of Philosophy Thesis, Department of Industrial and Operations Engineering, University of Michigan, USA.

Meller, R.D., and Bozer, Y.A. (1996). A New Simulated Annealing Algorithm for The Facility Layout Problem. International Journal of Production Research. 34(6), 1675-1692.

Meller, R.D., and Bozer, Y.A. (1997). Alternative approaches to solve the multi-floor facility layout problem. Journal of Manufacturing Systems. 16(3), 192-203.

Meller, R.D., and Gau, K.Y. (1996). The facility layout problem: recent and emerging trends and perspectives. Journal of Manufacturing Systems. 15(5), 351-366.

Meller, R.D., Chen, W., and Sherali, H.D. (2007). Applying the Sequence-pair Representation to Optimal Facility Layout designs. Operations Research Letters. 35(5), 651-659.

Meller, R.D., Narayanan, V., and Vance, P.H. (1999). Optimal Facility Layout Design. Operations Research Letters. 23(3-5), 117127.

Mendel, J. (1995). Fuzzy logic systems for engineering: a tutorial. Proceedings of the IEEE. 83(3), 345-377.

Merkle, D., Middendorf, M., and Schmeck, H. (2002). Ant colony optimization for resource-constrained project scheduling. IEEE Transaction on Evolutionary Computation. 6(4), 333-346.

105 Montreuil, B. (1990). A Modelling Framework for Integrating Layout Design and Flow Network Design. Proceedings of the Material Handling Research Colloqium. Hebron: 43-58.

Montreuil, B., and Ratliff, H.D. (1988). Optimizing the location of input and output stations within facilities layout. Engineering Costs and Production Economics. 14(3), 177-187.

Mouhoub, M., and Wang, Z. (2006). Ant Colony with Stochastic local Search for the Quadratic Assignment Problem. Proceedings of the 18th IEEE International Conference on Tools with Artificial Intelligence. 13-15 November. Washington D.C.: IEEE, 127 131.

Muther, R. (1955). Practical Plant Layout. New York: McGraw-Hill.

Osman, I.H., and Laporte, G. (1996). Metaheuristics: A bibliography. Annals of Operations Research. 63(5), 513-628.

Pilat, M.L., and White, T. (2002). Using Genetic Algorithms to optimize ACS-TSP. In M. Dorigo et al. (Eds): ANTS 2002. Lecture Notes in Computer Science. 2463, 282-287.

Qin, L., Luo, J., Chen, Z., Guo, J., Chen, L., and Pan, Y. (2006). Phelogenetic Tree Construction using Self Adaptive Ant Colony Algorithm. Proceedings of the First International Multi-Symposiums on Computer and Computational Sciences - Volume 1. 20-24 June. Zhejiang, China: 179-187.

Ramkumar, A.S., and Ponnambalam, S.G. (2006). Hybrid Ant Colony System for solving Quadratic Assignment Formulation of Machine Layout Problems. Proceedings IEEE Conference on Cybernetics and Intelligent Systems. 7-9 June. IEEE:1-5.

106 Randall, M. (2004). Near Parameter Free Ant Colony Optimisation. Lecture Notes in Computer Science. 3172, 374-381.

Ropke, S. and Pisinger, D. (2006). An Adaptive Large Neighborhood Search Heuristic for the Pickup and Delivery Problem with Time Windows. Transportation Science. 40, 455472.

Rosenblatt, M.J. (1986). The dynamics of plant layout. Management Science. 32(1), 76-86.

Sawai, H. and Kizu, S. (1998). Parameter-free Genetic Algorithm Inspired by Disparity Theory of Evolution. Proceedings of the 5th International Conference on Parallel Problem Solving from Nature. 27-30 September. Amsterdam, Holland: 702-711.

Scholz, D., Petrick, A., and Domschke, W. (2009). STaTS: A slicing tree and tabu search based heuristic for the unequal area facility layout problem. European Journal of Operational Research. 197(1), 166-178.

Shebanie II, C.R. (2004). An Integrated, Evolutionary Approach to Facility Layout and Detailed Design. Master of Engineering, Department of Industrial Engineering, University of Pittsburgh. Pittsburgh: USA.

Sherali, H.D., Fraticelli, B.M.P., and Meller, R.D. (2003). Enhanced Model Formulations for Optimal Facility Layout. Operations Research. 51(4), 629644.

Shmygelska, A., and Hoos, H.H. (2005). An ant colony optimisation algorithm for the 2D and 3D hydrophobic polar protein folding problem. Bioinformatics. 6(30), 1-22.

107 Singh, S.P., and Sharma R.R.K. (2005). A review of different approaches to the facility layout problem. International Journal of Advanced Manufacturing Technology. 30(5-6), 425-433.

Solimanpur, M., Vrat, P., and Shankar, R. (2004). Ant Colony Optimization Algorithm to the inter-cell layout problem in cellular manufacturing. European Journal of Operational Research. 157(3), 592-606.

Solimanpur, M., Vrat, P., and Shankar, R. (2005). An Ant algorithm for the single row layout problem in flexible manufacturing systems. Computers and Operations Research. 32(3), 583-598.

Sttzle, T., and Dorigo, M. (1999). ACO algorithms for the quadratic assignment problem. In D. Corne, M. Dorigo and F. Glover (eds) New Ideas in Optimization. New York: McGraw-Hill.

Sttzle, T., and Hoos, H.H. (2000). MAX MIN Ant system. Future Generation Computer Systems. 16(8), 889-914.

Syarif, A., Gen, M., and Yamashiro, M. (2004). Hybridized Parallel Genetic Algorithm for Facility Location Problem. Research Reports, Ashikaga Institute of Technology. 38, 65-71.

Talbi, E.G., Roux, O., Fonlupt, C. and Robillard, D. (2001). Parallel Ant Colonies for Quadratic Assignment Problem. Future Generation Computer Systems. 17(4), 441-449.

Tam, K.Y. (1992). A Simulated Annealing Algorithm for Allocating Space to Manufacturing Cells. International Journal of Production Research. 30(1), 63-87.

108 Tam, K.Y., and Li, S.G. (1991). A Hierarchical Approach to Facility Layout Problem. International Journal Production Research. 29(1), 165-184.

Tate, D.M., and Smith, A.E. (1995). Unequal-Area Facility Layout by Genetic Search. IIE Transactions. 27, 465-472.

Tompkins J.A., White J.A., Bozer J.A. and Tanchoco, J.M.A. (2003). Facilities Planning. John Wiley & Sons: New York.

Tretheway, S.J. and Foote, B.L. (1994). Automatic Computation and Drawing of Facility Layouts with Logical Aisle Structures. International Journal of Production Research. 32(7), 1545-1556.

Vakharia, A.J., and Chang, Y.L. (1997). Cell formation in group technology: a combinatorial search approach. International Journal of Production Research. 35(7), 20252043

van Camp, D.J. (1989). A Nonlinear Optimization Approach for Solving Facility Layout Problem. Master of Applied Science, Department of Industrial Engineering, University of Toronto, Canada.

van Camp, D.J., Carter, M.W., and Vannelli, A. (1991). A Nonlinear Optimization Approach for Solving Facility Layout Problem. European Journal of Operation Research. 57(2), 174-189.

Wang, P. Y., Wang, G. S., Song, Y. H., and Johns, A. T. (1996). Fuzzy logic controlled genetic algorithms. Proceedings of the Fifth IEEE International Conference on Fuzzy Systems. 8-11 September. New Orleans, LA, USA: IEEE, 2, 314-17.

Watanabe, I., and Matsui, S. (2003). Improving the Performance of ACO Algorithms by Adaptive Control of Candidate Set. Proceedings of Congress on

109 Evolutionary Computation (CEC '03). 8-12 December. Canberra, Australia: IEEE, 2, 1355-1362.

Wolpert, D.H. and Macready, W.G. (1997). No Free Lunch Theorems for Optimization. IEEE Transactions on Evolutionary Computation. 1(1), 67-82.

Wong, K.Y. and See, P.C. (2009). A New Minimum Pheromone Threshold Strategy (MPTS) for Max-Min Ant System. Applied Soft Computing. 9 (3), 882-888.

Xue, F., Sanderson, A.C., Bonissone, P., and Graves, R.J. (2005). Fuzzy Logic Controlled Multi-Objective Differential Evolution. Proceedings on the 14th IEEE International Conference on Fuzzy Systems. 22-25 May. Reno NV, USA: IEEE, 720-725.

Zadeh, L.A. (1988). Fuzzy Logic. IEEE Computer. 21(4), 83-93.

110

PUBLICATIONS

1. Wong, K.Y., Ahmad, R. and Komarudin. (2007). An evaluation of parameters tuning methods in metaheuristic algorithms. Proceedings of the Regional Conference on Advanced Processes and Systems in Manufacturing. Putrajaya, Malaysia, pp.41-48. 2. Wong, K.Y. and Komarudin. (2008). Parameter tuning for ant colony optimization: a review. Proceedings of the International Conference on Computer and Communication Engineering. Kuala Lumpur, Malaysia, pp.542-545. 3. Wong, K.Y. and Komarudin. (2008). Ant colony optimization in solving facility layout problems. Proceedings of the International Conference on Mechanical and Manufacturing Engineering. Johor Bahru, Malaysia. 4. Komarudin and Wong, K.Y. (2009). Applying Ant System for Solving Unequal Area Facility Layout Problems. Accepted for publication in European Journal of Operational Research. 5. Komarudin, Wong, K.Y., and See, P. C. (2009). Solving Facility Layout Problems using Flexible Bay Structure Representation and Ant System Algorithm. Submitted to Expert Systems with Applications. 6. Wong, K.Y. and Komarudin. (2010). Comparison of Techniques for Dealing with Empty Spaces in Unequal Area Facility Layout Problems. Accepted for publication in International Journal of Industrial and Systems Engineering, Vol.6, No.3.

111

APPENDICES

112 APPENDIX A Data Sets for UA-FLPs

A.1

Summary of UA-FLP Data Sets Table A.1 Summary of data for UA-FLPs

No

Problem Number of Facility size Common shape Reference set departments Width Height constraint Meller et al. (1998) max = 4 1 O7 7 8.54 13.00 Meller et al. (1998) max = 5 2 FO7 7 8.54 13.00 Meller et al. (1998) max = 5 3 FO8 8 11.31 13.00 4 O9 5 V10s 6 V10a 7 M11s 8 M11a 9 M15s 9 10 10 11 11 15 15 25 12 15 12 16 14 14 20 30 35 12.00 25.00 25.00 6.00 3.00 15.00 15.00 15.00 3.00 3.00 7.00 7.00 6.00 6.00 2.00 12.00 15.00 13.00 51.00 51.00 6.00 2.00 15.00 15.00 15.00 4.00 5.00 9.00 9.00 10.00 10.00 3.00 15.00 16.00 max = 5 lmin = 5 max = 5 lmin = 1 max = 5 lmin = 1 max = 5 max = 5 max = 5 max = 5 lmin = 1 lmin = 1 lmin = 1 lmin = 1 max = 1.75 max = 5 max = 4 Meller et al. (1998) van Camp (1989) van Camp (1989) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) Bozer et al. (1994) van Camp (1989) van Camp (1989) van Camp (1989) Tate and Smith (1995) van Camp (1989) Tate and Smith (1995) Armour and Buffa (1963) Liu and Meller (2007) Liu and Meller (2007)

10 M15a 11 M25 12 NUG12 13 NUG15 14 BA12 15 BA12TS 16 BA14 17 BA14TS 18 AB20 19 SC30 20 SC35

A.2

Data Set O7 Table A.2 Area requirement for problem set O7

Dept 1 2 3 4 5 6 7 Area 16 16 16 36 9 9 9 Maximum Aspect Ratio 4 4 4 4 4 4 4

113 Table A.3 Material flow matrix for problem set O7

From 1 1 2 2 3 3 4 4 5 6 To 4 7 4 7 4 7 5 6 7 7 Material Flow 5 1 3 1 2 1 4 4 2 1

A.3

Data Set FO7 Table A.4 Area requirement for problem set FO7
Dept 1 2 3 4 5 6 7 Area 16 16 16 36 9 9 9 Maximum Aspect Ratio 5 5 5 5 5 5 5

Table A.5 Material flow matrix for problem set FO7

From 1 2 3 4 5 6 To 2 3 4 5 6 7 Material Flow 1 1 1 1 1 1

A.4

Data Set FO8 Table A.6 Area requirement for problem set FO8
Dept 1 2 3 4 5 6 7 8 Area 16 16 16 36 36 9 9 9 Maximum Aspect Ratio 5 5 5 5 5 5 5 5

114

Table A.7 Material flow matrix for problem set FO8

From 1 2 3 4 5 6 7 To 2 3 4 5 6 7 8 Material Flow 1 1 1 1 1 1 1

A.5

Data Set O9 Table A.8 Area requirement for problem set O9

Dept 1 2 3 4 5 6 7 8 9 Area 16 16 16 36 36 9 9 9 9 Maximum Aspect Ratio 5 5 5 5 5 5 5 5 5

Table A.9 Material flow matrix for problem set O9

From 1 1 1 2 2 2 3 3 3 4 4 5 5 6 7 To 4 5 9 4 5 9 4 5 9 6 7 6 9 9 9 Material Flow 5 5 1 3 3 1 2 2 1 4 4 3 4 2 1

115 A.6 Data Set V10s Table A.10 Area requirement and material flow matrix for problem set V10s
Dept 1 2 3 4 5 6 7 8 9 10 1 2 0 3 0 0 4 5 6 7 8 9 10 Area Min Side 0 0 218 0 0 0 0 238 1 0 0 148 0 0 296 0 112 1 28 70 0 0 0 0 0 160 1 0 28 70 140 0 0 80 1 0 0 210 0 0 120 1 0 0 0 0 80 1 0 0 28 60 1 0 888 85 1 - 59.2 221 1 119 1

A.7

Data Set V10a Table A.11 Area requirement and material flow matrix for problem set V10a
Dept 1 2 3 4 5 6 7 8 9 10 1 2 0 3 0 0 4 5 6 7 8 9 10 Area 0 0 218 0 0 0 0 238 0 0 148 0 0 296 0 112 28 70 0 0 0 0 0 160 0 28 70 140 0 0 80 0 0 210 0 0 120 0 0 0 0 80 0 0 28 60 0 888 85 - 59.2 221 119 Max Aspect Ratio 5 5 5 5 5 5 5 5 5 5

A.8

Data Set M11s Table A.12 Area requirement and material flow matrix for problem set M11s
Dept 1 2 3 4 5 6 7 8 9 10 11 Area Min Side 1 0 10 0 0 140 90 20 0 40 0 0 3 1 2 0 0 10 0 0 0 0 0 0 0 0 2 1 4 1 3 0 0 0 10 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 4 5 1 5 0 10 0 0 0 0 40 0 0 20 0 2 1 6 0 0 10 0 0 0 0 0 20 0 0 3 1 7 0 0 0 0 0 0 0 10 0 0 0 4 1 8 0 0 0 0 0 0 0 0 0 0 11 5 1 9 0 0 0 0 0 0 0 0 0 20 0 1 1 10 0 0 0 0 0 0 0 0 0 0 20 1 1 11 146 0 0 0 0 0 0 0 0 0 0 6 1

116 A.9 Data Set M11a Table A.13 Area requirement and material flow matrix for problem set M11a
Dept 1 2 3 4 5 6 7 8 9 10 11 Area 1 0 10 0 0 140 90 20 0 40 0 0 3 2 0 0 10 0 0 0 0 0 0 0 0 2 4 3 0 0 0 10 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 4 5 5 0 10 0 0 0 0 40 0 0 20 0 2 6 0 0 10 0 0 0 0 0 20 0 0 3 7 0 0 0 0 0 0 0 10 0 0 0 4 8 0 0 0 0 0 0 0 0 0 0 11 5 9 0 0 0 0 0 0 0 0 0 20 0 1 10 0 0 0 0 0 0 0 0 0 0 20 1 11 146 0 0 0 0 0 0 0 0 0 0 6 Max Aspect Ratio
5 5 5 5 5 5 5 5 5 5 5

A.10

Data Set M15s

Table A.14 Area requirement and material flow matrix for problem set M15s
Dept 1 3 4 5 6 7 8 9 10 11 12 13 14 15 1 0 0 0 0 0 0 0 0 0 0 0 0 0 2 240 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 6 0 0 0 0 0 0 0 0 0 0 0 0 7 0 0 0 0 0 8 0 0 0 0 0 9 0 0 0 0 0 0 0 0 0 0 0 0 0 10 11 0 0 0 0 0 0 0 600 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 12 13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 14 0 0 0 0 0 0 0 0 0 0 0 15 Area Min side 15 10 9 7 9 25 25 15 10 25 10 15 6 19 25 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 240

0 1200

0 1200

0 600

0 480 0 480 0 0 0 0 0 0 0 0 0 0 0 0 0

0 120

0 600

0 480 0 480

0 600

0 600

0 10 50

0 25 40

0 25

0 40

0 20

117 A.11 Data Set M15a

Table A.15 Area requirement and material flow matrix for problem set M15a
Dept 1 3 4 5 6 7 8 9 10 11 12 13 14 15 1 0 0 0 0 0 0 0 0 0 0 0 0 0 2 240 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 6 0 0 0 0 0 0 0 0 0 0 0 0 7 0 0 0 0 0 8 0 0 0 0 0 9 0 0 0 0 0 0 0 0 0 0 0 0 0 10 11 0 0 0 0 0 0 0 600 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 12 13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 14 0 0 0 0 0 0 0 0 0 0 0 15 Area 15 10 9 7 9 25 25 15 10 25 10 15 6 19 25 0 0 0 0 0 0 0 0 0 0 0 0 Max aspect ratio 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 0 240

0 1200

0 1200

0 600

0 480 0 480 0 0 0 0 0 0 0 0 0 0 0 0 0

0 120

0 600

0 480 0 480

0 600

0 600

0 10 50

0 25 40

0 25

0 40

0 20

A.12

Data Set M25 Table A.16 Area requirement and material flow matrix for problem set M25

Dept 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

1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4

2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5

3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4

4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

5 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4

9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Area Max aspect ratio 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 2 1 5 3 1 3 5 3 2 1 3 2 4 5 2 4 1 5 3 4 1 4 5 2 4 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

0 20

0 20

0 10

0 40

0 10

0 20

0 40

0 20

0 40

0 40

0 80

0 20

0 40 0 20 0 0 0 0 0 0 0 0 0 0 0 0 40

0 20 0 10

0 16

0 10

0 10

0 16

0 16

118 A.13 Data Set NUG12 Table A.17 Area requirement for problem set NUG12
Dept Area Max aspect ratio 1 1 5 2 1 5 3 1 5 4 1 5 5 1 5 6 1 5 7 1 5 8 1 5 9 1 5 10 1 5 11 1 5 12 1 5

Table A.18 Material flow matrix for problem set NUG12

From 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 4 4 4 To 2 3 4 5 8 9 10 11 12 3 5 6 7 9 10 8 9 10 11 12 5 6 7 Material flow 5 2 4 1 6 2 1 1 1 3 2 2 2 4 5 5 5 2 2 2 5 2 2 From 4 4 4 5 5 5 5 6 6 6 6 6 7 7 7 7 7 8 9 9 10 11 To 8 11 12 6 10 11 12 7 8 9 10 11 8 9 10 11 12 11 11 12 11 12 Material flow 10 5 5 10 5 1 1 5 1 1 5 4 10 5 2 3 3 5 10 10 5 2

119 A.14 Data Set NUG15 Table A.19 Area requirement for problem set NUG15
Dept Area Max aspect Ratio 1 1 5 2 1 5 3 1 5 4 1 5 5 1 5 6 1 5 7 1 5 8 1 5 9 1 5 10 1 5 11 1 5 12 1 5 13 1 5 14 1 5 15 1 5

Table A.20 Material flow matrix for problem set NUG15

From 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 To 2 4 5 7 8 9 10 11 13 3 4 5 6 7 8 9 11 13 14 4 5 7 8 9 10 Material Flow 10 5 1 1 2 2 2 2 4 1 3 2 2 2 3 2 2 10 5 10 2 2 5 4 5 From 3 3 3 3 3 4 4 4 4 4 4 4 5 5 5 5 5 5 5 5 6 6 6 6 6 To 11 12 13 14 15 5 6 7 10 11 13 14 6 7 8 9 10 12 14 15 7 8 9 10 13 Material Flow 2 2 5 5 5 1 1 5 2 1 2 5 3 5 5 5 1 3 5 5 2 2 1 5 2 From 6 6 7 7 7 7 7 7 8 8 8 8 9 9 9 9 10 10 11 11 12 12 13 13 14 To 14 15 8 10 11 12 13 14 9 10 11 13 11 12 13 15 12 15 12 14 13 14 14 15 15 Material Flow 5 10 6 1 5 5 5 1 5 2 10 5 10 5 10 2 4 5 5 5 3 3 10 2 4

120 A.15 Data Set BA12 Table A.21 Area requirement for problem set BA12
Dept Area Min Side 1 2 3 4 5 6 7 8 9 10 11 12 9 8 10 6 4 3 3 4 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1

Table A.22 Material flow matrix for problem set BA12

Dept 1 2 3 4 5 6 7 8 9 10 11 12 1 2 288 3 180 240 4 54 54 120 5 72 72 80 72 6 180 24 0 18 12 7 27 48 60 18 12 18 8 72 160 120 48 64 24 0 9 36 16 60 24 16 6 6 16 10 0 64 0 48 16 12 6 16 4 11 0 8 0 12 4 3 3 16 4 2 12 9 16 30 0 8 3 6 4 2 2 2

A.16

Data Set BA12TS Table A.23 Area requirement for problem set BA12TS
Dept Area Min Side 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 9 8 10 6 4 3 3 4 2 2 2 2 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0

121

Table A.24 Material flow matrix for problem set BA12TS

Dept 1 2 3 4 5 6 7 8 9 10 11 12 1 2 288 3 180 240 4 54 54 120 5 72 72 80 72 6 180 24 0 18 12 7 27 48 60 18 12 18 8 72 160 120 48 64 24 0 9 36 16 60 24 16 6 6 16 10 0 64 0 48 16 12 6 16 4 11 0 8 0 12 4 3 3 16 4 2 12 9 16 30 0 8 3 6 4 2 2 2

A.17

Data Set BA14 Table A.25 Area requirement for problem set BA14
Dept Area Min Side 1 2 3 4 5 6 7 8 9 10 11 12 13 14 9 8 9 10 6 3 3 3 2 3 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0

Table A.26 Material flow matrix for problem set BA14

Dept 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2 72 3 162 72 4 90 80 45 5 108 0 54 30 6 27 48 27 0 18 7 0 0 27 30 0 9 8 0 48 27 30 18 9 9 9 18 32 0 20 12 0 12 6 10 27 0 27 0 18 0 9 9 6 11 18 16 0 20 24 6 6 0 4 6 12 0 8 9 10 0 6 3 3 6 3 2 13 0 0 18 10 0 6 0 0 2 6 0 4 14 0 0 0 0 0 0 0 0 0 0 0 0 0

122

A.18

Data Set BA14TS Table A.27 Area requirement for problem set BA14TS
Dept Area Min Side 1 2 3 4 5 6 7 8 9 10 11 12 13 14 9 8 9 10 6 3 3 3 2 3 2 1 1 3 1 1 1 1 1 1 1 1 1 1 1 1 0 0

Table A.28 Material flow matrix for problem set BA14TS

Dept 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1 2 72 3 162 72 4 90 80 45 5 108 0 54 30 6 27 48 27 0 18 7 0 0 27 30 0 9 8 0 48 27 30 18 9 9 9 18 32 0 20 12 0 12 6 10 27 0 27 0 18 0 9 9 6 11 18 16 0 20 24 6 6 0 4 6 12 0 8 9 10 0 6 3 3 6 3 2 13 0 0 18 10 0 6 0 0 2 6 0 4 14 0 0 0 0 0 0 0 0 0 0 0 0 0

123 A.19 Data Set AB20

Max aspect ratio 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75 1.75

Table A.29 Area requirement and material flow matrix for problem set AB20
Dept 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 18 12 0 0 0 0 0 0 10.4 11.2 0 0 12 0 0 0 0 0 0 2 18 9.6 244.5 7.8 0 139.5 0 12 13.5 0 0 0 0 0 0 0 0 69 0 3 12 9.6 0 0 22.1 0 0 31.5 39 0 0 0 130.5 0 0 0 0 136.5 0 4 0 244.5 0 10.8 57 75 0 23.4 0 0 14 0 0 0 0 0 15 157.5 0 5 0 7.8 0 10.8 0 22.5 13.5 0 15.6 0 0 0 0 13.5 0 0 0 0 0 6 0 0 22.1 57 0 61.5 0 0 0 0 4.5 0 0 0 0 0 10.5 0 0 7 0 139.5 0 75 22.5 61.5 240 0 18.7 0 0 0 9.6 0 0 0 16.5 0 37.5 8 0 0 0 0 13.5 0 240 0 0 0 0 6 0 0 0 0 0 75 334.5 9 0 12 31.5 23.4 0 0 0 0 0 0 0 0 75 0 0 75 0 0 0 10 10.4 13.5 39 0 15.6 0 18.7 0 0 3.6 120 0 186 19.2 0 0 0 52.5 0 11 11.2 0 0 0 0 0 0 0 0 3.6 22.5 0 30 9.6 0 0 0 0 0 12 0 0 0 14 0 4.5 0 0 0 120 22.5 0 0 16.5 0 150 0 84 0 13 0 0 0 0 0 0 0 6 0 0 0 0 80 10.4 60 0 0 0 0 14 12 0 130.5 0 0 0 9.6 0 75 186 30 0 80 97.5 0 0 9 0 0 15 0 0 0 0 13.5 0 0 0 0 19.2 9.6 16.5 10.4 97.5 0 52.5 0 0 0 16 0 0 0 0 0 0 0 0 0 0 225 0 60 0 0 120 0 0 0 17 0 0 0 0 0 0 0 0 75 0 0 150 0 0 52.5 120 0 75 0 18 0 0 0 15 0 10.5 16.5 0 0 0 0 0 0 9 0 0 0 46.5 0 19 0 69 136.5 157.5 0 0 0 75 0 52.5 0 84 0 0 0 0 75 46.5 0 20 0 0 0 0 0 0 37.5 334.5 0 0 0 0 0 0 0 0 0 0 0 Area 0.27 0.18 0.27 0.18 0.18 0.18 0.09 0.09 0.09 0.24 0.6 0.42 0.18 0.24 0.27 0.75 0.64 0.41 0.27 0.45

A.20

Data Set SC30 Table A.30 Area requirement for problem set SC30
Dept 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Area 3 4 4 16 4 5 2 3 5 6 2 24 5 3 11 6 2 8 4 5 4 3 1 3 Max Aspect Ratio 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Dept 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 Area 1 4 6 1 14 4 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Max Aspect Ratio 5 5 5 5 5 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

124 Table A.31 Material flow matrix for problem set SC30
From 1 1 1 1 1 1 2 2 2 2 2 2 4 5 5 6 6 6 6 6 6 7 8 9 9 To 24 25 26 27 28 30 6 19 20 21 22 23 3 11 12 5 7 8 10 12 17 8 9 11 12 Material From Flow 2.95 10 6.32 11 1.26 12 2.11 12 1.26 12 13.91 12 20.38 13 4.5 14 3.63 15 2.93 15 1.29 16 1.43 17 394.11 18 4.09 19 33.09 20 8.92 21 2.07 22 4.84 23 4.56 24 1.83 25 0.61 26 12.97 27 43.23 28 4.76 29 38.47 30 To 12 12 13 14 15 16 15 15 4 18 15 15 4 29 29 29 29 29 29 29 29 29 29 4 29 Material Flow 38.03 8.84 12.97 4.67 53.42 1.83 12.97 4.67 63.95 190.74 4.58 0.76 190.74 9.18 7.4 5.97 2.63 2.92 5.9 12.65 2.53 4.22 2.53 190.13 59.64

125 A.21 Data Set SC35 Table A.32 Area requirement for problem set SC35
Dept 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 Area 3 5 4 14 4 5 2 3 5 6 2 6 5 3 13 6 2 10 4 5 4 3 1 3 1 4 6 1 18 4 Max Aspect Ratio 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 Dept 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 Area 9 14 10 4 3 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 Max Aspect Ratio 4 4 4 4 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

126 Table A.33 Material flow matrix for problem set SC35
From 1 1 1 1 1 1 2 2 2 2 2 2 4 5 5 6 6 6 6 6 6 6 7 8 9 9 10 To 24 25 26 27 28 30 6 19 20 21 22 23 3 11 12 5 7 8 10 12 17 31 8 9 11 12 12 Material Flow 2.95 6.32 1.26 2.11 1.26 13.91 20.38 4.5 3.63 2.93 1.29 1.43 225.65 4.09 33.09 8.92 2.07 4.84 4.56 1.83 0.61 3.93 12.97 43.23 4.76 38.47 38.03 From 11 12 13 14 15 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 32 32 32 33 To 12 32 15 15 4 18 15 15 4 29 29 29 29 29 29 29 29 29 29 33 29 32 13 14 15 16 34 Material Flow 8.84 72.89 12.97 4.67 63.95 190.74 4.58 0.76 190.74 9.18 7.4 5.97 2.63 2.92 5.9 12.65 2.53 4.22 2.53 190.13 59.64 22.92 12.97 4.67 53.42 1.83 168.45

127 APPENDIX B Best layout obtained by the proposed AS algorithm

B.1

Best layout obtained for problem set O7

B.2

Best layout obtained for problem set FO7

128 B.3 Best layout obtained for problem set FO8

B.4

Best layout obtained for problem set O9

129 B.5 Best layout obtained for problem set V10s

B.6

Best layout obtained for problem set V10a

130 B.7 Best layout obtained for problem set M11s

B.8

Best layout obtained for problem set M11a

B.9

Best layout obtained for problem set M15s

131 B.10 Best layout obtained for problem set M15a

B.11

Best layout obtained for problem set M25

B.12

Best layout obtained for problem set NUG12

132 B.13 Best layout obtained for problem set NUG15

B.14

Best layout obtained for problem set BA12

133 B.15 Best layout obtained for problem set BA12 (mFBS)

B.16

Best layout obtained for problem set BA12TS

134 B.17 Best layout obtained for problem set BA14

B.18

Best layout obtained for problem set BA14 (mFBS)

B.19

Best layout obtained for problem set BA14TS

135 B.20 Best layout obtained for problem set AB20

B.21

Best layout obtained for problem set SC30

136 B.22 Best layout obtained for problem set SC35